text
stringlengths 47
469k
| meta
dict | domain
stringclasses 1
value |
|---|---|---|
---
abstract: 'We present a tight-binding theory of triangular graphene quantum dots (TGQD) with zigzag edge and broken sublattice symmetry in external magnetic field. The lateral size quantization opens an energy gap and broken sublattice symmetry results in a shell of degenerate states at the Fermi level. We derive a semi-analytical form for zero-energy states in a magnetic field and show that the shell remains degenerate in a magnetic field, in analogy to the 0th Landau level of bulk graphene. The magnetic field closes the energy gap and leads to the crossing of valence and conduction states with the zero-energy states, modulating the degeneracy of the shell. The closing of the gap with increasing magnetic field is present in all graphene quantum dot structures investigated irrespective of shape and edge termination.'
author:
- 'A. D. Güçlü'
- 'P. Potasz'
- 'P. Hawrylak'
title: 'Zero-energy states of graphene triangular quantum dots in a magnetic field'
---
INTRODUCTION
============
Graphene currently attracts considerable attention due to remarkable electronic and mechanical properties.[@Wallace+47; @Novoselov+Geim+04; @Novoselov+Geim+05; @Zhang+Tan+05; @Son+PRL+06; @Potemski+deHeer+06; @Geim+Novoselov+07; @Rycerz+Tworzydlo+07; @Xia+Mueller+09; @Mueller+Xia+10; @Neto+Guinea+09] When graphene is reduced to graphene nanostructures, new effects related to size-quantization and edges appear.[@Neto+Guinea+09; @Abergel+10; @Rozhkov+11] Considerable experimental effort has been made aiming at producing graphene nanostructures with desired shape and edges.[@Li+08; @Ponomarenko+08; @Ci+08; @You+08; @Schnez+09; @Ritter+09; @Jia+09; @Campos+09; @Neubeck+10; @Biro+10; @CruzSilva+10; @Yang+10; @Krauss+10; @Zhi+08; @Treier+10; @Mueller+10; @Morita+11; @Lu+11; @Chen+12] Among graphene nanostructures, nanoribbons and quantum dots are of particular interest. In graphene quantum dots, a size-dependent energy gap opens,[@Yamamoto+06; @Zhang+08; @Guclu+10] and its magnitude is determined by shape and edge termination. In graphene quantum dots with zigzag type edges, edge states with energy in the vicinity of the Fermi energy appear. [@Son+PRL+06; @Yamamoto+06; @NFD+96; @Fujita+96; @Son+06; @Ezawa+06; @Ezawa+07; @FRP+07; @AHM+08; @Wang+Yazyev+09; @Wimmer+10; @Potasz+10; @Voznyy+11] These edge states have significant effects on low-energy electronic properties such as a decrease of the energy gap compared to structures with armchair termination or, when combined with broken sublattice symmetry, a creation of the degenerate shell of zero-energy states in the middle of the energy gap. [@Yamamoto+06; @Ezawa+07; @FRP+07; @AHM+08; @Wang+Yazyev+09; @Guclu+09; @Wimmer+10; @Potasz+10; @Voznyy+11; @Ezawa10; @Wang+Meng+08] It was shown that the degenerate shell survives when various types of disorder are present in the system.[@Wimmer+10; @Potasz+10; @Voznyy+11; @Ezawa10]
The influence of an external magnetic field on the electronic properties of the graphene quantum dots was also studied. [@Yazyev10; @Chen+07; @Schnez+08; @Recher+07; @Abergel+08; @Wurm+08; @Guttinger+09; @Schnez+09; @Libisch+10; @Zhang+08; @Grujic+11; @Zarenia+11; @Bahamon+Pereira+09; @Romanovsky+11; @Romanovsky+12; @Potasz+09] The magnetic field plays the role of a tunable external parameter allowing to change electronic properties in a controllable way. Graphene quantum dots and rings with circular, square, hexagonal, triangular, and rhombus-shaped shapes with zigzag and armchair edges were investigated.[@Recher+07; @Abergel+08; @Wurm+08; @Schnez+09; @Bahamon+Pereira+09; @Libisch+10; @Grujic+11; @Zarenia+11; @Romanovsky+11] Triangular graphene quantum dots with reconstructed edges, consisting of a succession of pentagons and heptagons, were also considered.[@Romanovsky+12] The comparison between tight-binding and continuum model, the Dirac-Weyl equation, was analyzed for graphene quantum dots with different type of edges: zigzag, armchair, and infinite-mass boundary conditions.[@Grujic+11; @Zarenia+11; @Romanovsky+11] For a circular dot, good qualitative agreement between experiment and analytical model with infinite-mass boundary conditions was obtained.[@Schnez+08; @Schnez+09] Magneto-optical properties were also theoretically investigated.[@Zhang+08; @Grujic+11] The absorption spectra differ for hexagonal structures with armchair and zigzag edges due to different level structures and the oscillator strengths. A fast reduction of the energy gap with increasing magnetic field in zigzag hexagon in comparison with zigzag triangle was noted.[@Grujic+11; @Zarenia+11]
In this work, we present a tight-binding theory of triangular graphene quantum dots(TGQD) with zigzag edge and broken sublattice symmetry in external magnetic field. The lateral size quantization opens an energy gap and broken sublattice symmetry results in a shell of degenerate states at the Fermi level. Building on our previous work[@Potasz+10] we derive here a semi-analytical form for zero-energy states in a magnetic field and show that the shell remains degenerate at all magnetic fields perpendicular to the plane of the TGQD, in analogy to the 0th Landau level of bulk graphene. However, we find that the magnetic field closes the energy gap and leads to the crossing of valence and conduction states with the zero energy states, modulating the degeneracy of the shell. The closing of the gap with increasing magnetic field is present in all graphene quantum dot structures investigated irrespective of shape and edge termination.
The paper is organized as follows. In Sec. II, we present a brief outline of the tight-binding model with an incorporation of a perpendicular magnetic field. The analysis of the evolution of the energy spectra of TGQD, a derivation of the analytical form for eigenfunctions corresponding to zero-energy states, and a prediction of crossings of valence and conduction states with the zero energy Fermi level $E=0$ are included in Sec. III. In Sec. IV the energy gap in a magnetic field for GQDs with different shapes and edge termination is considered. The conclusions are presented in Sec. V.
MODEL
=====
We describe graphene quantum dots using the nearest-neighbor tight-binding model which has been successfully used to describe graphene [@Wallace+47] and applied to other graphene materials such as nanotubes, nanoribbons and quantum dots [@NFD+96; @Ezawa+06; @Yamamoto+06; @Ezawa+07; @FRP+07; @AHM+08; @Potasz+10; @Saito+98]. A perpendicular magnetic field can be incorporated by using Peierls substitution [@Peierls+33]. The Hamiltonian reads, $$\begin{aligned}
H_{TB}= t\sum_{\left\langle i,j\right\rangle,\sigma}
\left[ e^{i\varphiup_{ij}} a^\dagger_{i\sigma}b_{j\sigma}
+e^{-i\varphiup_{ij}} b^\dagger_{j\sigma}a_{i\sigma}\right],
\label{HTB}\end{aligned}$$ where $t$ is hopping integral, $a^\dagger_{i\sigma}$($b^\dagger_{i\sigma}$) and $a_{i\sigma}$($b_{i\sigma}$) are creation and annihilation operators on a site $i$ corresponding to sublattice A(B) of bipartite honeycomb lattice, $\left\langle i,j\right\rangle$ indicate summation over nearest-neighbors, and $\sigma$ is spin index. Hopping integral between nearest neighbors is $t=-2.8$ eV [@Neto+Guinea+09]. Under symmetric gauge, a vector potential ${\bf{A}}=B_z/2(-y,x,0)$, and $$\begin{aligned}
\varphiup_{ij}=2\pi\frac{e}{hc}\int_{r_i}^{r_j}{\bf A}d{\bf l}
=2\pi\frac{B_z}{2\phi_0}\left(x_iy_j-x_jy_i\right)
\label{phaseij}\end{aligned}$$ corresponds to a phase accumulated by electron going from site $i$ to $j$, which is equal to a magnetic flux going through area $S=\frac{x_iy_j-x_jy_i}{2}$ spanned by vectors ${\bf r}_i$ and ${\bf r}_j$, and $\phi_0=\frac{hc}{e}$ is magnetic flux quantum. The evolution of the energy spectrum in a magnetic field will be shown in units of the magnetic flux threading one benzene ring, $\phi/\phi_0=B_zS_0/\phi_0$, where $S_0=3\sqrt{3}a_0^2/2$ is benzene ring area with $a_0=1.42$ $\AA$.
ZIGZAG TRIANGULAR QUANTUM DOT IN A MAGNETIC FIELD
=================================================
The evolution of the energy spectrum
------------------------------------
We focus here on the effect of the magnetic field on the electronic properties of TGQDs, quantum dots with broken sublattice symmetry. We illustrate the energy spectrum and its evolution with increasing magnetic field on a TGQD with $N=97$ carbon atoms. Fig. \[fig:Fig1\] shows the energy spectrum and it’s evolution in the magnetic field obtained by numerical diagonalization of the Hamiltonian, Eq. (\[HTB\]).
At $B=0$ there are $N_{deg}=7$ degenerate states at zero energy or Fermi level. The number of states is equal to the difference between the number of $A$ and $B$ atoms [@Potasz+10]. The states belonging to the degenerate shell are primarily localized at the edge of the triangle and are entirely localized on one sublattice, say A, as shown in Fig. \[fig:Fig2\](c).
The evolution of the energy spectrum as a function of the magnetic field is shown on the right hand side of Fig. \[fig:Fig1\]. The spectrum is symmetric with respect to $E=0$ due to electron-hole symmetry. This symmetry is broken when hoppings to the second nearest neighbors in Hamiltonian,Eq. (\[HTB\]), are included. The highest valence state and the lowest conduction state with $E=\pm 1.57$ eV, which in the absence of the magnetic field are each doubly degenerate, split in the presence of a magnetic field. The state labeled by II from the valence band increases and the state labeled by I from the conduction band decreases its energy with increasing magnetic field, closing the energy gap. Around $\phi/\phi_0\simeq 0.11$ these states reach Fermi level at $E=0$.
The explanation of why the energy gap closes in a magnetic field can be found by considering Dirac Fermions in bulk graphene. [@Novoselov+Geim+05; @Zhang+Tan+05]. We focus on one of two Fermi points, say $K$ point. Following Refs. the energy spectrum of Dirac Hamiltonian in the presence of magnetic field is given by $$\begin{aligned}
E_n=\pm\sqrt{2\hbar v_FeB_z|n|/c},
\label{LLDirac}\end{aligned}$$ where $v_F$ is Fermi velocity, $c$ speed of light, and $n$ Landau level index. The $\pm$ sign corresponds to electron (hole) Landau levels. A unique property of the energy spectrum is the existence of the $n=0$ Landau level with energy $E=0$, constant for all magnetic fields. When the magnetic field is applied to graphene quantum dots, discrete energy levels evolve into the degenerate Landau Levels for Dirac Fermions. Thus, some levels have to evolve into the $0$-th Landau level, closing the energy gap as shown in Fig. \[fig:Fig1\]. Another feature of the $0$-th Landau level is that the wavefunctions are localized on only one sublattice, similar to the zero-energy states in TGQD[@Potasz+10].
We note in Fig. \[fig:Fig1\] that the zero-energy degenerate shell is immune to the magnetic field as is the $n=0$ Landau level. This is certainly different from electronic states in semiconductor quantum dots, where $\sim B^2$ dependence is observed. [@raymond+04]
These comments are now illustrated by examining wavefunctions of a TGQD in a magnetic field. We investigate the evolution of the probability density of the wavefunction corresponding to state I, bottom of the conduction band, from Fig. \[fig:Fig1\], and the total probability density of the zero-energy degenerate shell in a magnetic field. For state I, probability densities at low and high magnetic field values are shown in Fig. \[fig:Fig2\](a) and (b), respectively. We note that due to the electron-hole symmetry, an identical evolution for the state II from Fig. \[fig:Fig1\] (not shown here) occurs. Eigenfunctions of states with energy $-|E|$ and $+|E|$ differ only by a sign of a coefficient on sublattice $B$ indicated by filled circles in Fig. \[fig:Fig2\], giving identical electronic densities. For $\phi/\phi_0\simeq 0.01$, Fig. \[fig:Fig2\](a), the state I is mostly localized at the center of the dot. With increasing magnetic field, it starts to occupy the edge sites, shown for $\phi/\phi_0\simeq 0.08$. We note that for arbitrary magnetic field this state is equally shared over two sublattices, [*i.e*]{}, has $50\%$ sublattice content. In Fig. \[fig:Fig2\](c) and (d) the evolution of the total electronic density of the degenerate zero-energy shell is shown. The electronic density of the degenerate shell is obtained by summing over all $N_{deg}=7$ states. Initially, degenerate states are strongly localized on edges, shown in Fig. \[fig:Fig2\](c) for $\phi/\phi_0\simeq 0.01$. When the magnetic field increases, these states move slightly towards the center of the triangle, shown in Fig. \[fig:Fig2\](d) for $\phi/\phi_0\simeq 0.08$. We note that even in the presence of an external magnetic field states from the degenerate shell are still localized on only one type of atoms, sublattice A, indicated by open circles in Fig. \[fig:Fig2\].
Analytical solution for zero-energy states
------------------------------------------
Fig. \[fig:Fig1\] shows that numerical diagonalization of the tb-Hamiltonian gives the zero-energy states immune to external magnetic field. We will now prove this analytically. Our first goal is to show the existence of and find an expression for zero-energy eigenstates in the presence of a magnetic field. The zero energy states, if they exist, must be solutions of the singular eigenvalue problem, $$\begin{aligned}
H_{TB}\Psi=0,
\label{singul}\end{aligned}$$ where the Hamiltonian $H_{TB}$ is given by Eq. (\[HTB\]). There is no coupling between two sublattices and the solution can be written separately for $A$-type and $B$-type of atoms. We first focus on sublattice $A$ with an eigenfunction given by $$\begin{aligned}
|\Psi^{A}\rangle=\sum_j C_{j}a^\dagger_{j}|0\rangle=\sum_j C_{j}|\phi_{j}^{A}\rangle,
\label{psiA}\end{aligned}$$ where $C_{j}$ are expansion coefficients of eigenstates written in a basis of $p_z$ orbitals $\phi_{j}^{A}$ localized on $A$-type site $j$ for either spin state omitted in what follows. According to Eq. \[singul\], the coefficients $C_{j}$ corresponding to one type orbitals localized around the second type site $i$ obey $$\begin{aligned}
t\sum_{\left\langle i,j\right\rangle}C_{j}e^{i\varphiup_{ij}}=0,
\label{cond}\end{aligned}$$ where the summation is over $j$-th nearest neighbors of an atom $i$. In other words, the sum of coefficients multiplied by a phase $e^{i\varphiup_{ji}}$ gained by going from one type site $i$ to the other type site $j$ around each site $i$ must vanish. For the $i$-th $B$-type site plotted on the left in Fig. \[fig:Fig3\], Eq. (\[cond\]) gives $$\begin{aligned}
C_{j}e^{i\varphiup_{ij}}+C_{k}e^{i\varphiup_{ik}}+C_{l}e^{i\varphiup_{il}}=0,
\label{cond2}\end{aligned}$$ where phases $\varphiup_{ij},\varphiup_{ik},\varphiup_{il}$ are given by Eq. (\[phaseij\]). Using a fact that $\varphiup_{ik}=-\varphiup_{ki}$ for arbitrary $i$ and $k$, Eq. (\[cond2\]) can be written as $$\begin{aligned}
C_{l}=-\left(C_{j}e^{-i\varphiup_{jl}}+C_{k}e^{-i\varphiup_{kl}}\right),
\label{cond4}\end{aligned}$$ where $\varphiup_{jl}=\varphiup_{ji}+\varphiup_{il}$ and $\varphiup_{kl}=\varphiup_{ki}+\varphiup_{il}$ correspond to phase changes going from $A$-type sites $k$ to $j$, and $l$ to $j$, respectively, through $B$-type site $i$ (see right part in Fig. \[fig:Fig3\]). Thus, in analogy with the zero magnetic field case [@Potasz+10], a coefficient from a given row can be expressed as a sum of two coefficients from an upper lying row, $C_{j}$ and $C_{k}$ on the right in Fig. 3. The effect of the magnetic field is incorporated in the extra phase gained by going from a given site from an upper row of atoms to a lower one. For a reason which will become clear later, instead of using indices $i$, each $A$-type site will be labeled by two integer numbers, ${i}=\{n,m\}$. The first index, $n$, corresponds to an atom number in a given row counted from left to right, and the second one, $m$, corresponds to the row number. Let us illustrate our methodology on a hexagonal benzene ring with three auxiliary $A$-type atoms with indices $C_{0,0}$, $C_{2,0}$ and $C_{0,2}$, shown in Fig. \[fig:Fig4\](a). Eq. (\[cond4\]) can be used to obtain coefficients $C_{0,1}$ from $C_{0,0}$ and $C_{1,0}$, and $C_{1,1}$ from $C_{1,0}$ and $C_{2,0}$. Next, using $C_{0,1}$ and $C_{1,0}$, one obtains coefficient $C_{0,2}$, $$\begin{aligned}
C_{0,2}=C_{0,0}e^{-i\varphiup_{1}}+C_{1,0}\left(e^{-i\varphiup_{2}}+e^{-i\varphiup_{3}}\right)
+C_{2,0}e^{-i\varphiup_{4}},
\label{eq5}\end{aligned}$$ with phase changes $\varphiup_{i}$, $i=1,2,3,4$, shown as black arrows in Fig. \[fig:Fig4\](a). The paths related to phase changes $\varphiup_{i}$ go through intermediate atomic sites, e.g., for $\varphiup_{1}$ the path goes from a site $C_{0,0}$ to $C_{0,1}$ through an intermediate $B$-type atomic site, and next from a site $C_{0,1}$ to $C_{0,2}$ through connecting $B$-type atomic site. According to Eq. (\[eq5\]) and Fig. \[fig:Fig4\](a), there is one path connecting $C_{0,0}$ and $C_{0,2}$, one connecting $C_{2,0}$ and $C_{0,2}$, but there are two paths around a hexagonal benzene ring connecting coefficients $C_{1,0}$ and $C_{0,2}$. We have shown that the coefficient in the bottom, $C_{0,2}$, can be expressed as a linear combination of coefficients from the top row, $C_{n,0}$. We will now demonstrate that all coefficient in arbitrary size triangles can be expressed in terms of coefficients $C_{n,0}$.
In Fig. \[fig:Fig4\](b) a small triangle with $N_{ed}=2$ atoms on the one edge is plotted. Three auxiliary atoms with coefficients $C_{0,0}$, $C_{3,0}$, and $C_{0,3}$ were added.
The total number of atoms is $N=16$. In a similar way to the procedure used to obtain Eq. (\[eq5\]), a coefficient $C_{0,3}$ can be expressed as a sum of coefficients from the top. Here, from coefficients $C_{0,0}$ and $C_{3,0}$ to $C_{0,3}$ there is only one path for each coefficient, and three paths for each coefficient connecting $C_{1,0}$ to $C_{0,3}$, and $C_{2,0}$ to $C_{0,3}$. For transparency, only for the first two coefficients from the left ($C_{0,0}$ and $C_{1,0}$) paths are plotted in Fig. \[fig:Fig4\](b). The number of paths from a given site in the upper row of atoms to lower lying atomic sites corresponds to numbers from a Pascal triangle, $\{1,2,1\}$ for coefficient $C_{0,2}$, shown in Fig. \[fig:Fig4\](a), and $\{1,3,3,1\}$ for coefficient $C_{0,3}$, shown for the first two coefficients from the left in Fig. \[fig:Fig4\](b). The number of paths connecting a site $\{n,m\}$ with a site from the top $\{n+j,0\}$ can be described by binomial coefficient $N_{path}(n,m,n+j)={m \choose j}$, $0\leq
j\leq m$. The general form for an arbitrary coefficient expressed in coefficients from the top row can be written as $$\begin{aligned}
C_{n,m}=(-)^{m}\sum_{j=0}^{m}\sum_{i=1}^{m \choose j}C_{n+j,0}e^{-i\varphiup_{n+j}(i)},
\label{eq6}\end{aligned}$$ where two numbers $n$ and $m$ satisfy condition $0<n,m<N_{ed}+1$, and $\varphiup_{n+j}(i)$ is a path-dependent phase change from a site $\{n+j,0\}$ to $\{n,m\}$. One can note that in the absence of a magnetic field $\varphiup_{n+j}(i)=0$ and Eq. (\[eq6\]) reduces to Eq.(2) from Ref.[@Potasz+10].
The summation over all possible paths in Eq. (\[eq6\]) is not practical. We now show a way of reducing the number of paths to only one. We use the fact that a phase change corresponding to a closed path around a hexagon is by definition $\varphiup_{c}=2\pi\phi/\phi_0$. The sum of two exponential terms standing next to coefficient $C_{1,0}$ in Eq. (\[eq5\]) can be written as $$\begin{aligned}
\label{expon1}
e^{-i\varphiup_{2}}+e^{-i\varphiup_{3}}=\left(e^{i(\varphiup_{3}-\varphiup_{2})}+1\right)e^{-i\varphiup_{3}}
=\left(e^{2\pi i\frac{\phi}{\phi_0}}+1\right)e^{-i\varphiup_{3}},\end{aligned}$$ where $\varphiup_{3}-\varphiup_{2}=2\pi\phi/\phi_0$ is a closed path around a single hexagon, see Fig. \[fig:Fig4\](a). Similarly for three exponential terms corresponding to paths connecting $C_{10}$ and $C_{03}$, shown in Fig. \[fig:Fig4\](b), one can write $$\begin{aligned}
\nonumber
e^{-i\varphiup_{2}'}+e^{-i\varphiup_{3}'}+e^{-i\varphiup_{4}'}&=&
\left(e^{i(\varphiup_{4'}-\varphiup_{2'})}+e^{i(\varphiup_{4'}-\varphiup_{3'})}+1\right)e^{-i\varphiup_{4'}}\\
&=&\left(e^{2\pi i(2\frac{\phi}{\phi_0})}+e^{2\pi i\frac{\phi}{\phi_0}}+1\right)e^{-i\varphiup_{4'}},
\label{expon2}\end{aligned}$$ where $(\varphiup_{4'}-\varphiup_{2'})$ circles two hexagons and $(\varphiup_{4'}-\varphiup_{3'})$ only one, see Fig. \[fig:Fig4\](b). Note that phases $\varphiup_{3}$ in Eq. (\[expon1\]) and $\varphiup_{4'}$ in Eq. (\[expon2\]) correspond to paths going on the right edge of the triangle. The sum of exponential terms of type $e^{2\pi i(j\frac{\phi}{\phi_0})}$ with $j$-integer in Eq. (\[expon1\]) and Eq. (\[expon2\]) forms geometric series which can be written as $$\begin{aligned}
\sum_{j=0}^{k}e^{2\pi i(j\frac{\phi}{\phi_0})}=\frac{1-e^{2\pi i(k+1)\frac{\phi}{\phi_0}}}{1-e^{2\pi i\frac{\phi}{\phi_0}}},
\label{expon3}\end{aligned}$$ with $k$ determined by the number of encircled benzene rings, and $k+1={m
\choose j}$ is a number of paths connecting site $\{n+j,0\}$ to $\{n,m\}$, $k=1$ in Eq. (\[expon1\]) and $k=2$ in Eq. (\[expon2\]), see Fig. \[fig:Fig4\]. Using Eq. (\[expon3\]), the number of paths in Eq. (\[eq6\]) can be reduced to only one. Eq. (\[eq6\]) can be written as $$\begin{aligned}
C_{n,m}=(-)^{m}\sum_{j=0}^{m}C_{n+j,0}\frac{1-e^{2\pi i{m \choose j}\frac{\phi}{\phi_0}}}{1-e^{2\pi i\frac{\phi}{\phi_0}}}e^{-i\varphiup_{n+j}},
\label{eq7}\end{aligned}$$ where $\varphiup_{n+j}$ is the phase corresponding to the path on the right edge connecting site $\{n+j,0\}$ and $\{n,m\}$. The coefficients $C_{n,m}$ for all $A$-type atoms in the triangle are expressed as a linear combination of coefficients corresponding to atoms on one edge, i.e., $C_{j,0}$. There are $N_{ed}+2$ coefficients in an upper row of atoms, $C_{j,0}$, with $0<j<N_{ed}+1$, which gives $N_{ed}+2$ independent solutions. Applying three boundary conditions corresponding to auxiliary atoms, $C_{0,0}=C_{N_{ed}+1,0}=C_{0,N_{ed}+1}=0$, leaves only $N_{ed}-1$ solutions, which corresponds to the number of zero-energy states, similar to the result obtained in the absence of a magnetic field in Ref. . We note that the solutions given by Eq. (\[eq7\]) are smooth functions of magnetic field, and exist for any value of $\phi$. Thus they do not include zero-energy solutions corresponding to the crossing of conduction and valence states with $E=0$, e. g., for $\phi/\phi_0\simeq 0.11$ for the triangular dot with $N_{ed}=8$ and $N=97$ atoms, see Fig. \[fig:Fig1\]. We investigate this issue by analyzing $B$-type atoms.
Prediction of crossings of valence and conduction states with $E=0$
-------------------------------------------------------------------
We consider the solution of Eq. (\[singul\]) corresponding to wavefunction localized only on $B$-type atoms. In Fig. \[fig:Fig5\] the same structures as in Fig. \[fig:Fig4\] without auxiliary corner atoms are shown with coefficients assigned to $B$-type atoms. For simplicity, only one index for each coefficient is used.
According to Eq. \[cond\], for a benzene ring plotted in Fig. \[fig:Fig5\](a) we can write $$\begin{aligned}
\label{ap_eq1a}
C_{2}=-C_{1}e^{-i\varphiup_{12}},\\
\label{ap_eq1b}
C_{3}=-C_{2}e^{-i\varphiup_{23}},\\
C_{1}=-C_{3}e^{-i\varphiup_{31}},
\label{ap_eq1c}\end{aligned}$$ where phase changes from site $i$ to $j$, $\varphiup_{ij}$, are indicated in Fig. \[fig:Fig5\](a). Eq. (\[ap\_eq1a\]) can be substituted into Eq. (\[ap\_eq1b\]), and next Eq. (\[ap\_eq1b\]) into Eq. (\[ap\_eq1c\]), giving $$\begin{aligned}
C_{1}=C_{1}(-1)^3e^{-i(\varphiup_{12}+\varphiup_{23}+\varphiup_{31})},
\label{ap_eq2}\end{aligned}$$ which is satisfied for arbitrary $C_{1}$. Eq. \[ap\_eq2\] leads to a following condition $$\begin{aligned}
\varphiup_{12}+\varphiup_{23}+\varphiup_{31}+\pi=2\pi k,
\label{ap_eq4}\end{aligned}$$ with $k=0,\pm 1,\pm 2,...$. A phase change in Eq. (\[ap\_eq4\]) corresponds to a closed path around a single hexagon, $\varphiup_{12}+\varphiup_{23}+\varphiup_{31}=2\pi\phi/\phi_0$. A condition for crossing of a valence and conduction states with $E=0$ is $$\begin{aligned}
\phi/\phi_0=k-1/2.
\label{ap_eq7}\end{aligned}$$ In order to confirm validity of Eq. (\[ap\_eq7\]), we show the energy spectrum of a benzene ring as a function of a magnetic field in Fig. \[fig:Fig6\](a). The crossing of energy levels at $E=0$ occurs for $\phi/\phi_0=1/2$, in agreement with Eq. (\[ap\_eq7\]).
We carry out a similar derivation for triangular zigzag graphene quantum dot with $N=13$ carbon atoms and $N_{ed}=2$ atoms on the one edge, shown in Fig. \[fig:Fig5\](b). A coefficient from the left upper corner, $C_1$, determines a coefficient $C_2$, $$\begin{aligned}
\label{ap_eq8}
C_{2}=-C_{1}e^{-i\varphiup_{12}}.\end{aligned}$$ Next, a coefficient $C_3$ can be determined by a coefficient $C_2$ $$\begin{aligned}
\label{ap_eq9}
C_{3}=-C_{2}e^{-i\varphiup_{23}}\end{aligned}$$ and combining with Eq. (\[ap\_eq8\]) gives $$\begin{aligned}
\label{ap_eq10}
C_{3}=(-1)^2C_{1}e^{-i(\varphiup_{12}+\varphiup_{23})}=(-1)^2C_{1}e^{-i\varphiup_{13}} .\end{aligned}$$ Going in this way along the three edges of the triangle a closed loop, shown with black arrows in Fig. \[fig:Fig5\](b), can be created. In the
case of $N_{ed}=2$ shown in Fig. \[fig:Fig5\](b), one goes through all $B$-type coefficients, while in larger triangles one goes only through outer coefficients. Thus, all outer $B$-type coefficients can be expressed by one chosen coefficient, $C_1$ in this case. The loop from Fig. \[fig:Fig5\](b) can be written $$\begin{aligned}
\label{ap_eq11}
C_{1}=(-1)^6C_{1}e^{-i(\varphiup_{16}+\varphiup_{61})}=C_{1}e^{6\pi i-2\pi i(3\phi/\phi_0)},\end{aligned}$$ where the phase change $\varphiup_{16}=\sum_i^5\varphiup_{i,i+1}$, and we used a fact that the total phase change corresponds to a closed loop around three benzene rings, $\varphiup_{16}+\varphiup_{61}=2\pi
(3\phi/\phi_0)$. Eq. (\[ap\_eq11\]) gives a condition $$\begin{aligned}
\label{ap_eq14}
2k\pi=6\pi-6\pi\phi/\phi_0,\end{aligned}$$ and finally $$\begin{aligned}
\label{ap_eq15}
\phi/\phi_0=\frac{3-k}{3}.\end{aligned}$$ with $k=0,\pm 1,\pm 2,...$. Eq. (\[ap\_eq15\]) can be extended to different size triangles. The number of benzene rings in a triangle is $N_b=N_{ed}(N_{ed}+1)/2$, and Eq. (\[ap\_eq15\]) can be written as $$\begin{aligned}
\label{ap_eq15n}
\phi/\phi_0=\frac{3N_{ed}-2k}{N_{ed}\left(N_{ed}+1\right)}.\end{aligned}$$ For the triangle with $N_{ed}=2$, Eq. (\[ap\_eq15\]) predicts crossings for $\phi/\phi_0=0,\pm 1/3, \pm 2/3, 1,..$ but according to Fig. \[fig:Fig6\](b) there are no crossings for $\phi/\phi_0=0$ and $\phi/\phi_0=1$. This is related to an extra condition in the center of the triangle, for coefficients $C_2$, $C_4$, and $C_6$. Phase changes between these coefficients are indicated by white arrows in Fig. \[fig:Fig5\](b). We can write $$\begin{aligned}
\label{ap_eq16}
C_{6}=-\left(C_{2}e^{-i\varphiup_{26}}+C_{4}e^{-i\varphiup_{46}}\right)\end{aligned}$$ and also $$\begin{aligned}
\label{ap_eq17}
C_{6}=-C_{1}e^{-i(-\varphiup_{61})}=-C_{1}e^{i\varphiup_{61}},\\
\nonumber
C_{2}=-C_{1}e^{-i\varphiup_{12}},\\
\nonumber
C_{4}=(-1)^3C_{1}e^{-i\varphiup_{14}},\end{aligned}$$ where the phase change $\varphiup_{14}=\sum_i^3\varphiup_{i,i+1}$. Combining Eq. (\[ap\_eq16\]) and Eq. (\[ap\_eq17\]) we get $$\begin{aligned}
\label{ap_eq18}
-C_{1}e^{i\varphiup_{61}}=-\left(-C_{1}e^{-i(\varphiup_{12}+\varphiup_{26})}
+(-1)^3C_{1}e^{-i(\varphiup_{14}+\varphiup_{46})}\right),\end{aligned}$$ which gives $$\begin{aligned}
\label{ap_eq19}
-1=e^{-i(\varphiup_{12}+\varphiup_{26}+\varphiup_{61})}
+e^{-i(\varphiup_{14}+\varphiup_{46}+\varphiup_{61})}.\end{aligned}$$ With help of Fig. \[fig:Fig5\](b), we can notice $\varphiup_{12}+\varphiup_{26}+\varphiup_{61}=2\pi\phi/\phi_0$ and $\varphiup_{14}+\varphiup_{46}+\varphiup_{61}=2\pi(2\phi/\phi_0)$. Thus, we can write $$\begin{aligned}
\label{ap_eq21}
1+e^{-2\pi i\phi/\phi_0}+e^{-2\pi i(2\phi/\phi_0)}=0\end{aligned}$$ or using a sum of geometric series $$\begin{aligned}
\label{ap_eq22}
\frac{1-e^{-2\pi i(3\phi/\phi_0)}}{1-e^{-2\pi i\phi/\phi_0}}=0.\end{aligned}$$ Eq. (\[ap\_eq22\]) gives a solution for $-2\pi (3\phi/\phi_0)=2\pi k$, $k$-integer, and finally $\phi/\phi_0=-k/3$, but with an extra condition $\phi/\phi_0\neq l$, with $l=0,\pm 1, \pm 2,..$ due to a denominator. This is in agreement with Fig. \[fig:Fig6\](b). We note that for all triangles, the prediction of crossings of conduction and valence states with $E=0$ given by Eq. (\[ap\_eq15n\]) has to be supported by extra conditions from equations for coefficients from the center of the triangle. For example, for the triangle with $N_{ed}=8$, the first crossing occurs for $\phi/\phi_0=1/9$, while incomplete condition given by Eq. (\[ap\_eq15n\]) predicts the first crossing for $\phi/\phi_0=1/36$, and the fourth crossing for $\phi/\phi_0=1/9$.
An interesting prediction of Eq.(\[ap\_eq15\]) is that the zero energy crossing values of $\phi/\phi_0$ should scale as $\sim 1/N_{ed}$ for large $N_{ed}$. In order to check numerically the size dependence of the position of the first crossing, in Fig. \[fig:Fig7\] we show the energy gap as a function of $\phi/\phi_0$ for different $N_{ed}$ obtained by diagonalization of the tight-binding Hamiltonian. Strikingly, we find that the first crossing always occurs at $\phi/\phi_0=1/(N_{ed}+1)$ for all the values of $N_{ed}$ that we have looked at. This is consistent with Eq.(\[ap\_eq15\]) with $k=N_{ed}$. Extrapolating this result to larger structures, it would take a magnetic field value of $\sim 10$ Tesla for a quantum dot with $N_{ed}=4000$ to reach the first zero energy crossing.
However, for large quantum dots ($N_{ed}>100$, or linear size $L>25$ nm) it becomes increasingly difficult to pinpoint numerically the position of the zero energy crossing due to smallness of the energy gap around the crossing and numerical accuracy. Another quantity of interest is the width at half maximum (WHM) of the flux dependence of the energy gap. In Fig. \[fig:Fig8\] we plot the WHM as a function of $N_{ed}$. Unlike the first crossing point which scales as $N_{ed}^{-1}$, the WHM scales as $\sim N_{ed}^{-2}$ for large $N_{ed}$, thus much faster. In Fig. \[fig:Fig8\] the largest structure that we looked at has $N=161601$ atoms ($N_{ed}=401$, $L=98.6$ nm) for which the WHM occurs at a magnetic field value of $B=1.97$ Tesla.
SHAPE AND EDGE DEPENDENCE OF THE ENERGY GAP IN A MAGNETIC FIELD
===============================================================
We discussed above the magnetic field closing of the energy gap in triangular graphene quantum dots. In Fig. \[fig:Fig9\], we analyze the evolution of the energy gap in graphene quantum dots with different shapes and edges in a perpendicular magnetic field. The energy gaps as a function of a magnetic field obtained by diagonalizing Hamiltonian given by Eq. (\[HTB\]) are shown for three different types of quantum dots; zigzag triangle, zigzag hexagon, and armchair hexagon. All three structures have similar sizes, consisting of $N\simeq 600$ atoms with area $S\simeq 14$ nm$^2$. The energy gap corresponds to the difference between the energy of the lowest state from the empty conduction states and the highest state from the doubly occupied valence states. In the absence of magnetic field, the zigzag triangular graphene quantum dot has a significantly larger gap then for hexagonal armchair and zigzag dots as discussed in Ref.. The functional form of the gap closure of different types of structures has significant differences as well, as seen in Fig. \[fig:Fig9\]. When the magnetic field increases, the energy gap closes for all structures. Although the hexagonal zigzag structure has slightly smaller size, the gap decays fastest showing a different behavior than the $\sim N_{ed}^{-2}$ scaling shown earlier for the triangular zigzag structure. Moreover, after reaching a plateau close to zero ($\sim 10^{-8}$) the hexagonal zigzag quantum dot shows no more structures, [*i.e.*]{} no zero energy crossings, unlike the two other quantum dots. We note that for the hexagonal zigzag structure the gap comes from closure of the edge-like states (which have finite energies unlike the triangular zigzag structure). This shows that the zero crossings are characteristics of bulk-like states.\
CONCLUSIONS
===========
The electronic properties of triangular graphene quantum dots with zigzag edges and broken sublattice symmetry in the presence of perpendicular external magnetic field were described. It was shown that the degenerate shell of zero-energy states in the middle of the energy gap is immune to the magnetic field in analogy to the $n=0$ Landau level of bulk graphene. An analytical solution for zero-energy states in the magnetic field was derived. The energy gap was shown to close with increasing magnetic field, reaching zero at special values of the magnetic field. The gap closing was found independent of quantum dot size, shape, and edge termination.
ACKNOWLEDGMENT {#acknowledgment .unnumbered}
==============
The authors thank NSERC, the Canadian Institute for Advanced Research and TÜBITAK for support. P.P thanks for fellowship within “Mistrz” program from The Foundation for Polish Science.
[29]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, , , , , , , , ****, (). , , , , , , , , ****, (). , , , , ****, (). , , , ****, ().
, , , , , ****, ().
, ****, (). , , , ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , ****, (). , , , , , **** , ().
, , , , , **** , ().
, , , , , ****, ().
, , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, ****, ().
, , , , , , , , , , , , ****, ().
, , , , , ****, ().
, , , , , , , ****, ().
, ****, ().
, , , , , , , , , ****, ().
, , , , , , , , ****, ().
, , , , , , ****, ().
, ****, ().
, , , , , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , , ****, ().
X. Chen, S. Liu, L. Liu, X. Liu, X. Liu, and L. Wang, [Appl. Phys. Lett.]{} [**100**]{}, 163106 (2010).
, , , ****, ().
, , , **** , ().
, , , **** , ().
, , , , ****, (). , , , , ****, ().
, , , ****, ().
, **** , (). , **** , (). , ****, (). , , , **** , (). , , , , ****, ().
M. Wimmer, A. R. Akhmerov and F. Guinea, [*Phys. Rev. B*]{} [**82**]{} 045409 (2010).
, , , **** , ().
O. Voznyy, A. D. Güçlü, P. Potasz, and P. Hawrylak, [Phys. Rev. B]{} [**83**]{}, 165417 (2011).
M. Ezawa, Physica E [**42**]{}, 703 (2010).
, , , ****, (). , , , , , **** , ().
O. V. Yazyev, Rep. Prog. Phys. [**73**]{}, 056501 (2010).
, , , ****, ().
, , , , ****, ().
, , , , , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , , , ****, ().
, , , , ****, ().
, , , **** , ().
, , , **** , ().
, , , **** , ().
, , , .
, ****, ().
, , , ****, ().
J. W. McClure, [Phys. Rev.]{} [**104**]{}, 666 (1956).
S. Raymond, S. Studenikin, A. Sachrajda, Z. Wasilewski, S. J. Cheng, W. Sheng, P. Hawrylak, A. Babinski, M. Potemski, G. Ortner, and M. Bayer, [Phys. Rev. Lett.]{} [**92**]{}, 187402 (2004).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, a complete description of the channels $B \rightarrow V_1 V_2$ is given. Emphasis is put on the determination of the dynamical density matrix which elements are computed according to the Wilson operator product expansions entering into the formulation of the weak effective hamiltonian. 0.3cm Kinematical consequences related to the particular channel $B \rightarrow K^* {\rho}^0 (\omega)$ are described in details.'
---
--------------------
LHCb-PHYS-2001-041
PHYSICS
ADP-01-11/T446
PCCF-RI-01-04
--------------------
0.5 true cm
[[**$B^{0({\pm})}$ decays into two vector mesons** ]{}]{}\
1.0 cm [[**Physical motivations and a general method for simulations** ]{}]{} 1 true cm [ Z.J.Ajaltouni$^1$[^1], O.Leitner$^{1,2}$[^2], C.Rimbault$^1$[^3]]{}\
1 true cm
$~$
Introduction
============
In a previous note [@ZJA], an exhaustive study of the channel simulations:
$$B \to V_1 V_2, \ \ {\gamma} V, \ \ P V, \ \ P P,$$
($V = 1^-, P = 0^-)$ has been performed by stressing the helicity formalism and its consequences. General formulas have been established, notably those giving the final angular distributions in the case of the production of two vector mesons decaying into pseudoscalar mesons.
The squared modulus of the decay amplitude has the following form:
$$\begin{aligned}
|A|^2 \propto h_{\lambda,{\lambda}'} F_{\lambda,{\lambda}'}({\theta}_1) G_{\lambda,{\lambda}'}
({\theta}_2, \phi),\end{aligned}$$
where (summation over $\lambda ({\lambda}')$ is omitted):
- $h_{\lambda,{\lambda}'}$ is the matrix density element constructed from the weak effective hamiltonian $H_w^{eff}$ taken between the initial state ($B_0$) and the final state $f$.
- $F_{\lambda,{\lambda}'}({\theta}_1)$ and $G_{\lambda,{\lambda}'}({\theta}_2, \phi)$ are the matrix elements related to the decays $V_1 \rightarrow a_1 + b_1$ and $V_2 \rightarrow a_2 + b_2 $ respectively.
- ${\theta}_j$ is the polar angle of particle $a_j$ in the rest frame of the resonance $V_j$ while $\phi$ is the angular difference ${\phi}_2 - {\phi}_1$ , where ${\phi}_j$ is the polar angle of $a_j$ in $V_j$ rest frame.
$\lambda ({\lambda}')$ being the helicity state of the vector mesons; $ \lambda = -1, 0, +1$. 0.5cm
As it can be noticed, the essential parameters for the determination of the decay dynamics are the [*unknown matrix elements*]{} $h_{\lambda,{\lambda}'}$; while the two other ones, $F_{\lambda,{\lambda}'}({\theta}_1)$ and $G_{\lambda,{\lambda}'}({\theta}_2, \phi)$, are kinematic (or geometric) parameters because they are completely determined from the Wigner rotation matrices. The reader is referred to the note $99-051$ for a full kinematic description of the $B^0$ decay and the physical significance of the angles ${\theta}_{1,2}$ and $\phi$.
Before dealing with the mathematical determinations of the $h_{\lambda,{\lambda}'}$ elements, a simple justification of the two vector meson channel is given below.
Quantum numbers of the ${V_1}^0 {V_2}^0$ system
===============================================
In the case of two vector meson $B^0$ decay, the most interesting case is the one related to neutral mesons supplemented by the condition $ C|{{V^0}_i} \rangle = - |{{V^0}_i} \rangle $, where $C$ is the charge conjugation operator and $ {V^0}_i$ is a neutral vector meson [*eigenstate*]{} of $C$. Some examples of these channels are:
$${\rho}^0 {\rho}^0, \ \ J/{\Psi}{\rho}^0, \ \ J/{\Psi} {\Phi}, \ \ {\Phi}{\Phi} \dots$$
These vector mesons have, in addition, the [*parity*]{} quantum number equal to $-1$. Noticing that the total angular momentum of the $V^0_1 V^0_2$ system: $ \vec J = {\vec \ell} + {\vec S} = {\vec {s_B}}$ is equal to zero and because the total spin $ \vec S = {\vec {s_1}} + {\vec {s_2}} $, with $ s_1 = s_2 = 1$, the orbital angular momentum can have three different values: $ \ell = S = 0, 1, 2 $. 0.3cm
Thus, parity, charge conjugation and CP quantum numbers of the $ V^0_1 V^0_2 $ system can be computed:
$$P(V^0_1 V^0_2) = {(-1)^2} {(-1)^{\ell}}, \ \ \ C(V^0_1 V^0_2) = (-1)^2,$$
$$\Downarrow$$
$$CP(V^0_1 V^0_2) \ \ = \ \ {(-1)^{\ell}}.$$
0.5cm
We are led to the important result that the CP value of $V^0_1 V^0_2$ is a [**mixing**]{} of two different eigenvalues $+1$ and $-1$ whatever the initial state ($ B^0$ or ${\bar {B^0}} $) is. A direct consequence of this result is that [*CP symmetry is not an exact one*]{}.
0.3cm
The above relation does not hold for reactions involving a neutral $K^*$ like:
$$B^0_d \to K^{*0} {\rho}^0, \ \ \ J/{\Psi} K^{*0}\dots$$
because $K^{*0}$ and ${\bar K}^{*0}$ are two [*distinct particles*]{}; $ C|K^{*0} \rangle = |\bar {K^{*0}} \rangle \neq |K^{*0} \rangle $.
However, it is worth noticing two interesting features for channels with an intermediate resonance like $K^{*0} (\bar {K^{*0}})$:
$$K^{*0} \to K^+ {\pi}^-, \ \ \ \ K^0 {\pi}^0,$$
$$\bar {K^{*0}} \to K^- {\pi}^+, \ \ \ \ \bar {K^0} {\pi}^0.$$
The decay channels are in the ratio ${2/3}$ and ${1/3}$ respectively. On one hand, the sign of the charged kaon shows clearly the nature of the neutral $K^*$ from which it comes and consequently the [*flavour*]{} of the original $B^0 (\bar {B^0})$. So, a neutral $K^*$ decay is a direct way for [*$B^0$ flavour tagging*]{}. 0.5cm
On the other hand, when a neutral $K^{*0} (\bar {K^{*0}})$ decays into $K^0 (\bar {K^0}) {\pi}^0$, the neutral kaon $K^0 (\bar {K^0})$ is not the true physical particle, because approximately $50\%$ of the $K^0 (\bar {K^0})$ go into $K^0_S$ and $50\%$ into $K^0_L$ respectively and the true [*detectable*]{} particle is $K^0_S$ which goes to ${\pi}^+ {\pi}^-$. 0.3cm Thus, in the special channel:
$$B^0 ( \bar {B^0} ) \to K^{*0} (\bar {K^{*0}}) {\rho}^0,$$ $$\ \ \ \ \ \ \ \ \ \ \ \ \ \rightarrow K^0_S {\pi}^0,$$
tagging the original $B^0$ is no longer possible but, the $K^0_S {\pi}^0$ being a [*common final state*]{} to both $B^0$ and $\bar {B^0}$, the above relation $ CP = (-1)^{\ell}$ is still available [@RF].
In the following, emphasis will be put on the channels $K^{*0({\pm})}{\rho}^0{(\omega)}$ and the physical importance of the ${\rho}^0{(\omega)}$ mixing for the determination of [**CP**]{} violation.
${\rho}^0{(\omega)}$ mixing and its consequence
===============================================
It is well known from hadronic physics that the neutral isovector ${\rho}_8$ and the isosinglet ${\omega}_8$ mix together, leading to the “true” physical resonances ${\rho}^0$ and $\omega$. On the phenomenological level, this mixing is made possible because of the existence of a common final state to both ${\rho}^0$ and $\omega$ decays [@PDG]:
$${\rho}^0 \rightarrow {\pi}^+ {\pi}^-, \ \ \ \ \ (BR \approx 100\% ),$$ $${\omega} \rightarrow {\pi}^+ {\pi}^-, \ \ \ \ \ (BR \approx 2.2\% ).$$
In the same framework, it has been established that the $\pi \pi$ final state interaction provides a phase shift $\delta$ which reaches $90^{\circ}$ when the $\pi \pi$ invariant mass is at the $\omega$ pole ($M_{\omega} = 782 \ MeV$) [@CONNELL]. 0.5cm
This interesting physical property has important consequences in the case where a ${\rho^{0}}$ resonance is produced in some $B^{0 {\pm}}$ decays like:
$$B^0 \to K^{*0} {\rho}^0, \ \ \ \ (Fig.1)$$
$$B^+ \to K^{*+} {\rho}^0, \ \ B^- \to K^{*-} {\rho}^0, \ \ \ (Fig.2)$$
These decays require both tree (T) and penguin (P) diagrams. As it is emphasized in reference [@AWTHOMAS], the amplitude $ A$ and $\bar A$ respectively for $B^+$ and $B^-$ decays can be set in the following form: $$\begin{aligned}
A = A^T + A^P = A^T{\left( 1 + r \exp{(i {\delta})} \exp{(i {\phi})} \right)},\end{aligned}$$
$$\begin{aligned}
{\bar A} = {\bar A}^T + {\bar A}^P = A^T{\left( 1 + r \exp{(i{\delta})} \exp{(-i{\phi})}\right)},\end{aligned}$$
where: $$\begin{aligned}
r=\left|{\frac {A^P}{A^T}}\right|,\end{aligned}$$
$$\begin{aligned}
{\bar A}^T = A^T, \ \ {\bar A}^P = {|A^P|} {\exp{(i{\delta})} \exp{(-i{\phi})}}. \end{aligned}$$
0.3cm
Expressions of $A$ and $\bar A$ displayed above suppose that final state interactions (FSI) arise essentially from the penguin diagrams; this hypothesis is supported by the fact that, to order $G_F {\alpha}_s$ ($G_F$ and ${\alpha}_s$ are respectively the Fermi constant and the QCD fine-structure constant), the [*absorptive*]{} part of the transition amplitude is obtained from the penguin diagrams [@BANDERLYPKIN].
In the special case of ${\rho}^0 - \omega$ mixing, another hypothesis is made using more intuitive arguments: the phase shift due to the mixing is included in the FSI and it is [*predominating*]{} at the $\omega$ pole, justifying the above expressions of $A$ and $\bar A$ that the phase shift $\delta$ is principally the one generated by the ${\rho}^0 - \omega$ mixing.
By $CP$ transformation, the strong phase $\delta$ remains unchanged while the weak phase $\phi$, which is related to the CKM matrix elements, changes sign. Thus, the asymmetry parameter $a_{CP}^{dir}$ which can reveal [*direct CP violation*]{} can be deduced in the following way:
$$\begin{aligned}
a_{CP}^{dir} = {\frac {A^2 -{\bar A}^2}{A^2 +{\bar A}^2}} = \frac{-2 \ {\sin {\delta}} \ {\sin {\phi}}}{1 + r^2
+2r \ {\cos {\delta}} \ {\cos {\phi}}}. \end{aligned}$$
It is straightforward to notice that the parameter $a_{CP}^{dir}$ depends both on the strong phase [*and*]{} the weak phase and, consequently, the maximum value of $a_{CP}^{dir}$ can be reached if $\sin {\delta} = 1$, which allows us to state that the strong final state interaction (FSI) among pions coming from the ${\rho}^0 - \omega$ decays [*enhances*]{} the direct $CP$ violation in the vicinity of the resonance $\omega$ mass.
### Simulation of the ${\rho}^0 - \omega$ mixing {#simulation-of-the-rho0---omega-mixing .unnumbered}
A simple and phenomenological relation describing the amplitude of the ${\rho}^0 - \omega$ mixing is used for the Monte-carlo simulations [@LANGACKER]. In the ${\rho}^0$ Breit-Wigner, the (${\rho}^0$) propagator is replaced by the following one:
$$\begin{aligned}
{\cal A} = {\frac{1}{s_{\rho}}}+{\frac{T_{\omega}}{T_ {\rho}}} {\frac{{\Pi}_{\rho \omega}}{{s_{\rho}}{s_{\omega}}}},\end{aligned}$$
where
- $1/s_V = 1/(s - {M_V}^2 + i {{\Gamma}_V}{M_V}) $ is the $V$ resonance propagator, $M_V$ and ${\Gamma}_V$ being respectively the mass and the width of the resonance $V$.
- $T_{\omega}$ and $T_{\rho}$ are respectively the $\omega$ and $\rho$ production amplitudes.
- ${\Pi}_{\rho \omega}$ is the mixing parameter for which recent values come from $e^+ e^-$ annihilations:
$ \Re e {({\Pi}_{\rho \omega})} = -3500 \pm 300 \ \ MeV^2 $ and $ \Im m{({\Pi}_{\rho \omega})} = -300 \pm 300 \ \ MeV^2.$
0.5cm
Due to the same physical processes which enter into the production of the ${\rho}^0$ and $\omega$ resonances (they are both made out from $u {\bar u}$ and $d {\bar d}$ quark pairs with the same weight $1/2$), it seems natural to choose ${T_{\omega}}/{T_ {\rho}} = 1 $. So, the squared mass distribution of the $\pi \pi$ system becomes simplified and it is given by:
$${d\sigma}/{dm^2} \propto {|{\cal A}({{\rho^{0}} ({\omega})})|}^2,$$
where ${\cal A}$ is the amplitude of the two Breit-Wigner given above and $m$ is the $\pi \pi$ invariant mass.
In Figure $3$, are displayed the $\pi \pi$ invariant mass spectra for the $\rho^{0}$ Breit-Wigner and the ${\rho}^0 - {\omega}$ mixing respectively. Because of the very narrow $\omega$ width (${\Gamma}_{\omega} = 8.44 \ MeV $), we notice a high and narrow peak at the $\omega$ pole ($\approx 782 \ MeV $).
Dynamics of the $B \rightarrow V_1 V_2 $ decay
==============================================
The formalism describing the $B^{0 {(\pm)}}$ decay into two vector mesons is derived from the general formalism related to the hadronic weak decay of a heavy meson (or heavy quark). It is based on the new concepts introduced by the Heavy Quark Effective Theory ([**HQET**]{}) which involves additionnal symmetry due to the high mass of the heavy quark ($b$ or $c$ quark) [@NEUBERT]. Technical calculations require a weak effective hamiltonian, $H_w^{eff}$, by using the “Operator Product Expansions” ([**OPE**]{}) pioneered by Wilson and which involve field operators describing both [*tree*]{} and [*penguin*]{} diagrams, the last ones include both QCD and electroweak penguins (Figures 1 and 2).
The general form of $H_w^{eff}$ is given by:
$${H_w}^{eff} = {\frac {G_F}{\sqrt 2}} {{\sum}_{q=d,s} \Big(V_{ub}V_{uq}^*{(c_1
O_1 +c_2 O_2)}
-V_{tb}V_{tq}^* {{\sum}_{i=3}^{10} {c_iO_i}} \Big)},$$
where $c_i$ are the Wilson coefficients and $O_i$ are field operators with dimension $d \geq 4$; they are computed at an energy scale $\mu$ which is identified, here, with the $b$ quark mass $m_b$.
In the case of charmless $B$ decays, Wilson coefficients have been calculated by Buchalla et al [@BUCHALLA]. These coefficients represent the [*perturbative part*]{} of the weak hamiltonian, they are estimated by the Renormalization Group techniques and their values depend on the renormalization scheme which is used. Their physical significance is the [*weight*]{} of each field operator $O_i{(\mu)}$ entering in the weak hamiltonian $H_w^{eff}$. From reference [@DESHPANDE], the values of $c_i$ which have been computed at the energy scale $\mu = m_b $ are:
$$\begin{aligned}
\begin{array}{ll}
\vspace{0.5em}
c_{1}=-0.3125,\;\;\;\; c_{2}=1.1502, \\
\vspace{0.5em}
c_{3}=0.0174 ,\;\;\;\; c_{4}=-0.0373, \\
\vspace{0.5em}
c_{5}=0.0104 ,\;\;\;\; c_{6}=-0.0459, \\
\vspace{0.5em}
c_{7}=-1.050 \times 10^{-5}, \;\;\;\; c_{8}=3.839 \times 10^{-4},
\end{array}\end{aligned}$$ $$\begin{aligned}
\!\!\!\!\!\!\!\! \!\!\!\!\!\!\!c_{9}=-0.0101, \;\;\;\;
c_{10}=1.959 \times 10^{-3}. & &
$$
The first two coefficients, $c_1$ and $c_2$, are related to the [*tree*]{} diagrams and they show clearly their dominance with respect to the [*penguin*]{} ones. Coefficients $c_3 - c_6$ correspond to QCD penguin operators while $ c_7 - c_{10}$ are related to the EW ones.
However, those values of $c_i$ must be modified when renormalization of operator $O_i$ at one-loop order is taken into account.
0.3cm
Detailed expressions of operators $O_i{(\mu)}$ and their physical interpretation are given in reference [@IJMP].
0.3cm
Thus, a general form for the weak decay amplitude into a final state $f$ can be expressed like:
$$A(B^0 \rightarrow f) = \langle f|H_w^{eff}|B^0 \rangle = {\frac {G_F}{\sqrt 2}} {\sum_{i=1}^{10}}{\sum_{q= d,s}}
{{\lambda}_q}^i {c_i({\mu})}{\langle f|O_i({\mu})|B^0 \rangle},$$
where ${{\lambda}_q}^i$ is the product of two CKM matrix elements: $V_{ub}V_{uq}^*$ (for $i=1,2$) or ${(V_{tb}V_{tq}^*)}$ (for $i=3 ,\dots, 10$). 0.4cm
The hadronic matrix elements $ \langle f|O_i({\mu})|B^0 \rangle$ represent the [*non-perturbative contribution*]{} to the amplitude $A(B^0 \rightarrow f)$. Usually, they are estimated according to some specific models: Non Relativistic Quark Model (NRQM), Form Factor models (BSW) and especially the Lattice QCD calculations.
In the following, calculation of the hadronic matrix elements is performed in the framework of the BSW model [@BSW] from which form factors are derived by the knowledge of the hadronic wave functions for both initial and final states.
Determination of the density-matrix elements
============================================
The $B^0$ decay into two vector mesons requires the helicity formalism which has been intensively used in the previous paper [@ZJA]. To each vector meson (spin $1$) is assigned a set of three polarization 4-vectors defined in this way:
$${\epsilon}_1 = (0, \vec {{\epsilon}_1}), \ \ {\epsilon}_2 = (0, \vec {{\epsilon}_2}), \ \ {\epsilon}_3
={\left(|{\vec k}|/m , E {\hat {k}}/m \right)},$$
and verifying the following relations: $${{\epsilon}_i}^2 = -1, \ \ {\epsilon}_i \cdot {\epsilon}_j = 0, \ \ \mathrm{with} \ i \neq j,$$
where $m, E, \vec k$ are respectively the mass, the energy and the momentum of the vector meson; $\hat {k}$ is defined as the unit vector along the vector momentum, $\hat {k} = {\vec k}/{|\vec k|}$. 0.5 cm The three vectors $\vec {{\epsilon}_1}, \vec {{\epsilon}_2}$ and $\vec {{\epsilon}_3} = {E {\hat {k}}}/m$ form an orthogonal basis; ${\epsilon}_1$ and ${\epsilon}_2$ are called the [*transverse polarization*]{} vectors while $\vec {{\epsilon}_3}$ is the [*longitudinal polarization*]{} one.
From that basis, an [*helicity basis*]{} is defined according to:
$${\epsilon}(+) = \frac{\left({\epsilon}_1 + i {\epsilon}_2 \right)}{\sqrt 2}, \ \ {\epsilon}(-) =
\frac{\left({\epsilon}_1 - i {\epsilon}_2 \right)}{\sqrt 2}, \ \ {\epsilon}(0) = {\epsilon}_3.$$
These 4-vectors are [*eigenvectors*]{} of the helicity operator $\mathcal {H}$ with the eigenvalues $ \lambda = +1, -1$ and $0$ respectively. For a clear account of the helicity basis for a spin $1$ particle, the reader can consult the book of Dewitt-Smith [@DEWITT]. 0.5 cm
In the case of two vector mesons coming from the $B$ decay, their 4-momenta are defined in the $B$ rest frame and their corresponding polarization vectors are [*correlated*]{} because $\hat {k_1} = - \hat {k_2}$. For an explicit calculation of their spatial components, see the appendix A. 0.5cm
The weak hadronic amplitude is then decomposed on the helicity basis according to the general formalism developed by the authors BSW [@BSW]. This method allows one to obtain two interesting results:
$\bullet$ the contribution of the [*tree*]{} and [*penguin*]{} operators to the global amplitude via the helicity states.
$\bullet$ the total contribution of each helicity state.\
A way of illustrating this method is to study the channel: $B^0{(\bar B^0)} \rightarrow K^{*0} {(\bar K^{*0})} {\rho}^0$.
0.3cm
$(i)$ First of all, the mass of each resonance ($K^{*0}$ and ${\rho}^0$ ) is generated according to a relativistic Breit-Wigner:
$$\frac{d\sigma}{dM^2} \ \ = \ \ C \frac{\Gamma_R M_R}{{(M^2-{M^2_R})}^2 + {(\Gamma_R M_R)}^2},$$
$C$ being a normalization constant.
0.6cm
$(ii)$ The weak hadronic matrix element is expressed as the [**sum**]{} of three helicity matrix elements; each one of the form, $ H_{\lambda} = \langle V_1 V_2 | {H_w}^{eff} | B \rangle $, is defined by gathering all the Wilson coefficients of both tree and penguin operators. Linear combinations of those coefficients arise like: $c_{t1}^{\rho}, \ c_{p1}^{\rho}, \ $ and $c_{p2}^{\rho}$ (see Appendix B) and the [*helicity amplitude*]{} $H_ {\lambda}$ gets the following expression:
0.3cm
$$H_{\lambda} =\Big(V_{ub}V_{us}^{*}c_{t_{1}}^{\rho}-V_{tb}V_{ts}^{*}c_{p_{2}}^{\rho}\Big)\bigg\lbrace \beta_{1}\varepsilon_{\alpha \beta \gamma \delta}\epsilon_{K}^{*\alpha}(\lambda)\epsilon_{\rho}^{*\beta}(\lambda)P_{B}^{\gamma}P_{K}^{\delta}$$ $$+i\Big(\beta_{2}\epsilon_{K}^{*}(\lambda)\epsilon_{\rho}^{*}(\lambda) - \beta_{3}(\epsilon_{K}^{*}(\lambda).P_{B})(\epsilon_{\rho}^{*}(\lambda).P_{B})\Big)\bigg\rbrace$$ $$+ \Big(-V_{tb}V_{ts}^{*}c_{p_{1}}^{\rho}\Big)\bigg\lbrace \beta_{4}\varepsilon_{\alpha \beta \gamma \delta}\epsilon_{\rho}^{*\alpha}(\lambda)\epsilon_{K}^{*\beta}(\lambda)P_{B}^{\gamma}P_{\rho}^{\delta},$$ $$+i\Big(\beta_{5}\epsilon_{\rho}^{*}(\lambda)\epsilon_{K}^{*}(\lambda) - \beta_{6}(\epsilon_{\rho}^{*}(\lambda).P_{B})(\epsilon_{K}^{*}(\lambda).P_{B})\Big)\bigg\rbrace$$
0.5cm
with:
$\bullet$ $\varepsilon_{\alpha \beta \gamma \delta}$: antisymmetric tensor in the Minkowski space.\
$\bullet$ $\beta_{1,4}=\frac{G_{F}}{2} f_{\rho,{K}}m_{\rho,{K^*}}\frac{2}{m_{B}+m_{K^*,{\rho}}}V^{B\to K^*,{\rho}}(m^{2}_{\rho,{K^*}})$.\
$\bullet$ $\beta_{2,5}=\frac{G_{F}}{2} f_{\rho, K}m_{\rho, K^*}(m_{B}+m_{K^*,{\rho}})A_{1}^{B\to K^*,{\rho}}(m^{2}_{\rho, K^*})$.\
$\bullet$ $\beta_{3,6}=\frac{G_{F}}{2} f_{\rho, K}m_{\rho, K^*}\frac{2}{m_{B}+m_{K^*,{\rho}}}A_{2}^{B\to K^*,{\rho}}(m^{2}_{\rho, K^*})$.\
$\bullet$ $f_{K}$, $f_{\rho}$: respectively $K^{*0}$ and $\rho^{0}$ decay constants.\
$\bullet$ $V^{B\to K^*,{\rho}}$, $A_{i}^{B \to K^*,\rho}$: respectively Vector and Axial form factors (see Appendix C).\
$\bullet$ $\epsilon_{K, \rho}(\lambda)$: $K^{*0}$, $\rho^{0}$ polarization vectors expressed in the $B$ rest frame.
0.5cm
It is worth noticing that the tensorial terms which enter $H_{\lambda}$ become simplified in the $B$ rest frame because the $B$ 4-momentum is given by $P_B = {(m_b, {\vec 0})}$. Then, using the orthogonality properties of ${\epsilon}_j{(\lambda)}$, the helicity amplitude $H_{\lambda}$ acquires a much simpler expression than above:
$$H(\lambda) = iB(\lambda)(V_{ub}V_{us}^{*}c_{t_1}^{\rho}-V_{tb}V_{ts}^{*}c_{p_2}^{\rho})+
iC(\lambda)(-V_{tb}V_{ts}^{*}c_{p_1}^{\rho}),$$
0.6cm
with:
$$B(0)=\beta_{2}\frac{m_{B}^2 -(m_{K}^2 + m_{\rho}^2 )}{2m_{K}m_{\rho}} - \beta_{3}\frac{|\vec{p}|^{2}m_{B}^2 }{m_{K}m_{\rho}},$$
$$C(0)=\beta_{5}\frac{m_{B}^2 -(m_{K}^2 + m_{\rho}^2 )}{2m_{K}m_{\rho}} - \beta_{6}\frac{|\vec{p}|^{2}m_{B}^2}{m_{K}m_{\rho}},$$
$$B(\pm{1}) = \mp \beta_{1}m_{B}|\vec{p}| - \beta_{2},$$
$$C(\pm{1}) = \mp \beta_{4}m_{B}|\vec{p}| - \beta_{5},$$
$|{\vec{p}}|$ being the common momentum to $V_1$ and $V_2$ particles in the $B$ rest frame.
0.6 cm
$(iii)$ Expressing the CKM matrix elements according to Wolfenstein parametrization [@EPJC15]:
$$V_{CKM} = \left[
\begin{array}{ccc}
1-\frac{\lambda^2}{2} \ & \ \lambda \ & \
A\lambda^3(\rho-i\eta) \\
-\lambda \ & \ 1-\frac{\lambda^2}{2}\ & \ A\lambda^2
\\
A\lambda^3(1-\rho-i\eta)\ & \ -A\lambda^2 \ & \ 1 \\
\end{array}
\right] \ \ \ \ \ + \ \ O{({\lambda}^4),}$$
0.6cm
where we use[@ACHILLE]:
[ $A = 0.815$, $\lambda = 0.2205$: well known\
[**$0.09<\rho<0.254$, $0.323<\eta<0.442$.**]{}\
]{}
0.6cm
Taking into account the preceding relations, we arrive at the final form for the amplitudes $H_{\lambda}$:
[ $$H {0 \choose \pm{1}} = A\lambda^{2} \Bigg \lbrace \bigg \lbrack \Big(\eta \lambda^2 c_{t_1}^{\rho} - \Im{m(c_{p_2}^{\rho})}\Big)B {0 \choose \pm{1}} - \Im{m(c_{p_1}^{\rho})}C{0 \choose \pm{1}}\bigg \rbrack$$ $$+ \ i\ \bigg \lbrack \Big( \rho \lambda^{2}c_{t_1}^{\rho} + \Re{e(c_{p_2}^{\rho})}\Big)B{0 \choose\pm{1}} + \Re{e(c_{p_1}^{\rho})}C{0 \choose \pm{1}} \bigg \rbrack \Bigg \rbrace,$$ ]{}
from which the density-matrix elements $h_{{\lambda}, {\lambda}'}$ can be derived automatically;
$\Rightarrow h_{{\lambda}, {\lambda}'} = H_{\lambda} {H_{{\lambda}'}}^* $.
0.4cm
Due to the [*hermiticity*]{} of the matrix ($h_{{\lambda}, {\lambda}'}$), only six elements must be calculated and, furthermore, a normalization condition is applied: $$N \ {\left(h_{++} + h_{00} + h_{--}\right)} \ \ = \ \ 1,$$
($N$ being the normalization constant) which makes easier the comparison of the modulus of the different matrix elements.
0.6 cm
In the next histograms (Fig.4 - Fig.9) are displayed the spectra of $h_{{\lambda}, {\lambda}'}$ for different values of the Wolfenstein parameters $\rho$ and $\eta$. In our study, these spectra are obtained for the four couples of values: $(0.09, 0.323)$; $(0.09, 0.442)$; $(0.254, 0.323)$ and $(0.254, 0.442)$. But, due to the fact that some density matrix elements do not vary too much with $\rho$ and $\eta$, in most cases only the spectra corresponding to the first couple of values are shown. All the histograms correspond to a sample of $20000$ generated events.
0.4cm It is important to notice that large spectrum of values for $h_{{\lambda}, {\lambda}'}$ are obtained and [*not single ones*]{} because of the broad range of both the ${\rho}^0$ resonance mass and the common momentum $|{\vec p}|$ (see the analytical expressions of $B({\lambda})$ and $C({\lambda})$ given above).
- Whatever the values of $\rho$ and $\eta$ are, the dominant value of $h_{++} = {|H_{+1}|}^2$ is $\leq 10^{-2}$, numerical result which is proved too by complete analytical calculations. Thus the dominant polarization state is the [**longitudinal**]{} one because $h_{00} = {|H_{0}|}^2 \ \ \geq 60\%$, its mean value being around $85\%$ (Fig.4).
- Due to the tiny value of $|H_{+1}|$, the modulus of the non-diagonal elements $h_{+-} = {H_+}{H_-}^*$ and $h_{+0} = {H_+}{H_0}^*$ are usually smaller than $0.2$; while the modulus of $h_{-0} = {H_-}{H_0}^*$ can reach $0.5$ (Fig.5).
Fig.6 and Fig.7 display the variations of the diagonal matrix elements $h_{--}, h_{00}$ and $h_{++}$ with respect to the four sets of $\rho$ and $\eta$ values: it can be seen that $h_{++}$ has always a tiny value and $ h_{00}$ is always dominant. Other physical features appear: $h_{00}$ is very [*sensitive*]{} to the parameter $\eta$; its spectrum is rather wide for ${\eta} = 0.323$, while it is bounded between $0.8$ and $1.0$ for ${\eta} = 0.442$. For a fixed value of $\eta$, no noticeable variation with the parameter $\rho$ is seen. 0.3 cm Fig.8 shows the real and imaginary parts of the non-diagonal elements $h_{-0}, h_{+-}, h_{+0}$ respectively for $\rho = 0.09$ and $ \eta = 0.323$. It is worth noticing that both real and imaginary parts of $h_{+-}$ and $h_{+0}$ are too small and close to zero. 0.3 cm Due to the importance of $h_{+-}$ matrix element in the $\phi$ angle distribution (see Section 6), a full study of both real and imaginary parts of $h_{+-}$ has been done. Fig.9 shows the corresponding spectra according to the values of $\rho$ and $\eta$. It can be deduced that the real and imaginary parts have very similar distributions and both are dominated by small values ($\leq 0.05$).
Decays of vector mesons $V_1 \ V_2$ into two pseudoscalar mesons
=================================================================
The matrix elements derived above allow us to compute the [*degrees of polarization*]{} of each resonance like $K^*$ or ${\rho}^0$. The angular distributions of the pseudoscalar mesons in each $V_i$ rest frame depend on:\
$(i)$ the spin $1$ of the vector meson $V_i$.\
$(ii)$ the weight of each helicity state.\
$(iii)$ the correlations among the helicity states of the two vector mesons.
0.5cm
Complete analytical expression of the final angular distributions is the following one:
$$\begin{aligned}
{{d^3\Gamma}\over{d\cos\theta_1 d\cos\theta_2 d\phi}} & \propto &
{(h_{++} + h_{--})}{{\sin}^2{\theta_1}{\sin}^2{\theta_2}}/4 + {h_{00}{\cos}^2{\theta_1}{\cos}^2{\theta_2}} \nonumber
\\
& &\!\!\! \!\!\!\!\!\!\!\! \!\!\!\!\! \!\!\!\!\!\!\!\! \!\! \!\!\!\!\!\!\!\! \!\!\!\!\! + \left(\Re e{(h_{+0})}{\cos{\phi}} - \Im m{(h_{+0})}{\sin{\phi}} + \Re e{(h_{0-})}{\cos{\phi}} -
\Im m{(h_{0-})}{\sin{\phi}}\right){{\sin{2\theta_1}}{\sin{2\theta_2}}}/4 \nonumber \\
& & \!\!\!\!\!\! \!\! \!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\! \!\!\!\!\!\!\!\!\!\! \!\!\!\!\! +
\left(\Re e{(h_{+-})}{\cos{2\phi}} - \Im m{(h_{+-})}{\sin{2\phi}}\right){{\sin}^2{\theta_1}{\sin}^2{\theta_2}}/2.\end{aligned}$$
Angles ${\theta}_1$ , ${\theta}_2$ and $\phi$ have been defined in Section $1$. 0.5cm Explicit angular distributions for polar and azimuthal angles can be derived from the relation above. It is interesting to notice that, due to the pseudoscalar nature of the final particles, angles $\theta_1$ and $\theta_2$ have the [*same*]{} distributions:
0.4cm
- ${d{\sigma}}/{d{\cos{\theta_{1, 2}}}} \propto (3h_{00}-1)\cos^{2}{\theta_{1, 2}} + (1-h_{00}).$
- ${d{\sigma}}/{d{\phi}} \propto (1+2(\Re e {(h_{+-})}\cos{2\phi} - \Im m{(h_{+-})}\sin{2\phi})).$
0.4cm
In Fig.10 are displayed respectively the $\cos{\theta}$ distribution and the azimuthal angle $\phi$ one. Some comments on these curves are necessary:
- The $\cos{\theta}$ distribution is practically the same whatever the values of $\rho$ and $\eta$ are; no sensitivity to particular values of $\rho$ and $\eta$ is seen.
- As far as angle $\phi$ is concerned, its distribution depends on the matrix element $h_{+-}$. Despite the fact that $\Re e {(h_{+-})}$ and $\Im m{(h_{+-})}$ do not exhibit sensitive differences (see Fig.9), those parameters present some dependence upon $\rho$ and $\eta$: full curve corresponds to $\rho = 0.09, \eta = 0.323$; while dashed one is related to $\rho = 0.254, \eta = 0.442$. A visible discrepancy among these two curves is seen.
Perspectives and conclusion
===========================
$ \;\;\;\; \bullet $ Thanks to the HQET approach and the OPE formalism which is used, we have at our disposal rigorous and [*complete calculations*]{} of the dynamics of the $B^{0 {\pm}}$ decays into two vector mesons. This formalism is available for all charmless $B$ decays provided the spin of the intermediate resonance(s) is less or equal $1$; the only changes which must be taken into account are the $V_{CKM}$ matrix elements, the masses and the widths of the resonances involved in each decay. $ \bullet $ In the case of leptonic decay of one resonance, like $J/{\Psi} \rightarrow e^+ e^{-}, {\mu}^+
{\mu}^-$, the angular distributions are modified because of the spin $1/2$ final leptons; which require the use of other Wigner rotation matrices. Those calculations have been already done in our first paper [@ZJA]. $ \bullet$ In the case where a $(c {\bar c})$ bound state or a charmed meson is produced like:
$$B^0 \to J/{\Psi} {\rho}^{0}, D^* X (X = {\rho}^{0}, \omega, K^{*0}),$$
the Wilson coefficients involved in the effective hamiltonian have to be modified, but we do not expect big change with respect to the $c_i (c_i')$ coefficients used in the present paper. $ \bullet$ Other interesting consequences arise from this formalism: it can be easily extended to the numerous channels like: $B \to VP, \ PP \ $ where one or two pseudoscalar mesons ($P = 0^{-+}$) are produced directly from the $B$ decay. Because of the simple equality ${\lambda}(P) = {\lambda}(V) = 0$, the number of helicity states is reduced from $3$ to $1$. $ \bullet$ An important point which has been mentionned in the present note is the role of the ${\rho}^0 - {\omega}$ [*mixing*]{} and its consequence for the determination of the direct CPV parameter (Section $3$ and reference [@AWTHOMAS]). Tagging of $B^+$ and $B^-$ is made easy thanks to the $K^+$ and $K^-$ mesons coming from the cascade decays. In our opinion, we can also exploit all the angular distributions of the final particles (and their correlations) in order to detect an eventual discrepancy which can arise between the $B^+$ and $B^-$ decays respectively. However, a complete study of those channels and their simulations require the knowledge of the strong phase shift $\delta$ (mentionned in Section 3) according to the $\pi \pi$ invariant mass. Work is in progress. $ \bullet$ Those calculations and simulations can be implemented into [**SICBMC**]{}, the Monte-Carlo generator of the LHCb experiment, in order to perform afterwards a full analysis of the simulated channels.
### Acknowledgements {#acknowledgements .unnumbered}
The authors are very grateful to Dr P.Perret, leader of the LHCb Clermont-Ferrand team, for his advices and his suggestions.
One of us (Z.J.A.) is very indebted to Professor A.W.Thomas, Director of the Special Research Centre for the Subatomic Structure of Matter, for the very exciting and illuminating discussions he got with him about the QCD penguin diagrams and their importance in the evaluation of the $B^0$ decay width. 0.5cm This work was supported in part by the Australian Research Council and the University of Adelaide.
Appendix {#appendix .unnumbered}
========
Polarizations in $B^{0}_{d}$ rest frame
=======================================
$\;\;\;\;$ Momentum: $$\vec{k}_{K} = -\vec{k}_{\rho} = \vec{k} = \left(
\begin{array}{c}
k\sin\theta\cos\phi\\
k\sin\theta\sin\phi\\
k\cos\theta
\end{array}
\right),$$
where $\theta$ and $\phi$ are respectively polar and azimuthal angles of the produced $K^{*0}$.\
Longitudinal polarization:\
$$\epsilon_{K}(0) = \left(\frac{|\vec{k}|}{m_{K}},\frac{E_K}{m_K}\hat{k}\right), \ \ \ \epsilon_{\rho}(0) = \left(\frac{|\vec{k}|}{m_{\rho}},\frac{E_{\rho}}{m_{\rho}}(-\hat{k})\right).$$\
Tranversal polarizations : $$\vec{\epsilon}_{K}(1) = \left(
\begin{array}{c}
\cos\theta\cos\phi\\
\cos\theta\sin\phi\\
-\sin\theta
\end{array}
\right) = \vec{\epsilon}_{\rho}(1),$$
$$\vec{\epsilon}_{K}(2) = \left(
\begin{array}{c}
-\sin\phi\\
\cos\phi\\
0
\end{array}
\right) = -\vec{\epsilon}_{\rho}(2).$$
Helicity frame :\
$${\epsilon_{K}(+)}= \left(\epsilon(1) + i\epsilon(2)\right)/\sqrt{2}, \ \ \
{\epsilon_{K}(-)}= \left(\epsilon(1) - i\epsilon(2)\right)/\sqrt{2},$$\
$$\vec{\epsilon}_{K}(+) = \left(
\begin{array}{c}
\cos\theta\cos\phi-i\sin\phi\\
\cos\theta\sin\phi+i\cos\phi\\
-\sin\theta
\end{array}
\right) / \sqrt{2} = \vec{\epsilon}_{K}^{\ *}(-) = \vec{\epsilon}_{\rho}(-),$$
$$\vec{\epsilon}_{K}(-) = \left(
\begin{array}{c}
\cos\theta\cos\phi+i\sin\phi\\
\cos\theta\sin\phi-i\cos\phi\\
-\sin\theta
\end{array}
\right) / \sqrt{2} = \vec{\epsilon}_{K}^{\ *}(+) = \vec{\epsilon}_{\rho}(+).$$
Wilson’s coefficients
=====================
We use, in the case of the $\rho^{0}$ production, the following linear combinations of the effective Wilson coefficients:
$$\begin{aligned}
& & \! c_{t1}^{\rho}=c^{\prime}_{1} + \frac{c^{\prime}_{2}}{N_{c}}, \\
& & c_{p1}^{\rho}=-(c^{\prime}_{4} + \frac{c^{\prime}_{3}}{N_{c}})+\frac{1}{2}(c^{\prime}_{10}+\frac{c^{\prime}_9}{N_{c}}), \\
& & c_{p2}^{\rho}=\frac{3}{2}(c^{\prime}_7 + \frac{c^{\prime}_8}{N_{c}}+c^{\prime}_{9}+\frac{c^{\prime}_{10}}{N_{c}}),
$$
where $c_{t1}^{\rho}$ relative to tree diagram, $c_{pi}^{\rho}$ relative to penguin diagram and $0.98 <N_{c} <2.01$.\
When $q^{2}/m^{2}_{b} = 0.3$: $ { \hspace{1.5em}} c^{\prime}_{1} = -0.3125, \ \ \ c^{\prime}_{2} = 1.1502,$\
$ c^{\prime}_{3} = 2.443 \times 10^{-2} + 1.543 \times 10^{-3}i, \ \ \ c^{\prime}_{4} = -5.808 \times 10^{-2} - 4.628 \times 10^{-3}i,$\
$ c^{\prime}_{5} = 1.733 \times 10^{-2} + 1.543 \times 10^{-3}i, \ \ \ c^{\prime}_{6} = -6.668 \times 10^{-2} - 4.628 \times 10^{-3}i,$\
$ c^{\prime}_{7} = -1.435 \times 10^{-4} - 2.963 \times 10^{-5}i, \ \ \ c^{\prime}_{8} = 3.839 \times 10^{-4}, $\
$ c^{\prime}_{9} = -1.023 \times 10^{-2} - 2.963 \times 10^{-5}i, \ \ \ c^{\prime}_{10} = 1.959 \times 10^{-3}.$
When $q^2/m^2_b = 0.5$: $ { \hspace{1.5em}} c^{\prime}_{1} = -0.3125, \ \ \ c^{\prime}_{2} = 1.1502,$\
$ c^{\prime}_{3} = 2.120 \times 10^{-2} + 2.174 \times 10^{-3}i,\ \ \ c^{\prime}_{4} = -4.869 \times 10^{-2} - 1.552 \times 10^{-2}i,$\
$ c^{\prime}_{5} = 1.420 \times 10^{-2} + 5.174 \times 10^{-3}i,\ \ \ c^{\prime}_{6} = -5.729 \times 10^{-2} - 1.552 \times 10^{-2}i,$\
$ c^{\prime}_{7} = -8.340 \times 10^{-5} - 9.938 \times 10^{-5}i,\ \ \ c^{\prime}_{8} = 3.839 \times 10^{-4},$\
$ c^{\prime}_{9} = -1.017 \times 10^{-2} - 9.938 \times 10^{-5}i,\ \ \ c^{\prime}_{10} = 1.959 \times 10^{-3}.$
Form factors (BSW model)
========================
$ $ $V$ $A_1$ $A_2$
-------------- -------------------------------------------------------- -------------------------------------------------------- --------------------------------------------------------
$B\to K^* $ $\frac{0.369}{1-{m^2_{\rho}(GeV^2)}/{5.43^2(GeV^2)}} $ $\frac{0.328}{1-{m^2_{\rho}(GeV^2)}/{5.43^2}(GeV^2)} $ $\frac{0.331}{1-{m^2_{\rho}(GeV^2)}/{5.43^2}(GeV^2)} $
$B\to \rho $ $\frac{0.329}{1-{m^2_{K^*}(GeV^2)}/{5.32^2}(GeV^2)} $ $\frac{0.283}{1-{m^2_{K^*}(GeV^2)}/{5.32^2}(GeV^2)}$ $\frac{0.283}{1-{m^2_{K^*}(GeV^2)}/{5.32^2}(GeV^2)} $
For further details, see reference [@AWTHOMAS] and literature quoted therein.
[99]{}\
Z.J.Ajaltouni et al, LHCb 99-051, PHYSICS, December 1999. , private communication.\
Donoghue et al, Cambridge monographs on Particle Physics (1994).\
A.J.Buras, hep-ph/9806471\
(“Probing the Standard Model of Particle Interactions” , F.David and R.Gupta eds, Elsevier Science). Review of Particle Physics, [**EPJC C15**]{}, 2000.\
S.Gardner et al, Phys.Rev.Lett. [**80**]{}, 1834 (1998) and references therein.\
X.-H.Guo, O.Leitner and A.W.Thomas, Phys.Rev.D[**63**]{} (2001) 056012. Myron Bander et al, Phs.Rev.Lett. [**43**]{} (1979) 242.\
Harry Lipkin, Phy.Lett. [**B 415**]{} (1997) 186. P.Langacker, Phys.Rev.D20, (1979) 2983.\
M.Neubert, Physics Reports [**245**]{} (1994) 259-395.\
G.Buchalla et al, Reviews of Modern Physics, Vol.68, 1125 (October 1996). N.G.Deshpande et al, Phys.Rev.Lett. [**74**]{}, 26 (1995).\
R.Fleischer, International Journal of Modern Physics [**A12**]{}, 2459 (1997). M.Bauer, B.Stech and M.Wirbel, Z.Phys. [**C34**]{}, 103 (1987) ;\
M.Bauer et al, Z.Phys. [**C29**]{}, 637 (1985).\
De Wit and J.Smith, Nort-Holland (1986). Review of Particle Physics, [**EPJC C15**]{} (2000), 110. F.Parodi, P.Roudeau and A.Stocchi, Nuo.Cim. [**A112**]{} (1999) 833.
[^1]: ajaltouni@in2p3.fr
[^2]: oleitner@physics.adelaide.edu.au
[^3]: rimbault@clermont.in2p3.fr
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Contrast enhancement (CE) forensics has always been of concern to image forensics community. It can provide an effective tool for recovering image history and identifying tampered images. Although several CE forensic algorithms have been proposed, their robustness against some processing is still unsatisfactory, such as JPEG compression and anti-forensic attacks. In order to attenuate such deficiency, in this paper we first present a discriminability analysis of CE forensics in pixel and gray level histogram domains. Then, in such two domains, two end-to-end methods based on convolutional neural networks (P-CNN, H-CNN) are proposed to achieve robust CE forensics against pre-JPEG compression and anti-forensics attacks. Experimental results show that the proposed methods achieve much better performance than the state-of-the-art schemes for CE detection in the case of no other operation and comparable performance when pre-JPEG compression and anti-foresics attacks is used.'
address: |
Institute of Information Science, Beijing Jiaotong University, Beijing, China\
Beijing Key Laboratory of Advanced Information Science and Network Technology\
Communication University of China, Beijing, China
title: Robust Contrast Enhancement Forensics Using Pixel and Histogram Domain CNNs
---
Ł[[L]{}]{}
Contrast enhancement forensics, convolutional neural networks, peak/gap, pixel domain, histogram domain.
Introduction {#sec:intro}
============
As a simple yet efficient image processing operation, CE is typically used by malicious image attackers to eliminate inconsistent brightness for generating a visually imperceptible tampered images. CE detection algorithms play an important role in decision analysis for authenticity and integrity of digital images. Although some schemes have been proposed to detect contrast-enhanced images, the performance of such techniques is limited in the cases of pre-JPEG compression and anti-forensic attacks. Therefore, it is critical to develop robust and effective CE forensics algorithms.
With the efforts of researches in the past decade, a number of schemes \[1-9\] have been proposed to discriminate the contrast-enhanced images in uncompressed format. Stamm *et al.* \[1,2,3\] found that contrast enhancement would introduce the peaks and gaps into the image’s gray level histogram, which led to the specific high values in high-frequency components. Lin *et al.* \[6,7\] revealed that contrast enhancement would disturb the inter-channel correlation left by color image interpolation and measured such correlation to distinguish the original and enhanced images. Futhermore, in order to recover the image processing history, the algorithms \[10-13\] of estimating parameters for constrast-enhanced images are developed.
Despite good performance of the above algorithms, their robustness is unsatisfactory in some cases, such as the CE of JPEG images (pre-JPEG compression) and the occurrence of anti-forensic attacks \[14-19\]. The reason lies in that the fingerprint left by CE operation would be destroyed. Based on such a phenomenon, some researchers attempt to propose more robust CE forensic algorithms, which contains two major branches: overcoming pre-JPEG compression \[4\] and defensing anti-forensic attack \[9\]. Unfortunately, such methods can not address well both pre-JPEG compression and anti-forensic attacks. And to date there are no satisfactory solutions.
In this paper, we propose two robust CE detection algorithms based on convolutional neural networks (CNNs) to resist not only pre-JPEG compression but also anti-CE attacks. Firstly, discriminability analysis of CE forensics in pixel and histogram domains is presented. Then, inspired by the excellent performance of deep learning based techniques in various fields, we explore two types of CNNs architectures for CE forensics: pixel-domain CNNs (P-CNN) and histogram-domain CNNs (H-CNN). Especially for P-CNN, high-pass filter is used to reduce the affect of image contents and keep the data distribution balance cooperating with batch normalization \[21\]. Additionally, the width of architecture is experimentally designed to learn better feature representation for CE forensics. Besides, as a lower dimensional yet effective feature, the histogram with 256 dimensions is fed into CNNs for constructing H-CNN. Experimental results show that our proposed methods outperform the state-of-the-arts schemes in the case of uncompression and comparable performance in the cases of pre-JPEG compression, anti-forensics attack, and CE level variation.
Proposed Robust Algorithm for detecting Contrast Enhancement Images
===================================================================
The existing algorithms are not robust against pre-JPEG compression, anti-forensics attack and CE level variation. In this paper, two deep learning-based algorithms, data-driven framework, are proposed to detect contrast-enhanced images by auto-learning effective features from database: P-CNN and H-CNN. Specifically, their architectures are as shown in Fig.1.
Pixel-Domain Convolutional Neural Networks
------------------------------------------
As a common way of contrast enhancement, gamma correction can be found in many image-editing tools. In this paper, we mainly focus on the detection of gamma correlation, which is typically defined as, $$\label{key}
Y=[255(X/255)^\gamma ]\approx 255(T^\gamma)$$ where $X$ denotes an input and $Y$ represents the mapped value, $T=(X/255) \epsilon [0,1]$. In order to simplify the discussion, the mapped value, $Y$, is normalized: $$\label{key}
Z=Y/255\approx T^\gamma$$ where $Z\epsilon [0,1]$. As well known, gamma correction would lead to the nonlinear changes in pixel domain and introduce the peak/gap bins into histogram domain \[1-4\]. A number of handcrafted features are designed based on such phenomenons. In pixel domain, the difference between the original and enhanced images can be computed as follows, and the absolute value of difference is considered. $$\left\{
\begin{array}{lr}
D = Y-X=255(Z - T) \approx 255(T^\gamma-T), \gamma <1 \\
\\
D = X-Y=255(T - Z) \approx 255(T-T^\gamma), \gamma >1
\end{array}
\right.$$
It can be seen from (3) that the discriminability in pixel domain is related with pixel value (image contents), $T$, and parameter of gamma correction, $\gamma$. In order to describe such discriminability, the maximum of difference denoted by $D_{max}$ is considered. $D_{max}$ is obtained when partial derivative of $Z$ with respect to $T$ is equal to $1$. $$\label{key}
T_{D_{max}} = T_{\frac{\partial Z}{\partial T}=1}=(\frac{1}{\gamma })^{(\frac{1}{\gamma -1})}$$
$$\label{key}
D_{max} =
\left\{
\begin{array}{lr}
255[(\frac{1}{\gamma })^{\frac{\gamma }{\gamma -1}} - (\frac{1}{\gamma })^{\frac{1}{\gamma -1}}],\gamma <1 \\
\\
255[(\frac{1}{\gamma })^{\frac{1}{\gamma -1}} - (\frac{1}{\gamma })^{\frac{\gamma }{\gamma -1}}],\gamma >1
\end{array}
\right.$$
The curve of function of $D_{max}/255$ on $\gamma$ is shown in Fig.2. For the purposes of understanding, four groups of parameters are chosen in the following discussion: $\gamma=\{0.6,0.8,1.2,1.4\}$. It is easy to find that $D_{max_A}$($\gamma=0.6$)$=47.4045$ $>$ $D_{max_D}$($\gamma=1.4$)$=31.416$ $>$$D_{max_B}$($\gamma=0.8$)$=20.8896$ $>$$D_{max_C}$($\gamma=1.2$)$=17.0799$.
{width="4cm" height="3cm"}
Fortunately, in spite of the changes of discriminability in pixel domain, the difference in pixel domain could be learned by deep learning-based method. Inspired by it, the P-CNN is proposed to detect enhanced image. The design of P-CNN is as follows.
Firstly, the high-pass filter is added into the front-end of architecture to eliminate the interfere of image content. Another advantage of using high-pass filter could be that it accelerates training by cooperating with batch normalization. Because that the histogram of high-pass filtered images approximately follows the generalized Gaussian distribution, which is similar to batch normalization \[21\]. In particular, we experimentally find that the filter of the first-order difference along horizontal direction has better performance. $$\label{key}
I_{1} = H*I$$ where $H=[1,-1]$, $I$ is the input image, $I_{1}$ is the output of the first layer, ’\*’ represents the convolution operator.
Next, high-pass filtering layer are followed by four traditional convolutional layers. For each layer, there are four types of operations: convolution, batch normalization, ReLU and average pooling. The feature maps for each layer are 64, 16, 32, 128, respectively. The kernel size for convolutional and pooling operation is 3x3 with 1 stride, 5x5 with 2 strides. It should be pointed out that: 1) we experimentally find that the numbers of feature map for first convolutional layer is important for CE detection and it has better performance when the feature maps is 64. In other words, low-level feature would be more helpful; 2) instead of average pooling, the spatial pyramid pooling layer \[20\] is used in last convolutional layer to fuse multi-scale features. The convolutional layer is calculated as
$$\label{key}
I_{i}=
\left\{
\begin{array}{lr}
P(R(F(W_{i}*I_{i-1}+B_{i}))), i\epsilon (2,3,4)\\
S(R(F(W_{i}*I_{i-1}+B_{i}))), i=5
\end{array}
\right.$$
where $F, R, P ,S$ represents the batch normalization, ReLU, average pooling, and spatial pyramid pooling, respectively. For spatial pyramid pooling, three scales are chosen and lead to 2688 dimensional output.
In the end, the fully connected layer and softmax is followed by a multinomial logistic loss. The loss function is defined as, $$\label{key}
Loss=-log(\frac{e^{W^{j}I_{5}+B^{j}}}{\sum_{j=1}^{n}e^{W^{j}I_{5}+B^{j}}})$$ where $n$ is the number of classes and $j$ denotes the true label. In our experimental setup, Mini-batch Stochastic Gradient Descent is applied and the batch size is set as 120. The learning rate is initialized as 0.001, and scheduled to decrease 10% for every 10000 iterations. The max iterations is 100000. The momentum and weight\_decay are fixed to 0.9 and 0.0005, respectively.
Histogram-Domain Convolutional Neural Networks
----------------------------------------------
According to the report \[4\], the handcrafted feature based on histogram is also vulnerable. The peak and gap feature is easily destroyed by pre-JPEG compression and anti-forensic attacks. In order to detect the CE of JPEG compressed images, Cao *et al.* only used the numbers of gap bins as features. However, its performance for different gamma parameters is unstable and it does not work for anti-forensics attack, which could be caused by the unsteadiness of gap bins.
The reason why gamma correction could cause gap bins is that a strait range of values is projected to the wide one. For example, the values in the range $\rm \left [ 0,T_{D_{max}}^{} \right ], \gamma <1$ will be changed to the range of $\rm \left [ 0,T_{D_{max}}^{\gamma} \right ]$. Therefore, the probability of gap bins (zero bins) should be proportional to the ratio of wide range of values and corresponding strait range,
$$\label{key}
\begin{array}{lr}
P_{zero\_bin}\propto G\left ( r \right ) =
\left\{
\begin{array}{lr}
\frac{T_{D_{max}}^{\gamma }-T_{D_{max}}}{T_{D_{max}}} = \frac{1}{\gamma }-1,\gamma <1\\
\\
\frac{T_{D_{max}}-T_{D_{max}}^{\gamma }}{1-T_{D_{max}}} = \frac{(\frac{1}{\gamma })^{\frac{1}{\gamma -1}} - (\frac{1}{\gamma })^{\frac{\gamma }{\gamma -1}}}{1-(\frac{1}{\gamma })^{\frac{1}{\gamma -1}}},\gamma >1
\end{array}
\right.
\end{array}$$
It can be found that $G(0.6)>G(0.8)>G(1.4)>G(1.2)$ , which means that the numbers of gap bins is ranged among CE parameters. The statistical distribution of gap bins for the original and enhanced images with $\gamma=0.6,0.8,1.2,1.4$ is shown in Fig. 3. As can be seen that the numbers of gap bins for $\gamma=0.6, 0.8$ are larger than $\gamma=1.2,1.4$ and the overlapping parts with original images for $\gamma=0.6, 0.8$ are less than $\gamma=1.2,1.4$, which is consistent with the result of our theoretical analysis. Despite the instability of peak/gap bins, we believe that the effective feature could be auto-learned from histogram domain by using data-driven algorithm. Instead of designing features, the histogram-domain convolutional neural networks is constructed to achieve end-to-end self-learning detection. The H-CNN is proposed to self-learn better feature directly from histogram domain. In addition, as an input with low and fixed dimension, the histogram is suitable for convolutional neural networks. The architecture of H-CNN is shown in Fig 1 (b). Its input is the histogram of the image, namely a vector with 1x256 dimensions. Then, such an input layer is followed by two convolutional and three fully connected layers. The feature maps are 64, 64, 512, 1024, 2, respectively. Lastly, the softmax layer followed by a multinomial logistic loss is added to classify original and enhanced images. The parameters of convolutional layers and hyper-parameters are the same as the P-CNN.
Experimental Results
====================
In order to verify the validity of proposed methods, four groups of experiments are conducted: ORG VS P-CE, JPEG-ORG VS JPEG-CE, ORG VS Anti-CE, and JPEG-ORG VS JPEG-CE-Anti-CE, where ORG is original images in uncompressed format, JPEG-ORG represents original images in JPEG format, P-CE and JPEG-CE denote enhanced versions of ORG and JPEG-ORG, respectively, and Anti-CE and JPEG-CE-Anti-CE represent enhanced images with anti-forensics attack for P-CE, JPEG-CE, respectively. The BOSSBase \[22\] with 10000 images is chosen to construct the dataset. Firstly, the images are centrally cropped into 128x128 pixel patches as ORG. Then, JPEG compression with $Q=70, 50$ is carried out for ORG to build JPEG-ORG. Next, gamma correction with $\gamma=\{0.6, 0.8, 1.2, 1.4\}$ is implemented on ORG, JPEG-ORG to constitute P-CE and JPEG-CE. In the end, Anti-CE is produced by anti-forensics attacks \[12,14\] on P-CE. It should be noted that the reasons for our choice of pixe patch size are that 1) the detection for the images with lower resolution is much harder than higher resolution image; 2) 128x128 is a suitable size for tamper locating based on CE forensics; 3) our hardware configuration is limited. For each experiment, the training data, validation and testing data is 8000, 2000, 10000, respectively. The experiments about the proposed schemes are conducted on one GPU (NVIDIA TITAN X) with an open source framework of deep learning: Caffe \[23\].
Contrast Enhancement Detection For Contrast-Enhanced Images
-----------------------------------------------------------
The result for contrast-enhanced images in uncompressed format, is as shown in Table I. P-CNN is pixel-domain convolutional neural networks and H-CNN is histogram-domain convolutional neural networks. As seen from the Table 1, for Cao’s method, the detection accuracy for $\gamma=\{0.6, 0.8\}$ is much higher than one for $\gamma=\{1.2, 1.4\}$. The reason is that gap feature is unstable among CE parameters, which is consistent with our analysis in Section II.
Inspired by transfer learning techniques \[24\], we further improve performance of P-CNN by finetuning the model for $\gamma=\{0.8, 1.2, 1.4\}$ from the model for $\gamma=0.6$. P-CNN-FT achieve better performance than De Rosa’s and Cao’s methods and H-CNN have much better performance than the state-of-the-art schemes. It should be noted that the performance of H-CNN is better and more stable than the others. Such results demonstrated that the histogram domain feature should be effective for CE detection.
Method $\gamma=0.6$ $\gamma=0.8$ $\gamma=1.2$ $\gamma=1.4$
-------------- -------------- -------------- -------------- --------------
De Rosa\[9\] 94.02% 84.85% 78.37% 74.12%
Cao\[4\] 93.89% 93.90% 80.26% 81.40%
Li\[5\] 93.63% 89.48% 90.76% 93.44%
P-CNN 94.70% 89.00% 78.00% 86.00%
P-CNN \_FT 94.70% 90.00% 82.45% 88.00%
H-CNN **99.48%** **99.45%** **99.40%** **99.07%**
: The detection accuracy for contrast-enhanced images.[]{data-label="my-label"}
Robustness Against Pre-JPEG Compressed and Anti-Forensic Attacked Contrast-Enhanced Images
------------------------------------------------------------------------------------------
The performance of different methods for pre-JPEG compressed images with $Q=\{50, 70\}$ and anti-forensics attacked images are shown in Table 2,3,4 and 5. It can be seen from Table II that P-CNN and H-CNN have much higher detection accuracy than De Rosa’s and Cao’s methods and comparable performance with Li’s method. Besides, there is an interesting phenomenon that the performance of P-CNN has a significant improvement compared to P-CE detection. The reason may be attributed to that JPEG compression weakens the signal components in high frequence and the difference between original and enhanced images after JPEG compressing would be highlighted.
QF Method $\gamma=0.6$ $\gamma=0.8$ $\gamma=1.2$ $\gamma=1.4$
---- -------------- -------------- -------------- -------------- -------------- --
De Rosa\[9\] 81.50% 79.69% 75.16% 72.70%
Cao\[4\] 93.96% 93.75% 80.36% 81.57%
Li\[5\] 99.11% 98.59% 97.75% 98.43%
P-CNN 98.20% 98.25% 96.70% 97.30%
H-CNN **99.90%** **99.80%** **99.50%** **99.78%**
De Rosa\[9\] 83.99% 82.27% 77.47% 72.95%
Cao\[4\] 94.06% 93.77% 80.55% 81.56%
Li\[5\] 98.54% 97.42% 96.22% **97.79%**
P-CNN 98.60% 97.00% 95.70% 96.50%
H-CNN **98.86%** **99.03%** **98.27%** 97.68%
: The detection accuracy for JPEG compressed image with different QFs.[]{data-label="my-label"}
For anti-forensic attacks, Cao’s method does not work and there is a degradation in performance of H-CNN, especially, when anti-forensic method \[12\] is applied. Because that the anti-forensic attacks would conceal the peak/gap feature in histogram domain. In addition, the anti-forensics attacks based on histogram maybe has no or slight effect on pixel domain. Therefore, the P-CNN has best performance in this case. While the pre-compression and anti-forensic attack are put into together, as shown in Table 5, the proposed method have comparable with Li’ scheme.
In conclusion, De Rosa’s method is not robust for pre-JPEG compression and anti-forensics attack and Cao’s method is vulnerable for anti-forenisic attack. Furthermore, such prior algorithms are unstable in different gamma levels. Although Li’s method is better than previous works in the case of pre-JPEG compression and anti-forensic attack, its performance is unsatisfactory when no other operation is used. Comparing with the above schemes, the proposed P-CNN and H-CNN, achieve good robustness against pre-JPEG compression, anti-forenic attack, and CE level variation and H-CNN achieve much better performance in the case of no other operation.
Method $\gamma=0.6$ $\gamma=0.8$ $\gamma=1.2$ $\gamma=1.4$
-------------- -------------- -------------- -------------- --------------
De Rosa\[9\] 69.85% 66.03% 62.29% 64.42%
Cao\[4\] —– —– —– —–
Li\[5\] **99.57%** **99.38%** **99.33%** **99.51%**
P-CNN 98.6% 98.5% 97.8% 98%
H-CNN 98.82% 97.59% 97.57% 97.09%
: The detection accuracy in the case of anti-forensics attack \[14\].[]{data-label="my-label"}
Method $\gamma=0.6$ $\gamma=0.8$ $\gamma=1.2$ $\gamma=1.4$
-------------- -------------- -------------- -------------- --------------
De Rosa\[9\] 61.674% 58.826% 55.320% 59.329%
Cao\[4\] —– —– —– —–
Li\[5\] 96.3% 95.54% 95.72% 96.55%
P-CNN **97.9%** **96%** **96.5%** **96.55%**
H-CNN 88.77% 73.65% 74.85% 78.42%
: The detection accuracy in the case of anti-forensics attack \[12\].[]{data-label="my-label"}
QF Method $\gamma=0.6$ $\gamma=0.8$ $\gamma=1.2$ $\gamma=1.4$
---- -------------- -------------- -------------- -------------- -------------- --
De Rosa\[9\] 70.26% 67.85% 65.38% 66.52%
Cao\[4\] —– —– —– —–
Li\[5\] **99.9%** **99.9%** **99.9%** **99.9%**
P-CNN **99.9%** **99.9%** **99.9%** **99.9%**
H-CNN 99.45% 99.4% 99.2% 99.2%
De Rosa\[9\] 68.68% 65.61% 62.24% 63.93%
Cao\[4\] —– —– —– —–
Li\[5\] **99.9%** **99.9%** **99.9%** **99.9%**
P-CNN 99.8% 99.75% 99.55% 99.8%
H-CNN 97.35% 98.35% 97.8% 98.15%
: The detection accuracy for JPEG compressed image with different QFs and anti-forensics attack \[14\].[]{data-label="my-label"}
Effect of the scale of training data
------------------------------------
It is well known that the scale of data has an important effect on performance for deep-learning based method. In this part, we conducted experiments to evaluate the effect of the scale of data on performance of H-CNN and P-CNN. The images from BOSSBase are firstly cropped into 128x128 pixel patches with non-overlapping. Then these images are enhanced with $\gamma=0.6$. We randomly chose 80000 image pairs as test data and 5000, 20000, 40000, 80000 image pairs as training datas. Four groups of H-CNN, P-CNN are generated using above four training datas and the test data is same for these experiments. The result is as shown in Figure.4. It can be seen that the scale of training data has an slight effect on H-CNN with small parameters and the opposite happens for P-CNN. Therefore, larger scale of training data be beneficial to the performance of P-CNN and the performance of P-CNN would be improved by increasing training data.
{width="6cm" height="4cm"}
Conclusion
==========
The existing schemes for contrast enhancement forensics have an unsatisfactory performance, especially, in the cases of pre-JPEG comression and anti-forensic attacks. To deal with such problems, in this paper two robust CE forensics algorithms based on deep learning (H-CNN, P-CNN) are proposed. Such methods achieve end-to-end classification based on pixel and histogram domain. Experimental results show that our proposed H-CNN attains better performance than the state-of-the-art ones in the case of no other operation and proposed methods are robust against pre-JPEG compression, anti-forensic attack, and CE level variation.
[00]{}
M. C. Stamm and K. J. R. Liu, “Blind forensics of contrast enhancement in digital images,” in Proc. IEEE Int. Conf. on Image Processing, October 2008 M. Stamm and K. Liu, “Forensic detection of image manipulation using statistical intrinsic fingerprints,” Information Forensics and Security, IEEE Transactions on, vol. 5, no. 3, pp. 492–506, Sept 2010. M. C. Stamm and K. J. R. Liu, “Forensic estimation and reconstruction of contrast enhancement mapping,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, March 2010, pp. 1698–1701. G. Cao, Y. Zhao, R. Ni, and X. Li, “Contrast enhancement based forensics in digital images,” IEEE Transactions on Information Forensic and Security, vol. 9, pp. 515–525, March 2014 Li H, Luo W, Qiu X, et al. Identification of various image operations using residual-based features\[J\]. IEEE Transactions on Circuits and Systems for Video Technology, 2016. Lin X, Li C T, Hu Y. Exposing image forgery through the detection of contrast enhancement\[C\]//Image Processing (ICIP), 2013 20th IEEE International Conference on. IEEE, 2013: 4467-4471. Lin X, Wei X, Li C T. Two improved forensic methods of detecting contrast enhancement in digital images\[C\]//Media Watermarking, Security, and Forensics 2014. International Society for Optics and Photonics, 2014. Wen L, Qi H, Lyu S. Contrast Enhancement Estimation for Digital Image Forensics\[J\]. arXiv preprint arXiv:1706.03875, 2017. A. De Rosa, M. Fontani, M. Massai, A. Piva, and M. Barni, “Second-order statistics analysis to cope with contrast enhancement counter- forensics,” IEEE Signal Processing Letters, vol. 22, pp. 1132–1136, August 2015 H. Farid, “Blind Inverse Gamma Correction,” IEEE Transactions on Image Processing, vol. 10, no. 10, pp. 1428–1433, Oct 2001. A. C. Popescu and H. Farid, “Statistical Tools for Digital Forensics,” 6th Intl. Work. on Info. Hiding & LNCS, vol. 3200, pp. 128–147, May 2004 G. Cao, Y. Zhao, and R. Ni, “Forensic estimation of gamma correctionin digital images,” in Proc. IEEE Int. Conf. on Image Processing, Sept 2010, pp. 2097–2100 Wang P, Liu F, Yang C, et al. Parameter estimation of image gamma transformation based on zero-value histogram bin locations\[J\]. Signal Processing: Image Communication, 2018. M. Barni, M. Fontani, and B. Tondi, “A universal technique to hide traces of histogram-based image manipulations,” in Proc. of the ACM Workshop on Multimedia and Security, 2012, pp. 97–104. G. Cao, Y. Zhao, R. Ni, and H. Tian, “Anti-forensics of contrast enhancement in digital images,” in Proc. of the ACM Workshop on Multimedia and Security, 2010, pp. 25–34. C.-W. Kwok, O. C. Au, and S.-H. Chui, “Alternative anti-forensics method for contrast enhancement,” in Proc. of the Int. Conf. on Digital- Forensics and Watermarking, 2012, pp. 398–410. P. Comesana-Alfaro and F. Perez-Gonzalez, “Optimal counterforensics for histogram-based forensics,” in Proc. IEEE Int. Conf. on Acoustics, Speech and Signal Processing, May 2013, pp. 3048–3052. G. Cao, Y. Zhao, R. Ni, H. Tian, and L. Yu, “Attacking contrast enhancement forensics in digital images,” Science China Information Sciences, vol. 57, no. 5, pp. 1–13, 2014. Ravi H, Subramanyam A V, Emmanuel S. ACE–An Effective Anti-forensic Contrast Enhancement Technique\[J\]. IEEE Signal Processing Letters, 2016, 23(2): 212-216. He K, Zhang X, Ren S, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition\[J\]. IEEE transactions on pattern analysis and machine intelligence, 2015, 37(9): 1904-1916. Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift\[C\]//International conference on machine learning. 2015: 448-456. http://agents.fel.cvut.cz/stegodata/ An open source framework of deep learning: http://caffe.berkeleyvision.org/ Pan S J, Yang Q. A survey on transfer learning\[J\]. IEEE Transactions on knowledge and data engineering, 2010, 22(10): 1345-1359.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
=-0.5pt =6.0in =8.5in =0.2in =0.2in
**[A.Banerjee$^{1}$,A.Beesham$^{3}$,S.Chatterjee$^{1,2}$,A.A.Sen$^{1}$]{}**
$^{1}$Relativity and Cosmology Research Centre
Department of Physics,Jadavpur University
Calcutta-700032,India.
$^{2}$Department of Physics\
New Alipore Govt.College\
Calcutta-700053,India.\
$^{3}$Department of Applied Mathematics\
University of Zululand\
Kwadlanguzwa 3886\
South Africa.\
PACS NO:04.20.Jb,04.25.Dm,04.50.+h.\
Correspondence To : A.A.Sen.E-mail :anjan@juphys.ernet.in\
The well known monopole solution of Barriola and Vilenkin (BV) resulting from the breaking of a global SO(3) symmetry is extended in general relativity along with a zero mass scalar field and also in Brans-Dicke(BD) theory of gravity.In the case of BD theory , the behaviour of spacetime and other variables such as BD scalar field and the monopole energy density have been studied numerically.For monopole along with a zero mass scalar field , exact solutions are obtained and depending upon the choice of arbitary parameters , the solutions either reduce to the BV case or to a pure scalar field solution as special cases.It is interesting to note that unlike the BV case the global monopole in the BD theory does exert gravitational pull on a test particle moving in its spacetime.\
Introduction
============
Monopoles are point like topological objects that may arise during phase transitions in the early universe[@R1; @R2].Depending on the nature of the scalar field it can be shown that spontaneous symmetry breaking can give rise to such onjects which are nothing but the topological knots in the vacuum expectation value of the scalar field and most of their energy is concentrated in a small region near the monopole core.From the topological point of view they are formed in vacuum manifold when the latter contains surfaces which can not be continuously shrunk to a point.These monopoles have Goldstone fields with their energy density decreasing with distance as $r^{-2}$.These monopoles are found to have some interesting features in a sense that it exerts practically no gravitational force on its surrounding nonrelativistic matter but the spacetime around it has a deficit solid angle.
In a pioneering work Barriola and Vilenkin(BV) [@R3] showed the existence of such a monopole solution resulting from the breaking of global SO(3) symmetry of a triplet scalar field in a Schwarzschild background.Recently we [@R4] have also obtained a monopole solution in Kaluza-Klien spacetime which extends the earlier work of BV to its five dimensional analogue.
In the present work we make an attempt to study such a monopole in Brans-Dicke(BD) theory[@R5] of gravitation .The renewed interest of scalar tensor theory of gravitation is mainly due to two important theoretical developments in the study of early universe - one is the prediction of the dilaton field arising from the low energy limit of the string theory.The other is the recent theory of extended inflation which is beleived to have solved the fine tuning problem by slowing down the expansion rate of universe from the exponential to polynomial.
In view of serious difficulties in solving the monopole problem in the original version of the BD theory,an attempt is made in the second part of the paper to get the solution in the Dicke’s revised unit[@R6] where the field equations are identical with the Einstein’s field equations with the BD scalar field appearing as a “matter field” in the theory.But even in the revised version ,we have not as yet obtained the solution in a closed form.However we have been able , so far , to reduce the mathematical formalism to a single differential equation involving only a single variable namely the BD scalar field and once the scalar field is known , the metric components and the monopole energy density can be obtained from the other equations.We have,however,made a numerical study of the behaviour of the different variables and the results seem to be physically consistent.
In the third part of the paper we have studied the similar problem with the monopole interacting with a zero mass scalar field and have been able to obtain the exact solutions of the field equations.The BV monopole solutions are recovered from these solutions when the zero mass scalar field vanishes.On the other hand when the monopole field is switched off we get back the usual zero mass scalar field solution already existing in literature.
Our paper is organised as follows. In section 2 the field equations for a global monopole in Brans-Dicke theory in revised units are written, analysed, and their qualitative behaviour is studied numerically. In section 3 the same problem is addressed in Einstein’s theory in the presence of a zero mass scalar field for two distinctly different situations. The paper ends with a conclusion in section 4.
The Global Monopole In Branse Dicke Theory
==========================================
The gravitational field equation for a global monopole in Brans-Dicke theory written in Dicke’s revised unit[@R6] is in general $$G^{\mu}_{\nu} = - T^{\mu}_{\nu} - {{2\omega+3}\over{2}}{1\over{\phi^{2}}}[\phi^{,\mu}\phi_{,\nu}
-{1\over{2}}\delta^{\mu}_{\nu}\phi_{,\alpha}\phi^{,\alpha}]
\eqno{(1)}$$
Where $T^{\mu}_{\nu}$ is the energy momentum tensor due to a monopole field, $\phi$ is the B-D scalar field and $\omega$ is the B-D parameter.
Since the spacetime here is static and spherically symmetric, the metric is given in curvature coordinates in the form $$ds^{2} = e^{\nu}dt^{2}-e^{\beta}dr^{2}-r^{2}dr^{2}-r^{2}sin^{2}\theta d\Phi^{2}
\eqno{(2)}$$
where $\nu$ and $\beta$ are functions of r alone.
In the original work of Barriola and Vilenkin[@R3] there is no contribution to the energy momentum tensor from the B-D scalar field but the lagrangian is due to global SO(3) symmetry for a triplet scalar field, whose symmetry breaking gives rise to a global monopole. Under reasonable conditions the energy momentum tensor due to the monopole field outside the core now becomes [@R3] $$T^{t}_{t} = T^{r}_{r} = {\eta^{2}\over{r^{2}}}
\eqno{(3)}$$ $$T^{\theta}_{\theta} = T^{\Phi}_{\Phi} = 0
\eqno{(4)}$$
where $\eta$ is the symmetry breaking scale of the theory.
The above forms of $T^{\mu}_{\nu}$ are consistent with the Bianchi identity. One should note here that in the original version of the B-D theory[@R5] the relation $T^{\mu}_{\nu;\mu}=0$ is separately satisfied which again leads to the same expressions (3) and (4) for $T^{\mu}_{\nu}$. However, in the revised units, the B-D scalar field has additional contributions to the energy momentum tensor in the field equations and consequently the Bianchi identity in this case will certainly not lead to the identical expressions for energy momentum tensors as for monopole field only. We therefore prefer to write
$$T^{t}_{t} = T^{r}_{r} = \rho(r)
\eqno{(5)}$$ $$T^{\theta}_{\theta} = T^{\Phi}_{\Phi} = 0
\eqno{(6)}$$
where $\rho(r)$ is a function of r to be determined from the field equations.
With (2),(5) and (6), equation (1) yields explicitly the following relations: $$e^{-\beta}({1\over{r^{2}}}-{\beta^{'}\over{r}})-{1\over{r^{2}}}=-\rho-{k\over{2}}\psi^{'2}e^{-\beta}
\eqno{(7a)}$$ $$e^{-\beta}({1\over{r^{2}}}+{\nu^{'}\over{r}})-{1\over{r^{2}}}=-\rho+{k\over{2}}\psi^{'2}e^{-\beta}
\eqno{(7b)}$$ $$e^{-\beta}({\nu^{''}\over{2}}+{\nu^{'2}\over{4}}-{{\nu^{'}\beta^{'}}\over{4}}+{{\nu^{'}-\beta^{'}}\over{2r}})=-{k\over{2}}\psi^{'2}e^{-\beta}
\eqno{(7c)}$$
where $\psi=ln(\phi)$ and $k={{2\omega+3}\over{2}}$ and prime denotes differentiation with respect to r.
The wave equation for the BD scalar field is given by $$\psi^{''}+2{\psi^{'}\over{r}}+({{\nu^{'} -\beta^{'}}\over{2}})\psi^{'} = - \rho{e^{\beta}\over{k}}
\eqno{(8)}$$ By (7a)-(7b)-2x(7c) one gets $$\nu^{''}+2{\nu^{'}\over{r}}+{\nu^{'2}\over{2}}={{\nu^{'}\beta^{'}}\over{2}}
\eqno{(9)}$$
One may attempt to solve the eqn(9) for two different situations: (i)$\nu^{'}=0$, which means $\nu=constant$ and (ii)$\nu^{'}\neq 0$, which in turn leads to differential equation: $e^{\nu}\nu^{'2}={b^{2}\over{r^{4}}}e^{\beta}$, where $b$ is an arbitrary constant of integration.
In the first case one may choose $\nu=1$ without the loss of any generality, which in view of Bianchi identity relation immediately leads to the equation $$\rho^{'}+2{\rho\over{r}}+\rho\psi^{'}=0$$ The above equation has the first integral $$\rho r^{2}=a^{2}e^{-\psi}
\eqno{(10)}$$ where $a$ is another arbitrary integration constant.
In the second case that is for $\nu^{'}\neq 0$ it is extremely hard task to obtain any differntial equation which involves only a single variable. Hence we abandon this case in our present work and proceed with the other case, that is, $\nu^{'}=0$ for further calculations.
With $\nu^{'}=0$ one gets after some simple and straightforward calculation the following two equations: $$e^{\beta}={{({k\over{2}}\psi^{'2}r^{2}-1)}\over{(a^{2}e^{-\psi}-1)}}
\eqno{(11)}$$ $$(a^{2}-e^{\psi})\psi^{''}-{k\over{2}}r(a^{2}-e^{\psi})\psi^{'3}+{a^{2}\over{2}}\psi^{'2}+
{2\over{r}}(a^{2}-e^{\psi})\psi^{'}-{a^{2}\over{kr^{2}}}=0
\eqno{(12)}$$ One should note that the actual BD scalar field is $\phi=e^{\psi}$.
The equation(12) is the key equation to our subsquent analysis.If this equation can be integrated,the complete set of equations will be solvable,as in view of equation(10) and (11),$\rho$ and $e^{\beta}$ may be known in terms of $\psi$. However equation(12) is a highly nonlinear equation which may not be amenable to an analytic solution in a closed form.
Hence we try to solve the equation (12) numerically in order to estimate the behaviour of scalar field $\phi,e^{\beta}$ as well as the energy density of the monopole $\rho$ outside the core. We rewrite the equation(12) as, $${{{d^2}\phi}\over{dx^{2}}}=2.304{d\phi\over{dx}}+{{(a^{2}-2\phi)}\over{(2a^{2}-2\phi)\phi}}({d\phi\over{dx}})^{2}-
{0.434k\over{2\phi^{2}}}({d\phi\over{dx}})^{3}+{a^{2}\phi\over{0.434k(a^{2}-\phi)}}
\eqno{(13)}$$ where $x=log_{10}({\delta\over{r}})$ where $\delta$ = the core radius of the monopole, $\phi=e^{\psi}$ , the BD scalar field. The variables $e^{\beta}$ and $\rho$ take the form $$e^{\beta}={{0.095k({d\phi\over{dx}})^{2}-\phi^{2}}\over{a^{2}\phi-\phi}}
\eqno{(14)}$$ $$\rho={{{a^2}e^{4.608x}}\over{\phi}}
\eqno{(15)}$$
We have tried to solve the equation(13) numerically by fourth order Runge-Kutta method and plot $\phi,\phi^{'},\rho ,e^{\beta}$ against $r$.\
We have taken the initial condition that for a large distance from the monopole core i.e when $r$ takes a very large value, $\phi^{'} \rightarrow 0$ and $\phi$ approaches a very small value.We have taken the distance upto $10^{5}$ times the core radius $\delta$ of the monopole.
In figure(1) we have plotted both $\phi$ and $\phi^{'}$ and further in figure(2) and(3) we plot $e^{\beta},\rho$ respectively against $log_{10}(\delta/r)$.
The behaviour of $e^{\beta}$ in fig(2) shows that the spacetime outside the monopole core becomes flat for a large distance from the core.The fig(3) shows that the monopole energy density $\rho$ rapidly decreases with respect to the distance from the core.This is qualitatively consistent with the corresponding BV case where also the energy density of the monopole outside the core falls off with $r$.The behaviour of scalar field in fig(1) also shows that $\phi$ approaches a constant value and $\phi^{'}$ approaches zero for a large distance from the core.
**[Gravitational force in the field of a monopole]{}**
Before ending up this section we would like to point out an important property of the BD scalar field-the gravitational force acting on surrounding matter of the monopole.
In the original version of BD theory our metric (2) becomes $$ds^{2} = e^{-\psi}dt^{2}-e^{-\psi}(e^{\beta}dr^{2}+r^{2}d\theta^{2}+r^{2}sin^{2}\theta d\Phi^{2})
\eqno{(16)}$$ as $\nu=constant $ for our case and $\psi$ , $\beta$ are given by (13) and (14) respectively. The radial component of the acceleration acting on a test particle in the gravitational field of the monopole is given by: $$\dot{v^{1}}=v^{1}_{;0}v^{0}
\eqno{(17)}$$
Since for a comoving particle $v^{\mu}={1\over{\sqrt{g_{00}}}}\delta^{\mu}_{0}$ we have $v^{0}=e^{\psi/2}$ and $\dot{v^{1}}=-{1\over{2}}e^{(\psi-\beta)}\psi^{'}$. Assuming $\psi^{'}<0$ which is obvious from fig(1), the particle accelarates away in the radial direction in order to keep it at rest which implies that a monopole has an attractive influence on the test particle. This attractive force is due to the presence of the BD scalar field $\psi$.When $\psi^{'}=0$ the BD scalar field is absent and we have only the monopole remaining which however exerts no gravitational influence on the surrounding matter as shown by BV[@R3].
Global monopole interacting with a zero mass scalar field
=========================================================
In this section we have considered a monopole in the presence of a massless scalar field which also contribute to the energy momentum tensor in addition to that of the monopole field. Exact analytical solutions may be obtained in this case. However, for economy of space the details of mathematical steps are omitted and the main results are sketched below:\
In this case we take the metric as $$ds^{2}=e^{\nu}dt^{2}-e^{\beta}dr^{2}-r^{2}d\Omega^{2}
\eqno{(18)}$$ where $\nu$,$\beta$ are functions of $r$ only.\
Solving the field equations we get $$e^{\nu}=(1-\eta^{2}-2{M\over{r}})^{n}
\eqno{(19)}$$ $$e^{\beta}=(1-\eta^{2}-2{M\over{r}})^{n-2}
\eqno{(20)}$$ Where $M$ and $n$ are arbitary constant of integration and $\eta$ is the symmetry breaking scale for the monopole.\
The energy density of the monopole is given by $$\rho_{mon}={1\over{r^{2}}}[1-(1-\eta^{2})(1-\eta^{2}-2{M\over{r}})^{n-1}]
\eqno{(21)}$$ The zero mass scalar field is given by $$\psi^{'2}={{4M(n-1)}\over{(1-\eta^{2})r^{3}}}(1-\eta^{2}-2{M\over{r}})^{-1}
\eqno{(22)}$$ For $n=1$ the variation of the scalar field $\psi$ vanishes,and we get back the BV type monopole with $\rho_{mon}={\eta^{2}\over{r^{2}}}$\
On the other hand pure scalar field solution without the presence of a monopole field cannot be recovered in this form of metric, as is evident from the expressions (21) and (22).\
Here we take the metric in the form $$ds^{2}=e^{\nu}dt^{2}-e^{\beta}dr^{2}-r^{2}R^{2}d\Omega^{2}
\eqno{(23)}$$ where $\nu,\beta$ and $R$ are functions of $r$ only.\
The solutions are $$e^{\nu}=(1-\eta^{2}-2{M\over{r}})^{n}
\eqno{(24a)}$$ $$e^{\beta}=(1-\eta^{2}-2{M\over{r}})^{-n}
\eqno{(24b)}$$ $$R^{2}=(1-\eta^{2}-2{M\over{r}})^{1-n}
\eqno{(24c)}$$ $$\rho_{mon}={\eta^{2}\over{r^{2}}}(1-\eta^{2}-2{M\over{r}})^{n-1}
\eqno{(24d)}$$ $$\psi^{'}=\sqrt{2}{M\over{r^{2}}}\sqrt{n^{2}-1}(1-\eta^{2}-2{M\over{r}})^{-1}
\eqno{(24e)}$$
It is clear from the expressions (24d) and (24e) that when $n=1$, we have $\rho_{mon}=\eta^{2}/r^{2}$ and $\psi^{'}=0$, so that the BV solution is recovered. But on the other hand when we put $\eta^{2}=0$, we obtain $\rho_{mon}=0$ and $\psi^{'}\neq 0$ meaning that we have a pure massless scalar field distribution.
Conclusion
==========
In this work we have extended the monopole solution of BV to the scalar tensor theory. In BD theory we have not been able to get the analytical solution.So we have taken recourse to numerical methods to study the behaviour of BD scalar field.From this field ,the behaviour of metric coefficient and the energy density of the monopole are also found out.It is encouraging to point out that the nature of variation of BD scalar field,metric coefficient and also the energy density of the monopole are quite physically realistic in the sense that the spacetime is asymptotically flat.Also the monopole energy density outside the core decreases sharply with the distances.
For the sake of completeness ,we have also studied the similar problem with monopole interacting with a massless scalar field,and have obtained exact solutions in this case.Under suitable choice of the parameters in some cases our solutions reduce to either the BV solution or to that of a pure zero mass scalar field.
To end up, a final remark may be in order.It is well known that in Einstein’s theory,the BV type of monopole does not practically exert any gravitational pull on any surrounding matter.However our analysis shows that the monopole in BD theory does exert gravitational force on a test particle.So this observation is in striking contrast with the analogue in Einstein’s theory.
Acknowledgement
===============
S.C wishes to thank the University of Zululand,South Africa,for local hospitability where a part of the work was done.A.A.S.thanks the University Grants Commission, India for financial support.A.B. and S.C. also thank D.S.T.,India,for financial support.\
**[Figure Captions]{}**
1>In Fig1 , we have plotted the BD scalar field $\phi$ and its first derivative $\phi^{'}$ with respect to $log(\delta/r)$.
2>In Fig2 , we have plotted the metric coefficient $e^{\beta}$ with respect to $log(\delta/r)$.
3>In Fig3 , we have plotted the monopole energy density $\rho$ with respect to $log(\delta/r)$.\
[40]{} T.W.Kibble,J.Phys.A,[**[9]{}**]{},1387(1976). A.Vilenkin,Phys.Rep,[**[121]{}**]{},263(1985). M.Barriola,A.Vilenkin,Phys.Rev.Lett,[**[63]{}**]{},341,(1989). A.Banerjee,S.Chatterjee and A.A.Sen,Class.Quant.Grav,[**[13]{}**]{},3141(1996). C.Brans,R.H.Dicke,Phys.Rev,[**[124]{}**]{},925,(1961). R.H.Dicke,Phys.Rev,[**[125]{}**]{},2163,(1962).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
[**Cosmology with a Nonlinear Born-Infeld Type Scalar Field** ]{} 0.15 in H.Q.Lu[[^1]]{}\
[*Department of Physics,\
Shanghai University, Shanghai 200436, China*]{}\
0.5 in
Abstract
0.2 in
[ Recent many physicists suggest that the dark energy in the universe might result from the Born-Infeld(B-I) type scalar field of string theory. The universe of B-I type scalar field with potential can undergo a phase of accelerating expansion. The corresponding equation of state parameter lies in the range of $\displaystyle -1<\omega<-\frac{1}{3}$. The equation of state parameter of B-I type scalar field without potential lies in the range of $0\leq\omega\leq1$. We find that weak energy condition and strong energy condition are violated for phantom B-I type scalar field. The equation of state parameter lies in the range of $\omega<-1$.\
[**Keywords:**]{}Dark energy; Born-Infeld type scalar field; Phantom cosmology.\
[**PACS:**]{}98.80.Cq]{}
Introduction
============
Evidence that the universe is undergoing a phase of accelerating expansion at the present epoch continues to grow. This not only can be inferred by accelerating dynamics in high redshift surveys of Ia type supernovae\[1\], but also now it is independently implied from seven cosmic microwave background experiments(including the latest WMAP)\[2\], The favoured explanation for this behavior is that the universe is presently dominated by some form of dark energy density, contribution up to 70% of the critical energy density, with the remaining 30% comprised of clumpy baryonic and non-baryonic dark matter. One of the central questions in cosmology today is the origin of the dark energy. Many candidates for dark energy have been proposed so far to fit the current observations. Among these models, the most important one is a self-interacting scalar field with a potential and thereby acts as a negative pressure source, referred to as “quintessence”\[3\]. This paradigm has caused attentions because a wide class of models exhibits tracking behavior at late time, where the dynamics of the field becomes independent of its initial conditions in early universe. In principle, this may resolve the fine-tuning inherent problem in dark energy models purely based on a cosmological constant. The major difference among these models is that different model predicts dissimilar equation of state of the dark energy, thus different cosmology is predicted. Especially, for these models, the equations of state parameter are confined within the range of $\displaystyle -1<\omega=\frac{p}{q}<-\frac{1}{3}$, which can drive the conclusion of accelerating expansion of the universe. However, some analysis of the observation data hold that the range of the equation of state parameter may not always be greater than $-1$ . In fact, they can lie in the range of $-1.48<\omega<-0.72$ \[4\]. It is obvious that the equation of state of conventional quintessence models based on a scalar field with positive kinetic energy can not evolve into the range of $\omega<-1$, and therefore, some authors\[5\] investigated phantom field models that possess negative kinetic energy and can realize $\omega<-1$ in their evolution. It is true that the field theory with negative kinetic energy poses a challenge to the widely accepted energy condition and leads to a rapid vacuum decay\[6\], but it is still very important to study these models, in some sense that is phenomenologically interesting.
On the other hand, the role of tachyon field in string theory in cosmology has been widely studied\[5\]. It shows that the tachyon can be described by a Born-Infeld (B-I)type lagrangian resulting from string theory. It is clear that the lagrangian $\displaystyle\frac{1}{\eta}[1-\sqrt{1-\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}]-u(\varphi)$ is equivalent formally to the tachyon type lagrangian $-\displaystyle\frac{1}{\eta}\sqrt{1-g^{\mu\nu}\Phi_{,~\mu}\Phi_{,~\nu}}$ with a potential $\left[\displaystyle\frac{1}{\eta}-u(\varphi)\right]$, where re-scale the scalar field as $\Phi=(\eta)^{1/2}\varphi$. In this paper we consider cosmology of B-I type scalar field. The paper is organized as follows: In sec.2, we consider B-I type Lagrangian of scalar field without a potential. We obtain $0<\omega<1$. In sec.3, the B-I type Lagrangian of scalar field with a potential is considered. We find that potential $u(\varphi)$ is greater than $\displaystyle\frac{1}{\eta}$, and the kinetic energy of B-I scalar field is smaller than $\displaystyle\frac{1}{3\eta}$, thus we obtain $-1<\omega<-1/3$. In sec.4, the phantom with B-I Lagrangian is considered. We find that weak energy condition and strong energy condition are violated for phantom B-I type scalar field with potential. The equation of state parameter lies in the range $\omega<-1$. Sec.5 is Summary.
The Model with B-I Lagrangian $\displaystyle\frac{1}{\eta}[1-\sqrt{1-\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}]$
=====================================================================================================================
In 1934\[7\], Born and Infeld put forward a theory of non-linear electromagnetic field. The lagrangian density is $$\displaystyle L_{BI}=b^2\left[1-\sqrt{1-(\frac{1}{2b^2})F_{\mu\nu}F^{\mu\nu}}~\right]\eqno(2-1)$$ The lagrangian density for a B-I type scalar field is $$\displaystyle L_S=\frac{1}{\eta}\left[1-\sqrt{1-\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}~\right]\eqno(2-2)$$ Eq.(2-2) is equivalent to the tochyon lagrangian $[-V(\varphi)\sqrt{1-g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}+\Lambda]$ if $\displaystyle V(\varphi)=\frac{1}{\eta}$ and cosmological constant $\displaystyle \Lambda=\frac{1}{\eta}$ ($\displaystyle
\frac{1}{\eta}$ is two times as “critical” kinetic energy of $\varphi$ field). The lagrangian (2-2) possesses some interesting characteristics, it is exceptional in the sense that shock waves do not develop under smooth or continuous initial conditions and because nonsingular scalar field solution can be generated\[8\]. When $\eta \rightarrow 0$, by Taylor expansion, Eq.(2-2) approximates to the lagrangian of linear scalar field. $$\lim_{\eta\rightarrow 0}L_S=\frac{1}{2}g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}\eqno(2-3)$$ A quantum model of gravitation interacting with a lagrangian (2-2) of B-I type scalar has been considered by us. We obtained the Wheeler-Dewitt equation of B-I scalar field and found the wave function of the universe. An inflationary universe, with the largest possible vacuum energy and the largest interaction between the particles of B-I scalar field\[9\], is predicted. Next we consider classical cosmology. For the spatially homogeneous scalar field, Eq.(2-2) becomes $$L_S=\frac{1}{\eta}\left[1-\sqrt{1-\eta \dot{\varphi}^2}~\right]\eqno(2-4)$$ In the spatially flat Robertson-Walker metric $ds^2=dt^2-a^2(t)(dx^2+d^2y+d^2z)$, Einstein equation $G_{\mu\nu}=KT_{\mu\nu}$ can be written as $$\left(\frac{\dot{a}}{a}\right)^2=\frac{K}{3}T^0_0\eqno(2-5)$$ $$2\frac{\ddot{a}}{a}+\left(\frac{\dot{a}}{a}\right)^2=KT^1_1=KT^2_2=KT^3_3\eqno(2-6)$$ Substituting Eq.(2-5) into Eq.(2-6), we get $$\frac{\ddot{a}}{a}=-\frac{K}{3}(T^0_0-3T^1_1)\eqno(2-7)$$ where $$T^\mu_\nu=\frac{g^{\mu\rho}\varphi_{,~\nu}\varphi_{,~\rho}}{\sqrt{1-\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}}-\delta^\mu_\nu L_S \eqno(2-8)$$ The energy density $\rho_s=T^0_0$ and pressure $p_s=-T^i_i$ are defined as following: $$\rho_s=T^0_0=\frac{\dot{\varphi}^2}{1-\eta\dot{\varphi}^2}-L_s\eqno(2-9)$$ $$p_s=-T^i_i=\frac{1}{\eta}[1-\sqrt{1-\eta\dot{\varphi}^2}]\eqno(2-10)$$ where the upper index “.” denotes the derivative with respect to $t$.\
The equation of motion of scalar field $\varphi$ is $$\frac{1}{\sqrt{-g}}\frac{\partial}{\partial x^\nu}\left[\frac{\sqrt{-g}g^{\mu\nu}\varphi_{,~\mu}}{\sqrt{1-\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}}\right]=0\eqno(2-11)$$ The scalar field $\varphi$ only depends on $t$. From Eq.(2-11), we can obtain $$\dot{\varphi}=\frac{c}{\sqrt{a^6+\eta c^2}}\eqno(2-12)$$ where $c$ is integral constant. When $a(t)=0$, the kinetic energy $\dot{\varphi}^2=\displaystyle\frac{1}{\eta}$ is critical maximum. From Eqs.(2-5),(2-7),(2-9),(2-10),(2-12) we get $$\left(\frac{\dot{a}}{a}\right)^2=\frac{K}{3\eta}[\sqrt{1+\eta c^2a^{-6}}-1]\eqno(2-13)$$ $$\ddot{a}=-\frac{K}{3\eta}\left[2-\frac{2a^6-\eta c^2}{a^3\sqrt{a^6+\eta c^2}}\right]\eqno(2-14)$$ From Eq.(2-14) we can find that $\ddot{a}$ is always smaller than zero , no matter what the value of $a(t)$ is. When $a\rightarrow\infty$, then $\ddot{a}\rightarrow 0$. It shows that the universe starts with decelerated regime and gradually enters the zero acceleration. From Eq.(2-13) we get $$\dot{a}=\displaystyle\sqrt{\frac{K}{3\eta}\left[a^2\sqrt{1+\eta c^2 a^{-6}}-a^2~\right]}\eqno(2-15)$$ By Eq.(2-15), then we can find that the minimum of $a(t)$ is zero, assume $a(t)\mid_{t=0}=0$. When $a(t)\ll 1$, Eq.(2-15) can be approximated as $$\sqrt{a}~\dot{a}\approx\sqrt{\frac{Kc}{3}\sqrt{\eta}}\eqno(2-16)$$ $$a^{3/2}\approx\frac{3}{2}\sqrt{\frac{Kc}{3}\sqrt{\eta}}~t\eqno(2-17)$$ $$a(t)\sim t^{2/3}\eqno(2-18)$$ For the energy density $\rho_\varphi$ (2-9) and pressure $p_\varphi$ (2-10) of B-I scalar field, there is no violation of the strong energy condition. In universe with B-I scalar field without potential, there is no a phase of accelerating expansion.\
From Eqs.(2-9)(2-10)and (2-12), we have $$\omega=\frac{p_\varphi}{\rho_\varphi}=\frac{a^3}{\sqrt{a^6+\eta c^2}}\eqno(2-19)$$ and can see $$0\leq\omega\leq1\eqno(2-20)$$ In Sec.3, it is different that we find the universe accelerating expansion.
The Model with Lagrangian of $\displaystyle\frac{1}{\eta}[1-\sqrt{1-\eta
g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}]-u(\varphi)$
========================================================================
From Eq.(2-9) and lagrangian, we can get $$T^0_0=\rho_\varphi=u-\frac{1}{\eta}+\frac{1}{\eta\sqrt{1-\eta \dot{\varphi}^2}}\eqno(3-1)$$ $$-T^i_i=p_\varphi=\frac{1}{\eta}[1-\sqrt{1-\eta\dot{\varphi}^2}]-u\eqno(3-2)$$ When $u>0$, so we can always find $\rho_\varphi>0$.\
From Eqs.(3-1)and (3-2), we have $$\rho_\varphi+3p_\varphi=\frac{2}{\eta}-2u+\frac{3\eta\dot{\varphi}^2-2}{\eta\sqrt{1-\eta\dot{\varphi}^2}}\eqno(3-3)$$ When potential is greater than $\displaystyle\frac{1}{\eta}$ ($\displaystyle\frac{1}{\eta}$ is two times as “critical” kinetic energy of $\varphi$ field). And the kinetic energy of $\varphi$ field evolves to region of $\displaystyle\dot{\varphi}^2<\frac{2}{3\eta}$, we have $\rho_\varphi+3p_\varphi<0$ from Eq.(3-3).\
The universe undergoes a phase of accelerating expansion.\
From Eqs.(3-1)and (3-2), we also can get $$p_\varphi+\rho_\varphi=\frac{\dot{\varphi}^2}{\sqrt{1-\eta \dot{\varphi}^2}}>0\eqno(3-4)$$ Eq.(3-4) could be written as $$\omega=\frac{p_\varphi}{\rho_\varphi}>-1\eqno(3-5)$$ when $\dot{\varphi}$ approximation zero $p_\varphi=-\rho_\varphi$. The universe is dominated by the potential. It will undergo inflation phase. In next section, we consider the case that the kinetic energy term is negative.
The Model with Lagrangian of $\displaystyle\frac{1}{\eta}[1-\sqrt{1+\eta
g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}]-u(\varphi)$
========================================================================
In this section we consider the case that the kinetic energy term is negative. $$L=\frac{1}{\eta}[1-\sqrt{1+\eta
g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}]-u(\varphi)\eqno(4-1)$$ The Energy-moment tensor is $$\displaystyle T^\mu_\nu=-\frac{g^\mu_\nu \varphi_{,~\nu}\varphi_{,~\rho}}{\sqrt{1+\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}}-\delta^\mu_\nu L\eqno(4-2)$$ From Eq.(4-2), we have $$\rho=T^0_0=\frac{1}{\eta\sqrt{1+\eta \dot{\varphi}^2}}-\frac{1}{\eta}+u\eqno(4-3)$$ $$p=-T^i_i=\frac{1}{\eta}-\frac{\sqrt{1+\eta \dot{\varphi}^2}}{\eta}-u\eqno(4-4)$$ From Eqs.(4-3) and (4-4), we get $$\rho+p=-\frac{\dot{\varphi}^2}{\sqrt{1+\eta \dot{\varphi}^2}}\eqno(4-5)$$ It is clear that the equation of static $\omega<-1$ is completely confirmed by Eq.(4-5)and it accords with the recent analysis of observation data. We also can get $$\rho+3p=\frac{2}{\eta}\left[\frac{-2-3\eta \dot{\varphi}^2}{\sqrt{1+\eta \dot{\varphi}^2}}\right]-2u\eqno(4-6)$$ It is obvious that $\rho+3p<0$. Eq.(4-6)shows that the universe is undergoing a phase of accelerating expansion. The model of phantom B-I scalar field without potential $u(\varphi)$ is hard to understand. In this model we can always find $\displaystyle\rho=\frac{1}{\eta\sqrt{1+\eta\dot{\varphi}^2}}-\frac{1}{\eta}<0$ and $\displaystyle\left(\frac{\dot{a}}{a}\right)^2=\frac{K}{3}\rho<0$. It is unreasonable apparently. However, in the model of phantom B-I scalar field with potential $u(\varphi)$, when $\displaystyle
u(\varphi)>\frac{1}{\eta}-\frac{1}{\eta\sqrt{1+\eta\dot{\varphi}^2}}$, $\rho$ is always greater than zero. In phantom B-I scalar model with a potential $u(\varphi)$, we also find the strong and weak energy condition always failed from Eqs.(4-5)and (4-6).
We investigate the case of a specific simple example $u=u_0=const$ and $u_0-\displaystyle\frac{1}{\eta}=\frac{A}{\eta}(A>0)$. Eq.(4-3) becomes $$\displaystyle\rho=\frac{1}{\eta\sqrt{1+\eta\dot{\varphi}^2}}+
\frac{A}{\eta}\eqno(4-7)$$ It is clear that there is $\rho>0$ from Eq.(4-7). Substituting Eq.(4-7) into Einstein equation, we have $$\left(\frac{\dot{a}}{a}\right)^2=\frac{K}{3}\left[\frac{1}{\eta\sqrt{1+\eta\dot{\varphi}^2}}+\frac{A}{\eta}\right]\eqno(4-8)$$ The equation of motion of phantom field is $$\frac{1}{\sqrt{-g}}\frac{\partial}{\partial x^\nu}\left[\frac{\sqrt{-g}g^{\mu\nu}\varphi_{,~\mu}}{\sqrt{1+\eta g^{\mu\nu}\varphi_{,~\mu}\varphi_{,~\nu}}}\right]=0\eqno(4-9)$$ The field $\varphi$ only depends on $t$. We can obtain from Eq.(4-9) $$\dot{\varphi}=\frac{C}{\sqrt{a^6-\eta C^2}}\eqno(4-10)$$ where $C$ is integrate constant. Substituting Eq.(4-10) into Eq.(4-8), the Eq.(4-8)becomes $$\dot{a}=\displaystyle\sqrt{\frac{Ka^2}{3\eta}\left[\sqrt{1-\eta C^2a^{-6}}+A~\right]}\eqno(4-11)$$ The smallest $a_{\min}=(\eta C^2)^{1/6}$ from Eq.(4-11), the universe is non-singular. When the universe scalar approximation $a_{\min}$, Eq.(4-11) becomes $$\dot{a}=\sqrt{\frac{KA}{3\eta}}~a\eqno(4-12)$$ $$a\sim exp\left(\displaystyle \sqrt{\frac{KA}{3\eta}}~t\right)\eqno(4-13)$$ When $a\rightarrow\infty$, Eq.(4-11) becomes $$\dot{a}=\sqrt{\frac{K(A+1)}{3\eta}}~a\eqno(4-14)$$ $$a\sim exp\left(\displaystyle \sqrt{\frac{K(A+1)}{3\eta}}~t\right)\eqno(4-15)$$ In our phantom model with potential, the universe is always undergoing a phase of inflation and gradually enters the more accelerated expansion in late time.
Summary
=======
We considers cosmological solution of B-I type scalar field without the potential and come to the conclusion that the equation of state parameter $0<\omega<1$. However, in the dark energy models of canonical B-I scalar field with potential, the universe is undergoing a phase of accelerating expansion if the potential rolls down to the minimum which is greater than $\displaystyle\frac{1}{\eta}$ while scalar field evolves into region of $\displaystyle\dot{\varphi}^2<\frac{2}{3\eta}$. Correspondingly the equation of state parameter $\omega$ is always greater than $-1$. This model admits a late time attractor solution that leads to an equation of state $\omega=-1$. The lagrangian of B-I type scalar field with negative kinetic energy also is considered by us. In the phantom B-I scalar model, the universe is undergoing a phase of accelerating expansion and the equation of state parameter $\omega$ is always smaller than $-1$. It accords with the recent analysis of the observation that the equation of state parameter of dark energy might be smaller than $-1$.\
[**Acknowledgment**]{}\
This work was partially supported by National Nature Science Foundation of China.\
0.5in [**References**]{}
[^1]: $alberthq_-lu@hotmail.~com$
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider in this paper a two-pulse photon echo sequence as a potential quantum light storage protocol. It is widely believed that a two-pulse scheme should lead to very low efficiency and is then not relevant for this specific application. We show experimentally by using a Tm${}^{3+}$:YAG crystal that such a protocol is on contrary very efficient and even too efficient to be considered as a good quantum storage protocol. Our experimental work allows us to point out on one side the real limitations of this scheme and on the other side its benefits which can be a source of inspiration to conceive more promising procedures with rare-earth ion doped crystals.'
author:
- Jérôme Ruggiero
- 'Jean-Louis Le Gouët'
- Christoph Simon
- Thierry Chanelière
title: 'Why the two-pulse photon echo is not a good quantum memory protocol'
---
Introduction
============
The prospect of quantum light storage in solids motivates us to reconsider the interaction of light and matter at the single quantum level. Historically coherent transient phenomena appeared very active because they are primarily based on two-level system absorption. The community naturally focused on rare-earth ion doped crystals (REIC) instead of atomic vapors because the storage time can be long as well and is not limited by the atomic diffusion whatsoever. Recent proposals are of course inspired by previous realizations of classical light memories [@mossbergOL83] and all-optical processing more generally [@babbitAO94]. In this lineage recent progresses toward quantum storage are involving common physical ingredients. Because of the large inhomogeneous broadening, any light retrieval is intimately related to a dipole rephasing. Surprisingly the very well known conventional photon echo has not been considered as a promising quantum protocol despite impressive realizations in the classical domain. A major identified drawback is the low efficiency of the process [@moiseevPRL01; @moiseevJOB03]. This has been a commonly admitted idea since the pioneering studies of the 60’s [@hartmann64]. Although we roughly adhere to this statement, we would like to point out that high retrieval efficiency was already observed in a specific regime of two-pulse photon echo (2PE) [@cornishPRA98; @cornishOL00], which might pave the way to quantum light storage. Our experimental work consists of a clear observation of the large predicted efficiency in this regime we first need to precisely defined. Our analysis is also stressing clearly what are the physical ingredients which lead to this result. Our study is definitely placed in the prospect of a quantum memory. With respect to other protocols, we finally clarify the advantages and drawbacks of this technique that should be considered as a general tool for coherent manipulations.
As mentioned before, the investigation of classical light storage largely paved the way toward their quantum equivalent. REIC have shown interesting processing capabilities especially with all-optical control [@babbitAO94; @tian]. As derived from the conventional photon echo, these techniques are based on an optical manipulation of the coherences. Experiments largely benefit from the agility of the laser controlling the crystalline processor [@crozatier]. This convenience would be still appreciable for manipulation at the quantum level. The 2PE time-to-bandwidth product properties should also be emphasized. In the 2PE process, this parameter, critical for information processing applications, is not limited by the memory opacity, in contrast with the most promising quantum storage protocols involving REIC, namely the “stopped-light” approach [@EITlukin; @turukhin2002] or the “controlled reversible inhomogeneous broadening” (CRIB) procedure [@crib1; @alexander2006; @crib3; @hetet07]. Finally, contrary to the above storage protocols, doesn’t require any initial state preparation. A spectral selection within the inhomogeneous broadening is in a sense build-in because of the selective excitation of the first incoming pulse that we define as the signal. The 2PE has the singular advantage to rephase a random distribution of level shifts without any assumption on the source of inhomogeneity. By clarifying the characteristics of the 2PE, we aspire to a deeper understanding of the atomic coherences optical manipulation.
Since the observation [@hartmann64] and the interpretation [@hartmann66] of the photon echo, the quantitative comparison with the observed efficiency has been widely studied. At the basis of data processing application, the interest has been renewed relatively recently [@cornishPRA98; @mossbergPRA99]. A realistic approach of the problem usually requires a numerical resolution of the Bloch-Maxwell equations [@mossbergOC98]. This gives a solid interpretation of the experimental data [@mossbergPRA99]. As a matter of fact investigating the 2PE for quantum light storage is strongly simplifying the problem. As compared to the canonical 2PE where a $\pi/2$ excitation pulse is followed by a $\pi$ rephasing pulse, memory-like version of this scheme would first involve a very weak signal pulse. In that very specific case as pointed out by Tsang et al. [@cornishJosab03], one can derive an analytic solution for the efficiency. In this paper, we will first deduce these equations from a simple physical interpretation of the scheme. We will clearly specify the underlying assumptions to be verified in practice. We will then compare these calculations to the experiment. A detailed analysis of our observations allows us to conclude and place this work in the context of a quantum memory by comparison to other storage protocols.
Efficiency of the protocol
==========================
This subject has been covered by a wide range of literature [@cornishPRA98; @cornishOL00; @cornishJosab03; @mossbergPRA99; @mossbergOC98]. Nevertheless it is relatively easy to derive these equations based on simple physical arguments. The interaction of light pulses with our medium is well described by the Maxwell-Bloch equations assuming the slowly varying amplitude and the rotating wave approximations.
$$\label{MB}
\begin{array}{ll}
\partial_z\Omega(z,t)&=-\displaystyle\frac{\alpha}{2\pi}\int d\omega_{ab}\mathrm{v}(\omega_{ab};z,t)\\[0.4cm]
\partial_t\mathrm{u}(\omega_{ab};z,t)&=-\Delta \mathrm{v}(\omega_{ab};z,t)\\[0.3cm]
\partial_t\mathrm{v}(\omega_{ab};z,t)&=-\Omega(z,t)\mathrm{w}(\omega_{ab};z,t)+\Delta \mathrm{u}(\omega_{ab};z,t)\\[0.3cm]
\partial_t\mathrm{w}(\omega_{ab};z,t)&=\Omega(z,t)\mathrm{v}(\omega_{ab};z,t)
\end{array}$$
where $\Delta=\omega_{ab}-\omega_L$ is the detuning, $\Omega$ is the Rabi frequency of the field under consideration and $(\mathrm{u},\mathrm{v},\mathrm{w})$ the three components of the Bloch vector. The decay of the coherences and the population is assumed to be negligible. We have dropped the usual term $\partial_t\Omega(z,t)$ because realistically the spatial extension of the pulse is always much longer the length of our crystal.
To describe the broadest range of situation a numerical resolution of the system is usually necessary [@mossbergOC98]. This is not our approach. A sketch of the time sequence is depicted in Fig.\[fig:seqT\] and looks like any 2PE sequence. Nevertheless within the prospect of quantum storage, the signal and the echo are assumed to be weak. This greatly simplifies the description [@crisp] essentially because these two fields do not modify the population difference $\mathrm{w}(\omega_{ab};z,t)$ that is not time dependent anymore. This is the small area approximation where the Maxwell-Bloch system can be linearized [@crisp].
![Outline of the time sequence. The signal is first absorbed in the medium. After a time $t_{12}$ a rephasing pulse of large area A induces a build-up of the coherence at the time $2t_{12}$ and give rise to the photon echo. We address the problem in the weak signal and the echo regime, which shall be satisfied in quantum field conditions.[]{data-label="fig:seqT"}](seqT.eps){width="6cm"}
Calculation in the weak signal limit
------------------------------------
If the three pulses are well separated in time, one can consider them independently. The signal, the rephasing pulse and the echo are respectively centered on $t=0$, $t_{12}$ and $2t_{12}$. The incoming signal and outgoing echo Rabi frequencies are respectively denoted $\mathcal{S}$ and $\mathcal{E}$. The propagation of $\mathcal{S}$ is simply described by an absorption law if the atoms are initially in the ground state $\mathrm{w}(\omega_{ab};z,-\infty)=\mathrm{w}(\omega_{ab};z,0)=-1$ [@crisp]: $$\label{AbsS}
\partial_z\mathcal{S}(z,t)=\mathrm{w}(\omega_{ab};z,0) \frac{\alpha}{2} \mathcal{S}(z,t) = -\frac{\alpha}{2} \mathcal{S}(z,t)$$
The weak echo is expected to behave in a similar way, except the medium has been previously excited and modified by the signal and the rehasing pulse. Therefore the echo equation reads as: $$\label{AbsE}
\begin{array}{ll}
\partial_z\mathcal{E}(z,t)=&\mathrm{w}(\omega_{ab};z,2t_{12}) \frac{\alpha}{2} \mathcal{E}(z,t) \\
&-\displaystyle\frac{\alpha}{2\pi}\int d\omega_{ab}\mathrm{v_E}(\omega_{ab};z,t)
\end{array}$$ The coherence $\mathrm{v_E}$, resulting from interaction with the first two pulses, evolves freely within the time interval $t_{12} \rightarrow 2t_{12}$. The population $\mathrm{w}(\omega_{ab};z,2t_{12})$ has been affected by the rephasing pulse. We assume the rephasing pulse is much shorter than $\mathcal{S}$. Therefore the population is uniformly modified by the rephasing pulse all over the spectral interval initially excited by $\mathcal{S}$. We will see in \[sec:exp\] how this constraint is treated experimentally. A second benefit of this assumption is that the rephasing pulse can be consider as instantaneous (time $t_{12}$) and then uniformly modifying the coherences. The physical interpretation of Eq.\[AbsE\] is based on the generation of $\mathcal{E}$ by the macroscopic dipole that builds up when the coherences $\mathrm{v_E}$ get phased together. The growing field propagates through the medium characterized by the uniform population difference $\mathrm{w}(\omega_{ab};z,2t_{12})$.
We aim at reducing Eq.\[AbsE\] by expressing the atomic quantities in terms of the optical fields only. We shall be left with an equation of propagation for $\mathcal{E}$. The echo efficiency will be deduced from the solution of this equation. We first have to track the excitation of the coherences by the signal, then their modification by the rephasing pulse and finally their free evolution toward the echo emission. The problem is addressed locally, at position $z$. The signal $\mathcal{S}$ excites the atoms that initially all sit in the ground state: $$\begin{array}{ll}
\mathrm{u}(\omega_{ab};z,t)&+i\mathrm{v}(\omega_{ab};z,t)=\\[0.3cm]
&i\exp\left(i\Delta t\right) \displaystyle\int_{-\infty}^{t}\mathcal{S}(z,\tau)\exp\left(-i\Delta\tau\right)\mathrm{d}\tau
\end{array}$$ At a certain time $t$ between $0$ and $t_{12}$ the signal field is off. We then recognize the Fourier transform of $\mathcal{S}$ written $\widetilde {\mathcal{S}}$, the $\exp\left(i\Delta t\right)$ accounts for the free evolution during this interval. $$\label{UV}
\mathrm{u}(\omega_{ab};z,t)+i\mathrm{v}(\omega_{ab};z,t)= i\exp\left(i\Delta t\right) \widetilde {\mathcal{S}}\left(z,\Delta \right)$$ This expression represents the evolution of the coherence after the initial absorption process and before the rephasing pulse. Next, we calculate this strong pulse effect on the coherences (how $\mathrm{v_E}$ is related to $\mathrm{v}$) and on the population, which modifies the echo propagation (Eq.\[AbsE\]). This can be done analytically by integrating the Bloch-Maxwell equations (see ref. [@cornishJosab03]). Nevertheless the results are relatively intuitive at the end and can be derived from simple physical ingredients. This can be done first of all in the specific case of a $\pi$-rephasing pulse, the more general case of an area $A(z)$ for the strong pulse can be solved by introducing “by hand” geometrical factors. All over the spectral interval excited by the signal pulse, the rephasing pulse is assumed to behave as a $\pi$-pulse. This corresponds to a brief pulse assumption. The experimental fulfillment of this condition will be addressed in \[sec:exp\]. A $\pi$-pulse simply drives the Bloch vector by a rotation of $\pi$ around an equatorial axis. On the one hand, along the population axis it corresponds to an inversion from $-1$ to $\mathrm{w}(\omega_{ab};z,2t_{12})=1$ at the time $t_{12}$. On the other hand, it transforms the coherences $\mathrm{v}(\omega_{ab};z,t_{12})\rightarrow-\mathrm{v}(\omega_{ab};z,t_{12})$, $\mathrm{u}(\omega_{ab};z,t_{12})$ stays the same (complex conjugation of $\mathrm{u}+i\mathrm{v}$). Right after the rephasing pulse, Eq. \[UV\] becomes $$\label{UEVE}
\begin{array}{ll}
\mathrm{u_E}(\omega_{ab};z,t)&+i\mathrm{v_E}(\omega_{ab};z,t)=\\[0.3cm]
& i\exp\left(i\Delta \left(t-t_{12}\right)\right) \widetilde {\mathcal{S}}^*\left(z,\Delta \right)\exp\left(-i\Delta t_{12}\right)
\end{array}$$ where the complex conjugation sign $*$ accounts for the rephasing transformation. The coherences are freely evolving after $t_{12}$. One can now write the propagation equation of the echo (Eq.\[AbsE\]) by including the direct influence of the signal on $\mathrm{v_E}(\omega_{ab};z,t)$. The signal field being assumed to be a real number, one recognizes the time-reversed signal $\mathcal{S}(z,2t_{12}-t)$ whose $z$ dependency is given by an absorption law $\mathcal{S}(z,2t_{12}-t)=\mathcal{S}(0,2t_{12}-t)\exp\left(-\alpha z\right)$(Eq.\[AbsS\]) $$\label{EfinalPI}
\partial_z\mathcal{E}(z,t)=+\frac{\alpha}{2} \mathcal{E}(z,t)-\alpha \mathcal{S}(z,2t_{12}-t)$$ The signal field acts as a source and generates the echo that propagates in an inverted medium. This gives the equation of propagation for a $\pi$-rephasing pulse. It is now rather easy to account for an imperfect rephasing. More generally, a $A(z)$-area strong pulse drives an $A(z)$-rotation of the Bloch vector. The population is not fully inverted anymore: $\mathrm{w}(\omega_{ab};z,2t_{12})=-\cos A\left(z\right)$. The rotation of the coherences is also incomplete and limited to $\left[1-\cos\left(A\left(z\right)\right)\right]/2$ of its maximum value. This factors are purely geometrical and are interpreted as projections on the Bloch sphere. We finally get the general analytic expression for the efficiency. This expression has been previously derived by Tsang *et al.* [@cornishJosab03 Eq.(40)] by integrating the Bloch-Maxwell equations. Here we simply focus on the two crucial stages, the absorption of the signal on one side and the re-emission of the echo on the other side. The rephasing pulse in between is interpreted as an instantaneous manipulation of the Bloch vector. $$\label{Efinal}
\begin{array}{ll}
\displaystyle\partial_z\mathcal{E}(z,t)=&-\cos\left(A\left(z\right)\right)\frac{\alpha}{2} \mathcal{E}(z,t)\\[0.4cm]
&-\displaystyle\frac{1-\cos\left(A\left(z\right)\right)}{2}\alpha \mathcal{S}(z,2t_{12}-t)
\end{array}$$ There is an underlying assumption here: the rephasing pulse is very brief and is then fully covering the spectral range of excitation. As a consequence the echo is not deformed as compared to the signal, it is only time reversed. The $A(z)$-rotation on the Bloch sphere is also uniform and doesn’t depend on $\Delta$. The $z$-dependency of the area accounts for the propagation of the strong pulse itself. This is usually a complicated problem but in that case we are only interested in the propagation of the area. The result is remarkably simple and is given by the Area Theorem of McCall & Hahn [@mccall1967]. We don’t have to know the exact temporal shape through the propagation because the area is the relevant quantity for the rephasing pulse and is simply given by $$\label{Area}
\partial_z A\left(z\right)=-\frac{\alpha}{2}\;\sin A\left(z\right)$$ It can be solve analytically for a given $A\left(0\right)$. A straightforward integration of Eq.\[Efinal\] allows us to calculate the retrieval efficiency $\eta$ as a function of the optical thickness $\alpha L$ where $L$ is the length of the medium: $$\label{eta}
\begin{array}{ll}
\displaystyle\eta\left(A\left(0\right),\alpha L\right)&=\displaystyle\left(\frac{\mathcal{E}(L,t)}{\mathcal{S}(0,2t_{12}-t)}\right)^2\\[0.4cm]
&=\displaystyle\left(\frac{2\sinh\left(\alpha L\right)}{1+\exp\left(\alpha L\right)\cot^2\left(A\left(0\right)/2\right)}\right)^2
\end{array}$$ For different optical thickness $\alpha L$, we plot the efficiency as a function of the rephasing pulse area $A\left(0\right)$ in Fig.\[fig:PlotEtaVSAi\].
![Efficiency of the 2PE as a function of the rephasing pulse area $A\left(0\right)$ for different optical thickness $\alpha L$. Inset: for larger $\alpha L$, the efficiency is much larger than unity at the maximum rephasing.[]{data-label="fig:PlotEtaVSAi"}](PlotEtaVSAi.eps){width="8cm"}
We observe in that case that the efficiency is strongly depending on the optical thickness. When it is low, the efficiency is weak essentially because the signal is poorly absorbed. The efficiency is then a sinus-like function and the rephasing area directly accounts for an imperfect rotation on the Bloch sphere. On the other side, at large optical thickness, the efficiency can be much larger than unity but only on a narrow window around a $\pi$. This specific situation is certainly the most interesting because the signal in fully mapped into the medium (large optical thickness) and the efficiency is exceptionally large. The $\pi$-rephasing pulse have very particular properties that can be interpreted independently to physically understand this behavior.
Specificity of the $\pi$-rephasing pulse {#specPI}
----------------------------------------
The case of an exact $\pi$-rephasing pulse is relatively straightforward. More fundamentally it allows us to understand the expected efficiency at large optical thickness. We simply derived the propagation equation (\[EfinalPI\]) for the echo by assuming that locally, at position $z$, the area is exactly $\pi$. Now we have to examine the $\pi$-pulse propagation inside the medium. This is a very specific situation. According to Eq.\[Area\] a $\pi$-pulse preserves its initial area throughout the medium. Even if the energy is absorbed, the area is conserved as the pulse stretches temporally [@ruggiero]. Roughly speaking, if the pulse is elongated by a factor $r$, the amplitude (Rabi frequency) is reduced by $r$ to conserve the area and the energy decreases by a factor $r$. This alteration is a pure coherent propagation effect.
The propagation equation \[EfinalPI\] is then valid at any position $z$ and easily gives the efficiency. This expression is consistent with our general formula Eq. \[eta\]: $$\label{etaPi}
\eta\left(\pi,\alpha L\right)=\left[\exp\left(\alpha L/2\right)-\exp\left(-\alpha L/2\right)\right]^2$$ At large optical thickness, the efficiency is much larger than unity and grows exponentially. This is relatively counter-intuitive. The echo efficiency is generally observed to be low, which is usually assigned to absorption. As mentioned before the $\pi$-rephasing pulse retains its area along the propagation. In other word, the medium is completely inverted: the echo is emitted in an amplifying medium. This explains why the echo is gaining exponentially. Practically, the assumptions we made, such as 1-dimension, infinite plane wave geometry, will be difficult to satisfy. Any divergence from the ideal theoretical frame shall affect the echo efficiency.
The $\pi$-pulse propagation is not only unusual, it is also a singular solution of the Area Theorem. The $\pi$-solution is indeed not stable because any area slightly lower (or larger) than $\pi$ will decrease (or increase resp.) toward $0$ (or $2\pi$ resp.) [@mccall1967; @ruggiero]. Even so, the pulses with an area close to $\pi$ can propagate deeper inside the medium than a weak pulse. To see that we plot in Fig. \[fig:PlotLeff\] the penetration depth $L_{P}$ at which the incoming area is divided by $1/\sqrt{e}$ (the curve is symmetrized around $\pi$ to account for the deviation toward $2\pi$ of pulses larger than $\pi$).
![Penetration depth $L_{P}$ of a strong pulse in unit of $\alpha^{-1}$ as a function of the incoming area. Inset: Penetration of pulses with area of $\pi$ plus or minus 1%. For small area pulses, it is simply given by an absorption law $L_{P}=\alpha^{-1}$.[]{data-label="fig:PlotLeff"}](PlotLeff.eps){width="8cm"}
For small areas, this length is $\alpha^{-1}$ as expected [@crisp]. As one gets closer to $\pi$, the pulse can propagate deeper inside the medium. For example, if the incoming area is controlled at the 1% level (inset Fig. \[fig:PlotLeff\]), the penetration depth is larger than $7 \alpha^{-1}$. The curve is very narrow about $\pi$ which shows a high sensitivity for the propagation. Realistically a well-controlled $\pi$-pulse should induce a population inversion much deeper inside the medium than the absorption length $\alpha^{-1}$. If this depth is larger than the optical thickness of the medium, it will be fully inverted. However, the deeper the $\pi$-pulse propagates through the absorbing medium, the more it is stretched, since it has to keep a constant area while losing energy. The pulse bandwidth shrinks accordingly, making the pulse act as a $\pi$-pulse on a reduced spectral interval [@ruggiero].
Based on this analysis, we expect to observe two remarkable qualitative features. The efficiency should be very high at large optical thickness and strongly depending on the area of the rephasing pulse.
Experiments {#sec:exp}
===========
A 2PE experiment can be performed in any system were a transient phenomena can be observed. Here we use a thulium-doped yttrium aluminum garnet (YAG), cooled down to cryogenic temperature. The long optical coherence time makes it particularly attractive for quantum storage application. We will now briefly describe the experimental set-up and focus on the precautions we take to satisfy the assumptions introduced previously.
Experimental set-up
-------------------
Our 0.5% [Tm$^{3+}$:YAG ]{}crystal is immersed in liquid helium at 1.4K. The coherence time of the $^{3}$H$_{6}$(0) to $^{3}$H$_{4}$(0) transition is typically $T_2=50\mu s$ in these conditions. The crystal is oriented and cut in order to propagate along the \[1$\mathrm{\bar{1}}$0\] direction. Along this axis, the length is 5mm and the optical thickness $\alpha L=5$. The laser polarisation is parallel to \[111\] to maximize the Rabi frequency [@ConeNutAngle]. The laser system is operating at 793 nm, stabilized on a high-finesse Fabry-Perot cavity with the Pound-Drever-Hall technique (200Hz over 10ms) [@crozatier2004]. The laser is split in two independent beams. Temporal shaping is achieved by two acousto-optic modulators (AOM) controlled by a dual channel arbitrary waveform generator (Tektronix AWG520). Both beams are injected into two single mode fibers. Before recombination on a beamsplitter, we use expanders to independently manage their waists inside the crystal. After the sample, the signal is collected in a single mode fiber terminated by a photodiode.
The signal is supposed to mimic a weak quantum field, so this pulse should verify the small area approximation. The signal beam is in practice much weaker than the rephasing one. At the maximum, there is a 36dB power difference between the two. More precisely, the signal area is kept constant at 9% of $\pi$ and we vary the rephasing pulse area $A\left(0\right)$ from 0 to $~3\pi$.
Keeping the rephasing pulse significantly briefer than the signal pulse is the most stringent condition we have to satisfy. This is required to maintain a uniform coverage of the rephasing process over the signal excitation bandwidth. To do so, we use gaussian-shaped signal (duration $\mathrm{2.1\mu s}$) and rephasing pulses. On the one hand a gaussian pulse is spectrally narrower than a rectangular pulse with the same duration. On the other hand we observed a gaussian $\pi$-pulse undergoes less temporal streching than a gaussian one after propagation through an absorbing medium. We set the rephasing pulse $\approx2.5$ times shorter than the signal. This value is slightly fluctuating depending on the rephasing amplitude value. Changing the AOM driving power marginally impacts on the pulse shape. A much shorter pulse would be preferable but we are limited by the available power (few milliwatts) to ensure a significant area in a reasonable time.
One last point we ignored so far is the transverse dimension of the beams. To be consistent with the 1-dimension theory, the power of the rephasing beam should be constant over the spatial extension of the signal. The signal beam waist ($\mathrm{17 \mu m}$) is then chosen to be 2 times smaller than the waist of the rephasing beam ($\mathrm{35 \mu m}$). This is the same overlap argument we used in the spectral domain.
Results
-------
We perform a 2PE experiment in the beforementioned conditions (see Fig. \[fig:TimeSeqArticle\]). We pay special attention to an accurate calibration of the rephasing pulse area. We indeed first perform an optical nutation experiment to evaluate the exact Rabi frequency of the pulse. Comparing the signal and rephasing beam intensities, we estimate the area of the signal which is confirmed to be weak ($0.09\pi$).
![(color online) Time sequence for $1.1\pi$-rephasing pulse. The signal and the echo being very weak, we magnify their scale by a factor 150 (solid red line). The intensity is given as the square of the Rabi frequency. The signal and the rephasing pulse have a duration of $\mathrm{2.1\mu s}$ and $\mathrm{0.8\mu s}$ respectively (rms width of the gaussian). We also represent the rephasing pulse after transmission (dashed line in Arb. units).[]{data-label="fig:TimeSeqArticle"}](TimeSeqArticle.eps){width="8cm"}
As expected, we observe an echo at delay $t_{12}$ after the rephasing pulse (see Fig. \[fig:TimeSeqArticle\] ). We also carefully calibrate the efficiency. To do so, we shift the laser far from the absorption line (a few cm$^{-1}$). The corresponding intensity level represents the 100% reference line on the measurement detector. According to the efficiency definition given by Eq. \[eta\], the echo is assumed to exhibit the time-reversed temporal shape of the signal. This is not exactly the case experimentally as we shall discuss later. So we define the efficiency by comparing the maxima of the two pulses: $$\label{etaEXP}
\eta_\mathrm{exp}\left(A\left(0\right)\right)=\left(\frac{\max_t\left(\mathcal{E}(L,t)\right)}{\max_t\left(\mathcal{S}(0,t)\right)}\right)^2$$ By varying the incoming rephasing area, we obtain $\eta_\mathrm{exp}$ as a function of $A\left(0\right)$ (see Fig. \[fig:EffArticle\]).
![Efficiency of the echo retrieval as a function of the rephasing pulse area $A\left(0\right)$. The maximum (circle) is obtained for $A\left(0\right)=1.1\pi$ where the efficiency is 92%, this corresponds to the situation depicted in Fig. \[fig:TimeSeqArticle\].[]{data-label="fig:EffArticle"}](EffArticle.eps){width="8cm"}
The main source of uncertainty is due to the alignement and the spatial overlap of the beams in the crystal. To quantify it and derive error bars, we estimate the typical intensity variation of the rephasing beam (waist $\mathrm{35 \mu m}$) over a length corresponding to the signal waist ($\mathrm{17 \mu m}$). This is simply given by the direct comparison of two gaussian curves (11% in that case, which gives the horizontal error bars in Fig. \[fig:EffArticle\]).
Discussion
----------
The experimental results are qualitatively in good agreement with the expected efficiencies in Fig. \[fig:PlotEtaVSAi\]. The main features are indeed observed. First of all the maximum efficiency is obtained for a $1.1\pi$-area, which is consistent with $\pi$, within the error bar. We also observe a reincrease of the efficiency close to $3\pi$ after a minimum at $2\pi$. We certainly predict a maximum at any odd number of $\pi$, but we are experimentally limited by the available laser power. Secondly, the curve is peaked in the vicinity of $\pi$ and cannot be fitted by a sinus-shape oscillation. This is also expected (see Fig. \[fig:PlotEtaVSAi\]) and is due to the large optical thickness of the sample $\alpha L\simeq5$. Finally, the maximum efficiency is relatively high 0.92 for $A\left(0\right)=1.1\pi$ (circle in Fig. \[fig:TimeSeqArticle\]). Although far below the $\eta\left(\pi\right)=146$ predicted value (Eq. \[etaPi\]), this result demonstrates a highly efficient 2PE.
We can invoke many reasons to explain the discrepancy between the measured and the predicted efficiency values. (i) The first obvious one is the total duration of the time sequence. The echo is indeed decaying exponentially because of the coherence lifetime $T_2=50\mu s$ which has been completely neglected in our treatment. In our case $t_{12}=\mathrm{25\mu s}$ so the echo is reduced by a factor $\exp(4t_{12}/T_2)=7.4$. Without this decay, the efficiency would be much larger than unity. In section \[sec:QM\] we shall see why $t_{12}$ is chosen to be long for this experiment. (ii) Another limitation is certainly due to the duration of the rephasing pulse. With this pulse 2.5 times shorter than the signal, the incoming pulse spectral overlap is rather good. However, propagation through the sample strongly stretches the rephasing pulse, as expected from discussion in Sec. \[specPI\], and as observed in Fig. \[fig:TimeSeqArticle\]. The pulse cannot be considered as much briefer than the incoming signal all the way through the sample. An observable proof of this effect is the retrieval time of the echo (Fig. \[fig:TimeSeqArticle\]). The retrieval should be centered on $t=2t_{12}=\mathrm{50\mu s}$. We clearly see that the echo is delayed by few microseconds. Indeed, because of stretching the rephasing pulse is no longer centered at a delay $t_{12}$ from the signal. The retrieval time is shifted accordingly. Since one of the assumption of our model is not fully verified, we then expect an efficiency reduction. (iii) In the spatial domain, the same argument is also valid. The signal is tightly focused ($\mathrm{17 \mu m})$ to ensure that its waist is smaller that the rephasing beam. The associated confocal parameter is typically two time shorter than the crystal length. So the rephasing beam does not overlap the signal uniformly all along the propagation. In other words, the rephasing area is not constant in the transverse direction. This should reduce the efficiency and broaden the peak around $\pi$ (convolution effect).
Based on this analysis, we believe our model contains all the physical ingredients to explain qualitatively the experimental results. We have given three probable explanations to interpret the quantitative discrepancy with the predicted values.
Relevance for quantum memory application {#sec:QM}
========================================
Our experiment has been performed in the classical domain using weak small-area pulses. It tells us however what should be the limitations in the quantum domain.
An obvious one is already present in our experiment. Since the rephasing pulse stretches while propagating through the sample, it gains a trailing tail that is not negligible as compared to the echo amplitude. As we can see in Fig. \[fig:TimeSeqArticle\] at $t=\mathrm{45\mu s}$ the pulse tail is falling slightly before the echo comes out. That’s the reason why we cannot make $t_{12}$ shorter, otherwise the echo would be submerged. With only few photons in the signal, this effect would be disastrous. As already discussed, the strong pulse distortion is not an artefact. This is a coherent propagation effect [@ruggiero], thus a fundamental limitation. This should not be confused with the noise induced by the fluorescence, which will be another limitation at the few photons level.
The rephasing process is inherently associated with a population inversion. The decay has been neglected in our model. In practice the medium excitation will be followed by spontaneous emission. The fidelity of the 2PE as a quantum memory protocol is fundamentally limited by fluorescence. This can already be understood within the framework of the Dicke model [@dicke], i.e. without taking propagation effects into account. Consider an ensemble of two-level systems, where the two states are denoted $|g\rangle_k$ and $|e\rangle_k$ for the $k$-th system. Note that we are interested in the case where the transition energy for the $g$ to $e$ transition is slightly different for different systems (inhomogeneous broadening).
We will compare the case where the input to be stored is a single photon to the case where there is no input (i.e. where the input state is the vacuum). The initial state of the atomic ensemble is $|\psi^0\rangle=|g\rangle_1
|g\rangle_2 ... |g\rangle_N$. For a vacuum input, this state remains of course unchanged. It is then transformed to $|\psi^N\rangle=|e\rangle_1 |e\rangle_2 ... |e\rangle_N$ by the $\pi$-pulse. For a single photon input, absorption of the photon creates a state of the form $$\label{psi1}
|\psi^{1}\rangle=\frac{1}{\sqrt{N}}\left( |e\rangle_1 |g\rangle_2 ... |g\rangle_N +...+ |g\rangle_1 |g\rangle_2 ... |e\rangle_N\right)$$ which contains a single atomic excitation. (Propagation effects would lead to the coefficients of the $N$ terms not being all the same.) The $\pi$-pulse transforms this state into $$\label{psiminus1}
|\psi^{N-1}\rangle=\frac{1}{\sqrt{N}}\left( |g\rangle_1 |e\rangle_2 ... |e\rangle_N +...+ |e\rangle_1 |e\rangle_2 ... |g\rangle_N\right)$$ which has $N-1$ atomic excitations.
In an inhomogeneously broadened system, the various terms in Eqs. (\[psi1\],\[psiminus1\]) will acquire different phases depending on the transition energies of the various atoms. However, at the time of the echo all terms will be in phase. Emission from the state $|\psi^{N-1}\rangle$ gives rise to the echo signal corresponding to a single-photon input, whereas emission from the state $|\psi^N\rangle$ corresponds to a vacuum input and thus defines the noise background due to fluorescence. The photon emission probability for a state $|\psi\rangle$ is proportional to $||J_-|\psi\rangle||^2$, where $J_-=\sum_{k=1}^N |g\rangle_k \langle e|_k$. This is due to the fact that the interaction Hamiltonian between the atomic ensemble and the relevant mode $a$ of the electro-magnetic field (corresponding to emission in the direction of phase matching) is proportional to $a^{\dagger} J_-+h.c.$.
Following Ref. [@dicke] it is easy to see that $||J_-|\psi^{N-1}\rangle||^2=2(N-1)$ and $||J_-|\psi^{N}\rangle||^2=N$. As a consequence, the probability to emit a photon at the echo time is only twice as large for a single-photon input as for no input at all, corresponding to a signal-to-noise ratio of one. This severely limits the achievable fidelity of quantum state storage.
Finally, from what we have shown is this article, we can conclude that the efficiency is actually too high for quantum memory application. We indeed observed a maximum 92% retrieval but the efficiency can be much larger than unity with optimized conditions [@cornishOL00]. This amplification due to the medium inversion is precisely a propagation effect that is not considered in the beforementioned Dicke model. This is a key ingredient to interpret our experimental results. In quantum optics terms, the statistics of the field will be modified: for one photon coming in, more than one will come out. Because the medium is inverted, it actes as a gain medium and modifies the quantum field and then again reduces the fidelity [@GainMed]. In that sense, the 2PE is also too efficient to be a good quantum protocol.
Those reasons are three fundamental limitations that we expect in the quantum regime. Even if the 2PE suffers from drastic drawbacks, it should be considered with attention. It is not only a historical example that helps us to understand rephasing phenomena. It has the unique ability to rephase atoms with randomly distributed level shifts, whatever the distribution structure [@crib3]. This a major difference as compared to CRIB (Controlled Reversible Inhomogeneous Broadening) [@crib1; @hetet07; @crib3; @alexander2006] where the first initial step is a spectral tailoring of the distribution. There is not preparation of the system in the 2PE. The feat performed by the CRIB protocol is its capability to produce a rephasing of the coherence without any population inversion. As a consequence the equations of the 2PE (Eq. \[EfinalPI\]) and the CRIB [@sangouard] are remarkably similar except that a minus sign accounts for the population inversion in the propagation equation. One can finally wonder if an optical manipulation would achieve a rephasing in the ground state as the CRIB does. This is *a priori* not possible because even a complex optical sequence will be decomposed with rotations on the Bloch sphere, on contrary the CRIB protocol can be interpreted as a planar symmetry (detuning sign reversal). These two are then intrinsically and fundamentally different.
We have here listed the limitations of the 2PE when considered as a potential quantum storage protocol. We pointed out the pulse deformation than can be a technical issue when using strong pulse. The two other limitations are directly and fundamentally related to the medium inversion induced by the optical rephasing operation. On one side the spontaneous emission will produce a noise comparable to the retrieved signal and then deteriorate the storage fidelity. On the other side the inversion will make the medium amplifying, which mainly explains the large predicted and observed efficiency. A larger than one efficiency is also associated with a fidelity reduction for quantum fields.
Conclusion
==========
In this article, we study the 2PE efficiency within the context of quantum storage. In this framework, we experimentally observe large efficiencies that are well explained by a simple model. Our calculations are based on a physical analysis of this specific situation. The experimental set-up has been devised to verify the underlying assumptions of the model. In 2PE, rephasing goes along with population inversion. This is a crucial ingredient of this protocol. The emitted echo is then widely amplified and can be stronger than the incoming signal. We have observed this effect. We finally conclude by analysing the potential extension of this work at low light level. We have pointed out the inherent limitations of the process.
By clarifying the physics involved in the very well-known two-pulse photon echo, we more generally tackle the problem of using strong light pulses for rephasing purposes. Our study should then be considered as a tool for the conception of new quantum storage protocols.
[99]{}
N. Carlson, L. Rothberg, A. Yodh, W. Babbitt, and T. Mossberg, Opt. Lett. **8**, 483 (1983). W. R. Babbitt and J. A. Bell, Appl. Opt. **33**, 1538 (1994).
S. A. Moiseev and S. Kröll, Phys. Rev. Lett. **87**, 173601 (2001).
S. A. Moiseev, V. F. Tarasov and B. S. Ham, J. Opt. B: Quantum Semiclass. Opt. **5**, S497 (2003).
I. D. Abella , N. A. Kurnit, and S. R. Hartmann, Phys. Rev. Lett. **13**, 567 (1964)
C. Sjaarda Cornish, W. R. Babbitt, and L. Tsang, Opt. Lett. **25**, 1276 (2000). M. Azadeh, C. Sjaarda Cornish, W. R. Babbitt, and L. Tsang, Phys. Rev. A **57**, 4662 (1998). M. Tian, F. Grelet, I. Lorgeré, J.-P. Galaup, J.-L. Le Gouët, J. Opt. Soc. Am. B **16**, 74 (1999).
V. Crozatier, G. Gorju, J.-L. Le Gouët, F. Bretenaker and I. Lorgeré, Optics Letters **31**, 3264 (2006)
M. Fleischhauer and M. D. Lukin, Phys. Rev. A **65**, 022314, 2002.
A. V. Turukhin, V. S. Sudarshanam, M. S. Shahriar, J. A. Musser, B. S. Ham and P. R. Hemmer, Phys. Rev. Lett. **88**, 023602 (2002).
M. Nilsson and S. Kröll, Opt. Commun. **247**, 393 (2005).
A. L. Alexander, J. J. Longdell, M. J. Sellars and N. B. Manson, Phys. Rev. Lett. **96**, 043602 (2006).
B. Kraus, W. Tittel, N. Gisin , M. Nilsson, S. Kroll, J.I. Cirac, Phys. Rev. A **73**, 020302(R) (2006).
G. Hétet, J. J. Longdell, A. L. Alexander, P. K. Lam and M. J. Sellars, Phys. Rev. Lett. **100**, 023601 (2008).
I. D. Abella , N. A. Kurnit, and S. R. Hartmann, Phys. Rev. **141**, 391 (1966)
T. Wang, C. Greiner, J. R. Bochinski, and T. W. Mossberg, Phys. Rev. A **60**, R757 (1999). T. Wang, C. Greiner, and T. W. Mossberg, Opt. Commun. **153**, 309 (1998). L. Tsang, C. Sjaarda Cornish, and W. R. Babbitt, J. Opt. Soc. Am. B,**20** 379 (2003) M. D. Crisp, , Phys. Rev. A **1**, 1604 (1970).
S. L. McCall , E. L. Hahn, Phys. Rev. Lett. [**18**]{}, (1967) 908;
J. Ruggiero, T. Chanelière, J.-L. Le Gouët, submitted to Phys. Rev. A.
Y. Sun, G. M. Wang, R. L. Cone, R. W. Equall, M. J. M. Leask, Phys. Rev. B **62**, 15443 (2000).
V. Crozatier, F. de Seze, L. Haals, F. Bretenaker, I. Lorgeré and J.-L. Le Gouët, Opt. Comm. [**241**]{}, 203 (2004)
R.H. Dicke, Phys. Rev. [**93**]{}, 99 (1954).
C. Simon, G. Weihs, and A. Zeilinger , Phys. Rev. Lett., [**84**]{}, 2993 (2000). N. Sangouard, C. Simon, M. Afzelius and N. Gisin, Phys. Rev. A **75**, 032327 (2007).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The large scale properties of spatiotemporal chaos in the 2d Kuramoto-Sivashinsky equation are studied using an explicit coarse graining scheme. A set of intermediate equations are obtained. They describe interactions between the small scale (e.g., cellular) structures and the hydrodynamic degrees of freedom. Possible forms of the effective large scale hydrodynamics are constructed and examined. Although a number of different universality classes are allowed by symmetry, numerical results support the simplest scenario, that being the KPZ universality class.
PACS numbers: 87.10.+e, 87.22.Jb, 07.50.Qx, 02.50.Ey
address: |
$^{1}$ Center for Computational Science, Boston University, Boston MA 02215.\
$^{2}$Department of Mathematics, University of Pittsburgh, Pittsburgh PA 15260.\
$^{3}$Department of Physics, University of California at San Diego, La Jolla, CA 92093-0319
author:
- 'Bruce Boghosian$^{1}$, Carson C. Chow$^{2,*}$, and Terence Hwa$^{3}$'
title: 'Hydrodynamics of the Kuramoto-Sivashinsky Equation in Two Dimensions'
---
[2]{}
A major goal in the study of spatiotemporal chaos (STC) [@cross93] is to obtain quantitative connections between the chaotic dynamics of a system at small scales and the apparent stochastic behavior at large scales. The Kuramoto-Sivashinsky (KS) equation [@ksrefs] $$\partial_t h = -\nabla^2 h - \nabla^4h +(\nabla h)^2
\label{KS}$$ has been used as a paradigm in efforts to elucidate the micro-macro connections [@yakhot81; @zaleski].
The qualitative behavior of the KS equation is quite simple. Cellular structures are generated at scales of the order $\ell_0 =
2\sqrt{2}\pi$ due to the linear instability. These cells then interact chaotically with each other via the nonlinear spatial coupling to form the STC steady state at scales much larger than $\ell_0$. The characterization of the STC state has been studied extensively in one spatial dimension [@yakhot81; @zaleski; @sneppen92; @ch95]. It was conjectured early by Yakhot [@yakhot81], based partially on symmetry grounds, that the large scale behavior of the one dimensional KS (1d-KS) equation is equivalent to that of the 1d noisy Burgers equation, also known as the KPZ equation [@kpz]. This conjecture has since been validated by detailed numerical studies [@zaleski; @sneppen92]. More recently, an explicit coarse-graining procedure was used by Chow and Hwa [@ch95] to derive a set of coupled effective equations describing the interaction between the chaotic cellular dynamics and the long wavelength fluctuations of the $h$-field. From this description, the large scale (KPZ-like) behavior of the 1d-KS system can be predicted quantitatively from the knowledge of various response functions at the ‘mesoscopic scale’ of several $\ell_0$’s.
The behavior of the 2d-KS equation is not as well understood. The simplest scenario is the generalization of Yakhot’s conjecture to 2d, with the large scale behavior described by the 2d-KPZ equation, $$\partial_t h = \nu \nabla^2 h + \frac{\lambda}{2} (\nabla h)^2 +
\eta({{\boldmath{r}}},t), \label{kpz}$$ where $\nu > 0 $ can be interpreted as a stabilizing ‘surface tension’ for the height profile $h$, and $\eta$ a [*stochastic*]{} noise with $\langle
\eta({{\boldmath{r}}},t)\eta({{\boldmath{r}}}',t')\rangle =2D\delta^2({{\boldmath{r}}}-{{\boldmath{r}}}')\delta(t-t')$. For $\nu
>0$, the asymptotic scaling properties of Eq. (\[kpz\]) are described by ‘strong-coupling’ behavior with algebraic (rather than logarithmic) scaling in the roughness of $h$ and super-diffusive dynamics [@kpzrev]. The length scale at which the asymptotic regime is reached is given by ${\ell_\times}\sim e^{8\pi/g}$, where $g\equiv \lambda^2 D/\nu^3$. At scales below ${\ell_\times}$, the effect of the nonlinear term in (\[kpz\]) can be accounted for adequately via perturbation theory. The system behaves in this ‘weak-coupling’ regime as a linear stochastic diffusion equation with additive logarithmic corrections [@tang].
Previous studies of the 2d-KS equation [@proc92a; @proc92b; @jaya93] found behavior consistent with linear diffusion with logarithmic corrections but had different interpretations. Jayaprakash [*et al.*]{} [@jaya93] performed a numerical analysis akin to Zaleski’s on the 1d-KS [@zaleski], and concluded that their results were consistent with the weak-coupling regime of the 2d-KPZ equation, with (in principle) a crossover to strong-coupling beyond a length of ${\ell_\times}\approx 10^{26}\ell_0$, for $g=0.4$. Procaccia [*et al.*]{} [@proc92a; @proc92b] used a comparative Dyson-Wyld diagrammatic analysis of the two equations to argue that 2d-KS and 2d-KPZ [*cannot*]{} belong to the same universality class [@note2]. They maintained instead that the asymptotic behavior of the 2d-KS equation is described by a ‘nonlocal’ solution, consisting of diffusion with [*multiplicative*]{} logarithmic corrections [@note1]. We feel that the ensuing debate [@comments] failed to rule out either interpretation.
It is very difficult to distinguish between the above two scenarios numerically, as one must resolve different forms of logarithmic corrections to the (already logarithmic) correlation function of the linear diffusion equation. Theoretically, there is no [*a priori*]{} reason why simple symmetry considerations such as Yakhot’s should be valid in two and higher dimensions. Unlike in 1d where there are only scalar density fluctuations, the 2d case is complicated because three or more large-$k$ modes can couple and contribute to low-$k$ fluctuations. Such nonlocal interactions in $k$ may not be adequately accounted for in the type of analysis performed in Refs. [@zaleski; @jaya93], which numerically impose KPZ dynamics and then test for self-consistency.
In this study, we perform a systematic symmetry analysis, taking into account the possibility of large-k coupling. Specifically, we extend the coarse graining procedure of Ref. [@ch95] to two dimensions to derive a set of coupled equations describing the local arrangement of cells, and study their effect on the macroscopic dynamics of the $h$-field. The resulting behavior depends crucially on the small scale arrangement of the cells. In the simplest case, the strong-coupling 2d-KPZ behavior is recovered. Nevertheless, more complicated behaviors are allowed if the microscopic cellular arrangement exhibits [*spontaneous rotational symmetry breaking*]{}. A number of possible scenarios are listed for this case. To determine which of the allowed scenarios is selected by the 2d-KS equation, we performed numerical measurements of the cellular dynamics at the mesoscopic scale of 4 to 16 $\ell_0$’s. Our results disfavor the occurrence of the more exotic scenarios, leaving the strong-coupling 2d-KPZ behavior as the most likely possibility.
As in 1d, we coarse grain over a region of size $L\times L$, where $L$ is several times the typical cellular size $\ell_0$. $h({{\boldmath{r}}},t)$ is separated into fast cellular modes $h_>$ and slow long wavelength modes $h_<$. Inserting $h({{\boldmath{r}}},t)=h_<({{\boldmath{r}}},t)+h_>({{\boldmath{r}}},t)$ into Eq. (\[KS\]), we obtain the following equations for the fast and slow modes: $$\begin{aligned}
&\partial_t h_> &= -\nabla^2 h_> - \nabla^4 h_> +(\nabla
h_>)^2_> + 2 (\nabla h_> \cdot \nabla h_<)
\label{fast}\\
&\partial_t h_< &=-\nabla^2h_< + (\nabla h_<)^2 + w({{\boldmath{r}}},t)
+ O(\nabla^4h_<).
\label{slowdyn}\end{aligned}$$ where $w({{\boldmath{r}}},t) \equiv (\nabla h_>)^2_<$ is the only contribution of the fast modes on the dynamics of $h_<$. It can be interpreted as the ‘drift rate’ of $h_<$ over a regime of $L\times L$ centered at ${{\boldmath{r}}}$. To specify the dynamics of $h_<$, it is necessary to obtain the dynamics of $w$ from the fast mode equation (\[fast\]). Due to the structure of the nonlinear term, we must consider the tensor ${{\sf W}}$, with elements $W_{ij}= 2\ \overline{\partial_i h_>\cdot\partial_j h_>}$ where the over-line denotes a spatial average over the coarse-graining scale $L$. It is convenient to introduce the curvature tensor ${{\sf K}}$, with elements $K_{ij}= 2 \partial_i\partial_j h_<$. In this notation, $w= \frac12{\rm Tr}\, {{\sf W}}$ and $\kappa \equiv \nabla^2 h_< = \frac12 {\rm Tr}\, {{\sf K}}$. Taking the time derivative of $W_{ij}$ and using (\[fast\]), we obtain $$\partial_t {{\sf W}}= {{\sf F}}\left[{{\sf W}}\right] + {{\sf W}}\cdot {{\sf K}}+ {{\sf K}}\cdot {{\sf W}}\label{Wdyn}$$ where ${{\sf F}}\left[{{\sf W}}\right]$ contains purely fast mode dynamics and will be described shortly. The forms of the last two terms in Eq. (\[Wdyn\]) are fixed by the Galilean invariance of the KS equation and are exact.
Equation (\[Wdyn\]) can be made more transparent by rewriting the two tensors as ${{\sf W}}= w \cdot {{\sf 1}}+ {\widetilde{w}}\cdot {{\sf Q}}(\phi)$ and ${{\sf K}}= \kappa \cdot {{\sf 1}}+ {\widetilde{\kappa}}\cdot {{\sf Q}}(\theta)$, where ${{\sf 1}}$ is the identity matrix and ${{\sf Q}}(\alpha)$ is a unit [*traceless*]{} matrix, represented by an angle $\alpha$, e.g., $Q_{12}(\alpha)=Q_{21}(\alpha)=\sin(2\alpha)$ and $Q_{11}(\alpha)=-Q_{22}(\alpha)=\cos(2\alpha)$. Adopting vector notation ${\vec{\psi}}= ({\widetilde{w}}\cos 2\phi, {\widetilde{w}}\sin 2\phi)$ and ${\vec{\chi}}=({\widetilde{\kappa}}\cos 2\theta,{\widetilde{\kappa}}\sin 2\theta)$, Eq. (\[Wdyn\]) can be rewritten as $$\begin{aligned}
\partial_t w&=& f[w] + 2\kappa w + 2{\vec{\chi}}\cdot {\vec{\psi}}\label{wdyn2}\\ \partial_t {\vec{\psi}}&=&
{\vec{\varphi}}\left[{\vec{\psi}}\right] + \kappa {\vec{\psi}}+ w {\vec{\chi}},
\label{psidyn}\end{aligned}$$ to leading order, with $f$ and ${\vec{\varphi}}$ obtained from the appropriate decomposition of ${{\sf F}}$.
Eqs. (\[wdyn2\]) and (\[psidyn\]), together with the slow mode equation (\[slowdyn\]), form a closed set of coarse grained equations which specifies the dynamics of $h_<$ once the effective forms of the small scale dynamics, i.e., $f$ and ${\vec{\varphi}}$ are given. These equations are constructed from symmetry considerations, and can be regarded as the more complete generalization of Yakhot’s conjecture for two dimensions. We first discuss the physical meaning of the coarse grained variables appearing in ${{\sf W}}$ and ${{\sf K}}$.
The tensor ${{\sf K}}$ describes the local curvature of the slow modes $h_<$. With ${\widetilde{\kappa}}=0$, we have a symmetric paraboloid — a ‘valley’ if $\kappa>0$ or a ‘hill’ if $\kappa<0$. With $\kappa=0$, we have a ‘saddle’, with ${\widetilde{\kappa}}$ and $\theta$ specifying the strength and the orientation. The tensor ${{\sf W}}$ characterizes the local [*packing*]{} of the cells. As in the 1d case [@ch95], $w$ gives the local cell density. The traceless component of ${{\sf W}}$ describes the local [*anisotropy*]{} in cell packing. Fig. \[fig1\](a), shows an example of an arrangement of cells where ${\vec{\psi}}$ is pointed in the $y$-direction. Fluctuations in the anisotropic part of the curvature ${{\sf K}}$ will affect the local cell packing. For example, the cell density at the bottom of a valley will be higher and a saddle configuration in $h_<$ will induce anisotropy. Eqs. (\[wdyn2\]) and (\[psidyn\]) describe these effects of curvature quantitatively, much like the relation between stress and strain in elastic systems. Cell packing in turn influences the slow mode dynamics via the $w$ term in Eq. (\[slowdyn\]). The anisotropic parts of ${{\sf K}}$ and ${{\sf W}}$ are invariant upon a rotation by $180^\circ$; see Fig. \[fig1\](a). Thus, we can view the vector field ${\vec{\psi}}({{\boldmath{r}}},t)$ as a ‘nematic’ order parameter describing the local cellular orientation, and ${\vec{\chi}}$ as an applied field biasing ${\vec{\psi}}$ towards a specific orientation.
If we turn off the applied field $\kappa$ and ${\vec{\chi}}$ in Eqs. (\[wdyn2\]) and (\[psidyn\]), we have $\partial_t
w = f[w]$ and $\partial_t {\vec{\psi}}= {\vec{\varphi}}\left[{\vec{\psi}}\right]$ for each coarse-grained region; thus $f$ and ${\vec{\varphi}}$ describe the small scale dynamics. Even for a coarse-grained region of a few $\ell_0$’s, the small scale dynamics of $h$ are already [*chaotic*]{}. The fields $w(t)$ and ${\vec{\psi}}(t)$ are ‘projections’ of this small scale chaotic dynamics. They can be quantitatively characterized numerically as we will present shortly. Before doing so, we first construct some possible scenarios.
We expect that the $h$-field has on average a finite drift rate, i.e., a finite time-averaged value of $w$. The simplest dynamics of $w$ is then $\partial_t w({{\boldmath{r}}},t) = f[w] = -\alpha(w-w_0) + \xi({{\boldmath{r}}},t)$, where $\xi$ is a stochastic forcing mimicking the chaotic small scale dynamics, and $w_0$ is a constant. This yields $w({{\boldmath{r}}},t\to\infty) \to w_0$. The behavior of ${\vec{\psi}}$ is less straightforward. In the simplest scenario, we have $\partial_t {\vec{\psi}}= -{\widetilde{\alpha}}{\vec{\psi}}+ {\vec{\zeta}}({{\boldmath{r}}},t)$ to leading order, with ${\vec{\zeta}}$ being a vectorial stochastic forcing. Eq. (\[psidyn\]) then yields (in the hydrodynamic limit) ${\vec{\psi}}\simeq (2w_0/{\widetilde{\alpha}}) {\vec{\chi}}$, where we took the asymptotic result $w=w_0$ and assumed that the typical curvature $\kappa$ is small. Note that in this scenario, the cellular orientation [*passively*]{} follows the curvature. In particular, there is no orientational anisotropy on average if there is no external forcing. Inserting this result and $f[w]$ into (\[wdyn2\]), we find in the hydrodynamic limit $$w \simeq w_0 +\frac{2w_0}{\alpha}\nabla^2 h_< + \frac{\xi}{\alpha}
+ O((\partial_i\partial_j h_<)^2).
\label{w}$$ Substituting (\[w\]) into (\[slowdyn\]) yields an equation for $h_<$ of the KPZ form (\[kpz\]) to leading order, with $\nu = (2w_0/\alpha)-1$ and $\eta({{\boldmath{r}}},t)= \xi({{\boldmath{r}}},t)/\alpha$. Dynamics of the KPZ universality class will be obtained if $\nu > 0$ and the noise $\xi$ is uncorrelated between different coarse-graining regions.
Unlike the constant $\alpha$ however, there is no a priori reason why the constant ${\widetilde{\alpha}}$ cannot be negative. This would be the case if the microscopic chaotic dynamics has a preference for the [*spontaneous*]{} breaking of local isotropy. If ${\widetilde{\alpha}}\le 0$, then the dynamics of ${\vec{\psi}}$ would be more complicated. Higher order terms, e.g., $|{\vec{\psi}}|^2{\vec{\psi}}$, will be needed for stability. The minimal equation for (\[psidyn\]) becomes $$\partial_t {\vec{\psi}}= -{\widetilde{\alpha}}{\vec{\psi}}-{\widetilde{\beta}}\left|{\vec{\psi}}\right|^2 {\vec{\psi}}+
\gamma\nabla^2{\vec{\psi}}+ w_0 {\vec{\chi}}+ {\vec{\zeta}}({{\boldmath{r}}},t)
\label{psifast}$$ where ${\widetilde{\beta}}$ is a positive constant, and the $\gamma$ term describes the coupling of neighboring coarse grained regions. Eq. (\[psifast\]) describes the relaxational dynamics of a nematic liquid crystal under an applied ‘field’ ${\vec{\chi}}$. Its behavior depends crucially on the dynamics of the phase field, $\phi$, which is the Goldstone mode associated with symmetry breaking. The latter in turn depends on the parameters of Eq. (\[psifast\]), particularly the coupling constant $\gamma$ and the amplitude of the noise ${\vec{\zeta}}$. The possibilities along with the effects on $h_<$ are:
Case i) If the noise ${\vec{\zeta}}$ dominates over the spatial coupling $\gamma$, then the local anisotropy will be destroyed at large scales due to the proliferation of topological defects (disclinations) in $\phi$. Isotropy is restored and the KPZ universality class is recovered.
Case ii) If the spatial coupling is large, then the direction of ${\vec{\psi}}$ may ‘phase-lock’ with the direction of ${\vec{\chi}}$, as manifested by $\langle(\theta-\phi)^2\rangle \ll 1$. Solving for the steady state of $w$ in this case gives $w\simeq w_0 + \frac{w_0}{\alpha} \kappa + \frac{{\widetilde{w}}_0}{\alpha}{\widetilde{\kappa}}$, leading to a slow mode equation which is explicitly [*not*]{} KPZ-like since ${\widetilde{\kappa}}= \left[(h_{xx}-h_{yy})^2+4h_{xy}\right]^{1/2}$. (In the KPZ case, ${\widetilde{\kappa}}$ comes in at second order and is presumed irrelevant; see Eq. (\[w\])).
Case iii) For intermediate parameters, there may exist a ‘spin wave’ phase characterized by $\langle\phi({{\boldmath{r}}})\phi(0)\rangle = a \log|{{\boldmath{r}}}|$. Here, ‘spin wave’ fluctuations would add a long range component to the effective KPZ noise, since $\langle\cos(\phi(r)-\phi(0))\rangle\sim
r^{-a}$. For sufficiently small $a$, it would yield dynamics that are not in the KPZ universality class.
To distinguish between these scenarios, we numerical measured the ‘fast mode’ dynamics with simulations on systems of size $L\times L$ with $L$ ranging from 32 to 128 with periodic boundary conditions. (A single cell had a length $\ell_0\approx 8.9$.) We used a simple spatial discretization scheme with an Euler time step of .02. We first checked for spontaneous symmetry breaking by examining the distribution of ${\vec{\psi}}$. We measured ${\vec{\psi}}$ spatially averaged over a system of size $L=32$, sampling at time intervals of $t=4$ over a total period of $t=4.8\times 10^5$. The distribution of $|{\vec{\psi}}|={\widetilde{w}}$ (normalized by the cylindrical area) is shown in Fig. \[fig1\](b). The Gaussian shape (within $\sim 3\%$), indicates a lack of spontaneous symmetry breaking strongly supporting the KPZ scenario.
We next measured the response of ${{\sf W}}$ to an imposed curvature tensor ${{\sf K}}$ using methods similar to those described in Ref. [@ch95]. In an STC steady state, we abruptly turned on a forcing term in Eq. (\[fast\]) of the form $h_<(x,y)=c k^{-2}\sin k_x x \sin k_y y$, with $k_x=k_y=\pi/2 L$, for a range of amplitudes $c$. This configuration yielded eight separate overlapping regions of size $L/2\times L/2$: four hill and valley regions, and four saddle regions with orientations $\theta= \pi/4$ and $\theta =
3\pi/4$. To account for the spatial dependence of the curvature ${{\sf K}}$ within each forcing region, we simply took the imposed curvature to be the average curvature. This gave $\kappa=\overline{\nabla^2 h_<} =(4/\pi^2)c$. We zeroed the first three Fourier modes of $h_>$ after each time step to remove the nonlinear slow mode response to the forcing (see Ref. [@ch95]).
We found that $w$ only responded to $\kappa$ and ${\vec{\psi}}$ only responded to ${\vec{\chi}}$, further validating the KPZ scenario. Fig. \[fig:scalar\](a) shows an example of the time dependent response of the drift rate $w$ to the forcing. The saturated amplitude $A_L(c)$ was a linear function of $c$ as seen in Fig \[fig:scalar\](b). In accordance with Eq. (\[w\]), the ratio $2w_0/\alpha$ was computed from the slope, from which we obtained $\nu=14.9\pm 0.5$ for $L=64$ and $\nu = 14.5 \pm 0.5$ for $L=128$. The averaged response times were $\alpha^{-1}= 3.3\pm 0.2$ for $L=64$, and $\alpha^{-1}= 4\pm 0.2$ for $L=128$. Similar behaviors were obtained in the hyperbolic forcing region. We verified that ${\vec{\psi}}\propto{\vec{\chi}}$, and found that ${\widetilde{\alpha}}^{-1}= 2.0\pm 0.1$ for $L=64$, and ${\widetilde{\alpha}}^{-1}= 3.0\pm 0.3$ for $L=128$.
To characterize the effective stochasticity, we measured the two-point correlation function $C_L(\tau)=\langle
(w(t+\tau)- w(t))^2\rangle$ of the drift rate $w$ for systems of sizes ranging from $L=32$ to $L=128$. The relevant quantity is the correlator $D_L(\tau)=[C_L(\infty)-C_L(\tau)]/2$ (see Fig. \[fig:noise\](a)). For short-range correlated noise, the effective noise amplitude $D=L^2\int_0^\infty D_L(\tau) d\tau$ is expected to be independent of $L$. Fig. \[fig:noise\](b) shows measurements of $D$, with the average $D=89\pm 5$. The numerical values of the effective parameters $D$ and $\nu$ extracted using our coarse-graining scheme are in reasonable agreement with that of Ref. [@jaya93].
In summary, we have performed an analysis of the 2d-KS equation at the mesoscopic scale of several cell sizes. By using an explicit coarse graining scheme, we constructed various possibilities for the effective equation of motion for the slow modes $h_<$. We found, as pointed out before in [@proc92a; @proc92b], that the KPZ universality class is not the only possibility for the 2d-KS equation. Since our analysis is confined to the mesoscopic (large-k) limit, the various scenarios obtained are all nonperturbative and nonlocal in $k$. The more interesting scenarios involve spontaneous breaking of rotational symmetry. Although we found no symmetry breaking for the KS equation numerically, our analysis indicates that such solutions are allowed by symmetry, and may occur in other KS-like system. For the KS equation proper, we conclude that it belongs to the KPZ universality class. However, for all practical purposes, the behavior is well described by a stochastic diffusion equation with logarithmic corrections.
This research is supported in part by NIH grant K01 MH1058 (CC), ONR grant N00014-95-1-1002 (TH), and the A.P. Sloan Foundation (CC,TH).
[99]{}
To whom correspondences should be addressed.
M.C. Cross and P.C. Hohenberg, Rev. Mod. Phys. [ ]{}[**65**]{}, 851 (1993).
G.I. Sivashinsky and D.M. Michelson, Prog. Theor. Phys. [**63**]{}, 2112 (1980); Y. Kuramoto and T. Tsuzuki, Prog. Theor. Phys. [**54**]{}, 687 (1975).
V. Yakhot, Phys. Rev. A [**24**]{}, 642 (1981).
S. Zaleski, Physica D [**34**]{}, 427 (1989).
K. Sneppen [*et al.*]{}, Phys. Rev. A [**46**]{}, R7351 (1992).
C.C. Chow and T. Hwa, Physica D [**84**]{}, 494 (1995).
D. Forster, D.R. Nelson, and M.J. Stephen, Phys. Rev. A [**16**]{}, 732 (1977); M. Kardar, G. Parisi, and Y.-C. Zhang, Phys. Rev. Lett. [**56**]{}, 889 (1986).
J. Krug and H. Spohn, in [*Solids far from equilibrium*]{}, C. Godreche ed. Cambridge University Press.
T. Nattermann and L.-H. Tang, Phys. Rev. A [**45**]{}, 7156 (1992).
I. Procaccia[*et al.*]{}, Phys. Rev. A [**46**]{}, 3220 (1992).
V. S. L’vov and I. Procaccia, Phys. Rev. Lett. [**69**]{}, 3543 (1992).
C. Jayaprakash, F. Hayot, and R. Pandit, Phys. Rev. Lett. [**71**]{}, 12 (1993).
Their argument is contingent upon an equality (Eq. (23) of Ref. [@proc92a]) relating the difference of two integration constants (called $C_1$ and $C_2$) and the bare coefficient of the diffusion term. The nonlocal solution is tenable if the equality is satisfied. They claim that the constants, which come from Wyld diagrammatic calculations, are determined [*uniquely*]{} by the $(\nabla h)^2$nonlinearity and therefore the [*same*]{} for both the KS and KPZ equations. However, the calculation involved integration of the ‘full’ response and correlation functions over the [*entire*]{} range of $k$. We note that these functions should be different for large values of $k$’s where the microscopic dynamics matter. Consequently the constants $C_1$ and $C_2$ need [*not*]{} be the same for the KS and KPZ equations.
The nonlocal solution of Ref. [@proc92a] has the same form as that of the $g=0$ fixed point of the 2d-KPZ equation. For the latter, multiplicative logarithmic correction to diffusion arises from the ‘asymptotic freedom’ of the system as $g \to 0^-$. The conjectured exponents of the logarithmic correction factor follow naturally from the nonrenormalization of $\nu$ of the KPZ equation to all orders in $\lambda$.
V. L’vov and I. Procaccia, Phys. Rev. Lett. [**72**]{}, 307 (1994); C. Jayaprakash [*et al.*]{}, [*ibid*]{}, 308 (1994).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We show that every block of category $\O$ for the general linear Lie superalgebra $\mathfrak{gl}_{m|n}(\operatorname{{\Bbbk}})$ is equivalent to some corresponding block of category $\O$ for the queer Lie superalgebra $\q_{m+n}(\operatorname{{\Bbbk}})$. This implies the truth of the Kazhdan-Lusztig conjecture for the so-called type A blocks of category $\O$ for the queer Lie superalgebra as formulated by Cheng, Kwon and Wang.'
author:
- Jonathan Brundan and Nicholas Davidson
title: 'Type A blocks of super category $\O$'
---
[^1]
Introduction
============
In this article, we study the analog of the BGG category $\O$ for the Lie superalgebra $\mathfrak{q}_n(\operatorname{{\Bbbk}})$. Recent work of Chen [@C] has reduced most questions about this category just to the study of three particular types of block, which we refer to here as the type A, type B and type C blocks. Type B blocks (which correspond to integral weights) were investigated already by the first author in [@B1], leading to a Kazhdan-Lusztig conjecture for characters of irreducibles in such blocks in terms of certain canonical bases for the quantum group of type B$_\infty$. In [@CKW], Cheng, Kwon and Wang formulated analogous conjectures for the type A blocks (defined below) and the type C blocks (which correspond to half-integral weights) in terms of canonical bases of quantum groups of types A$_\infty$ and C$_\infty$, respectively.
The main goal of the article is to prove the Cheng-Kwon-Wang conjecture for type A blocks ([@CKW Conjecture 5.14]). To do this, we use some tools from higher representation theory to establish an equivalence of categories between the type A blocks of category $\mathcal O$ for the Lie superalgebra $\mathfrak{q}_n(\operatorname{{\Bbbk}})$ and integral blocks of category $\mathcal O$ for a general linear Lie superalgebra. This reduces the Cheng-Kwon-Wang conjecture for type A blocks to the Kazhdan-Lusztig conjecture of [@B0], which was proved already in [@CLW; @BLW].
Regarding the types B and C conjectures, Tsuchioka discovered in 2010 that the type B canonical bases considered in [@B1] fail to satisfy appropriate positivity properties, so that the conjecture from [@B1] is certainly false. Moreover, after the first version of [@CKW] appeared, Tsuchioka pointed out similar issues with the type C canonical bases studied in [@CKW], so that the Cheng-Kwon-Wang conjecture for type C blocks as formulated in the first version of their article ([@CKW Conjecture 5.10]) also seems likely to be incorrect.
In fact, the techniques developed in this article can be applied also to the study of the type C blocks. This will be spelled out in a sequel to this paper [@BDC]. In this sequel, we prove a modified version of the Cheng-Kwon-Wang conjecture for type C blocks: one needs to replace Lusztig’s canonical basis with Webster’s “orthodox basis” arising from the indecomposable projective modules of the tensor product algebras of [@Web $\S$4]. This modified conjecture was proposed independently by Cheng, Kwon and Wang in a revision of their article ([@CKW Conjecture 5.12]). It is not as satisfactory as the situation for type A blocks, since there is no elementary algorithm to compute Webster’s basis explicitly (unlike the canonical basis). Also in the sequel, we will prove [@CKW Conjecture 5.13], and settle [@CKW Question 5.1] by identifying the category of finite-dimensional half-integer weight representations of $\q_n(\operatorname{{\Bbbk}})$ with a previously known highest weight category (as suggested by [@CK Remark 6.7]).
There is more to be said about type B blocks too; in fact, these are the most intriguing of all. Whereas the types A and C blocks carry the additional structure of tensor product categorifications in the sense of [@LW; @BLW] for the infinite rank Kac-Moody algebras of types A$_\infty$ and C$_\infty$, respectively, the type B blocks produce an example of a tensor product categorification of [*odd*]{} type B$_\infty$, i.e. one needs a [*super*]{} Kac-Moody 2-category in the sense of [@BE2]. This will be developed in subsequent work by the second author.
In the remainder of the introduction, we are going to formulate our main result for type A blocks in more detail. To do this, we first briefly recall some basic notions of superalgebra. Let $\operatorname{{\Bbbk}}$ be a ground field which is algebraically closed of characteristic zero, and fix a choice of $\sqrt{-1}\in\operatorname{{\Bbbk}}$. We adopt the language of [@BE Definition 1.1]:
- A [*supercategory*]{} is a category enriched in the symmetric monoidal category of vector superspaces, i.e. the category of $\operatorname{\mathbb{Z}}/2$-graded vector spaces over $\operatorname{{\Bbbk}}$ with morphisms that are parity-preserving linear maps.
- Any morphism in a supercategory decomposes uniquely into an even and an odd morphism as $f = f_\0 + f_\1$. A [*superfunctor*]{} between supercategories means a $\operatorname{{\Bbbk}}$-linear functor which preserves the parities of morphisms.
- For superfunctors $F, G : \C \to \operatorname{\mathcal{D}}$, a *supernatural transformation* $\eta :
F \Rightarrow G$ is a family of morphisms $\eta_M = \eta_{M,\0} + \eta_{M,\1} : FM \to GM$ for each $M\in\ob \C$, such that $\eta_{N,p} \circ Ff = (-1)^{|f|p} Gf \circ \eta_{M,p}$ for every homogeneous morphism $f:M \rightarrow N$ in $\C$ and each $p \in \operatorname{\mathbb{Z}}/2$.
For any supercategory $\C$, the [*Clifford twist*]{} $\C^{\CT}$ is the supercategory whose objects are pairs $(X, \phi)$ for $X \in \ob \C$ and $\phi \in \operatorname{End}_\C(X)_\1$ with $\phi^2 = \operatorname{{\operatorname{id}}}$, and whose morphisms $f:(X,\phi) \rightarrow (X',\phi')$ are morphisms $f:X \rightarrow X'$ in $\C$ such that $f_p \circ \phi = (-1)^p \phi'\circ f_p$ for each $p \in \operatorname{\mathbb{Z}}/2$. One can also take Clifford twists of superfunctors and supernatural transformations (details omitted), so that $\CT$ is actually a 2-superfunctor from the 2-supercategory of supercategories to itself in the sense of [@BE Definition 2.2]. The following basic lemma is a variation on [@KKT Lemma 2.3].
[**Lemma.**]{}
*Suppose $\C$ is a supercategory such that*
- $\C$ is additive;
- $\C$ is $\Pi$-complete, i.e. every object of $\C$ is the target of an odd isomorphism;
- all even idempotents split.
Then the supercategories $\C$ and $(\C^\CT)^\CT$ are superequivalent.
Note that objects in the supercategory $(\C^\CT)^\CT$ consist of triples $(X, \phi, \psi)$ for $X \in \ob \C$ and $\phi, \psi \in \operatorname{End}_\C(X)_\1$ such that $\phi^2 = \psi^2 = \operatorname{{\operatorname{id}}}$ and $\phi \circ \psi = - \psi\circ
\phi$. Morphisms $f:(X, \phi, \psi) \rightarrow (X', \phi',\psi')$ in $(\C^\CT)^\CT$ are morphisms $f:X \rightarrow X'$ in $\C$ such that $f_p \circ \phi = (-1)^p \phi'\circ f_p$, and $f_p \circ \psi = (-1)^p \psi' \circ f_p$ for each $p \in \operatorname{\mathbb{Z}}/2$.
Define a superfunctor $F:\C \rightarrow (\C^\CT)^\CT$ as follows. On an object $X \in \ob \C$, let $F X := (X \oplus \Pi X, \phi, \psi)$ where $$\phi =\left(\begin{array}{cc}0 & \zeta_X \\ \zeta_X^{-1} &
0\end{array}\right),
\qquad
\psi = \left(\begin{array}{cc}0 & -\sqrt{-1}\zeta_X \\
\sqrt{-1}\zeta_X^{-1} & 0\end{array}\right).$$ On a homogeneous morphism $f:X \rightarrow X'$, we let $Ff:FX \rightarrow FX'$ be the morphism defined by the matrix $\left(
\begin{array}{rr}
f&0\\
0&\Pi f
\end{array}\right)$, where $\Pi f:\Pi X \rightarrow \Pi X'$ denotes $(-1)^{|f|} \zeta_{X'}^{-1} \circ f \circ \zeta_X$. We show that $F$ is a superequivalence by checking that it is full, faithful and evenly dense (see [@BE]). It is obviously faithful. To see that it is full, take an arbitrary homogeneous morphism $f:F X \rightarrow F X'$ in $(\C^\CT)^\CT$. Viewing $f$ as a $2 \times 2$ matrix $\left(\begin{array}{ll}f_{11}&f_{12}\\f_{21}&f_{22}\end{array}\right)$ of morphisms in $\C$, we need to show that $f_{12} = f_{21} = 0$ and $f_{22} = \Pi f_{11}$. This follows easily on considering the matrix entries in the identities $\phi' \circ f = (-1)^{|f|} f \circ
\phi$ and $\psi' \circ f = (-1)^{|f|} f \circ
\psi$.
Finally, to check that $F$ is evenly dense, we take any object $(X,
\phi, \psi) \in \ob (\C^\CT)^\CT$, and must show that it is evenly isomorphic to an object in the image of $F$. Let $$e_1 := \frac{1-\sqrt{-1}\phi \circ \psi}{2},
\qquad
e_2 := \frac{1+\sqrt{-1}\phi \circ \psi}{2}.$$ These are mutually orthogonal idempotents summing to the identity in $\operatorname{End}_\C(X)_\0$. Hence, we may decompose $X$ as $X = X_1 \oplus X_2$ with $X_i$ being the image of $e_i$. We then have that $\phi = e_2 \circ \phi \circ e_1 + e_1 \circ \phi
\circ e_2$, and similarly for $\psi$. Now we observe that $$e_2 \circ \psi \circ e_1 = e_2 \circ \psi =
\frac{\psi+\sqrt{-1}\phi}{2}
= \sqrt{-1} \,\phi \circ e_1 = \sqrt{-1} \,e_2 \circ \phi \circ e_1.$$ Similarly, $e_1 \circ \psi \circ e_2 = - \sqrt{-1} \,e_1 \circ \phi
\circ e_2$. The map $e_2 \circ \phi \circ e_1 \circ \zeta_{X_1}$ is an even isomorphism $\Pi X_1 \stackrel{\sim}{\rightarrow} X_2$, hence, $X = X_1 \oplus X_2 \cong X_1 \oplus \Pi X_1 = F X_1$. Under this isomorphism, $\phi = e_2 \circ \phi \circ e_1 + e_1 \circ
\phi \circ e_2$ corresponds to the matrix $
\left(\begin{array}{cc}0 & \zeta_{X_1} \\ \zeta_{X_1}^{-1} &
0\end{array}\right)$. Similarly, $\psi = \sqrt{-1}\,
e_2 \circ \phi \circ e_1 - \sqrt{-1}\, e_1 \circ
\phi \circ e_2$ corresponds to $
\left(\begin{array}{cc}0 & -\sqrt{-1}\zeta_{X_1} \\ \sqrt{-1}\zeta_{X_1}^{-1} &
0\end{array}\right)$. This verifies that $(X,\phi,\psi)$ is evenly isomorphic to $F X_1$.
For example, suppose that $A$ is a locally unital superalgebra, i.e. an associative superalgebra $A = A_\0 \oplus A_\1$ equipped with a distinguished collection $\{1_x\:|\:x \in X\}$ of mutually orthogonal even idempotents such that $A = \bigoplus_{x, y \in X} 1_y A 1_x$. Then there is a supercategory $A\operatorname{\text{-}smod}$ consisting of finite-dimensional left $A$-supermodules $M$ which are locally unital in the sense that $M = \bigoplus_{x \in X} 1_x M$. Even morphisms in $A\operatorname{\text{-}smod}$ are parity-preserving linear maps such that $f(av) = a f(v)$ for all $a \in A, v \in M$; odd morphisms are parity-reversing linear maps such that $f(av) = (-1)^{|a|} a f(v)$ for homogeneous $a$. There is an obvious isomorphism between the Clifford twist $A\operatorname{\text{-}smod}^\CT$ of this supercategory and the supercategory $A\otimes C_1 \operatorname{\text{-}smod}$, where $C_1$ denotes the rank one Clifford superalgebra generated by an odd involution $c$, and $A \otimes C_1$ is the usual braided tensor product of superalgebras. Hence, $(A\operatorname{\text{-}smod}^\CT)^\CT$ is isomorphic to $A \otimes C_2 \operatorname{\text{-}smod}$ where $C_2 := C_1 \otimes C_1$ is the rank two Clifford superalgebra generated by $c_1 := c \otimes 1$ and $c_2
:= 1 \otimes c$. In this situation, the above lemma is obvious as $A \otimes C_2$ is isomorphic to the matrix superalgebra $M_{1|1}(A)$, which is Morita superequivalent to $A$.
Now fix $n \geq 1$ and let $\mathfrak{g} = \g_\0\oplus\g_\1$ be the Lie superalgebra $\q_n(\operatorname{{\Bbbk}})$. Recall this is the subalgebra of the general linear Lie superalgebra $\mathfrak{gl}_{n|n}(\operatorname{{\Bbbk}})$ consisting of all matrices of block form $$\label{blockform}
\left( \begin{array}{c|c} A & B \\\hline B & A \end{array} \right).$$ Let $\mathfrak{b}$ (resp. $\mathfrak{h}$) be the standard Borel (resp. Cartan) subalgebra of $\mathfrak{g}$ consisting of all matrices (\[blockform\]) in which $A$ and $B$ are upper triangular (resp. diagonal). Let $\mathfrak{t} := \mathfrak{h}_{\0}$. We let $\delta_1,\dots,\delta_n$ be the basis for $\mathfrak{t}^*$ such that $\delta_i$ picks out the $i$th diagonal entry of the matrix $A$. Fix also a [sign sequence]{} $\bsigma = (\sigma_1,\dots,\sigma_n)$ with each $\sigma_r \in \{\pm \}$, and a scalar $z \in \operatorname{{\Bbbk}}$ such that $2z \notin \operatorname{\mathbb{Z}}$. [**We stress that all of our subsequent notation depends implicitly on these choices.**]{}
It will be convenient to index certain weights in $\mathfrak{t}^*$ by the set $\mathbf{B} := \operatorname{\mathbb{Z}}^n$ via the following [*weight dictionary*]{}: for $\bb = (b_1,\dots,b_n) \in \operatorname{\mathbf{B}}$ let $$\label{wtdict1}
\lambda_\bb := \sum_{r=1}^n \lambda_{\bb,r} \delta_r
\quad\text{where}\quad \lambda_{\bb,r} := \sigma_r(z+b_r).$$ We let $\sO$ be the category of all $\mathfrak{g}$-supermodules $M$ satisfying the following properties:
- $M$ is finitely generated as a $\mathfrak{g}$-supermodule;
- $M$ is locally finite-dimensional over $\mathfrak{b}$;
- $M$ is semisimple over $\t$ with all weights of the form $\lambda_\bb$ for $\bb \in \operatorname{\mathbf{B}}$.
Morphisms in $\sO$ are arbitrary (not necessarily even) $\mathfrak{g}$-supermodule homomorphisms, so that it is a supercategory. It also admits a parity switching functor $\Pi$. The [*type A blocks*]{} mentioned earlier are the blocks of $\sO$ for all possible choices of $\bsigma$ and $z$.
For each $\bb \in \operatorname{\mathbf{B}}$, there is an irreducible supermodule $L(\bb) \in \ob \sO$ of highest weight $\lambda_\bb$. Note the highest weight space of $L(\bb)$ is not one-dimensional: it is some sort of irreducible Clifford supermodule over the Cartan subalgebra $\h$. Every irreducible supermodule in $\sO$ is isomorphic to $L(\bb)$ for a unique $\bb \in \operatorname{\mathbf{B}}$ via a homogeneous (but not necessarily even) isomorphism. If $n$ is odd, $L(\bb)$ is of type $\ttQ$, i.e., $L(\bb)$ is evenly isomorphic to its parity flip $\Pi L(\bb)$. When $n$ is even, the irreducible $L(\bb)$ is of type $\ttM$, and we should explain how to distinguish it from its parity flip. For each $i \in \operatorname{\mathbb{Z}}$, we fix a choice $\sqrt{z+i}$ of a square root of $z+i$, then set $\sqrt{-(z+i)} := (-1)^{i} \sqrt{-1} \sqrt{z+i}$. The key point about this is that $$\label{violin}
\sqrt{-(z+i)} \sqrt{-(z+i+1)} = \sqrt{z+i}\sqrt{z+i+1}$$ for each $i \in \operatorname{\mathbb{Z}}$. Let $d_r' \in \g_\1$ be the matrix of the form (\[blockform\]) such that $A =
0$ and $B$ is the $r$th diagonal matrix unit. Then, for even $n$, we assume that $L(\bb)$ is chosen so that $d_1' \cdots d_n'$ acts on any even highest weight vector by the scalar $(\sqrt{-1})^{n/2} \sqrt{\lambda_{\bb,1}}
\cdots
\sqrt{\lambda_{\bb,n}}$. This determines $L(\bb)$ uniquely up to even isomorphism.
Turning our attention to the category on the other side of our main equivalence, let $\mathfrak{g}'$ be the general linear Lie superalgebra consisting of $n \times n$ matrices under the supercommutator, with $\operatorname{\mathbb{Z}}/2$-grading defined by declaring that the $rs$-matrix unit is even if $\sigma_r = \sigma_s$ and odd if $\sigma_r \neq
\sigma_s$. Let $\mathfrak{b}'$ (resp. $\mathfrak{t}'$) be the standard Borel (resp. Cartan) subalgebra consisting of upper triangular (resp. diagonal) matrices in $\mathfrak{g}'$. As before, we let $\delta_1',\dots,\delta_n'$ be the basis for $(\mathfrak{t}')^*$ defined by the diagonal coordinate functions. We introduce another [*weight dictionary*]{} (which in this setting is some “signed $\rho$-shift”): for $\bb \in \operatorname{\mathbf{B}}$, let $$\label{wtdict}
\lambda_\bb' :=
\sum_{r=1}^n \lambda_{\bb,r}' \delta_r'
\quad\text{where}\quad
\lambda_{\bb,r}' :=
\sigma_r \left(b_r
+ \sigma_1 1 +\cdots+\sigma_{r-1} 1 +\half(\sigma_r 1-1)\right).$$ Let $\sO'$ be the supercategory of $\g'$-supermodules $M'$ such that
- $M'$ is finitely generated as a $\g'$-supermodule;
- $M'$ is locally finite-dimensional over $\b'$;
- $M'$ is semisimple over $\t'$ with all weights of the form $\lambda'_\bb$ for $\bb \in \operatorname{\mathbf{B}}$.
Note $\sO'$ is the sum of all of the blocks of the usual category $\O$ for $\g'$ corresponding to integral weights of $\t'$. For each $\bb \in \operatorname{\mathbf{B}}$, there is a unique (up to even isomorphism) irreducible supermodule $L'(\bb) \in \ob \sO'$ generated by a homogeneous highest weight vector of weight $\lambda'_\bb$ and parity $
\sum_{\sigma_r = -} \lambda'_{\bb,r} \pmod{2}.
$
[**Main Theorem.**]{} [*If $n$ is even then there is a superequivalence $\mathbb{E}:\sO \rightarrow \sO'$ such that $\mathbb{E} L(\bb)$ is evenly isomorphic to $L'(\bb)$ for each $\bb
\in \operatorname{\mathbf{B}}$. If $n$ is odd then there is a superequivalence $\mathbb{E}:\sO \rightarrow (\sO')^\CT$ such that $\mathbb{E} L(\bb)$ is evenly isomorphic to $(L'(\bb)\oplus \Pi L'(\bb), \phi)$ for either of the two choices of odd involution $\phi$.* ]{}
If $\C$ is any $\operatorname{{\Bbbk}}$-linear category, we let $\C \oplus \Pi \C$ be the supercategory whose objects are formal direct sums $V_1 \oplus \Pi V_2$ for $V_1, V_2 \in \ob \C$, with morphisms $V_1 \oplus \Pi V_2 \rightarrow W_1 \oplus \Pi W_2$ being matrices of the form $f =
\left( \begin{array}{cc} f_{11} & f_{12} \\ f_{21} & f_{22} \end{array} \right)$ for $f_{ij} \in \operatorname{Hom}_\C(V_j,W_i)$. The $\operatorname{\mathbb{Z}}/2$-grading is defined so that $f_\0 =
\left( \begin{array}{cc} f_{11} & 0 \\ 0 & f_{22} \end{array} \right)$ and $f_\1 = \left( \begin{array}{cc} 0 & f_{12} \\f_{21} & 0 \end{array}
\right)$. For example, if $\C$ is the category $A\operatorname{\text{-}mod}$ of finite-dimensional locally unital modules over some locally unital algebra $A$, then $\C \oplus \Pi \C$ may be identified with the category $A\operatorname{\text{-}smod}$, viewing $A$ as a purely even superalgebra.
It was noticed originally in [@B0] that the category $\sO'$ can be decomposed in this way: let $\O'$ be full subcategory of $\sO'$ consisting of all $\g'$-supermodules whose $\lambda'_\bb$-weight space is concentrated in parity $\sum_{\sigma_r = -} \lambda'_{\bb,r} \pmod{2}$ for each $\bb \in \operatorname{\mathbf{B}}$; obviously, there are no non-zero odd morphisms between objects of $\O'$. Then $\sO'$ decomposes as $\sO' = \O' \oplus \Pi \O'$. Moreover, $\O'$ is a highest weight category with irreducible objects $\{L'(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$ indexed by the set $\operatorname{\mathbf{B}}$ as above. In fact, $\O'$ equivalent to $A\operatorname{\text{-}mod}$ for a locally unital algebra $A$ such that the left ideals $A
1_x$ and right ideals $1_x A$ are finite-dimensional for all distinguished idempotents $1_x \in A$. Although not needed here, the results of [@BLW] imply further that the algebra $A$ may be equipped with a $\operatorname{\mathbb{Z}}$-grading making it into a (locally unital) [*Koszul algebra*]{}; this leads to the definition of a graded analog of the category $\O'$ similar in spirit to Soergel’s graded lift of classical category $\O$ as in e.g. [@BGS].
Combining these remarks with our Main Theorem, we deduce:
- For even $n$, the category $\sO$ decomposes as $\sO = \O \oplus \Pi \O$, where $\O$ is the Serre subcategory generated by the irreducible supermodules $\{L(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$ introduced above (but [*not*]{} their parity flips). Moreover, $\O$ is equivalent to $\O'$, hence, to the category $A\operatorname{\text{-}mod}$ where $A$ is the Koszul algebra just introduced.
- For odd $n$, $\sO$ is superequivalent to $A \otimes C_1 \operatorname{\text{-}smod}$, viewing $A$ as a purely even superalgebra. This implies that the underlying category $\underline{\sO}$ consisting of the same objects as $\sO$ but only its even morphisms is equivalent to $A \operatorname{\text{-}mod}$, hence, to $\O'$.
As already mentioned, the Kazhdan-Lusztig conjecture for $\sO$ formulated in [@CKW] follows immediately from this discussion together with the Kazhdan-Lusztig conjecture for $\O'$ proved in [@CLW; @BLW].
There is also a parabolic analog of our Main Theorem. Let $\nu = (\nu_1,\dots,\nu_l)$ be a composition of $n$ with $\sigma_r = \sigma_s$ for all $\nu_1+\cdots+\nu_{k-1}+1 \leq
r < s \leq \nu_1+\cdots+\nu_k$ and $k=1,\dots,l$. Let $\mathfrak{p}_\nu$ be the corresponding standard parabolic subalgebra of $\g$, i.e. the matrices $A$ and $B$ in (\[blockform\]) are block upper triangular with diagonal blocks of shape $\nu$. Let $\sO_\nu$ be the corresponding parabolic analog of the category $\sO$, i.e. it is the full subcategory of $\sO$ consisting of all supermodules that are locally finite-dimensional over $\mathfrak{p}_\nu$. Similarly, there is a standard parabolic subalgebra $\mathfrak{p}'_\nu$ of $\g'$ consisting of block upper triangular matrices of shape $\nu$, and we let $\sO'_\nu$ be the analogously defined parabolic subcategory of $\sO'$. Various special cases of the following corollary for maximal parabolics/two-part compositions $\nu$ were known before; see [@C $\S$4] and [@CC].
[**Corollary.**]{} [*If $n$ is even then $\sO_\nu$ is superequivalent to $\sO_\nu'$. If $n$ is odd then $\sO_\nu$ is superequivalent to $(\sO_\nu')^\CT$.* ]{}
This follows from our Main Theorem on observing that $\sO_\nu$ and $\sO_\nu'$ may be defined equivalently as the Serre subcategories of $\sO$ and $\sO'$ generated by the irreducible supermodules $\{L(\bb), \Pi L(\bb)\}$ and $\{L'(\bb), \Pi L'(\bb)\}$, respectively, for $\bb \in \operatorname{\mathbf{B}}$ such that the following hold for $r \notin\{\nu_1,\nu_1+\nu_2,\dots,\nu_1+\cdots+\nu_l\}$:
- if $\sigma_r = +$ then $b_r > b_{r+1}$;
- if $\sigma_r =
-$ then $b_r < b_{r+1}$.
(This assertion is a well-known consequence of the construction of parabolic Verma supermodules in $\sO$ and $\sO'$, respectively; see e.g. [@M].)
In order to prove the Main Theorem, we will exploit the following powerful theorem established in [@BLW]: the category $\O'$ defined above is the unique (up to strongly equivariant equivalence) $\mathfrak{sl}_\infty$-tensor product categorification of the module $$V^{\otimes \bsigma} := V^{\sigma_1} \otimes
\cdots \otimes V^{\sigma_n},$$ where $V^{+}$ denotes the natural $\mathfrak{sl}_\infty$-module and $V^{-}$ denotes its dual. Hence, for even $n$, it suffices to show that the category $\sO$ decomposes as $\O \oplus \Pi \O$ for some $\mathfrak{sl}_\infty$-tensor product categorification $\O$ of $V^{\otimes \bsigma}$. For odd $n$, we show instead that $\sO^{\CT}$ decomposes as $\O \oplus \Pi \O$ for some $\mathfrak{sl}_\infty$-tensor product categorification $\O$ of $V^{\otimes \bsigma}$, hence, $\sO^\CT$ is superequivalent to $\sO'$. The Main Theorem then follows on taking Clifford twists, using also the lemma formulated above. In both the even and odd cases, our argument relies crucially also on an application of the main result of [@KKT].
We remark finally that the more general parabolic supercategories $\sO_\nu$ may be used to give new realizations of the more general sorts of $\mathfrak{sl}_\infty$-tensor product categorifications considered in [@BLW] involving exterior powers of $V^{\pm}$.
[*Acknowledgements.*]{} We thank the authors of [@CKW] for several helpful discussions at the KIAS conference “Categorical Representation Theory and Combinatorics" in December 2015, and Shunsuke Tsuchioka for sharing his computer-generated counterexamples to the type B and C conjectures.
Verma supermodules {#s2}
==================
We continue with $n \geq 1$, $\bsigma = (\sigma_1,\dots,\sigma_n) \in
\{\pm 1\}^n$, and $z \in \operatorname{{\Bbbk}}$ with $2z \notin \operatorname{\mathbb{Z}}$. Let $m := \lceil n/2 \rceil$, so that $n=2m$ or $2m-1$. Also set $$\label{jdef}
I := \operatorname{\mathbb{Z}},
\qquad
J := \left\{\pm \sqrt{z+i}\sqrt{z+i+1}\:\Big|\:i \in I\right\},$$ where the square roots are as chosen in the introduction.
In this paragraph, we work with the Lie superalgebra $\widehat\g := \mathfrak{gl}_{2m|2m}(\operatorname{{\Bbbk}})$ in order to introduce some coordinates. Let $\widehat{U}$ be the natural $\widehat\g$-supermodule with standard basis $u_1,\dots,u_{4m}$. Write $x_{r,s}$ for the $rs$-matrix unit in $\widehat\g$, so $x_{r,s} u_t = \delta_{s,t} u_r$. We denote the odd basis vectors $u_{2m+1}, \dots, u_{4m}$ instead by $u_1',\dots,u_{2m}'$. For $1 \leq r,s \leq 2m$, we set $$\begin{aligned}
\label{cough1}
e_{r,s} &:= x_{r,s} + x_{2m+r,2m+s},
&
e'_{r,s} &:= x_{r,2m+s}+x_{2m+r,s},\\\label{cough2}
f_{r,s} &:= x_{r,s} - x_{2m+r,2m+s},
&
f'_{r,s} &:= x_{r,2m+s} - x_{2m+r,s}.\\\intertext{Also let}
d_r &:= e_{r,r},&
d'_r &:= e_{r,r}'.\end{aligned}$$ Then we have that $$\begin{aligned}
\label{f1}
e_{r,s} u_t &= \delta_{s,t} u_r,
&
e_{r,s} u_t' &= \delta_{s,t} u_r', & e_{r,s}' u_t&= \delta_{s,t}
u_r', & e_{r,s}' u_t' &= \delta_{s,t}
u_r,\\ \label{f2}
f_{r,s} u_t &= \delta_{s,t} u_r,
&
f_{r,s} u_t' &= -\delta_{s,t} u_r', & f_{r,s}' u_t&= -\delta_{s,t} u_r', & f_{r,s}' u_t' &= \delta_{s,t}
u_r.\end{aligned}$$ Finally let $\widehat{U}^*$ be the dual supermodule to $\widehat{U}$, with basis $\phi_1,\dots,\phi_{2m}, \phi_1',\dots,\phi_{2m}'$ that is dual to the basis $u_1,\dots,u_{2m}, u_1',\dots,u_{2m}'$. We have that $$\begin{aligned}
\label{f3}
e_{r,s} \phi_t &= - \delta_{r,t} \phi_s, & e_{r,s} \phi_t' &= -\delta_{r,t} \phi_s',&
e_{r,s}' \phi_t &= - \delta_{r,t} \phi'_s, & e_{r,s}' \phi'_t &=
\delta_{r,t} \phi_s,\\\label{f4}
f_{r,s} \phi_t &= -\delta_{r,t} \phi_s, & f_{r,s} \phi_t' &= \delta_{r,t} \phi_s',
&
f_{r,s}' \phi_t &= -\delta_{r,t} \phi'_s, & f_{r,s}' \phi'_t &=
-\delta_{r,t} \phi_s. \end{aligned}$$
When $n$ is even, we continue with $\g, \b$ and $\h$ as in the introduction, so $\g$ is the subalgebra of $\widehat\g$ spanned by $\{e_{r,s}, e_{r,s}'\:|\:1 \leq r,s \leq n\}$, while $\h$ has basis $\{d_r, d_r'\:|\:1 \leq r \leq 2m\}$. However, when $n$ is odd, it is convenient to change some of this notation. The point of doing this is to unify our treatment of even and odd $n$ as much as possible in the remainder of the article. So, if $n$ is odd, we henceforth [*redefine*]{} $\g, \b$ and $\h$ as follows:
- $\g$ denotes the Lie superalgebra $\mathfrak{q}_n(\operatorname{{\Bbbk}}) \oplus
\mathfrak{q}_1(\operatorname{{\Bbbk}})$, which we identify with the subalgebra of $\widehat{\g}$ spanned by $\{e_{r,s}, e_{r,s}'\:|\:1 \leq r,s \leq n\} \sqcup
\{d_{2m}, d_{2m}'\}$.
- $\b$ is the Borel subalgebra spanned by $\{e_{r,s}, e_{r,s}'\:|\:1 \leq r \leq s \leq n\} \sqcup
\{d_{2m}, d_{2m}'\}$;
- $\h$ is the Cartan subalgebra spanned by $\{d_r, d_r'\:|\:1 \leq r \leq 2m\}$.
In both the even and the odd cases, the subspaces $U \subseteq \widehat{U}$ and $U^* \subseteq \widehat{U}^*$ spanned by $u_1,\dots,u_n, u_1',\dots, u_n'$ and $\phi_1,\dots,\phi_n, \phi_1',\dots,\phi_n'$, respectively, may be viewed as $\g$-supermodules. Also set $\t := \h_\0$ and let $\delta_1,\dots,\delta_{2m}$ be the basis for $\t^*$ that is dual to the basis $d_1,\dots,d_{2m}$ for $\t$.
For $\bb \in \operatorname{\mathbf{B}}$, we define $\lambda_\bb$ according to (\[wtdict1\]) if $n$ is even, but redefine it in the odd case as $$\label{wtdict2}
\lambda_\bb := \sum_{r=1}^n \lambda_{\bb,r} \delta_r + \delta_{2m}
\quad\text{where}\quad
\lambda_{\bb,r} := \sigma_r(z+b_r).$$ We also introduce the tuple $\bd_r \in \operatorname{\mathbf{B}}$ which has $1$ as its $r$th entry and $0$ in all other places, so that $$\label{handier}
\lambda_\bb \pm \delta_r
= \lambda_{\bb \pm \sigma_r \bd_r}.$$ Then we define $\sO$ exactly as we did in the introduction but using the current choices for $\g,
\b, \t$ and $\lambda_\bb$. This is exactly the same category as in the introduction when $n$ is even, but when $n$ is odd our new version of $\sO$ is superequivalent to the Clifford twist $\sO^\CT$ of the supercategory from the introduction. Indeed, if $M$ is a supermodule in our new $\sO$, the restriction of $M$ to the subalgebra $\q_n(\operatorname{{\Bbbk}})$, equipped with the odd involution defined by the action of $d_{2m}'$, gives an object of the Clifford twist of the supercategory from before.
We proceed to define some irreducible $\h$-supermodules $\{\u(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$. Let $C_2$ be the rank 2 Clifford superalgebra with odd generators $c_1,c_2$ subject to the relations $c_1^2 = c_2^2 = 1, c_1 c_2 = - c_2 c_1$. Let $\u$ be the irreducible $C_2$-supermodule on basis $v, v'$ with $v$ even and $v'$ odd, and action defined by $$c_1 v = v',
\quad
c_1 v' = v,
\quad
c_2 v = \sqrt{-1} v',
\quad
c_2 v' = -
\sqrt{-1} v.$$ Then, for $\bb \in \operatorname{\mathbf{B}}$, we set $\u(\bb) := \u^{\otimes m}$. For $1 \leq r \leq n$, we let $d_r$ act by the scalar $\lambda_{\bb,r}$ and $d_r'$ act by left multiplication by $\sqrt{\lambda_{\bb,r}} \,\operatorname{{\operatorname{id}}}^{\otimes (s-1)}\otimes c_{r+1-2s}
\otimes \operatorname{{\operatorname{id}}}^{ \otimes (m-s)}$ where $s := \lfloor r/2 \rfloor$ (and we are using the usual superalgebra sign rules). In the odd case, we also need to define the actions of $d_{2m}$ and $d_{2m}'$: these are the identity and the odd involution $\operatorname{{\operatorname{id}}}^{\otimes(m-1)} \otimes c_2$, respectively. In all cases, $\u(\bb)$ is an irreducible $\h$-supermodule of type $\ttM$, and its $\t$-weight is $\lambda_\bb$. Moreover, by construction, $d_1'\cdots d_{2m}'$ acts on any even (resp. odd) vector in $\u(\bb)$ as $c_\bb$ (resp. $-c_\bb$), where $$\label{y1}
c_\bb
:=
(\sqrt{-1})^m \sqrt{\la_{\bb,1}}\cdots \sqrt{\la_{\bb,n}}.$$ The signs here distinguish $\u(\bb)$ from its parity flip. The following is well known; e.g. see [@CW $\S$1.5.4].
\[projectivity\] For $\bb \in \operatorname{\mathbf{B}}$, any $\h$-supermodule that is semisimple of weight $\lambda_\bb$ over $\t$ decomposes as a direct sum of copies of the supermodules $\u(\bb)$ and $\Pi \u(\bb)$.
We can identify $\h$-supermodules that are semisimple of weight $\lambda_\bb$ over $\t$ with supermodules over the Clifford superalgebra $C_{2m} := C_2^{\otimes m}$, so that $d_r'\:(r=1,\dots,n)$ acts in the same way as $\sqrt{\lambda_{\bb,r}} \,
\operatorname{{\operatorname{id}}}^{\otimes (s-1)}\otimes c_{r+1-2s} \otimes \operatorname{{\operatorname{id}}}^{\otimes (m-s)}$ where $s := \lfloor r/2 \rfloor$, and in the odd case $d_{2m}'$ acts as $\operatorname{{\operatorname{id}}}^{\otimes(m-1)} \otimes c_2$. The lemma then follows since $C_{2m}$ is simple, indeed, it is isomorphic to the matrix superalgebra $M_{2^{n-1}|2^{n-1}}(\operatorname{{\Bbbk}})$.
Let $\underline{\sO}$ denote the [*underlying category*]{} consisting of the same objects as $\sO$ but only the even morphisms. This is obviously an Abelian category. In order to parametrize its irreducible objects explicitly, we introduce the *Verma supermodule* $M(\bb)$ for $\bb \in \operatorname{\mathbf{B}}$ by setting $$M(\bb) := U(\g) \otimes_{U(\b)} \u(\bb),$$ where we are viewing $\u(\bb)$ as a $\b$-supermodule by inflating along the surjection $\b \twoheadrightarrow \h$. The weight $\lambda_\bb$ is the highest weight of $M(\bb)$ in the usual *dominance order* on $\t^*$, i.e. $\lambda \leq \mu$ if and only if $\mu-\lambda\in
\bigoplus_{r=1}^{n-1} \operatorname{\mathbb{N}}(\delta_r-\delta_{r+1})$. Note also that we can distinguish $M(\bb)$ from its parity flip in the same way as for $\u(\bb)$: the element $d_1'\cdots d_{2m}'$ acts on any even (resp. odd) vector in the highest weight space $M(\bb)_{\lambda_\bb}$ as the scalar $c_\bb$ (resp. $-c_\bb$).
As usual, the Verma supermodule $M(\bb)$ has a unique irreducible quotient denoted $L(\bb)$. Thus, $L(\bb)$ is an irreducible $\g$-supermodule of highest weight $\lambda_\bb$, and the action of $d_1' \cdots
d_{2m}'$ on its highest weight space distinguishes it from its parity flip. The irreducible supermodules $\{L(\bb), \Pi L(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$ give a complete set of pairwise inequivalent irreducible supermodules in $\underline{\sO}$. The endomorphism algebras of these objects are all one-dimensional, so they are irreducibles of type $\mathtt{M}$. Moreover, by a standard argument involving restricting to the underlying even Lie algebra as in [@B2 Lemma 7.3], we get that $\underline{\sO}$ is a [Schurian category]{} in the following sense (cf. [@BLW $\S$2.1]):
\[schurcat\] A [*Schurian category*]{} is a $\operatorname{{\Bbbk}}$-linear Abelian category in which all objects have finite length, there are enough projectives and injectives, and the endomorphism algebras of irreducible objects are all one-dimensional.
Let $x^T$ denote the usual transpose of a matrix $x \in \widehat{\g}$. This induces an antiautomorphism of $\g$, i.e. we have that $[x,y]^T = [y^T, x^T]$. Given $M \in \ob \sO$, we can view the direct sum $\bigoplus_{\bb \in \operatorname{\mathbf{B}}} M_{\lambda_\bb}^*$ of the linear duals of the weight spaces of $M$ as a $\g$-supermodule with action defined by $(x f)(v) := f(x^T v)$. Let $M^\star$ be the object of $\sO$ obtained from this by applying also the parity switching functor $\Pi^m$. Making the obvious definition on morphisms, this gives us a contravariant superequivalence $\star : \sO \to \sO$. We have incorporated the parity flip into this definition in order to get the following lemma; this is also checked in [@F Lemma 7] or [@C Lemma 4.7] but we include the proof to consolidate our parity conventions.
\[star\] For $\bb \in \operatorname{\mathbf{B}}$, we have that $L(\bb)^\star \cong
L(\bb)$ via an even isomorphism.
By weight considerations, we either have that $L(\bb)^\star$ is evenly isomorphic to $L(\bb)$ or to $\Pi L(\bb)$. To show that the former holds, take an even highest weight vector $f \in L(\bb)^\star$. We must show that $d_1'\cdots d_{2m}' f = c_\bb f$ (rather than $-c_\bb f$). Remembering the twist by $\Pi^m$ in our definition of $\star$, there is a highest weight vector $v \in L(\bb)$ of parity $m\pmod{2}$ such that $f(v) = 1$. Then we get that $$(d_1' \cdots d_{2m}' f)(v) =
f(d_{2m}' \cdots d_1' v)
= (-1)^m f(d_1'\cdots d_{2m}' v) = c_\bb f(v).$$ Hence, $d_1'\cdots d_{2m}' f = c_\bb f$.
Let $P(\bb)$ be a projective cover of $L(\bb)$ in $\underline{\sO}$. There are even epimorphisms $P(\bb) \twoheadrightarrow M(\bb)
\twoheadrightarrow L(\bb)$. Applying $\star$, we deduce that there are even monomorphisms $L(\bb) \hookrightarrow M(\bb)^\star \hookrightarrow P(\bb)^\star$. The supermodule $P(\bb)^\star$ is an injective hull of $L(\bb)$, while $M(\bb)^\star$ is the [*dual Verma supermodule*]{}. The following lemma is well known; it follows from central character considerations (e.g. see [@CW Theorem 2.48]) plus the universal property of Verma supermodules.
\[typicaldominant\] Suppose that $\lambda_{\bb}$ is dominant and typical, by which we mean that the following hold for all $1 \leq r < s \leq
n$:
- if $\sigma_r = \sigma_s$ then $\lambda_{\bb,r} \geq \lambda_{\bb,s}$;
- if $\sigma_r \neq \sigma_s$ then $\lambda_{\bb,r}+\lambda_{\bb,s} \neq 0$.
Then $M(\bb) = P(\bb)$.
Let $\sO^\Delta$ be the full subcategory of $\sO$ consisting of all supermodules possessing a Verma flag, i.e. for which there is a filtration $0 = M_0 \subset \cdots \subset M_l = M$ with sections $M_k / M_{k-1}$ that are evenly isomorphic to $M(\bb)$’s or $\Pi M(\bb)$’s for $\bb \in \operatorname{\mathbf{B}}$. Since the classes of all $M(\bb)$ and $\Pi M(\bb)$ are linearly independent in the Grothendieck group of $\underline{\sO}$, the multiplicities $(M:M(\bb))$ and $(M:\Pi M(\bb))$ of $M(\bb)$ and $\Pi M(\bb)$ in a Verma flag of $M$ are independent of the particular choice of flag.
The following lemma follows from the general theory developed in [@B2 $\S$4]. We include a self-contained proof here in order to make this article independent of [@B2]. Later on, we will also give a self-contained proof of another fundamental fact established in [@B2], namely, that the projective supermodules $P(\ba)$ have Verma flags; cf. Theorem \[mainsplitthm\].
\[hom\] For $M \in \ob\sO^\Delta$ and $\bb \in \operatorname{\mathbf{B}}$, we have that $$\begin{aligned}
(M:M(\bb)) &= \dim \operatorname{Hom}_{\sO}(M, M(\bb)^\star)_\0,\\
(M:\Pi M(\bb)) &= \dim \operatorname{Hom}_{\sO}(M, M(\bb)^\star)_\1.\end{aligned}$$ Also, any direct summand of $M \in \ob \sO^\Delta$ possesses a Verma flag.
The first part of the lemma follows by induction on the length of the Verma flag, using the following two observations: for all $\ba,\bb \in \operatorname{\mathbf{B}}$ we have that
- $\operatorname{Hom}_{\sO}(M(\ba), M(\bb)^\star)$ is zero if $\ba \neq \bb$, and it is one-dimensional of even parity if $\ba =\bb$;
- $\operatorname{Ext}^1_{\sO}(M(\ba), M(\bb)^\star) = 0$.
To check these, for the first one, we use the universal property of $M(\ba)$ to see that $\operatorname{Hom}_{\sO}(M(\ba), M(\bb)^\star)$ is zero unless $\lambda_\ba \leq \lambda_\bb$. Similarly, on applying $\star$, it is zero unless $\lambda_\bb \leq \lambda_\ba$. Hence, we may assume that $\ba = \bb$. Finally, any non-zero homomorphism $M(\ba) \rightarrow M(\ba)^\star$ must send the head to the socle, so $\operatorname{Hom}_{\sO}(M(\ba), M(\ba)^\star)$ is evenly isomorphic to $\operatorname{Hom}_{\sO}(L(\ba), L(\ba)^\star)$, which is one-dimensional and even thanks to Lemma \[star\]. For the second property, we must show that all short exact sequences in $\underline{\sO}$ of the form $0 \to M(\ba)^\star \to M \to
M(\bb) \to 0$ or $\Pi M(\ba)^\star \to M \to
M(\bb) \to
0$ split. Either $\lambda_{\ba}$ or $\lambda_{\bb}$ is a maximal weight of $M$. In the latter case, using also Lemma \[projectivity\], we can use the universal property of $M(\bb)$ to construct a splitting of $M \twoheadrightarrow M(\bb)$. In the former case, we apply $\star$, the resulting short exact sequence splits as before, and then we dualize again.
The final statement of the lemma may be proved by mimicking the argument for semisimple Lie algebras from [@H §3.2].
Special projective superfunctors {#s3}
================================
Next, we investigate the superfunctors $U \otimes -$ and $U^* \otimes -$ defined by tensoring with the $\g$-supermodules $U$ and $U^*$ introduced in the previous section. They clearly preserve the properties of being finitely generated over $\g$, locally finite-dimensional over $\b$, and semisimple over $\t$. Since the $\t$-weights of $U$ and $U^*$ are $\delta_1,\dots,\delta_n$ and $-\delta_1,\dots,-\delta_n$, respectively, and using (\[handier\]), we get for each $M \in \ob \sO$ that all weights of $U \otimes M$ and $U^* \otimes M$ are of the form $\lambda_\bb$ for $\bb \in \operatorname{\mathbf{B}}$. Hence, these superfunctors send objects of $\sO$ to objects of $\sO$, i.e. we have defined $$\label{sfdef}
\sF := U \otimes -:\sO \rightarrow \sO,
\qquad
\sE := U^* \otimes -:\sO \rightarrow \sO.$$ Recalling the notation (\[cough1\])–(\[cough2\]), let $$\label{omegadef}
\omega :=
\sum_{r,s=1}^n \left(f_{r,s}\otimes e_{s,r} -
f_{r,s}'\otimes e_{s,r}'\right)
\in U(\widehat{\g}) \otimes U(\g).$$ This is essentially the same tensor as in [@HKS (7.2.1)]. Left multiplication by $\omega$ (resp. by $-\omega$) defines a linear map $x_M:U\otimes M \rightarrow U \otimes M$ (resp. $x_M^*:U^* \otimes M \rightarrow U^* \otimes M$) for each $\g$-supermodule $M$. The following lemma shows that these maps define a pair of even supernatural transformations $$\label{xdef}
x:\sF \Rightarrow \sF,
\qquad
x^*:\sE \Rightarrow \sE.$$ This is also established in the proof of [@HKS Theorem 7.4.1], but we give a self-contained argument since it explains the origin of these maps.
The linear maps $x_M$ and $x_M^*$ just defined are even $\g$-supermodule homomorphisms.
The odd element $$\label{oddz}
f' := \sum_{t=1}^n f_{t,t}'
\in U(\widehat\g)$$ supercommutes with the elements of $U(\g)$. Hence, $f' \otimes 1 \in U(\widehat{\g}) \otimes U(\g)$ supercommutes with the image of the comultiplication $\Delta:U(\g) \rightarrow U(\g) \otimes U(\g) \subset U(\widehat\g) \otimes U(\g)$. The odd Casimir tensor $$\Omega' :=
\sum_{r,s=1}^n
\left(e_{r,s} \otimes e_{s,r}' - e_{r,s}' \otimes e_{s,r}\right)
\in U(\g) \otimes U(\g)$$ also supercommutes with the image of $\Delta$. Hence, the even tensor $$\Omega := \Omega' (f' \otimes 1)
=
-\sum_{r,s,t=1}^n \left(e_{r,s} f_{t,t}'\otimes e_{s,r}' + e_{r,s}' f_{t,t}'\otimes
e_{s,r} \right)
\in U(\widehat\g) \otimes U(\g)$$ commutes with the image of $\Delta$. Consequently, left multiplication by $\Omega$ defines even $\g$-supermodule endomorphisms $x_M:U \otimes M \rightarrow U \otimes M$ and $x_M^*:U^* \otimes M\rightarrow U^* \otimes M$. It remains to observe that these endomorphisms agree with the linear maps defined by left multiplication by $\omega$ and $-\omega$, respectively. Indeed, by a calculation using (\[f1\])–(\[f4\]), the elements $e_{r,s} f'_{t,t}$ and $e_{r,s}' f_{t,t}'$ of $U(\widehat\g)$ act on vectors in $U$ (resp. $U^*$) in the same way as $\delta_{s,t}
f_{r,s}'$ and $-\delta_{s,t} f_{r,s}$ (resp. $-\delta_{r,t} f_{r,s}'$ and $\delta_{r,t} f_{r,s}$), respectively.
\[eigenvalues\] Suppose that $\bb \in \operatorname{\mathbf{B}}$ and let $M := M(\bb)$.
1. There is a filtration $$0 = M_0 \subset M_1 \subset \cdots \subset M_n =
U \otimes M$$ with $M_{t} / M_{t-1} \cong M(\bb + \sigma_t \bd_t) \oplus \Pi M(\bb +
\sigma_t \bd_t)$ for each $t=1,\dots,n$. The endomorphism $x_{M}$ preserves this filtration, and the induced endomorphism of $M_t / M_{t-1}$ is diagonalizable with exactly two eigenvalues $\pm \sqrt{\lambda_{\bb, t}}\sqrt{\lambda_{\bb,t}+1}$. Its $\sqrt{\lambda_{\bb, t}}\sqrt{\lambda_{\bb,t} + 1}$-eigenspace is evenly isomorphic to $M(\bb + \sigma_t \bd_t)$, while the other eigenspace is evenly isomorphic to $\Pi M(\bb + \sigma_t \bd_t)$.
2. There is a filtration $$0 = M^{n} \subset \cdots \subset M^1 \subset M^0 =
U^*
\otimes M$$ with $M^{t-1} / M^{t} \cong M(\bb - \sigma_t \bd_t) \oplus \Pi M(\bb -
\sigma_t \bd_t)$ for each $t=1,\dots,n$. The endomorphism $x^*_{M}$ preserves this filtration, and the induced endomorphism of $M^{t-1} / M^{t}$ is diagonalizable with exactly two eigenvalues $\pm \sqrt{\lambda_{\bb, t}}\sqrt{\lambda_{\bb,t}-1}$. Its $\sqrt{\lambda_{\bb, t}}\sqrt{\lambda_{\bb,t} - 1}$-eigenspace is evenly isomorphic to $M(\bb - \sigma_t \bd_t)$, while the other eigenspace is evenly isomorphic to $\Pi M(\bb - \sigma_t \bd_t)$.
\(1) The filtration is constructed in [@B1 Lemma 4.3.7], as follows. By the tensor identity $$U \otimes M = U \otimes (U(\g) \otimes_{U(\b)} \u(\bb))
\cong U(\g) \otimes_{U(\b)} (U \otimes \u(\bb)).$$ As a $\b$-supermodule, $U$ has a filtration $0 = U_0 \subset U_1 \subset \cdots\subset U_n = U$ in which the section $U_t / U_{t-1}$ is spanned by the images of $u_t$ and $u_t'$. Let $M_t$ be the submodule of $U \otimes M$ that maps to $U(\g) \otimes_{U(\b)} (U_t \otimes \u(\bb))$ under this isomorphism.
Now fix $t \in \{1,\dots,n\}$. Let $v_1,\dots,v_k$ be a basis for the even highest weight space $M(\bb)_{\lambda_\bb,\0}$, so that $d_t' v_1,\dots,d_t' v_k$ is a basis for $M(\bb)_{\lambda_\bb,\1}$. The subquotient $M_t / M_{t-1} \cong U(\g) \otimes_{U(\b)} (U_t / U_{t-1} \otimes \u(\bb))$ is generated by the images of the vectors $\{u_t\otimes v_i, u_t' \otimes v_i,
u_t \otimes d_t'v_i, u_t' \otimes d_t'v_i\:|\:
i=1,\dots,k\}$, which by weight considerations span a $\b$-supermodule isomorphic to $\u(\bb + \sigma_t\bd_t) \oplus \Pi \u(\bb + \sigma_t\bd_t)$. Hence, $$M_t / M_{t-1} \cong M(\bb+\sigma_t \bd_t) \oplus \Pi M(\bb+\sigma_t \bd_t).$$ The action of $f_{r,s} \otimes e_{s,r} - f_{r,s}' \otimes e_{s,r}'$ on any of $u_t \otimes v_i, u_t' \otimes v_i, u_t \otimes d_t'v_i$ or $u_t'\otimes d_t'v_i$ is zero unless $r \leq s =t$, and if $r < s = t$ then it sends these vectors into $M_{t-1}$. Therefore, $x_{M}$ preserves the filtration. Moreover, this argument shows that it acts on the highest weight space of the quotient $M_t / M_{t-1}$ in the same way as $x_t := f_{t,t}\otimes d_t - f_{t,t}'\otimes d'_t$.
Now consider the purely even subspace $S_{i,t}$ of $M_t / M_{t-1}$ with basis given by the images of $u_t \otimes v_i, u_t' \otimes d_t'v_i$. Recalling that $d_t$ acts on $v_i$ and on $d_t' v_i$ by $\lambda_{\bb,t}$, and that $(d_t')^2 = d_t$, it is straightforward to check that the matrix of the endomorphism $x_t$ of $S_{i,t}$ in the given basis is equal to $$A := \left( \begin{array}{cc} \lambda_{\bb,t} & \lambda_{\bb,t} \\
1 & -
\lambda_{\bb,t} \end{array}
\right).$$ Also recall from our construction of $\u(\bb)$ that $d_1'\cdots d_{2m}'$ acts on $v_i$ as the scalar $c_\bb$ from (\[y1\]), and it acts on $d_t' v_i$ as $-c_{\bb}$. Using this, another calculation shows that $d_1'\cdots d_{2m}'$ acts on $S_{i,t}$ as the matrix $\frac{c_\bb}{\lambda_{\bb,t}} A$. Similarly, on the purely odd subspace $S_{i,t}'$ with basis given by the images of $u_t' \otimes v_i, u_t \otimes d_t'v_i$, $x_t$ has matrix $-A$ and $d_1'\cdots d_{2m}'$ has matrix $-\frac{c_\bb}{\lambda_{\bb,t}} A$.
Since the matrix $A$ has eigenvalues $\pm
\sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t}+1}$, the calculation made in the previous paragraph implies that $x_t$ is diagonalizable on $M_t / M_{t-1}$ with exactly these eigenvalues. Moreover on any even highest weight vector in its $
\sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t}+1}$-eigenspace, we get that $d_1'\cdots d_{2m}'$ acts as $$\begin{aligned}
\frac{c_{\bb}}{\lambda_{\bb,t}} \sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t} + 1}
&= c_{\bb+\sigma_t \bd_t}.\end{aligned}$$ This implies that the $\sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t}+1}$-eigenspace is evenly isomorphic to $M(\bb+\sigma_t \bd_t)$. Similarly, the $-\sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t}+1}$-eigenspace is evenly isomorphic to $\Pi M(\bb+\sigma_t \bd_t)$.
\(2) Similar.
\[minpoly\] For $M \in \ob\sO$, all roots of the minimal polynomials of $x_M$ and $x_M^*$ (computed in the finite dimensional superalgebras $\operatorname{End}_{\sO}(\sF\,M)$ and $\operatorname{End}_{\sO}(\sE\,M)$) belong to the set $J$ from (\[jdef\]).
This is immediate from the theorem in case $M$ is a Verma supermodule. We may then deduce that it is true for all irreducibles, hence, for any $M
\in \ob \sO$.
Corollary \[minpoly\] implies that we can decompose $$\label{cc}
\sF = \bigoplus_{j \in J} \sF_j,
\qquad
\sE = \bigoplus_{j \in J} \sE_j,$$ where $\sF_j$ (resp. $\sE_j$) is the subfunctor of $\sF$ (resp. $\sE$) defined by letting $\sF_j\, M$ (resp. $\sE_j\, M$) be the generalized $j$-eigenspace of $x_M$ (resp. $x_M^*$) for each $M \in \ob\sO$. Recall that $I$ denotes $\operatorname{\mathbb{Z}}$. For $i \in I$, we define the [*$i$-signature*]{} of $\bb \in \operatorname{\mathbf{B}}$ to be the $n$-tuple $\isig(\bb) = (\isig_1(\bb),\dots,\isig_n(\bb)) \in\{\mathtt{e}, \mathtt{f}, \bullet\}^n$ with $$\label{sigdef}
\isig_t(\bb) :=
\left\{
\begin{array}{ll}
\mathtt{f}&\text{if either }
\sigma_t = +\text{ and }b_t = i,
\text{ or }
\sigma_t = -\text{ and }b_t = i+1,\\
\mathtt{e}&\text{if either }
\sigma_t = + \text{ and }b_t = i+1,
\text{ or }
\sigma_t = -\text{ and }b_t = i,\\
\bullet&\text{otherwise.}
\end{array}
\right.$$
\[maintfthm\] Given $\bb \in \operatorname{\mathbf{B}}$ and $i \in I$, let $j := \sqrt{z+i}\sqrt{z+i+1}$. Then:
- $\sF_j \,M(\bb)$ has a multiplicity-free filtration with sections that are evenly isomorphic to the Verma supermodules $$\{M(\bb+\sigma_t \bd_t)\:|\:\text{for }1 \leq t \leq n\text{ such that }\isig_t(\bb) =
\mathtt{f}\},$$ appearing from bottom to top in order of increasing index $t$.
- $\sE_j \,M(\bb)$ has a multiplicity-free filtration with sections that are evenly isomorphic to the Verma supermodules $$\{M(\bb-\sigma_t \bd_t)\:|\:\text{for }1 \leq t \leq n\text{ such that
}\isig_t(\bb) = \mathtt{e}\},$$ appearing from top to bottom in order of increasing index $t$.
\(1) It is immediate from Lemma \[eigenvalues\] that $\sF_j \,M(\bb)$ has a multiplicity-free filtration with sections that are evenly isomorphic to the supermodules $M(\bb+\sigma_t \bd_t)$ for $t=1,\dots,n$ such that $\sqrt{\lambda_{\bb,t}}\sqrt{\lambda_{\bb,t}+1} = j$. Squaring both sides, this equation implies that $\left(\lambda_{\bb,t}+\half\right)^2 = \left(z+i+\half\right)^2$. Hence, $$\lambda_{\bb,t} = \sigma_t(z+b_t) =
-\half \pm \left(z+i+\half\right).$$ We deduce either that $\sigma_t = +$ and $b_t = i$, or $\sigma_t = -$ and $b_t = i+1$. Since we squared our original equation, it remains to check that we do indeed get solutions to that in both cases. This is clear in the case that $\sigma_t =
+$, and it follows in the case that $\sigma_t = -$ using also (\[violin\]).
\(2) Similar.
Using Theorem \[hkst\] below, one can show that there are odd supernatural isomorphisms $c:\sF_{j}\stackrel{\sim}{\Rightarrow} \sF_{-j}$ and $c^*:\sE_{j}\stackrel{\sim}{\Rightarrow} \sE_{-j}$ for each $j \in J$. One consequence (which could be checked directly right away) is that there is another version of Theorem \[maintfthm\], in which one takes $j := -\sqrt{z+i}\sqrt{z+i+1}$ and replaces the Verma supermodules $M(\bb \pm \sigma_t \bd_t)$ in the statement by their parity flips.
The superfunctors $\sF$ and $\sE$ are both left and right adjoint to each other via some canonical (even) adjunctions. The adjunction making $(\sE, \sF)$ into an adjoint pair is induced by the linear maps $$\eps:U^* \otimes U \rightarrow \operatorname{{\Bbbk}},\:
\phi \otimes u \mapsto \phi(u),
\qquad
\eta:\operatorname{{\Bbbk}}\rightarrow U \otimes U^*,\:
1 \mapsto \sum_{r=1}^n (u_r \otimes \phi_r + u_r' \otimes \phi_r').$$ Thus, the unit of adunction $c:1 \Rightarrow \sF \,\sE$ is defined on supermodule $M$ by the map $c_M:M \stackrel{\operatorname{can}}{\longrightarrow}
\operatorname{{\Bbbk}}\otimes M \stackrel{\eta \otimes \operatorname{{\operatorname{id}}}}{\longrightarrow} U \otimes
U^* \otimes M$, and the counit of adjunction $d:\sE\, \sF \Rightarrow 1$ is defined by $d_M:U^* \otimes U \otimes M
\stackrel{\eps\otimes\operatorname{{\operatorname{id}}}}{\longrightarrow} \operatorname{{\Bbbk}}\otimes M
\stackrel{\operatorname{can}}{\longrightarrow} M$. Similarly, the adjunction making $(\sF, \sE)$ into an adjoint pair is induced by the linear maps $$U \otimes U^* \rightarrow \operatorname{{\Bbbk}},\,
u \otimes \phi \mapsto (-1)^{|\phi||u|}\phi(u),
\quad
\operatorname{{\Bbbk}}\rightarrow U^* \otimes U,\,
1 \mapsto \sum_{r=1}^n (\phi_r \otimes u_r - \phi_r' \otimes u_r').$$ The following lemma implies that these adjunctions restrict to adjunctions making $(\sF_j, \sE_j)$ and $(\sE_j, \sF_j)$ into adjoint pairs for each $j \in J$. It follows that all of these superfunctors send projectives to projectives, and they are all exact, i.e. they preserve short exact sequences in $\underline{\sO}$.
The supernatural transformation $x^*:\sE \Rightarrow \sE$ is both the left and right mate of $x:\sF \Rightarrow \sF$ with respect to the canonical adjunctions defined above.
We just explain how to check that $x^*$ is the left mate of $x$ with respect to the adjunction $(\sE, \sF)$; the argument for right mate is similar. We need to show for each $M \in \ob \sO$ that the composition $$U^* \otimes M \stackrel{\operatorname{id} \otimes c_{M}}{\longrightarrow} U^* \otimes U
\otimes U^* \otimes M
\stackrel{\operatorname{id}\otimes x_{U^* \otimes M}}{\longrightarrow}
U^* \otimes U \otimes U^* \otimes M \stackrel{d_{U^* \otimes M}}{\longrightarrow} U^* \otimes M$$ is equal to $x_M^*:U^*\otimes M \rightarrow U^* \otimes M$. Recall for this that $x_M^*$ is defined by left multiplication by $\sum_{r,s=1}^n \left(f_{r,s}' \otimes e_{s,r}' -
f_{r,s}\otimes e_{s,r}\right)$, while $x_{U^* \otimes M}$ is defined by left multiplication by $\sum_{r,s=1}^n (f_{r,s} \otimes e_{s,r}\otimes 1 + f_{r,s} \otimes 1 \otimes e_{s,r}
- f_{r,s}' \otimes e_{s,r}'\otimes 1 - f_{r,s}' \otimes 1 \otimes
e_{s,r}')$. Now one computes the effect of both maps on homogeneous vectors of the form $\phi_t\otimes v$ and $\phi_t' \otimes v$ using (\[f1\])–(\[f4\]).
Bruhat order {#s4}
============
Consider the Dynkin diagram $
{\begin{picture}(113, 15)\put(18,2){\circle{4}}\put(37,2){\circle{4}}\put(56,2){\circle{4}}\put(75, 2){\circle{4}}\put(94, 2){\circle{4}}\put(20, 2){\line(1, 0){15.5}}\put(39, 2){\line(1, 0){15.5}}\put(58, 2){\line(1, 0){15.5}}\put(77, 2){\line(1, 0){15.5}}\put(98, 2){\line(1, 0){1}}\put(101, 2){\line(1, 0){1}}\put(104, 2){\line(1, 0){1}}\put(107, 2){\line(1, 0){1}}\put(110, 2){\line(1, 0){1}}\put(4, 2){\line(1, 0){1}}\put(1, 2){\line(1, 0){1}}\put(7, 2){\line(1, 0){1}}\put(10, 2){\line(1, 0){1}}\put(13, 2){\line(1, 0){1}}\put(15, 8){\makebox(0, 0)[b]{$_{{-2}}$}}\put(34, 8){\makebox(0, 0)[b]{$_{{-1}}$}}\put(56, 8){\makebox(0, 0)[b]{$_{0}$}}\put(75, 8){\makebox(0, 0)[b]{$_{1}$}}\put(94, 8){\makebox(0, 0)[b]{$_{{2}}$}}\end{picture}}
$ whose vertices are indexed by the totally ordered set $I = \operatorname{\mathbb{Z}}$. We denote the associated Kac-Moody algebra by $\operatorname{\mathfrak{sl}}_{\infty}$. This is the Lie algebra of traceless, finitely-supported complex matrices whose rows and columns are indexed by $I$. It is generated by the matrix units $f_i :=e_{i+1, i}$ and $e_i := e_{i, i+1}$ for $i \in I$. The *natural representation* $V^{+}$ of $\mathfrak{sl}_{\infty}$ is the module of column vectors with standard basis $\{ v^{+}_i\:|\: i \in I \}$. We also need the [*dual natural representation*]{} $V^-$ with basis $\{v_i^-\:|\:i \in I\}$. The action of the Chevalley generators on these bases is given by $$\begin{aligned}
\label{e}
e_i v^{+}_j &= \delta_{i+1,j} v^{+}_i,
&
e_i v^{-}_j &= \delta_{i,j} v^{-}_{i+1},\\
f_i v^{+}_j &= \delta_{i,j} v^{+}_{i+1},
& f_i v^{-}_j &= \delta_{i+1,j} v^{-}_i.
\label{f}\end{aligned}$$ The tensor product $V^{\otimes \bsigma} := V^{\sigma_1} \otimes \cdots \otimes V^{\sigma_n}$ has monomial basis $\{v_\bb\:|\:\bb \in \operatorname{\mathbf{B}}\}$ defined from $v_\bb := v^{\sigma_1}_{b_1} \otimes \cdots \otimes v^{\sigma_n}_{ b_n
}$. Recalling (\[sigdef\]), the Chevalley generators act on these monomials by $$\begin{aligned}
\label{haha}
f_i v_{\bb} &= \sum_{\substack{1 \leq t \leq n \\ \isig_t(\bb)
=\mathtt{f}}} v_{\bb+\sigma_t \bd_t},
&
e_i v_{\bb} &= \sum_{\substack{1 \leq t \leq n\\ \isig_t(\bb) =
\mathtt{e}}} v_{\bb-\sigma_t \bd_t}.\end{aligned}$$ This should be compared with Theorem \[maintfthm\], which already makes some connection between the endofunctors $\sF_j, \sE_j$ of $\sO$ and the $\mathfrak{sl}_\infty$-module $V^{\otimes \bsigma}$.
We next introduce an important partial order $\succeq$ on $\operatorname{\mathbf{B}}$, which we call the [Bruhat order]{} It is closely related to the [inverse dominance order]{} of [@LW Definition 3.2], which comes from Lusztig’s construction of tensor products of based modules [@Lubook $\S$27.3]. The root system of $\mathfrak{sl}_\infty$ has *weight lattice* $P := \bigoplus_{i \in I} \operatorname{\mathbb{Z}}\omega_i$ where $\omega_i$ is the [$i$th fundamental weight]{}. For $i \in I$, we set $$\eps_i := \omega_i - \omega_{i-1},
\qquad
\alpha_i := \eps_i - \eps_{i+1}.
$$ We identify $\eps_i$ with the weight of the vector $v_i^+$ in the $\mathfrak{sl}_\infty$-module $V^+$. Then, $v_i^- \in V^-$ is of weight $-\eps_i$. For $\bb \in \operatorname{\mathbf{B}}$, let $$\operatorname{\mathbf{wt}}(\bb) = (\operatorname{wt}_1(\bb),\dots,\operatorname{wt}_n(\bb)) \in P^n$$ be the $n$-tuple of weights defined from $\operatorname{wt}_r(\bb) := \sigma_r \eps_{b_r}$, so that $v_\bb \in V^{\otimes\bsigma}$ is of weight $|\operatorname{\mathbf{wt}}(\bb)| := \operatorname{wt}_1(\bb)+\cdots+\operatorname{wt}_n(\bb) \in P$. Because the weight spaces of $V^\pm$ are all one-dimensional, the map $\operatorname{\mathbf{B}}\rightarrow P^n, \bb \mapsto \operatorname{\mathbf{wt}}(\bb)$ is injective.
\[bruhatdef\] Let $\trianglelefteq$ denote the [dominance order]{} on $P$, so $\beta \trianglelefteq \gamma\Leftrightarrow\gamma - \beta \in \bigoplus_{i \in
I} \operatorname{\mathbb{N}}\alpha_i$. The [*inverse dominance order*]{} on $P^n$ is the partial order defined by declaring that $(\beta_1,\dots,\beta_n) \succeq (\gamma_1,\dots,\gamma_n)$ if and only if $$\beta_1+\cdots+\beta_s \trianglelefteq \gamma_1+\cdots+\gamma_s,$$ for each $s=1,\dots,n$, with the inequality being an equality when $s=n$. Finally, define the [*Bruhat order*]{} $\succeq$ on $\operatorname{\mathbf{B}}$ by $\ba \succeq \bb \Leftrightarrow \operatorname{\mathbf{wt}}(\ba) \succeq
\operatorname{\mathbf{wt}}(\bb)$.
Our first lemma makes the definition of the Bruhat order more explicit. Using it, one can check in particular that $\ba \succeq \bb$ implies that $\lambda_\ba \geq
\lambda_\bb$ in the dominance order on $\mathfrak{t}^*$; cf. [@BLW Lemma 3.4].
\[eleme\] For $\ba \in \operatorname{\mathbf{B}}, i \in I$ and $1 \leq s \leq n$, we let $$N_{[1,s]}(\ba,i)
:=
\#\{1 \leq r \leq s\:|\:a_r > i, \sigma_r = +\}
-
\#\{1 \leq r \leq s\:|\:a_r > i, \sigma_r = -\}.$$ Then we have that $\ba \succeq \bb$ if and only if
- $N_{[1,n]}(\ba, i) = N_{[1,n]}(\bb, i)$ for all $i \in I$;
- $N_{[1,s]}(\ba, i) \geq N_{[1,s]}(\bb, i)$ for all $i \in I$ and $s = 1,\dots,n-1$.
This is a special case of [@BLW Lemma 2.17].
\[not\] Assume that $\ba \succeq \bb$ and $\isig_r(\ba) = \isig_n(\bb) =
\mathtt{f}$ for some $i \in I$ and $1 \leq r \leq n$. Then $\ba + \sigma_r \bd_r \succeq \bb + \sigma_n \bd_n$, with equality if and only if $\ba = \bb$ and $r=n$.
We use the conditions from Lemma \[eleme\]. For [*either* ]{} $j \neq i$ and $1 \leq s \leq n$, [*or*]{} $j = i$ and $1 \leq s < r$, we have that $$N_{[1,s]}(\ba+\sigma_r \bd_r,j)
=
N_{[1,s]}(\ba,j)
\geq
N_{[1,s]}(\bb,j)
=
N_{[1,s]}(\bb+\sigma_n \bd_n,j).$$ For $r \leq s < n$, we have that $$N_{[1,s]}(\ba+\sigma_r \bd_r, i)
=
N_{[1,s]}(\ba,i)+1
\geq
N_{[1,s]}(\bb,i)+1
>
N_{[1,s]}(\bb,i)
=
N_{[1,s]}(\bb+\sigma_n \bd_n,i).$$ Finally, $
N_{[1,n]}(\ba+\sigma_r \bd_r,i)
=
N_{[1,n]}(\ba,i)+1
=
N_{[1,n]}(\bb,i)+1
=
N_{[1,n]}(\bb+\sigma_n \bd_n,i).
$
To prepare for the next lemma, suppose that we are given $\bb \in \operatorname{\mathbf{B}}$. Define $\ba \in \operatorname{\mathbf{B}}$ by setting $a_1 := b_1$, then inductively defining $a_s$ for $s=2,\dots,n$ as follows.
- If $\sigma_s = +$ then $a_s$ is the greatest integer such that $a_s \leq b_s$, and the following hold for all $1 \leq r < s$:
- if $\sigma_r=+$ then $a_s < a_r$;
- if $\sigma_r=-$ then $a_s < b_r$.
- $a_s \leq a_r$ if $\sigma_r=+$ and $a_r = b_r$;
$a_s$ is the smallest of the following: $b_s$; $b_r-1$ for all $1 \leq r < s$ with $\sigma_r = -$; $a_r-1$ for all $1 \leq r < s$ with $\sigma_r = +$ and $a_r \neq b_r$; $a_r$ for all $1 \leq r < s$ with $\sigma_r = +$ and $a_r = b_r$.
- If $\sigma_s = -$ then $a_s$ is the smallest integer such that $a_s \geq b_s$, and the following hold for all $1 \leq r < s$:
- if $\sigma_r=-$ then $a_s > a_r$;
- if $\sigma_r=+$ then $a_s > b_r$.
If $\sigma_s = -$ then $a_s$ is the greatest of the following: $b_s$; $b_r+1$ for all $1 \leq r < s$ with $\sigma_r = +$; $a_r+1$ for all $1 \leq r < s$ with $\sigma_r = -$ and $a_r \neq b_r$; $a_r$ for all $1 \leq r < s$ with $\sigma_r = -$ and $a_r = b_r$.
Also define a monomial $X = X_n \cdots X_2$ in the Chevalley generators $\{f_i\:|\:i \in I\}$ by setting $$X_r := \left\{
\begin{array}{ll}
f_{b_r-1} \cdots f_{a_r+1} f_{a_r}&\text{if $\sigma_r = +$,}\\
f_{b_r} \cdots f_{a_r-2} f_{a_r-1}&\text{if $\sigma_r = -$,}
\end{array}\right.$$ for each $r=2,\dots,n$.
If $\bsigma = (+,+,-,+,-,-)$ and $\bb = (3,4,3,4,3,4)$, then $\ba = (3,2,5,1,6,7)$ and $X = (f_4 f_5 f_6) (f_3 f_4 f_5) (f_3 f_2
f_1) (f_3 f_4) (f_3 f_2)$.
\[construction\] In the above notation, we have that $X v_\ba = v_\bb +
(\text{a sum of $v_\bc$'s
for $\bc \succ \bb$})$.
We proceed by induction on $n$, the result being trivial in case $n=1$. For $n > 1$, let $\bar\bsigma :=
(\sigma_1,\dots,\sigma_{n-1})$, $\bar\ba := (a_1,\dots,a_{n-1})$, $\bar\bb
:= (b_1,\dots,b_{n-1})$ and $\bar X := X_{n-1} \cdots X_2$. Applying the induction hypothesis in the $\mathfrak{sl}_\infty$-module $V^{\otimes \bar\bsigma}$, we get that $$\bar X v_{\bar\ba} = v_{\bar\bb} +
(\text{a sum of $v_{\bar\bc}$'s
for $\bar\bc \succ \bar\bb$}).$$ Now we observe that if $f_i$ is a Chevalley generator appearing in one of the monomials $X_r$ for $r < n$, then $f_i v^{\sigma_n}_{a_n} = 0$. This follows from the definitions: if $\sigma_n = +$ we must show that $i \neq a_n$, which follows as $i \geq a_r > a_n$ if $\sigma_r = +$ or $i \geq b_r > a_n$ if $\sigma_r=-$; if $\sigma_n = -$ we must show that $i \neq a_n-1$, which follows as $i < b_r < a_n$ if $\sigma_r = +$ or $i < a_r < a_n$ if $\sigma_r=-$. Hence, letting $\tilde \bb := (b_1,\dots,b_{n-1}, a_n)$, we deduce that $$\bar X v_{\ba} = v_{\tilde\bb} +
(\text{a sum of $v_{\bc}$'s
for $\bc \succ \tilde\bb$}).$$ Finally we act with $X_n$, which sends $v^{\sigma_n}_{a_n}$ to $v^{\sigma_n}_{b_n}$, and apply Lemma \[not\].
\[mainsplitthm\] For every $\bb \in \operatorname{\mathbf{B}}$, the indecomposable projective supermodule $P(\bb)$ has a Verma flag with top section evenly isomorphic to $M(\bb)$ and other sections evenly isomorphic to $M(\bc)$’s for $\bc \in \operatorname{\mathbf{B}}$ with $\bc \succ \bb$.
Let notation be as in Lemma \[construction\]. Let $i_1,\dots,i_l \in I$ be defined so that $X$ is the monomial $f_{i_l} \cdots f_{i_2} f_{i_1}$. Let $j_k := \sqrt{z+i_k}\sqrt{z+i_k+1}$ for each $k$ and consider the supermodule $$P := \sF_{j_l} \cdots \sF_{j_2} \sF_{j_1} M(\ba).$$ For each $1 \leq r < s \leq n$, we have that $a_r > a_s$ if $\sigma_r = \sigma_s = +$, $a_r < a_s$ if $\sigma_r = \sigma_s = -$, and $a_r \neq a_s$ if $\sigma_r \neq \sigma_s$. This implies that the weight $\lambda_\ba$ is typical and dominant, hence $M(\ba)$ is projective by Lemma \[typicaldominant\]. Since each $\sF_j$ sends projectives to projectives, we deduce that $P$ is projective. Since the combinatorics of (\[haha\]) matches that of Theorem \[maintfthm\], we can reinterpret Lemma \[construction\] as saying that $P$ has a Verma flag with one section evenly isomorphic to $M(\bb)$ and all other sections evenly isomorphic to $M(\bc)$’s for $\bc \succ
\bb$. In fact, the unique section isomorphic to $M(\bb)$ appears at the top of this Verma flag, thanks the order of the sections arising from Theorem \[maintfthm\](1). Hence, $P$ has a summand evenly isomorphic to $P(\bb)$, and it just remains to apply Lemma \[hom\].
\[bruhat\] For $\bb \in \operatorname{\mathbf{B}}$, we have that $[M(\bb):L(\bb)] = 1$. All other composition factors of $M(\bb)$ are evenly isomorphic to $L(\bc)$’s for $\bc \prec \bb$.
This follows from Theorem \[mainsplitthm\] and the following analog of [*BGG reciprocity*]{}: for $\ba, \bb \in \operatorname{\mathbf{B}}$, we have that $$\begin{aligned}
[M(\bb):L(\ba)]
&= \dim \operatorname{Hom}_{\sO}(P(\ba), M(\bb)^\star)_\0
=
(P(\ba):M(\bb)),\\
[M(\bb):\Pi L(\ba)]
&= \dim \operatorname{Hom}_{\sO}(P(\ba), M(\bb)^\star)_\1
=
(P(\ba):\Pi M(\bb)).\end{aligned}$$ The various equalities here follow from Lemmas \[hom\] and \[star\].
\[split\] For any $\bb \in \operatorname{\mathbf{B}}$, every irreducible subquotient of the indecomposable projective $P(\bb)$ is evenly isomorphic to $L(\ba)$ for $\ba \in \operatorname{\mathbf{B}}$ with $|\operatorname{\mathbf{wt}}(\ba)| = |\operatorname{\mathbf{wt}}(\bb)|$.
By Theorem \[mainsplitthm\], $P(\bb)$ has a Verma flag with sections $M(\bc)$ for $\bc
\succeq \bb$. By Corollary \[bruhat\], the composition factors of $M(\bc)$ are $L(\ba)$’s for $\ba \preceq \bc$. Hence, every irreducible subquotient of $P(\bb)$ is evenly isomorphic to $L(\ba)$ for $\ba \in \operatorname{\mathbf{B}}$ such that $\ba \preceq \bc
\succeq \bb$ for some $\bc$. This condition implies that $|\operatorname{\mathbf{wt}}(\ba)| = |\operatorname{\mathbf{wt}}(\bb)|$.
Weak categorical action {#s5}
=======================
Let $\O$ be the Serre subcategory of $\sO$ generated by $
\{L(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\},$ i.e. it is the full subcategory of $\sO$ consisting of all supermodules whose composition factors are evenly isomorphic to $L(\bb)$’s for $\bb \in \operatorname{\mathbf{B}}$. Since each $L(\bb)$ is of type $\mathtt{M}$, there are no non-zero odd morphisms between objects of $\O$. Because of this, we forget the $\operatorname{\mathbb{Z}}/2$-grading and simply view $\O$ as a $\operatorname{{\Bbbk}}$-linear category rather than a supercategory.
\[itsplits\] We have that $\sO = \O \oplus \Pi \O$ in the sense defined in the introduction.
Let $\Pi \O$ be the Serre subcategory of $\sO$ generated by $\{\Pi L(\ba)\:|\:\ba \in \operatorname{\mathbf{B}}\}.$ By Corollary \[split\], all even extensions between $\Pi L(\ba)$ and $L(\bb)$ are split. Hence, every supermodule in $\sO$ decomposes uniquely as a direct sum of an object of $\O$ and an object of $\Pi \O$. The result follows.
For typical blocks, Theorem \[itsplits\] has a more direct proof exploiting the action of the anticenter of $U(\g)$; see [@F $\S$3.1].
In order to state our next theorem, we briefly recall the following definition due to Cline, Parshall and Scott [@CPS]:
\[hwdef\] A [*highest weight category*]{} is a Schurian category $\C$ in the sense of Definition \[schurcat\], together with an interval-finite poset $(\Lambda, \leq)$ indexing a complete set of irreducible objects $\{L(\lambda)\:|\:\lambda \in \Lambda\}$, subject to the following axiom. For each $\lambda \in \Lambda$, let $P(\lambda)$ be a projective cover of $L(\lambda)$ in $\C$. Define the [*standard object*]{} $\Delta(\lambda)$ to be the largest quotient of $P(\lambda)$ such that $[\Delta(\lambda):L(\lambda)] = 1$ and $[\Delta(\lambda):L(\mu)] = 0$ for $\mu \not\leq \lambda$. Then we require that $P(\lambda)$ has a filtration with top section isomorphic to $\Delta(\lambda)$ and all other sections of the form $\Delta(\mu)$ for $\mu > \lambda$.
\[itshw\] The category $\O$ is a highest weight category with weight poset $(\operatorname{\mathbf{B}}, \preceq)$. Its standard objects are the Verma supermodules $\{M(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$.
It is clear that $\O$ is a Schurian category with isomorphism classes of irreducible objects represented by $\{L(\bb)\:|\:\bb \in \operatorname{\mathbf{B}}\}$. By Theorem \[mainsplitthm\], $P(\bb)$ has a Verma flag with $M(\bb)$ at the top and other sections that are evenly isomorphic to $M(\bc)$’s for $\bc \succ
\bb$. It just remains to observe that the Verma supermodules $M(\bb)$ coincide with the standard objects $\Delta(\bb)$. This follows using the filtration just described plus Corollary \[bruhat\].
By Lemma \[star\], the duality $\star$ on $\sO$ restricts to a duality $\star:\O \rightarrow \O$ fixing isomorphism classes of irreducible objects.
Next, take $i \in I$ and set $j := \sqrt{z+i}\sqrt{z+i+1}$. Theorem \[maintfthm\] implies that the exact functors $\sF_j$ and $\sE_j$ send the standard objects in $\O$ to objects of $\O$ with a Verma flag. Hence, they send arbitrary objects in $\O$ to objects of $\O$. Thus, their restrictions define endofunctors $$\label{them}
F_i := \sF_j|_{\O}:\O\rightarrow\O,
\quad
E_i := \sE_j|_{\O}:\O \rightarrow \O.$$ Again, these functors are both left and right adjoint to each other. Let $\O^\Delta$ be the full subcategory of $\O$ consisting of all objects possessing a Verma flag. This is an exact subcategory of $\O$. Its complexified Grothendieck group $\operatorname{\mathbb{C}}\otimes_{\operatorname{\mathbb{Z}}} K_0(\O^\Delta)$ has basis $\{[M(\bb)]\:|\:\bb \in \operatorname{\mathbf{B}}\}$.
\[itacts\] For each $i \in I$, the functors $F_i$ and $E_i$ are exact endofunctors of $\O^\Delta$. Moreover, if we identify $\operatorname{\mathbb{C}}\otimes_{\operatorname{\mathbb{Z}}} K_0(\O^\Delta)$ with $V^{\otimes \bsigma}$ so $[M(\bb)] \leftrightarrow v_\bb$ for each $\bb \in \operatorname{\mathbf{B}}$, then the induced endomorphisms $[F_i]$ and $[E_i]$ of the Grothendieck group act in the same way as the Chevalley generators $f_i$ and $e_i$ of $\mathfrak{sl}_\infty$.
Compare Theorem \[maintfthm\] with (\[haha\]).
Thus, we have constructed a highest weight category $\O$ with weight poset $(\operatorname{\mathbf{B}},\preceq)$, and equipped it with a weak categorical action of the Lie algebra $\mathfrak{sl}_\infty$ in the sense of [@CR; @Rou].
Strong categorical action {#s6}
=========================
In this section, we upgrade the weak categorical action of $\mathfrak{sl}_\infty$ on $\mathcal O$ constructed so far to a strong categorical action. For the following definition, we represent morphisms in a strict monoidal category via the usual string calculus, adopting the same conventions for horizontal and vertical composition as [@KL1].
\[qhdef\] The *quiver Hecke category* of type $\mathfrak{sl}_\infty$ is the strict $\operatorname{{\Bbbk}}$-linear monoidal category $\QH$ with objects generated by the set $I$ from (\[jdef\]), and morphisms generated by $\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkred] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.05) {$\color{darkred}\bullet$};
\node at (0.08,-.25) {$\scriptstyle{i}$};
\end{tikzpicture}
}:i \rightarrow i$ and $\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkred] (0.18,-.15) to (-0.18,.3);
\draw[-,thick,darkred] (-0.18,-.15) to (0.18,.3);
\node at (-0.18,-.25) {$\scriptstyle{i_2}$};
\node at (0.18,-.25) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}:i_2 \otimes i_1 \rightarrow i_1 \otimes i_2$, subject to the following relations: $$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-,thick,darkred] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkred] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{i_2}$};
\node at (0.25,-.28) {$\scriptstyle{i_1}$};
\node at (-0.14,0.42) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
-
\mathord{
\begin{tikzpicture}[baseline =2]
\draw[-,thick,darkred] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkred] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{i_2}$};
\node at (0.25,-.28) {$\scriptstyle{i_1}$};
\node at (0.14,-0.02) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
&=
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-,thick,darkred] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkred] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{i_2}$};
\node at (0.25,-.28) {$\scriptstyle{i_1}$};
\node at (-0.13,-0.02) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
-\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-,thick,darkred] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkred] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{i_2}$};
\node at (0.25,-.28) {$\scriptstyle{i_1}$};
\node at (0.14,0.42) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
=
\left\{
\begin{array}{ll}
\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkred] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkred] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_2}$};
\node at (0.08,-.4) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}
& \text{if $i_1 = i_2$,}\\
\:\:\:0 & \text{if $i_1 \neq i_2$;}\\
\end{array}
\right. \end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 8]
\draw[-,thick,darkred] (0.28,.4) to[out=90,in=-90] (-0.28,1.1);
\draw[-,thick,darkred] (-0.28,.4) to[out=90,in=-90] (0.28,1.1);
\draw[-,thick,darkred] (0.28,-.3) to[out=90,in=-90] (-0.28,.4);
\draw[-,thick,darkred] (-0.28,-.3) to[out=90,in=-90] (0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_2}$};
\node at (0.28,-.4) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}
&=
\left\{
\begin{array}{ll}
\:\:\:0&\text{if $i_1 = i_2$,}\\
(i_2-i_1)
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkred] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkred] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_2}$};
\node at (0.08,-.4) {$\scriptstyle{i_1}$};
\node at (0.08,0.05) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
+ (i_1-i_2)
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkred] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkred] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_2}$};
\node at (0.08,-.4) {$\scriptstyle{i_1}$};
\node at (-0.28,0.05) {$\color{darkred}\bullet$};
\end{tikzpicture}
}
&\text{if $|i_1 - i_2| = 1$,}\\
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkred] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkred] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_2}$};
\node at (0.08,-.4) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}&\text{if $|i_1 - i_2| > 1$;}\\
\end{array}
\right. \end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-,thick,darkred] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkred] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkred] (0,-.4) to[out=90,in=-90] (-.45,0.2);
\draw[-,thick,darkred] (-0.45,0.2) to[out=90,in=-90] (0,0.8);
\node at (-0.45,-.5) {$\scriptstyle{i_3}$};
\node at (0,-.5) {$\scriptstyle{i_2}$};
\node at (0.45,-.5) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}
\!\!-
\!\!\!
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-,thick,darkred] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkred] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkred] (0,-.4) to[out=90,in=-90] (.45,0.2);
\draw[-,thick,darkred] (0.45,0.2) to[out=90,in=-90] (0,0.8);
\node at (-0.45,-.5) {$\scriptstyle{i_3}$};
\node at (0,-.5) {$\scriptstyle{i_2}$};
\node at (0.45,-.5) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}
&=
\left\{
\begin{array}{ll}
(i_2 - i_1)
\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkred] (0.44,-.3) to (0.44,.4);
\draw[-,thick,darkred] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkred] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{i_3}$};
\node at (0.08,-.4) {$\scriptstyle{i_2}$};
\node at (0.44,-.4) {$\scriptstyle{i_1}$};
\end{tikzpicture}
}
&
\text{if $i_1 = i_3$ and $|i_1 - i_2|= 1$,}\\
\:\:\:0 &\text{otherwise.}
\end{array}
\right.\end{aligned}$$ Let $I^d$ denote the set of words $\bi = i_d\cdots i_1$ of length $d$ in the alphabet $I$, and identify $\bi \in I^d$ with the object $i_d \otimes \cdots \otimes i_1 \in \ob \QH$. Then, the locally unital algebra $$\label{recla}
QH_d := \bigoplus_{\bi,\bi' \in I^d}
\operatorname{Hom}_{\QH}(\bi,\bi')$$ is the [*quiver Hecke algebra*]{} of type $\mathfrak{sl}_\infty$ defined originally by Khovanov and Lauda [@KL1] and Rouquier [@Rou].
Recall for a $\operatorname{{\Bbbk}}$-linear category $\mathcal{C}$ that there is an associated strict $\operatorname{{\Bbbk}}$-linear monoidal category $\mathcal{E}nd(\mathcal{C})$ consisting of $\operatorname{{\Bbbk}}$-linear endofunctors and natural transformations. The remainder of the section will be devoted to the proof of the following theorem.
\[itlifts\] There is a strict monoidal functor $\Phi:\QH \rightarrow \mathcal{E}nd(\O)$ sending the generating objects $i \in I$ to the endofunctors $F_i$ from (\[them\]). Moreover, for all $M \in \ob \O$ and $i \in I$, the endomorphism $F_i M \rightarrow F_i M$ defined by the natural transformation $\Phi\Big(
\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkred] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.05) {$\color{darkred}\bullet$};
\node at (0.08,-.25) {$\scriptstyle{i}$};
\end{tikzpicture}
}
\Big)$ is nilpotent.
In order to construct $\Phi$, we need to pass through two intermediate objects $\AHC$, the (degenerate) affine Hecke-Clifford supercategory, and $\QHC$, which is a certain quiver Hecke-Clifford supercategory in the sense of [@KKT]. Both $\AHC$ and $\QHC$ are examples of (strict) [*monoidal supercategories*]{}, meaning that they are supercategories equipped with a monoidal product in an appropriate enriched sense. We refer the reader to the introduction of [@BE] for the precise definition, just recalling that morphisms in a monoidal supercategory satisfy the [*super interchange law*]{} rather than the usual interchange law of a monoidal category: in terms of the string calculus as in [@BE] we have that $$\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkpurple] (0.08,-.4) to (0.08,-.23);
\draw[-,thick,darkpurple] (0.08,.4) to (0.08,.03);
\draw[thick,darkpurple] (0.08,-0.1) circle (4pt);
\node at (0.08,-0.1) {\color{darkpurple}$\scriptstyle{g}$};
\draw[-,thick,darkpurple] (-.8,-.4) to (-.8,-.03);
\draw[-,thick,darkpurple] (-.8,.4) to (-.8,.23);
\draw[thick,darkpurple] (-.8,0.1) circle (4pt);
\node at (-.8,.1) {\color{darkpurple}$\scriptstyle{f}$};
\end{tikzpicture}
}
\quad=\quad
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkpurple] (0.08,-.4) to (0.08,-.13);
\draw[-,thick,darkpurple] (0.08,.4) to (0.08,.13);
\draw[thick,darkpurple] (0.08,0) circle (4pt);
\node at (0.08,0) {\color{darkpurple}$\scriptstyle{g}$};
\draw[-,thick,darkpurple] (-.8,-.4) to (-.8,-.13);
\draw[-,thick,darkpurple] (-.8,.4) to (-.8,.13);
\draw[thick,darkpurple] (-.8,0) circle (4pt);
\node at (-.8,0) {\color{darkpurple}$\scriptstyle{f}$};
\end{tikzpicture}
}
\quad=\quad
(-1)^{|f||g|}\:
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkpurple] (0.08,-.4) to (0.08,-.03);
\draw[-,thick,darkpurple] (0.08,.4) to (0.08,.23);
\draw[thick,darkpurple] (0.08,0.1) circle (4pt);
\node at (0.08,0.1) {\color{darkpurple}$\scriptstyle{g}$};
\draw[-,thick,darkpurple] (-.8,-.4) to (-.8,-.23);
\draw[-,thick,darkpurple] (-.8,.4) to (-.8,.03);
\draw[thick,darkpurple] (-.8,-0.1) circle (4pt);
\node at (-.8,-.1) {\color{darkpurple}$\scriptstyle{f}$};
\end{tikzpicture}
}$$ for homogeneous morphisms $f$ and $g$ of parities $|f|$ and $|g|$, respectively.
The (degenerate) [*affine Hecke-Clifford supercategory*]{} $\AHC$ is the strict monoidal supercategory with a single generating object $1$, even generating morphisms $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.08) {$\color{darkblue}\bullet$};
\end{tikzpicture}
}:1 \rightarrow 1$ and $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.18,-.15) to (-0.18,.3);
\draw[-,thick,darkblue] (-0.18,-.15) to (0.18,.3);
\end{tikzpicture}
}:1 \otimes 1 \rightarrow 1 \otimes 1$, and an odd generating morphism $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.08) {$\color{darkblue}\circ$};
\end{tikzpicture}
}:1 \rightarrow 1$. These are subject to the following relations: $$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-, thick,darkblue] (0, 0.6) to (0, -0.3);
\node at (0, -0.00) {$\color{darkblue} \circ$};
\node at (0, 0.3) {$\color{darkblue} \bullet$};
\end{tikzpicture}
} &=
-
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-, thick,darkblue] (0, 0.6) to (0, -0.3);
\node at (0, -0.0) {$\color{darkblue} \bullet$};
\node at (0, 0.3) {$\color{darkblue} \circ$};
\end{tikzpicture}
}\:,
&\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-, thick,darkblue] (0, 0.6) to (0, -0.3);
\node at (0, -0.0) {$\color{darkblue} \circ$};
\node at (0, 0.3) {$\color{darkblue} \circ$};
\end{tikzpicture}
}
&=
\mathord{
\begin{tikzpicture}[baseline = 2]
\draw[-, thick,darkblue] (0, 0.6) to (0, -0.3);
\end{tikzpicture}
}\:,
&
\mathord{
\begin{tikzpicture}[baseline = 8.5]
\draw[-,thick,darkblue] (0.2,.4) to[out=90,in=-90] (-0.2,.9);
\draw[-,thick,darkblue] (-0.2,.4) to[out=90,in=-90] (0.2,.9);
\draw[-,thick,darkblue] (0.2,-.1) to[out=90,in=-90] (-0.2,.4);
\draw[-,thick,darkblue] (-0.2,-.1) to[out=90,in=-90] (0.2,.4);
\end{tikzpicture}}
&=
\mathord{
\begin{tikzpicture}[baseline = 8.5]
\draw[-,thick,darkblue] (0.1,-.1) to (0.1,.9);
\draw[-,thick,darkblue] (-0.3,-.1) to (-0.3,.9);
\end{tikzpicture}
}\:,\end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkblue] (0.25,-.2) to (-0.25,.6);
\node at (-0.14,0.42) {$\color{darkblue}\circ$};
\end{tikzpicture}
}
&=
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkblue] (0.25,-.2) to (-0.25,.6);
\node at (0.14,-0.02) {$\color{darkblue}\circ$};
\end{tikzpicture}
}
&\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkblue] (0.25,-.2) to (-0.25,.6);
\node at (-0.14,0.42) {$\color{darkblue}\bullet$};
\end{tikzpicture}
}
-
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkblue] (0.25,-.2) to (-0.25,.6);
\node at (0.14,-0.02) {$\color{darkblue}\bullet$};
\end{tikzpicture}
}
&=
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.08,-.2) to (0.08,.6);
\draw[-,thick,darkblue] (-0.28,-.2) to (-0.28,.6);
\end{tikzpicture}
}
- \!
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkblue] (0.08,-.2) to (0.08,.6);
\draw[-,thick,darkblue] (-0.28,-.2) to (-0.28,.6);
\node at (0.08, 0.2) {$\color{darkblue}\circ$};
\node at (-0.28, 0.2){$\color{darkblue}\circ$};
\end{tikzpicture}
}\:,
&\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-,thick,darkblue] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkblue] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkblue] (0,-.4) to[out=90,in=-90] (-.45,0.2);
\draw[-,thick,darkblue] (-0.45,0.2) to[out=90,in=-90] (0,0.8);
\end{tikzpicture}
}
&=
\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-,thick,darkblue] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkblue] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkblue] (0,-.4) to[out=90,in=-90] (.45,0.2);
\draw[-,thick,darkblue] (0.45,0.2) to[out=90,in=-90] (0,0.8);
\end{tikzpicture}
}\,.\end{aligned}$$ Denoting the object $1^{\otimes d} \in \ob \AHC$ simply by $d$, the (degenerate) [*affine Hecke-Clifford superalgebra*]{} is the superalgebra $$\label{train1}
AHC_d := \operatorname{End}_{\AHC}(d).$$ This was introduced originally by Nazarov [@N $\S$3].
For a supercategory $\C$, we write $\mathcal{E}nd(\C)$ for the strict monoidal supercategory consisting of superfunctors and supernatural transformations.
\[hkst\] There is a strict monoidal superfunctor $\Psi:\AHC \rightarrow \mathcal{E}nd(\sO)$ sending the generating object $1$ to the endofunctor $\sF = U \otimes -$ from (\[sfdef\]), and the generating morphisms $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.08) {$\color{darkblue}\bullet$};
\end{tikzpicture}
}$, $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.08) {$\color{darkblue}\circ$};
\end{tikzpicture}
}$ and $\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkblue] (0.18,-.15) to (-0.18,.3);
\draw[-,thick,darkblue] (-0.18,-.15) to (0.18,.3);
\end{tikzpicture}
}$ to the supernatural transformations $x, c$ and $t$ which are defined on $M \in \ob \sO$ as follows:
- $x_M:U \otimes M \rightarrow U \otimes M$ is left multiplication by the tensor $\omega$ from (\[omegadef\]);
- $c_M: U \otimes M \rightarrow U \otimes M$ is left multiplication by $\sqrt{-1} \,f' \otimes 1$ for $f'$ as in (\[oddz\]);
- $t_M:U \otimes U \otimes M \rightarrow U \otimes U \otimes M$ sends $u \otimes v \otimes m \mapsto (-1)^{|u||v|} v \otimes u
\otimes m$.
This an elementary check of relations, similar to the one made in the proof of [@HKS Theorem 7.4.1].
\[qhcdef\] The *quiver Hecke-Clifford supercategory* of type $\mathfrak{sl}_\infty$ is the monoidal supercategory $\QHC$ with objects generated by the set $J$ from (\[jdef\]), even generating morphisms $\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkgreen] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.05) {$\color{darkgreen}\bullet$};
\node at (0.08,-.25) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}:j_1 \rightarrow j_1$ and $\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkgreen] (0.18,-.15) to (-0.18,.3);
\draw[-,thick,darkgreen] (-0.18,-.15) to (0.18,.3);
\node at (-0.18,-.25) {$\scriptstyle{j_2}$};
\node at (0.18,-.25) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}:j_2 \otimes j_1 \rightarrow j_1 \otimes j_2$, and odd generating morphisms $\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkgreen] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.05) {$\color{darkgreen}\circ$};
\node at (0.08,-.25) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}:j_1 \rightarrow -j_1$, for all $j_1,j_2 \in J$. These are subject to the following relations: $$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-, thick,darkgreen] (0, 0.6) to (0, -0.2);
\node at (0, 0.05) {$\color{darkgreen} \bullet$};
\node at (0, 0.35) {$\color{darkgreen} \circ$};
\node at (-0.01,-.35) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
&=
-
\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-, thick,darkgreen] (0, 0.6) to (0, -0.2);
\node at (0, 0.05) {$\color{darkgreen} \circ$};
\node at (0, 0.35) {$\color{darkgreen} \bullet$};
\node at (-0.01,-.35) {$\scriptstyle{j_1}$};
\end{tikzpicture}
},
\qquad
\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-, thick,darkgreen] (0, 0.6) to (0, -0.2);
\node at (0, 0.05) {$\color{darkgreen} \circ$};
\node at (0, 0.35) {$\color{darkgreen} \circ$};
\node at (-0.01,-.35) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
=
\mathord{
\begin{tikzpicture}[baseline = 3]
\draw[-, thick,darkgreen] (0, 0.6) to (0, -0.2);
\node at (-0.01,-.35) {$\scriptstyle{j_1}$};
\end{tikzpicture}
},
\qquad
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.13,-0.02) {$\color{darkgreen}\circ$};
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
=
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (0.14,0.42) {$\color{darkgreen}\circ$};
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\end{tikzpicture}
},
\qquad
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.13,0.42) {$\color{darkgreen}\circ$};
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\end{tikzpicture}
} =
\mathord{
\begin{tikzpicture}[baseline = 4]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (0.14,-0.02) {$\color{darkgreen}\circ$};
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\end{tikzpicture}
},\end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\node at (-0.13,-0.02) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
-\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\node at (0.14,0.42) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
&=
\left\{
\begin{array}{cl}
\mathord{
\begin{tikzpicture}[baseline = -5]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
& \text{if $j_1 = j_2$,} \\
\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28, 0.06){$\color{darkgreen}\circ$};
\node at (0.08, 0.06){$\color{darkgreen}\circ$};
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
& \text{if $j_1= - j_2$,} \\
0 & \text{otherwise;}
\end{array}
\right.\end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\node at (-0.13,0.42) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
-
\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.25,.6) to (-0.25,-.2);
\draw[-,thick,darkgreen] (0.25,-.2) to (-0.25,.6);
\node at (-0.25,-.28) {$\scriptstyle{j_2}$};
\node at (0.25,-.28) {$\scriptstyle{j_1}$};
\node at (0.15,-0.02) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
&=
\left\{
\begin{array}{cl}
\mathord{
\begin{tikzpicture}[baseline = -5]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
& \text{if $j_1 = j_2$,} \\
\!\!\!\!-\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28, 0.06){$\color{darkgreen}\circ$};
\node at (0.08, 0.06){$\color{darkgreen}\circ$};
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
& \text{if $j_1= - j_2$,} \\
0 & \text{otherwise;}
\end{array}
\right.\end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 7]
\draw[-,thick,darkgreen] (0.28,.4) to[out=90,in=-90] (-0.28,1.1);
\draw[-,thick,darkgreen] (-0.28,.4) to[out=90,in=-90] (0.28,1.1);
\draw[-,thick,darkgreen] (0.28,-.3) to[out=90,in=-90] (-0.28,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to[out=90,in=-90] (0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.28,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
&=
\left\{
\begin{array}{ll}
\:\:\:\:0&\text{if $i_1 = i_2$,}\\
\mathord{
\kappa_1 (i_1-i_2)
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\node at (0.08,0.05) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
+
\kappa_2(i_2-i_1)
\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\node at (-0.28,0.05) {$\color{darkgreen}\bullet$};
\end{tikzpicture}
}
&\text{if $|i_1 - i_2| = 1$,}\\
\mathord{
\begin{tikzpicture}[baseline = 0]
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_2}$};
\node at (0.08,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}&\text{if $|i_1 - i_2| > 1$;}\\
\end{array}
\right. \end{aligned}$$$$\begin{aligned}
\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkgreen] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkgreen] (0,-.4) to[out=90,in=-90] (-.45,0.2);
\draw[-,thick,darkgreen] (-0.45,0.2) to[out=90,in=-90] (0,0.8);
\node at (-0.45,-.5) {$\scriptstyle{j_3}$};
\node at (0,-.5) {$\scriptstyle{j_2}$};
\node at (0.45,-.5) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
-
\mathord{
\begin{tikzpicture}[baseline = 1]
\draw[-,thick,darkgreen] (0.45,.8) to (-0.45,-.4);
\draw[-,thick,darkgreen] (0.45,-.4) to (-0.45,.8);
\draw[-,thick,darkgreen] (0,-.4) to[out=90,in=-90] (.45,0.2);
\draw[-,thick,darkgreen] (0.45,0.2) to[out=90,in=-90] (0,0.8);
\node at (-0.45,-.5) {$\scriptstyle{j_3}$};
\node at (0,-.5) {$\scriptstyle{j_2}$};
\node at (0.45,-.5) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
\!&=
\left\{
\begin{array}{ll}
\kappa_1(i_1 - i_2)
\mathord{
\begin{tikzpicture}[baseline = -1]
\draw[-,thick,darkgreen] (0.44,-.3) to (0.44,.4);
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28,-.4) {$\scriptstyle{j_3}$};
\node at (0.08,-.4) {$\scriptstyle{j_2}$};
\node at (0.44,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
&
\text{if $j_1 = j_3$ and $|i_1 - i_2|= 1$,}\\
\kappa_1 (i_2 - i_1)
\mathord{
\begin{tikzpicture}[baseline =-1]
\draw[-,thick,darkgreen] (0.44,-.3) to (0.44,.4);
\draw[-,thick,darkgreen] (0.08,-.3) to (0.08,.4);
\draw[-,thick,darkgreen] (-0.28,-.3) to (-0.28,.4);
\node at (-0.28, 0.06) {$\color{darkgreen} \circ$};
\node at (0.44, 0.06) {$\color{darkgreen} \circ$};
\node at (-0.28,-.4) {$\scriptstyle{j_3}$};
\node at (0.08,-.4) {$\scriptstyle{j_2}$};
\node at (0.44,-.4) {$\scriptstyle{j_1}$};
\end{tikzpicture}
}
&\text{if $j_1=-j_3$ and $|i_1 - i_2| = 1$,}
\\
\:\:\:\:0 &\text{otherwise.}
\end{array}
\right.\end{aligned}$$ In the above, we have adopted the convention given $j_r \in J$ that $i_r \in I$ and $\kappa_r \in \{\pm 1\}$ are defined from $j_r = \kappa_r \sqrt{z+i_r}\sqrt{z + i_r +1}$. Identifying the word $\bj = j_d\cdots j_1 \in J^d$ with $j_d \otimes\cdots\otimes j_1 \in \ob \QHC$, the [*quiver Hecke-Clifford superalgebra*]{} is the locally unital algebra $$\label{train2}
QHC_d :=
\bigoplus_{\bj,\bj' \in J^d}
\operatorname{Hom}_{\QHC}(\bj,\bj').$$ This is exactly as in [@KKT Definition 3.5] in the special case of the $\mathfrak{sl}_\infty$-quiver.
Now we are going to exploit a remarkable isomorphism between certain completions $\widehat{AHC}_d$ and $\widehat{QHC}_d$ of the superalgebras $AHC_d$ and $QHC_d$ from (\[train1\]) and (\[train2\]), which was constructed in [@KKT]. To define these, we need some further notation.
Numbering strands of a diagram by $1,\dots,d$ from right to left, $AHC_d$ is generated by its elements $x_r, c_r\:(1 \leq r \leq d)$ and $t_r\:(1 \leq r < d)$ corresponding to the closed dot on the $r$th strand, the open dot on the $r$th strand, and the crossing of the $r$th and $(r+1)$th strands, respectively. Let $HC_d := S_d \ltimes C_d$ be the [*Sergeev superalgebra*]{}, that is, the smash product of the symmetric group $S_d$ with basic transpositions $t_1,\dots,t_{d-1}$ acting on the Clifford superalgebra $C_d$ on generators $c_1,\dots,c_d$. Let $A_d$ denote the purely even polynomial superalgebra $\operatorname{{\Bbbk}}[x_1,\dots,x_d]$. By the basis theorem for $AHC_d$ established in [@BK $\S$2-k], the natural multiplication map gives a superspace isomorphism $HC_d \otimes A_d \stackrel{\sim}{\rightarrow}
AHC_d$. Transporting the multiplication on $AHC_d$ to $HC_d\otimes A_d$ via this isomorphism, the following describe how to commute a polynomial $f \in A_d$ past the generators of $HC_d$: $$\begin{aligned}
(1 \otimes f) (c_r \otimes 1) &= c_r \otimes c_r(f),\\
(1 \otimes f) (t_r \otimes 1) &= t_r \otimes t_r(f) + 1 \otimes \partial_r(f)
+c_r c_{r+1} \otimes \tilde\partial_r(f),\end{aligned}$$ for operators $c_r, t_r, \partial_r,\tilde\partial_r:A_d \rightarrow
A_d$ such that
- $t_r$ is the automorphism that interchanges $x_r$ and $x_{r+1}$ and fixes all other generators;
- $c_r$ is the automorphism that sends $x_r \mapsto -x_r$ and fixes all other generators;
- $\partial_r$ is the Demazure operator $\partial_r(f) :=
\frac{t_r(f)-f}{x_r - x_{r+1}};
$
- $\tilde{\partial}_r$ is the [twisted Demazure operator]{} $c_{r+1} \circ \partial_r \circ c_{r}$, so $\tilde\partial_r(f) =
\frac{t_r(f) - c_{r+1}(c_r(f))}{x_r+x_{r+1}}.$
Given a tuple $\mu = (\mu_i)_{i \in I}$ of non-negative integers all but finitely many of which are zero, the quotient superalgebra $$AHC_d(\mu) := AHC_d \Big/ \Big\langle
\prod_{i \in I} \left(x_1^2 - (z+i)(z+i+1)\right)^{\mu_i}
\Big\rangle$$ is a (degenerate) [*cyclotomic Hecke-Clifford superalgebra*]{} in the sense of [@BK $\S$3.e]. It is finite dimensional. Moreover, all roots of the minimal polynomials of all $x_r \in AHC_d(\mu)$ belong to the set $J$. It follows for each $\bj = j_d \cdots j_1$ in the set $J^d$ of words of length $d$ in letters $J$ that there is an idempotent $1_\bj \in AHC_d(\mu)$ defined by the projection onto the simultaneous generalized eigenspaces for $x_1,\dots,x_d$ with eigenvalues $j_1,\dots,j_d$, respectively. Moreover, we have that $$AHC_d(\mu) = \bigoplus_{\bj,\bj' \in J^d} 1_{\bj'} AHC_d(\mu) 1_\bj.$$ If $\mu \leq \mu'$, i.e. $\mu_i \leq \mu_i'$ for all $i$, there is a canonical surjection $AHC_d(\mu') \twoheadrightarrow AHC_d(\mu)$ sending $x_r,c_r,t_r,1_\bj \in AHC_d(\mu')$ to the elements of $AHC_d(\mu)$ with the same names. Let $$\label{phew}
\widehat{AHC}_d := \varprojlim_{\mu} AHC_d(\mu)$$ be the inverse limit of this system of superalgebras taken in the category of locally unital superalgebras with distinguished idempotents indexed by $J^d$. Using the basis theorem for the cyclotomic quotients $AHC_d(\mu)$ from [@BK $\S$3-e], one can identify $\widehat{AHC}_d$ with the completion defined in [@KKT Definition 5.3][^2]. In particular, letting $$\widehat{A}_d := \bigoplus_{\bj \in J^d}
\operatorname{{\Bbbk}}[[x_1-j_1,\dots,x_d-j_d]]1_\bj,$$ there is a superspace isomorphism $HC_d \otimes \widehat{A}_d \stackrel{\sim}{\rightarrow}
\widehat{AHC}_d$ induced by the obvious multiplication maps $HC_d \otimes \widehat{A}_d \twoheadrightarrow AHC_d(\mu)$ for all $\mu$. The multiplication on $HC_d\otimes \widehat{A}_d$ corresponding to the one on $\widehat{AHC}_d$ via this isomorphism has the following properties for all $f \in \widehat{A}_d$: $$\begin{aligned}
(1 \otimes f 1_{\bj}) (c_r \otimes 1_{\bj'})
&= c_r \otimes c_r(f) 1_{c_r(\bj)} 1_{\bj'},\\\notag
(1 \otimes f 1_{\bj}) (t_r \otimes 1_{\bj'}) &= t_r \otimes t_r(f) 1_{t_r(\bj)}1_\bj'
+ 1 \otimes \frac{t_r(f) 1_{t_r(\bj)}
- f 1_\bj}{x_r-x_{r+1}}1_{\bj'} \\
&+c_r c_{r+1} \otimes \frac{
t_r(f) 1_{t_r(\bj)}- c_{r+1}(c_{r}(f)) 1_{c_{r+1}(c_r(\bj))}}{x_r + x_{r+1}} 1_{\bj'}.
\label{fr}\end{aligned}$$ The fractions on the right hand side of (\[fr\]) make sense: in the first, $(x_r - x_{r+1}) 1_{\bj'}$ is invertible unless $j'_r = j'_{r+1}$, in which case the expression equals $\partial_r(f) 1_\bj 1_{\bj'}$; the second is fine when $j'_r \neq -j'_{r+1}$ as then $(x_r+x_{r+1}) 1_{\bj'}$ is invertible, while if $j'_r = -j'_{r+1}$ it equals $\tilde\partial_r(f) 1_{t_r(\bj)}1_{\bj'}$.
Similarly, there is a completion $\widehat{QHC}_d$ of $QHC_d$. To introduce this, we denote the elements of $QHC_d 1_\bj$ defined by an open dot on the $r$th strand, a closed dot on the $r$th strand and a crossing of the $r$th and $(r+1)$th strands by $\gamma_r 1_\bj, \xi_r 1_\bj$ and $\tau_r 1_\bj$, respectively. For $\mu = (\mu_i)_{i \in I}$ as above, we define the [*cyclotomic quiver Hecke-Clifford superalgebra*]{} $$QHC_d(\mu) := QHC_d \Big/ \left\langle
\xi_1^{2\mu_{i}}1_\bj \:\Big|\:\bj \in J^d
, i \in I
\text{ with }j_1^2 = (z+i)(z+i+1)
\right\rangle.$$ Using the relations, it is easy to see that the images of all $\xi_r 1_\bj$ are nilpotent in $QHC_d(\mu)$. Then we set $$\widehat{QHC}_d := \varprojlim_\mu QHC_d(\mu),$$ taking the inverse limit once again in the category of locally unital superalgebras with distinguished idempotents indexed by $J^d$. The obvious locally unital homomorphisms $QHC_d
\otimes_{\operatorname{{\Bbbk}}[\xi_1,\dots,\xi_d]} \operatorname{{\Bbbk}}[[\xi_1,\dots,\xi_d]]
\twoheadrightarrow QHC_d(\mu)$ for each $\mu$ induce a surjective homomorphism $$QHC_d
\otimes_{\operatorname{{\Bbbk}}[\xi_1,\dots,\xi_d]} \operatorname{{\Bbbk}}[[\xi_1,\dots,\xi_d]]
\rightarrow
\widehat{QHC}_d.$$ This map is actually an isomorphism, as may be deduced using the basis theorem for $QHC_d$ from [@KKT Corollary 3.9] plus the observation that the image of any non-zero element $u \in QHC_d$ is non-zero in $QHC_d(\mu)$ for sufficiently large $\mu$; the latter assertion follows by elementary considerations involving the natural $\operatorname{\mathbb{Z}}$-grading on $QHC_d$. Consequently, $\widehat{QHC}_d$ is isomorphic to the completion introduced in a slightly different way in [@KKT Definition 3.16]. Moreover, there is a locally unital embedding $QHC_d \hookrightarrow \widehat{QHC}_d$.
At last, we are ready to state the crucial theorem from [@KKT]. We need this only in the special situation of [@KKT $\S$5.2(i)(a)], but emphasize that the results obtained in [@KKT] are substantially more general. In particular, for us, all elements of the set $I$ are even in the sense of [@KKT $\S$3.5], so that we do not need the more general quiver Hecke [*super*]{}algebras of [@KKT].
\[kkttheorem\] There is a superalgebra isomorphism $\widehat{QHC}_d \stackrel{\sim}{\rightarrow} \widehat{AHC}_d$ such that $$1_\bj \mapsto 1_\bj,\quad
\gamma_r 1_\bj \mapsto c_r 1_\bj,\quad
\xi_r 1_\bj \mapsto y_r 1_\bj,\quad
\tau_r 1_\bj \mapsto t_r g_r 1_\bj
+ f_r 1_\bj
+
c_r c_{r+1} \tilde f_r 1_\bj,$$ for all $\bj \in J^d$ and $r$. Here, $y_r \in \operatorname{{\Bbbk}}[[x_r-j_r]]$ and $g_r, f_r, \tilde f_r \in \operatorname{{\Bbbk}}[[x_r-j_r,x_{r+1}-j_{r+1}]]$ are the power series determined uniquely by the following: $$\begin{aligned}
j_r &= \kappa_r \sqrt{z+i_r}\sqrt{z+i_r +1}\text{ for $i_r \in I$ and $\kappa_r \in \{\pm\}$},\\
y_r &=\kappa_r \left(\sqrt{x_r^2+{\textstyle\frac{1}{4}}} - \left(z+i_r+\half\right)\right) \in (x_r-j_r),\\
p_r &= \frac{(x_r^2-x_{r+1}^2)^2}{2 (x_r^2+x_{r+1}^2)-(x_r^2-x_{r+1}^2)^2},\\
g_r &=
\left\{
\begin{array}{ll}
-1&\text{if $i_r < i_{r+1}$,}\\
p_r
\left(\kappa_r y_r - \kappa_{r+1} y_{r+1}\right)
&\text{if $i_r = i_{r+1}+1$,}\\
p_r&\text{if $i_r > i_{r+1}+1$,}\\
\frac{\sqrt{p_r}}{y_r-y_{r+1}}
\in \frac{x_r-x_{r+1}}{y_r-y_{r+1}} + (x_r-x_{r+1})
&\text{if $j_r = j_{r+1}$,}\\
\frac{\sqrt{p_r}}{y_r+y_{r+1}}
\in \frac{x_r+x_{r+1}}{y_r+y_{r+1}} + (x_r+x_{r+1})
&\text{if $j_r = -j_{r+1}$;}\\
\end{array}\right.\\
f_r &= \frac{g_r}{x_r-x_{r+1}} -
\frac{\delta_{j_r,j_{r+1}}}{y_r-y_{r+1}},\qquad
\tilde f_r = \frac{g_r}{x_r+x_{r+1}} -
\frac{\delta_{j_r,-j_{r+1}}}{y_r+y_{r+1}}.\end{aligned}$$ (All of this notation depends implicitly on $\bj$.)
This is a special case of [@KKT Theorem 5.4]. To help the reader to translate between our notation and that of [@KKT], we note that the set $J$ in [@KKT] is the same as our set $J$, but the set $I$ there is $\tilde I := \{j^2\:|\:j \in J\}$, which is different from our $I$. We have made various other choices as stipulated in [@KKT] in order to produce concrete formulae: we have taken the functions $\eps:J \rightarrow \{0,1\}$ and $h:\tilde I
\rightarrow \operatorname{{\Bbbk}}$ from [@KKT (5.7)] so that $\eps(j) = (1-\kappa)/2$ and $h(j^2) = z+i+\half$ for $J\ni j = \kappa \sqrt{z+i}\sqrt{z+i+1}$; for [@KKT (5.11)] we took $G_{j_r, j_{r+1}}$ (our $g_r$) to be $-1$ when $i_r < i_{r+1}$. The fact that $g_r, f_r$ and $\tilde f_r$ are all well-defined elements of $\operatorname{{\Bbbk}}[[x_r-j_r,x_{r+1}-j_{r+1}]]$ is justified by [@KKT Lemma 5.5]. Note also that the ambiguous square roots appearing in the formulae for $y_r$ and $g_r$ are uniquely determined by the containments we have specified.
Notes: Our $J$ is the same as the one in [@KKT], but $I$ in [@KKT] is $\{j^2\:|\:j \in J\}$. Also in [@KKT], a key role is played by the algebra $A = \operatorname{{\Bbbk}}[x,\lambda] / (x^2-\lambda^2+\frac{1}{4})$. I’m writing instead $y = \lambda - h$ and my $x$ is their $x-j$. So $A = \operatorname{{\Bbbk}}[x,y] / ((x+j)^2 - (y+h)^2 + \frac{1}{4})$. Then they form the completion at $I = (x,y)$ (in my notation) and observe this is naturally $\operatorname{{\Bbbk}}[[x]]$ and $\operatorname{{\Bbbk}}[[y]]$. The formula for $y$ in terms of $x$ is the above.
For $\bi = i_d \cdots i_1 \in I^d$, let $F_\bi := F_{i_d} \cdots
F_{i_1}:\O\rightarrow\O$. The usual vertical composition of natural transformations makes the vector space $$NT_d := \bigoplus_{\bi,\bi' \in I^d} \operatorname{Hom}(F_\bi, F_{\bi'})$$ into a locally unital algebra with distinguished idempotents $\{1_\bi\:|\:\bi \in I^d\}$ arising from the identity endomorphisms of each $F_\bi$. Also horizontal composition of natural transformations defines homomorphisms $a_{d_2,d_1}:NT_{d_2} \otimes NT_{d_1} \rightarrow NT_{d_2+d_1}$ for all $d_1,d_2 \geq 0$. Recalling (\[recla\]), the data of a strict monoidal functor $\Phi:\QH \rightarrow
\mathcal{E}nd(\O)$ sending $i$ to $F_i$ is just the same as a family of locally unital algebra homomorphisms $\Phi_d:
QH_d \rightarrow NT_d$ for all $d \geq 0$, such that $1_\bi \mapsto 1_\bi$ for each $\bi \in I^d$ and $$\label{analog}
a_{d_2,d_1} \circ \Phi_{d_2} \otimes \Phi_{d_1} = \Phi_{d_2+d_1} \circ
b_{d_2,d_1}$$ for all $d_1,d_2 \geq 0$, where $b_{d_2,d_1}:QH_{d_2} \otimes QH_{d_1} \rightarrow QH_{d_2+d_1}$ is the obvious embedding defined by horizontal concatenation of diagrams.
To construct $\Phi_d$, we start from the monoidal superfunctor $\Psi$ from Theorem \[hkst\]. This induces superalgebra homomorphisms $\Psi_d:AHC_d \rightarrow \operatorname{End}(\sF^d)$ for all $d \geq 0$, where $\operatorname{End}(\sF^d)$ denotes supernatural endomorphisms of $\sF^d:\sO \rightarrow \sO$. For each $M \in \ob \sO$, Corollary \[minpoly\] implies that $\operatorname{ev}_M \circ \Psi_d : AHC_d \rightarrow \operatorname{End}_{\sO}(\sF^d \,M)$ factors through all sufficiently large cyclotomic quotients $AHC_d(\mu)$. Hence, $\Psi_d$ extends uniquely to a locally unital superalgebra homomorphism $\widehat{\Psi}_d:\widehat{AHC}_d \rightarrow SNT_d$, where $$SNT_d := \bigoplus_{\bj, \bj' \in J^d} \operatorname{Hom}(\sF_\bj, \sF_{\bj'}) \subset \operatorname{End}(\sF^d)$$ and $\sF_{\bj} := \sF_{j_d}\cdots \sF_{j_1}$. Composing $\widehat{\Psi}_d$ with the isomorphism from Theorem \[kkttheorem\] and the inclusion $QHC_d \hookrightarrow \widehat{QHC}_d$, we obtain a locally unital superalgebra homomorphism $\Theta_d:QHC_d \rightarrow SNT_d$. It is obvious from Definitions \[qhdef\] and \[qhcdef\] that there is a locally unital algebra homomorphism $\inc:QH_d \rightarrow (QHC_d)_\0$ sending the idempotent $1_\bi$ to $1_\bj$ for $\bj$ with $j_r := \sqrt{z+i_r}\sqrt{z+i_r+1}$, and taking the elements of $QH_d 1_\bi$ defined by the dot on the $r$th strand and the crossing of the $r$th and $(r+1)$th strands to $\xi_r 1_\bj$ and $\tau_r 1_\bj$, respectively. Also, recalling (\[them\]), restriction from $\sO$ to $\O$ defines a homomorphism $$\pr:\displaystyle\bigoplus_{\bj, \bj' \in J_+^d} 1_{\bj'} (SNT_d)_\0 1_\bj
\rightarrow NT_d$$ where $J_+ := \left\{\sqrt{z+i}\sqrt{z+i+1}\:\big|\:i \in I\right\} \subset J$. Then the composition $\pr \circ \Theta_d \circ \inc$ gives us the desired locally unital homomorphism $\Phi_d:QH_d \rightarrow NT_d
$ sending $1_\bi \mapsto 1_\bi$ for each $\bi \in I^d$. It just remains to observe that the property (\[analog\]) is satisfied, and that $\Phi_d(x_r 1_\bi)_M$ is nilpotent for each $r$, $\bi \in I^d$ and $M \in \ob\O$. These things follow from the explicit formulae in Theorems \[hkst\] and \[kkttheorem\] plus Corollary \[minpoly\] once again.
Proof of the Main Theorem
=========================
Everything is now in place for us to be able to prove the Main Theorem from the introduction. Recall $\bsigma = (\sigma_1,\dots,\sigma_n)$ is a sign sequence, and $V^{\otimes \bsigma}$ denotes the $\mathfrak{sl}_\infty$-module $V^{\sigma_1}\otimes\cdots\otimes V^{\sigma_n}$. The following is a special case of [@BLW Definition 2.10], which reformulated [@LW Definiton 3.2] for tensor products of minuscule representations; it may be helpful to recall Definitions \[bruhatdef\], \[hwdef\] and \[qhdef\] at this point.
\[tpcdef\] An $\mathfrak{sl}_\infty$-[*tensor product categorification*]{} of $V^{\otimes \bsigma}$ is the following data:
- a highest weight category $\mathcal C$ with weight poset $(\operatorname{\mathbf{B}}, \preceq)$;
- adjoint pairs $(F_i, E_i)$ of endofunctors of $\mathcal C$ for each $i
\in I$;
- a strict monoidal functor $\Phi:\mathcal{QH} \rightarrow
\mathcal{E}nd(\C)$ with $\Phi(i) = F_i$ for each $i \in I$.
We impose the following additional axioms for all $i
\in I$, $\bb \in \operatorname{\mathbf{B}}$ and $M \in \ob \C$:
- $E_i$ is isomorphic to a left adjoint of $F_i$;
- $F_i \Delta(\bb)$ has a $\Delta$-flag with sections $\{\Delta(\bb+\sigma_t \bd_t)\:|\: 1 \leq t \leq n,
\isig_t(\bb) = \mathtt{f}\}$;
- $E_i \Delta(\bb)$ has a $\Delta$-flag with sections $\{\Delta(\bb-\sigma_t \bd_t)\:|\:1 \leq t \leq n, \isig_t(\bb) =
\mathtt{e}\}$;
- the endomorphism $\Phi\Big(
\mathord{
\begin{tikzpicture}[baseline = -2]
\draw[-,thick,darkred] (0.08,-.15) to (0.08,.3);
\node at (0.08,0.05) {$\color{darkred}\bullet$};
\node at (0.08,-.25) {$\scriptstyle{i}$};
\end{tikzpicture}
}
\Big)_M:
F_i M\rightarrow F_i M$ is nilpotent.
Theorems \[itsplits\], \[itshw\], \[itacts\] and \[itlifts\] together imply:
\[typea\] The supercategory $\sO$ defined in section \[s2\] splits as $\O \oplus \Pi \O$, with $\O$ admitting all of the additional structure needed to make it into an $\mathfrak{sl}_\infty$-tensor product categorification of $V^{\otimes \bsigma}$.
As we already mentioned in the introduction, to complete the proof of our Main Theorem, we just need to appeal to the following results from [@BLW]:
The supercategory $\sO'$ from the introduction decomposes as $\O'
\oplus \Pi \O'$, with $\O'$ admitting the structure of an $\mathfrak{sl}_\infty$-tensor product categorification of $V^{\otimes \bsigma}$.
This is a special case of [@BLW Theorem 3.10].
Any two $\mathfrak{sl}_\infty$-tensor product categorifications of $V^{\otimes \bsigma}$ are strongly equivariantly equivalent (in the sense of [@LW Definition 3.1]).
This is a special case of [@BLW Theorem 2.12].
We get at once that the categories $\O$ and $\O'$ are equivalent, hence so too are $\sO$ and $\sO'$. This already proves the Main Theorem from the introduction in the case that $n$ is even. When $n$ is odd, one also needs to apply the Lemma from the introduction to see that the supercategory $\sO$ in the statement of the Main Theorem is the Clifford twist of the supercategory $\sO$ being studied here.
Canonical basis {#cbs}
===============
Combining our Main Theorem with the results of [@CLW; @BLW], it follows that the composition multiplicities $[M(\ba):L(\bb)]$ can be obtained by evaluating certain parabolic Kazhdan-Lusztig polynomials at $q=1$. In this section, we explain a simple algorithm to compute these polynomials explicitly. The algorithm is similar in spirit to the algorithm explained in [@B0 $\S$2-j], but actually the variant here is both easier to implement and a little faster. It also has the advantage of working for arbitrary sign sequences $\bsigma$, whereas the approach in [@B0] only makes sense for [*normally-ordered*]{} $\bsigma$’s, i.e. ones in which all $+$’s preceed all $-$’s. (But note that one can easily transition between different sign sequences as explained in [@CL $\S$5].)
We first need to introduce the quantum analog of the $\mathfrak{sl}_\infty$-module $V^{\otimes\bsigma}$. We will use similar notation to before, but decorated with dots to indicate $q$-analogs. Consider the generic quantized enveloping algebra $U_q \mathfrak{sl}_\infty$ over the field $\operatorname{\mathbb{Q}}(q)$. This has standard generators $\{\dot e_i, \dot f_i, \dot k_i, \dot k_i^{-1}\:|\:i
\in I\}$. We work with the comultiplication $\Delta$ defined from $$\label{comult}
\Delta(\dot{f}_i) =
1 \otimes \dot{f}_i + \dot{f}_i \otimes \dot{k}_i,
\quad
\Delta(\dot{e}_i) =
\dot{k}_i^{-1} \otimes \dot{e}_i + \dot{e}_i \otimes 1,
\quad
\Delta(\dot{k}_i) = \dot{k}_i \otimes \dot{k}_i.$$ We have the natural $U_q \mathfrak{sl}_\infty$-module $\dot V^+$ on basis $\{\dot v^+_i\:|\:i \in I\}$, and its dual $\dot V^-$ on basis $\{\dot v^-_i\:|\:i \in I\}$. The Chevalley generators $\dot f_i$ and $\dot e_i$ act on these basis vectors by exactly the same formulae (\[e\])–(\[f\]) as at $q=1$, and also $$\dot k_i \dot v^+_j = q^{\delta_{i,j}-\delta_{i+1,j}} \dot v^+_j,
\qquad
\dot k_i \dot v^-_j = q^{\delta_{i+1,j}-\delta_{i,j}}\dot v^-_j.$$ Then we form the tensor space $\dot V^{\otimes \bsigma} := \dot V^{\sigma_1}
\otimes\cdots\otimes \dot V^{\sigma_n}$, which is a $U_q \mathfrak{sl}_\infty$-module with its monomial basis $\left\{\dot v_\bb := \dot v_{b_1}^{\sigma_1} \otimes\cdots\otimes
\dot v_{b_n}^{\sigma_n}\:\big|\:\bb \in \operatorname{\mathbf{B}}\right\}$.
Next we pass to a certain completion $\widehat{V}^{\otimes\bsigma}$. Recall the inverse dominance order $\succeq$ on $P^n$ from Definition \[bruhatdef\]. We define $\widehat{V}^{\otimes\bsigma}$ to be the $\operatorname{\mathbb{Q}}(q)$-vector space consisting of formal linear combinations of the form $\sum_{\bb \in \operatorname{\mathbf{B}}} p_\bb(q) \dot v_\bb$ for rational functions $p_\bb(q) \in \operatorname{\mathbb{Q}}(q)$, such that the [*support*]{} $\{\bb \in \operatorname{\mathbf{B}}\:|\:p_\bb(q) \neq 0\}$ is contained in a finite union of sets of the form $
\{\bb \in \operatorname{\mathbf{B}}\:|\:\operatorname{\mathbf{wt}}(\bb) \succeq \bbeta\}
$ for $\bbeta \in P^n$.
The action of $U_q
\mathfrak{sl}_\infty$ on $\dot V^{\otimes\bsigma}$ extends to a well-defined action on $\widehat{V}^{\otimes\bsigma}$ such that $u \big(\sum_{\bb \in \operatorname{\mathbf{B}}} p_\bb(q) \dot v_\bb\big) =
\sum_{\bb \in \operatorname{\mathbf{B}}} p_\bb(q) \,u \dot v_\bb$ for every $u \in U_q \mathfrak{sl}_\infty$. Moreover, $\widehat{V}^{\otimes \bsigma}$ splits as the direct sum of its weight spaces.
For the first assertion, we need to show that the expression $\sum_{\bb \in \operatorname{\mathbf{B}}} p_\bb(q) \,u \dot v_\bb$ satisfies the condition on its support required to belong to $\widehat{V}^{\otimes\bsigma}$. It suffices to check this for $u \in \{\dot f_i, \dot e_i\:|\:i \in I\}$. If $\operatorname{\mathbf{wt}}(\bb) \succeq (\beta_1,\dots,\beta_n)$ and $v_\ba$ appears with non-zero coefficient in the expansion of $\dot e_i v_\bb$ (resp. $\dot f_i v_\bb$), then $\operatorname{\mathbf{wt}}(\ba)$ is equal to $\operatorname{\mathbf{wt}}(\bb)$ with $\alpha_i$ added (resp. subtracted) from one of its entries. Hence, $\operatorname{\mathbf{wt}}(\ba) \succeq (\beta_1+\alpha_i,\beta_2,\dots,\beta_n)$ (resp. $\operatorname{\mathbf{wt}}(\ba) \succeq
(\beta_1,\dots,\beta_{n-1},\beta_n-\alpha_i)$). This is all that is needed.
The fact that $\widehat{V}^{\otimes\bsigma}$ is a direct sum of its weight spaces follows because all $v_\bb$ with $\operatorname{\mathbf{wt}}(\bb) \succeq (\beta_1,\dots,\beta_n)$ are of the same weight $\beta_1+\cdots+\beta_n$.
The completion $\widehat{V}^{\otimes \bsigma}$ admits a [*bar involution*]{} $\psi: \widehat{V}^{\otimes\bsigma}
\rightarrow
\widehat{V}^{\otimes\bsigma}$ which is anti-linear with respect to the field automorphism $\operatorname{\mathbb{Q}}(q) \rightarrow \operatorname{\mathbb{Q}}(q), q \mapsto q^{-1}$. To define $\psi$, let $\Theta$ be Lusztig’s quasi-$R$-matrix from [@Lubook Theorem 4.1.2]; note for this due to our different choice of $\Delta$ compared to [@Lubook] that Lusztig’s $v$ is our $q^{-1}$ (and his $E_i, F_i, K_i$ are our $\dot e_i, \dot f_i, \dot
k_i^{-1}$). We proceed by induction on $n$, setting $\psi(\dot v_i^+) = \dot
v_i^+$ and $\psi(\dot v_i^-) = \dot v_i^-$ in case $n=1$. For $n > 1$, let $\bar\bsigma$ and $\bar\bb$ denote the $(n-1)$-tuples $(\sigma_1,\dots,\sigma_{n-1})$ and $(b_1,\dots,b_{n-1})$, respectively. Assuming that the analog $\bar\psi$ of $\psi$ on the space $\widehat{V}^{\otimes\bar\bsigma}$ has already been defined by induction, we define $\psi$ on $\widehat{V}^{\otimes\bsigma}$ by setting $$\label{bardef}
\psi\left({\sum_{\bb \in \operatorname{\mathbf{B}}}} p_\bb(q) \dot v_\bb\right) :=
{\sum_{\bb \in \operatorname{\mathbf{B}}}} p_\bb(q^{-1})
\,\Theta\!\left(\bar\psi(\dot v_{\bar\bb}\right) \otimes \dot
v_{b_n}^{\sigma_n}).$$
The antilinear map $\psi$ defined by (\[bardef\]) is a well-defined involution of $\widehat{V}^{\otimes\bsigma}$ preserving all weight spaces and commuting with the actions of $\dot f_i, \dot e_i$ for all $i \in I$. Moreover, $\psi(\dot v_\bb)$ is equal to $\dot v_\bb$ plus a (possibly infinite) $\operatorname{\mathbb{Z}}[q,q^{-1}]$-linear combination of $\dot v_\ba$’s for $\ba \succ
\bb$.
Recall that $\Theta$ is a formal sum of terms $\Theta_\beta$ for $\beta \in \bigoplus_{i \in I} \operatorname{\mathbb{N}}\alpha_i$, with $\Theta_0 = 1$ and $\Theta_\beta \in (U^-_q \mathfrak{sl}_\infty)_{-\beta} \otimes (U_q^+
\mathfrak{sl}_\infty)_\beta$. The only monomials in the generators of $U_q^+\mathfrak{sl}_\infty$ that are non-zero on $\dot v_{j}^{\sigma_n}$ are of the form $\dot e_i \dot e_{i+1}\cdots \dot e_{j-1}$ for $i \leq j$ if $\sigma_n=+$ (resp. the form $\dot e_{i-1} \dot e_{i-2} \cdots \dot e_{j}$ for $i \geq j$ if $\sigma_n=-$). Using also the integrality of the quasi-$R$-matrix from [@Lubook Corollary 24.1.6] (or a direct calculation from [@Lubook Theorem 4.1.2(b)]), it follows for any $v \in \widehat{V}^{\otimes\bar\bsigma}$ that $$\label{rhs}
\Theta\left(v \otimes \dot
v_{j}^{\sigma_n}\right)
= v \otimes \dot
v^{\sigma_n}_{j}+
\sum_i \left(\Theta_{i,j} v\right) \otimes \dot v^{\sigma_n}_i$$ summing over $i < j$ if $\sigma_n=+$ (resp. $i > j$ if $\sigma_n=-$), for $\Theta_{i,j}$’s that are $\operatorname{\mathbb{Z}}[q,q^{-1}]$-linear combinations of monomials obtained by multiplying the generators $\dot f_i, \dot f_{i+1},\dots, \dot f_{j-1}$ (resp. $\dot f_{i-1}, \dot f_{i-2}, \dots, \dot f_{j}$) together in some order. By induction, $\bar\psi(\dot v_{\bar\bb})$ equals $\dot v_{\bar\bb}$ plus a $\operatorname{\mathbb{Z}}[q,q^{-1}]$-linear combination of $\dot v_{\bar\ba}$’s for $\bar\ba \succ \bar\bb$. Combining these two statements, we deduce that $$\Theta\left(\bar\psi(\dot v_{\bar\bb}\right) \otimes \dot
v_{b_n}^{\sigma_n})
=
\dot v_{\bb}+\text{(a $\operatorname{\mathbb{Z}}[q,q^{-1}]$-linear combination of
$\dot v_\ba$'s for $\ba \succ \bb$)}.$$ This shows that the formula (\[bardef\]) makes $\psi(\dot v_\bb)$ into a well-defined element of $\widehat{V}^{\otimes\bsigma}$ of the desired form. The formula (\[bardef\]) also makes sense for arbitrary sums $\sum_{\bb \in \operatorname{\mathbf{B}}} p_\bb(q) \dot
v_\bb$ due to the interval-finiteness of the inverse dominance ordering on $P^n$. Finally, to see that $\psi$ commutes with the actions of all $\dot f_i$ and $\dot e_i$, and that it is an involution, one argues as in [@Lubook $\S$27.3.1].
Now we are in a position to apply “Lusztig’s Lemma” as in the proof of [@Lubook Theorem 27.3.2] to deduce for each $\bb \in \operatorname{\mathbf{B}}$ that there is a unique vector $\dot c_\bb \in \widehat{V}^{\otimes\bsigma}$ such that
- $\psi(\dot c_\bb)=
\dot c_\bb$;
- $\dot c_\bb = \dot v_\bb + \text{a (possibly infinite)
$q\operatorname{\mathbb{Z}}[q]$-linear combination of $\dot v_\ba$'s
for $\ba \succ \bb$}$.
This defines the [*canonical basis*]{} $\{\dot c_\bb\:|\:\bb \in \operatorname{\mathbf{B}}\}$. It is known (but non-trivial) that each $\dot c_\bb$ is always a [*finite*]{} sum of $\dot v_\ba$’s, i.e. $\dot c_\bb \in \dot V^{\otimes \bsigma}$ before completion. Moreover, the polynomials $d_{\ba,\bb}(q)$ arising from the expansion $$\dot c_\bb = {\sum_{\ba\in \operatorname{\mathbf{B}}}} d_{\ba,\bb}(q)
\dot v_\ba$$ are some finite type A parabolic Kazhdan-Lusztig polynomials (suitably normalized). All of these statements have a natural representation theoretic explanation discussed in detail in [@BLW $\S$5.9]. In particular, the results of [@BLW] (or [@CLW]) imply the following.
\[dnt\] Under the identification of $\mathbb{C} \otimes_{\mathbb{Z}} K_0(\O^\Delta)$ with $V^{\otimes \bsigma}$ from Theorem \[itacts\], $[P(\bb)]$ corresponds to the specialization $c_\bb$ of the canonical basis element $\dot c_\bb$ at $q=1$. Equivalently, $(P(\bb):M(\ba)) = [M(\ba):L(\bb)] = d_{\ba,\bb}(1)$ for each $\ba,\bb \in \operatorname{\mathbf{B}}$.
We are ready to explain our new algorithm to compute $\dot c_\bb$. We proceed by induction on $n$. In case $n=1$, we have that $\dot c_\bb = \dot v_\bb$ always. If $n > 1$, we first compute $\dot c_{\bar\bb} \in \widehat{V}^{\otimes\bar\bsigma}$. It is a linear combination of finitely many $\dot v_{\bar\ba}$’s for $\bar\ba \succeq \bar\bb$. Then we define $j \in I$ as follows.
- If $\sigma_n=+$ then $j$ is the greatest integer such that $j \leq b_n$, and the following hold for all $1 \leq r < n$ and all tuples $\bar\ba = (a_1,\dots,a_{n-1})$ such that $\dot v_{\bar\ba}$ occurs in the expansion of $\dot c_{\bar\bb}$:
- if $\sigma_r=+$ then $j \leq a_r$;
- if $\sigma_r=-$ then $j < a_r$.
- If $\sigma_n=-$ then $j$ is the smallest integer such that $j \geq b_n$, and the following hold for all $1 \leq r < n$ and all tuples $\bar\ba = (a_1,\dots,a_{n-1})$ such that $\dot v_{\bar\ba}$ occurs from the expansion of $\dot c_{\bar\bb}$:
- if $\sigma_r=-$ then $j \geq a_r$;
- if $\sigma_r=+$ then $j > a_r$.
\[lastofyear\] In the above notation, we have that $\Theta\left(\dot c_{\bar \bb} \otimes \dot
v_j^{\sigma_n}\right) = \dot c_{\bar\bb} \otimes \dot v_j^{\sigma_n}$.
By (\[rhs\]), we have that $\Theta\left(\dot c_{\bar \bb} \otimes \dot
v_j^{\sigma_n}\right) = \dot c_{\bar\bb} \otimes \dot v_j^{\sigma_n}
+ \sum_i \left(\Theta_{i,j} \dot c_{\bar\bb} \right) \otimes \dot v_i^{\sigma_n}$ summing over $i < j$ if $\sigma_n=+$ (resp. $i > j$ if $\sigma_n=-$), where $\Theta_{i,j}$ is a linear combination of non-trivial monomials in the generators $\dot f_{j-1}, \dot f_{j-2},\dots, \dot f_i$ (resp. $\dot f_j, \dot
f_{j+1},\dots, \dot f_{i-1}$). By the definition of $j$, all of these generators act as zero on $\dot c_{\bar\bb}$.
Lemma \[lastofyear\] shows that the vector $\dot c_{\bar \bb} \otimes \dot v_j^{\sigma_n}\in
\widehat{V}^{\otimes\bsigma}$ is fixed by $\psi$. Hence, so too is $X\left(\dot c_{\bar \bb} \otimes \dot v_j^{\sigma_n}\right)$ where $$X :=
\left\{
\begin{array}{ll}
\dot f_{b_n-1} \cdots \dot f_{j+1} \dot f_{j}&\text{if $\sigma_n = +$,}\\
\dot f_{b_n} \cdots \dot f_{j-2} \dot f_{j-1}&\text{if $\sigma_n = -$.}
\end{array}\right.$$ By Lemma \[not\], this new vector equals $\dot v_\bb$ plus a $\operatorname{\mathbb{Z}}[q,q^{-1}]$-linear combination of $\dot
v_{\ba}$’s for $\ba \succ \bb$. If all but its leading coefficient lie in $q \operatorname{\mathbb{Z}}[q]$, it is already the desired vector $\dot c_{\bb}$. Otherwise, one picks $\ba
\succ \bb$ minimal so that the $\dot v_\ba$-coefficient is not in $q
\operatorname{\mathbb{Z}}[q]$, then subtracts a bar-invariant multiple of the recursively computed vector $\dot c_\ba$ to remedy this defficiency. Continuing in this way, we finally obtain a bar-invariant vector with all of the required properties to be $\dot c_{\bb}$.
The algorithm just described has been implemented in [Gap]{}, and is available at [http://pages.uoregon.edu/brundan/papers/A.gap]{}. It is not obvious to us that it terminates in finite time for every $\bb \in \operatorname{\mathbf{B}}$. Based on examples, we believe that this is indeed the case.
Suppose $\bsigma = (+,+,-,-)$ and $\bb = (1,2,2,1)$. By induction, we have $\dot c_{(1,2,2)} =
\dot v_{(1,2,2)} + q \dot
v_{(2,1,2)} + q \dot v_{(1,3,3)} + q^2 \dot v_{(3,1,3)}$, so take $j=4$. We compute $$\begin{gathered}
\dot f_1 \dot f_2 \dot f_3
(\dot v_{(1,2,2,4)} + q \dot
v_{(2,1,2,4)} + q \dot v_{(1,3,3,4)} + q^2 \dot v_{(3,1,3,4)})=
\dot v_{(1, 2, 2, 1)}
+q\dot v_{(1, 2, 1, 2)}+
q\dot v_{(2, 1, 2, 1)}
\\+q\dot v_{(1, 4, 1, 4)}
+ q\dot v_{(1, 3, 3, 1)}
+q^2\dot v_{(2, 1, 1, 2)}
+q^2\dot v_{(3, 1, 3, 1)}+
q^2 \dot v_{(2, 3, 3, 2)}
+q^2 \dot v_{(4, 1, 1, 4)}
\\\:+
q^2\dot v_{(2, 4, 2, 4)}
+ (1+q^2) \dot v_{(1, 3, 1, 3)}
+ q^3\dot v_{(3, 2, 3, 2)}
+q^3\dot v_{(4, 2, 2, 4)}
+(q+q^3)\dot v_{(2, 3, 2, 3 )}\\+
(q+q^3) \dot v_{(3, 1, 1, 3)}
+ (q+q^3)
\dot v_{(2, 2, 2, 2)}
+(q^2+q^4) \dot v_{(3, 2, 2, 3)}.\end{gathered}$$ Then we subtract the recursively computed $$\begin{gathered}
\dot c_{(1,3,1,3)}=
\dot v_{(1, 3, 1, 3)}
+
q \dot v_{(3, 1, 1, 3)}
+
q
\dot v_{(2, 3, 2, 3)}
+
q
\dot v_{(1, 4, 1, 4)}\\
+q^2 \dot v_{(3, 2, 2, 3)}
+q^2 \dot v_{(4, 1, 1, 4)} + q^2 \dot v_{(2, 4, 2, 4)}
+
q^3 \dot v_{(4, 2, 2, 4)}\end{gathered}$$ to get $\dot c_{(1,2,2,1)}$.
[CKW]{} J. Brundan, Kazhdan-Lusztig polynomials and character formulae for the Lie superalgebra $\mathfrak{gl}(m|n)$, [*J. Amer. Math. Soc.*]{} [**16**]{} (2003), 185–231.
J. Brundan, Kazhdan-Lusztig polynomials and character formulae for the Lie superalgebra $\q(n)$, [*Advances Math.*]{} [**182**]{} (2004), 28–77.
J. Brundan, Tilting modules for Lie superalgebras, [*Commun. Alg.*]{} [**32**]{} (2004), 2251–2268.
J. Brundan, Representations of the general linear Lie superalgebra in the BGG category $\mathcal{O}$, in: “Developments and Retrospectives in Lie Theory: Algebraic Methods,” eds. G. Mason et al., [Developments in Mathematics]{} [**38**]{}, Springer, 2014, pp. 71–98.
J. Brundan and N. Davidson, Type C blocks of super category $\O$, in preparation.
J. Brundan and A. Ellis, Monoidal supercategories; .
J. Brundan and A. Ellis, Super Kac-Moody 2-categories, in preparation.
J. Brundan and A. Kleshchev, Hecke-Clifford superalgebras, crystals of type $A_{2\ell}^{(2)}$ and modular branching rules for $\widehat{S}_n$, [*Represent. Theory*]{} [**5**]{} (2001), 317-403.
J. Brundan, I. Losev, and B. Webster, Tensor product categorifications and the super Kazhdan-Lusztig conjecture, to appear in [*IMRN*]{}.
A. Beilinson, V. Ginzburg and W. Soergel, Koszul duality patterns in representation theory, [*J. Amer. Math. Soc.*]{} [**9**]{} (1996), 473–527.
B. Cao and N. Lam, An inversion formula for some Fock spaces; .
C.-W. Chen, Reduction method for representations of queer Lie superalgebras, [*J. Math. Phys.*]{} [**57**]{} 051703 (2016).
C.-W. Chen and S.-J. Cheng, Quantum group of type $A$ and representations of queer Lie superalgebra; .
S.-J. Cheng and J.-H. Kwon, Finite-dimensional half-integer weight modules over queer Lie superalgebras, [*Commun. Math. Phys.*]{} [**346**]{} (2016), 945–965.
S.-J. Cheng, J.-H. Kwon, and W. Wang, Character formulae for queer Lie superalgebras and canonical bases of types A/C, to appear in [*Commun. Math. Phys.*]{}.
S.-J. Cheng, N. Lam and W. Wang, Brundan-Kazhdan-Lusztig conjecture for general linear Lie superalgebras, [*Duke Math. J.*]{} [**164**]{} (2015), 617–695.
S.-J. Cheng and W. Wang, *Dualities and Representations of Lie Superalgebras,* Graduate Studies in Mathematics 144, AMS, 2012.
J. Chuang and R. Rouquier, Derived equivalences for symmetric groups and $\mathfrak{sl}_2$-categorification, [*Ann. Math.*]{} [**167**]{} (2008), 245–298.
E. Cline, B. Parshall and L. Scott, Finite dimensional algebras and highest weight categories, [*J. Reine Angew. Math.*]{} [**391**]{} (1988), 85–99.
A. Frisk, Typical blocks of the category $\O$ for the queer Lie superalgebra, [*J. Algebra Appl.*]{} [**6**]{} (2007), 731–778.
J. Humphreys, [*Representations of Semisimple Lie Algebras in the BGG Category $\mathcal O$*]{}, Graduate Studies in Mathematics 94, AMS, 2008.
S.-J. Kang, M. Kashiwara and S. Tsuchioka, Quiver Hecke superalgebras, [*J. Reine Angew. Math.*]{} [**711**]{} (2016), 1–54.
M. Khovanov and A. Lauda, A diagrammatic approach to categorification of quantum groups I, [*Represent. Theory*]{} [**13**]{} (2009), 309–347.
D. Hill, J. Kujawa and J. Sussan, Degenerate affine Hecke-Clifford algebras and type Q Lie superalgebras, [*Math. Zeit.*]{} [**268**]{} (2011), 1091–1158.
I. Losev and B. Webster, On uniqueness of tensor products of irreducible categorifications, [*Selecta Math.*]{} [**21**]{} (2015), 345–377.
G. Lusztig, [*Introduction to Quantum Groups*]{}, Birkhäuser, 1993.
V. Mazorchuk, Parabolic category $\mathcal O$ for classical Lie superalgebras, in: [*Advances in Lie Superalgebras*]{}, M. Gorelik and P. Papi (eds.), Springer INdAM Series 7, Springer, 2014, pp. 149–166.
M. Nazarov, Young’s symmetrizers for projective representations of the symmetric group, [*Advances Math.*]{} [**127**]{} (1997), 190–257.
R. Rouquier, 2-Kac-Moody algebras; .
B. Webster, Knot invariants and higher representation theory, to appear in [ *Mem. Amer. Math. Soc.*]{}.
[^1]: Research supported in part by NSF grant DMS-1161094.
[^2]: Note there is a sign error in [@KKT (5.5)]: it should read $-C_a C_{a+1}\dots$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We call CPE metrics the critical points of the total scalar curvature functional restricted to the space of metrics with constant scalar curvature of unitary volume. In this short note, we give a necessary and sufficient condition for a CPE metric to be Eisntein in therms of $\sigma_2$-singular spaces introduced in [@silvaandrade2018]. Such a result improves our understanding about CPE metrics and Besse’s conjecture (see [@besse2007einstein]) with a new geometric point of view.'
author:
- 'Maria Andrade[^1]'
bibliography:
- 'main.bib'
title: 'On the interplay between CPE metrics and $\sigma_2$-singular spaces.'
---
[**Keywords:**]{} Total scalar curvature, Critical point metric, Einstein metric, $\sigma_2$-curvature.
[**MSC:**]{} 53C20, 53C21, 53C24, 53C25.
Introduction
============
Scalar curvature appears in the study of Einstein metrics, since this is a type of curvature closed related with the Ricci tensor. Recall that a compact Riemannian manifold is said to be Einstein if the Ricci tensor is multiple of the metric $g,$ i.e., $Ric_g=\lambda g,$ where $\lambda:M\to \mathbb{R}.$ In others words, $(M^n, g)$ is Einstein if its traceless tensor $$\mathring{Ric_g}=Ric_g-\dfrac{R_g}{n}g$$ is identically zero, where $Ric_g$ and $R_g$ are Ricci and scalar curvatures, respectively.
Let $(M^n,g)$ be an $n$-dimensional closed (compact without boundary) oriented manifold with $n\geq 3,$ $ \mathcal{M}$ be the Riemannian metric space and $S_2(M)$ be the space of symmetric 2-tensors on $M.$ Fischer and Marsden [@fischer1975deformations] consider the scalar curvature map $\mathcal{R}:\mathcal{M}\to C^{\infty}$ which associates to each metric $g\in\mathcal{M}$ its scalar curvature. If $\gamma_g$ is the linearization of the map $\mathcal{R}$ and $\gamma^*_g$ is its $L^2$-formal adjoint, then they proved that $$\begin{aligned}
\gamma_gh=-\Delta_g\operatorname{tr}_gh+\delta^2_gh-\langle Ric_g,h\rangle\end{aligned}$$ and $$\begin{aligned}
\gamma_g^*f=\nabla^2_gf-(\Delta_gf)g-fRic_g\end{aligned}$$ where $\delta_g=-div_g,$ $h\in S_2(M)$ and $\nabla^2_g$ is the Hessian form on $M^n.$ The Einstein-Hilbert functional $\mathcal{S}:\mathcal{M}\to\mathcal{R}$ is defined by $$\begin{aligned}
\mathcal{S}(g)=\displaystyle\int_MR_gdv_g,\end{aligned}$$
It is well known that the solution of the Yamabe problem shows that any compact Manifold $M^n$ admits a Riemannian metric with constant scalar curvature. In particular, the set $\mathcal{C}=\{g\in \mathcal{M}; R_g \ \text{is constant}\}\neq \emptyset.$ Thus, we can consider $\mathcal{M}_1=\{g\in \mathcal{M}; R_g\in \mathcal{C}\ \text{and $vol_g(M)=1$}\}\neq\emptyset.$
Besse [@besse2007einstein] conjectured that the critical points of the total scalar curvature functional $\mathcal{S}$ restrited to $\mathcal{M}_1$ are Einstein. More precisely, the Euler-Lagrangian equation of Hilbert-Eisntein action restricted to $\mathcal{M}_1$ may be written as the following critical point equation (CPE) $$\begin{aligned}
\label{CPE}
\gamma_g^*f=\nabla^2_gf-(\Delta_gf)g-fRic_g=\mathring{Ric_g}.\end{aligned}$$
In this setting, it is interesting to study the critical points of the restriction of the Einstein-Hilbert functional to $\mathcal{M}_1.$ Following the notations developed in [@barros2014critical] and [@neto2015note] we consider the following definition.
A CPE metric is a triple $(M^n,g, f),$ where $(M^n,g)$ is a compact oriented Riemannian manifold of dimension $n\geq 3$ with constant scalar curvature and volume 1 and $f$ is a smooth potencial satisfying equation .
The Besse’s conjecture (or CPE conjeture) can be rewritten as
\[besseconjec\] A CPE metric is always Einstein.
Note that the equation $\eqref{CPE}$ is equivalent to $$\begin{aligned}
\label{eqcpe1}
\mathring{Ric_g}=\nabla^2_gf-\left(Ric_g-\dfrac{R_g}{n-1}\right)f\end{aligned}$$ and $$\begin{aligned}
\label{eqcpen}
(1+f)\mathring{Ric_g}=\nabla^2_gf+\dfrac{R_gf}{n(n-1)}g\end{aligned}$$ Observe that if $f$ is a constant function and satisfies the equation (\[eqcpen\]) then $f=0$ and this implies that $(M^n,g)$ is Einstein. Moreover, if $(M^n,g,f)$ is a CPE metric with $f$ non-constant function, then the set $$B=\{x\in M^n/f(x)=-1\}$$ has zero $n$-dimensional measure (see [@hwang2013three] and [@neto2015note]). Thus, to prove that CPE metric is Einstein is equivalent to show that $(g,f)$ satisfies the equation $$\begin{aligned}
\label{eq5}
\nabla^2_gf+\dfrac{R_gf}{n(n-1)}g=0\end{aligned}$$ where $f$ is not a constant function. This implies that $$\begin{aligned}
\label{eqlaplacian}
\Delta_gf=-\dfrac{R_g}{n-1}f.\end{aligned}$$ Now, multiplying the equation by $f$ and integrating over $M,$ we obtain $$\dfrac{R_g}{n-1}\displaystyle\int_Mf^2dv_g=-\displaystyle\int_Mf\Delta_gfdv_g=\displaystyle\int_M|\nabla_gf|^2dv_g>0$$ since $f$ is not a constant function, we get that $R_g>0.$
If $f$ satisfies the equation , then $(M^n, g)$ is isometric to $\mathbb{S}^n(r)$ where $r=\left(\frac{R_g}{n(n-1)}\right)^{1/2}$ (see [@tashiro1965complete]). In this way, the Conjecture \[besseconjec\] can be rewritten changing Einstein property by the manifold being the canonical sphere.
Hwang, [@hwang2013three], has proved the Besse’s conjecture when $n=3$ and under hypothesis that $ker (\gamma^*_g)\neq\{0\}.$ Barros and Ribeiro Jr., [@barros2014critical], proved that the Besse’s conjecture is true for $4$-dimensional half locally conformally flat manifolds. More recently, Neto (see [@neto2015note]) proved a necessary and suficient condition for a CPE metric to be Eisntein in terms of the potential function.
In this note, we give a necessary and suficient condition for a CPE metric to be Eisntein for $n\geq 3,$ improving the understanding about CPE metrics and Besse’s conjecture with a different geometric point of view. More precisely, we prove the following result.
\[main\] Let $(M^n,g,f)$ be an $n$-dimensional CPE metric, $n\geq 3$. $(M^n,g)$ is Eisntein if and only if $ker \Lambda^*_g\neq \{0\},$ where $\Lambda_g:S_2(M)\to C^{\infty}(M)$ is the linearization of the $\sigma_2$-curvature and $\Lambda^*_g$ is the $L^2$-formal adjoint of the operator $\Lambda_g.$
An immmediate consequence is the following
\[cmain\] Let $(M^n,g)$ be a compact, oriented $n$-dimensional manifold, with $n\geq 3$. Let $(g,f)$ be a non-trivial solution of . If $ker \Lambda_g^*\neq \{0\}$, then $(M^n,g)$ is isometric to the standard sphere $\mathbb{S}^n.$
Background and proofs
=====================
We start this section recalling some important results about $\sigma_2$-curvature. Then we prove the main result.\
Let $(M^n,g)$ be an $n$-dimensional Riemanian manifold, $n\geq 3$. The $\sigma_2$-curvature, which we will denotes by $\sigma_2(g),$ is as a nonlinear map $\sigma_2:\mathcal{M}\to C^{\infty}(M),$ defined as the second elementary symmetric function of the eigenvalue of the Schouter tensor $A_g=Ric_g-\frac{R_g}{2(n-1)}g$. In this case, we obtain that
$$\label{eq000}
\sigma_2(g)=-\dfrac{1}{2}|Ric_g|^2+\dfrac{n}{8(n-1)}R_g^2.$$
Motivated by works of Fischer and Marsden ([@fischer1975deformations]) and Lin and Yuan ([@lin2016deformations]), in [@silvaandrade2018] was proved that the linearization of the $\sigma_2$-curvature at the metric $g$, $$\Lambda_g:S_2(M)\to C^{\infty}(M),$$ is given by $$\begin{aligned}
\Lambda_g(h) & = & \frac{1}{2}\left\langle Ric_g,\Delta_gh+\nabla^2tr_gh+2\delta_g^*\delta_g h+2\mathring{R}(h)\right\rangle\\
& & -\frac{n}{4(n-1)}R_g\left(\Delta_g tr_gh-\delta_g^2h+\langle Ric,h\rangle\right),
\end{aligned}$$ where $\delta_g^*$ is the $L^2$-formal adjoint of $\delta_g$ and $\mathring{R}(h)_{ij}=g^{kl}g^{st}R_{kijs}h_{lt}.$
Thus, its $L^2$-formal adjoint, $\Lambda_g^*:C^{\infty}(M)\to S_2(M)$, is $$\begin{aligned}
\begin{array}{rcl}
\Lambda_g^*(f) & = & \displaystyle\frac{1}{2}\Delta_g(fRic_g)+\frac{1}{2}\delta_g^2(fRic_g)g+\delta_g^*\delta_g(fRic_g)+f\mathring{R}(Ric_g)\\
& & -\displaystyle\frac{n}{4(n-1)}\left(\Delta_g(fR_g)g-\nabla^2(fR_g)+fR_gRic_g\right).
\end{array}\end{aligned}$$
This implies that $$\label{eq003}
tr_g\Lambda_g^*(f)=\frac{2-n}{4}R_g\Delta_gf+\frac{n-2}{2}\langle \nabla^2f,Ric_g\rangle-2\sigma_2(g)f.$$ Note that, $$\begin{aligned}
\label{eq004}
\Lambda_g^*(1)&=&\frac{1}{2}\Delta_gRic_g-\frac{1}{4(n-1)}(\Delta_gR_g)g+\frac{2-n}{4(n-1)}\nabla^2R_g\nonumber\\
&+&\mathring{R}(Ric_g)-\frac{n}{4(n-1)}R_gRic_g
\end{aligned}$$
Then, by (\[eq003\]) and (\[eq004\]) we obtain $$\label{eqn005}
tr_g\Lambda_g^*(1)=-2\sigma_2(g)$$ and $$\label{eqn006}
div_g \Lambda_g^*(1)=-\frac{1}{2}d\sigma_2(g).$$
The relations (\[eqn005\]) and (\[eqn006\]) are similar to the relations between the Ricci tensor and the scalar curvature, namely $R_g=tr_gRic_g$ and $div_gRic_g=\frac{1}{2}dR_g$.
In [@silvaandrade2018] was introduced the notion of $\sigma_2$-singular space, which has the $L^2$-formal adjoint of the linearization of the $\sigma_2$-curvature map with nontrivial kernel, and under certain hypotheses it was proved rigidity and others results. More precisely,
\[def001\] A complete Riemannian manifold $(M,g)$ is $\sigma_2$-singular if $$ker \Lambda_g^*\not=\{0\},$$ where $\Lambda_g^*: C^{\infty}(M)\to S_2(M)$ is the $L^2$-formal adjoint of $\Lambda_g$. We will call the triple $(M,g,f)$ as a $\sigma_2$-singular space if $f$ is a nontrivial function in $ker \Lambda_g^*.$
\[isometricsphere\] Let $(M^n,g,f)$ be a closed $\sigma_2$-singular Einstein manifold with positive $\sigma_2$-curvature. Then $(M^n,g)$ is isometric to the round sphere with radius $r=\left(\frac{n(n-1)}{R_g}\right)^{\frac{1}{2}}$ and $f$ is an eigenfunction of the Laplacian associated to the first eigenvalue $\frac{R_g}{n-1}$ on $\mathbb{S}^n(r)$. Hence dim ker$\Lambda_g^*=n+1$ and $\displaystyle\int_M fdv_g=0.$
The next Lemma is crucial for our result.
\[mainlema\] Let $(M^n,g,f)$ be an $n$-dimensional CPE metric, then $$\operatorname{tr}\Lambda_g^*(f)=\left(\dfrac{n-2+nf}{2}\right)|\mathring{Ric}|^2.$$
Since $(M^n,g,f)$ is an $n$-dimensional CPE metric, then $f$ satisfies the equation $$\begin{aligned}
\label{eqcpe}
\nabla^2_gf=\mathring{Ric_g}-\left(Ric_g-\dfrac{R_g}{n-1}\right)f.\end{aligned}$$ Thus, by equations , and , we get $$\begin{aligned}
\operatorname{tr}_g\Lambda_g^*(f)&=&\frac{2-n}{4}R_g\left(\dfrac{-R_g}{(n-1)}\right)f+\frac{n-2}{2}\left< \mathring{Ric_g}-\left(Ric_g-\dfrac{R_g}{n-1}\right)f,Ric_g\right>\\
&+&\left(|Ric_g|^2-\dfrac{n}{4(n-1)}R_g^2\right)f\\
&=&\dfrac{n-2}{2}|\mathring{Ric_g}|^2+\dfrac{n}{2}\left(|Ric_g|^2-\dfrac{R_g^2}{n}\right)f\\
&=&\left(\dfrac{n-2+nf}{2}\right)|\mathring{Ric_g}|^2.\end{aligned}$$ This prove the result.
(Theorem \[main\]) First, we assume that $(M^n,g,f)$ is a CPE metric. Now, we suppose that $(M^n,g)$ is Eisntein, then by Lemma 2 in [@silvaandrade2018] we obtain that $$\label{eq0002}
\Lambda_g^*(f)=\frac{R_g(n-2)^2}{4n(n-1)}\left(\nabla^2f-(\Delta_gf)g-\frac{R_g}{n}fg\right).$$ So, by equation , we get $$\label{eq0002}
\Lambda_g^*(f)=\frac{R_g(n-2)^2}{4n(n-1)}\left(\nabla^2f+\frac{R_g}{n(n-1)}fg\right).$$ By hipothesis $(M^n,g)$ is Einstein, then using this in the equation we obtain that the equation is satisfied. Thus, by expression in we conclude that $f\in ker\Lambda_g^*.$
Conversely, we assume that $(M^n,g,f)$ is a $\sigma_2$-singular space, i.e, $\Lambda_g^*(f)=0,$ in particular $\operatorname{tr}\Lambda_g^*(f)=0.$ Since $f$ is not a constant function and $n\geq 3$, Lemma \[mainlema\] implies that $|\mathring{Ric_g}|^2=0,$ thus $(M^n,g)$ is Einstein.\
(Corollary \[cmain\]) Let $(M^n,g, f)$ be an $n$-dimensional CPE metric. If $(M^n,g)$ is a $\sigma_2$-singular space, then by Theorem \[main\], we obtain that $(M^n,g)$ is Einstein, and in this case $\sigma_2=\dfrac{(n-2)^2}{8n(n-1)}R_g^2>0.$ Thus, by Theorem \[isometricsphere\] $(M^n,g)$ is isometric to the round sphere with radius $r=\left(\frac{n(n-1)}{R_g}\right)^{\frac{1}{2}}$ and $f$ is an eigenfunction of the Laplacian associated to the first eigenvalue $\frac{R_g}{n-1}$ on $\mathbb{S}^n(r)$ with dim ker $\Lambda_g^*=n+1$ and $\displaystyle\int_M fdv_g=0.$
Acknowledgement {#acknowledgement .unnumbered}
===============
Maria Andrade thanks Ezequiel Barbosa, Pietro da Silva and Almir Silva Santos for their support and comments for this work. This work was done when Maria Andrade was visiting the Federal University of Minas Gerais in Belo Horizonte, supported by the grant PNPD/CAPES in 2019 and 2020. Maria Andrade would like to express her appreciation to its sponsorship and hospitality.
Universidade Federal de Minas Gerais (UFMG), Departamento de Matemática, Caixa Postal 702, 30123-970, Belo Horizonte, MG, Brasil and Universidade Federal de Sergipe, Departamento de Matemática, 49100-000, São Cristóvão, SE, Brasil.
E-mail adresss: maria@mat.ufs.br
[^1]: The author was partially supported by PNPD/CAPES/Brazil.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The study of genome rearrangement has many flavours, but they all are somehow tied to edit distances on variations of a multi-graph called the breakpoint graph. We study a weighted 2-break distance on Eulerian 2-edge-colored multi-graphs, which generalizes weighted versions of several Double Cut and Join problems, including those on genomes with unequal gene content. We affirm the connection between cycle decompositions and edit scenarios first discovered with the [[<span style="font-variant:small-caps;">Sorting By Reversals</span>]{}]{}problem. Using this we show that the problem of finding a parsimonious scenario of minimum cost on an Eulerian 2-edge-colored multi-graph – with a general cost function for 2-breaks – can be solved by decomposing the problem into independent instances on simple alternating cycles. For breakpoint graphs, and a more constrained cost function, based on coloring the vertices, we give a polynomial-time algorithm for finding a parsimonious 2-break scenario of minimum cost, while showing that finding a non-parsimonious 2-break scenario of minimum cost is NP-Hard.'
author:
- Pijus Simonaitis
- Annie Chateau
- 'Krister M. Swenson'
bibliography:
- 'cpm2018.bib'
title: 'A framework for cost-constrained genome rearrangement under Double Cut and Join'
---
Introduction
============
Edit distance problems are a mainstay in computer science. On strings, edit distances have roots in computational linguistics, and are at the heart of many approximate string matching algorithms for signal processing and text retrieval [@navarro2001guided]. The Levenshtein distance is the classic example, which asks for a minimum number of insertions, deletions, or substitutions of characters needed to transform one string into another. On graphs, the first edit distance to be considered is an analogue to the Levenshtein distance; insertion and deletion of vertices and edges are allowed, along with vertex and edge label substitution [@sanfeliu1983distance]. Graph edit distances have become an important tool in pattern recognition, which is at the heart of modern image processing and computer vision [@bunkegraph; @gao2010survey].
From early on, there were major applications for string edit distances in computational biology, mainly due to the linear nature of DNA, RNA, and protein molecules [@needleman1970general; @sankoff1972matching; @kruskal1983time; @gusfield1997algorithms]. More recently, graph edit distances have found a role in the comparison of gene regulatory networks [@mcgrane2016biological]. In the next subsection we outline the pervasive role of edit distances in genome rearrangements.
In this paper we address a graph edit distance on Eulerian 2-edge-colored multi-graphs, that is, a multi-graph with black and gray edges such that every vertex is incident to the same number of black and gray edges. When we say *graph*, we will mean Eulerian 2-edge-colored multi-graph. The edit operations that we consider are the *k-break*, where $k$ black edges are replaced such that the degree of all vertices in the graph are conserved. For 2-breaks, we generalize the edit distance problem to consider costs on the operations, posed as the [[<span style="font-variant:small-caps;">Minimum Cost Parsimonious Scenario</span>]{}]{}() problem. This problem asks for a minimum-length scenario of 2-breaks transforming the set of black edges into the set of gray edges.
Our main contribution is a clean formalism that facilitates simple proofs for edit distances on 2-breaks with costs. While weighted edit operations have been considered in the past, to our knowledge, this is the first study of weighted 2-breaks in this general setting [@benderGHHPSS2008; @Farnoud(Hassanzadeh):2012:SPC:2331863.2332291; @Swenson2016; @simonaitis_et_al:LIPIcs:2017:7660].
In Section \[decompositionOFscenario\] we show that a k-break scenario on a graph $G$ partitions $G$ into a set of Eulerian subgraphs, such that no k-break operates on edges from different subgraphs. This decomposition theorem allows us to show a strong link between [[<span style="font-variant:small-caps;">Maximum Alternating Cycle Decomposition</span>]{}]{}and the length of a parsimonious 2-break scenario for a graph. It is also the cornerstone for Section \[generalmlps\], showing that can be computed on a graph $G$ if there is a method for computing on a circle. By *circle* we mean a graph where all vertices are incident to exactly one gray and one black edge (Section \[generalmlps\]).
These results are general in the sense that the variants of the breakpoint graph typically used for genome rearrangement problems are all specific instance of the Eulerian 2-edge-colored multi-graph. Section \[secEditDistancesAndRearrangements\] gives such examples.
Section \[colorcost\] is dedicated to a specific cost function that depends on a coloring of the vertices: if a 2-break replaces edges $(a,b)$ and $(c,d)$ with $(a,d)$ and $(c,b)$, then the cost is zero if $a$ and $c$ have the same color, or $b$ and $d$ have the same color, otherwise it is of cost 1. Our tool for reasoning with this cost function is called the *color-merged graph*, which is obtained from a graph by merging all vertices of the same color. Theorem \[mcsnphard\] states that it is NP-Hard to compute [[<span style="font-variant:small-caps;">Minimum Cost Scenario</span>]{}]{}, even for circles. After establishing the necessary links between cycle decompositions in a graph and its merged graph (Section \[junctiondecomposition\]), we show how to compute on a circle (Sections \[circlecost\] and \[secPolyMCPS\]). Finally, we show that [[<span style="font-variant:small-caps;">Minimum Cost Parsimonious Scenario</span>]{}]{}for the colored cost function is computable in $O(n^4)$ time.
Edit distances and breakpoint graphs {#secEditDistancesAndRearrangements}
------------------------------------
In the area of gene order comparison through gene rearrangements, both the string and graph edit distances play a central role [@fertin2009combinatorics]. The typical genome rearrangement problem is an edit distance problem on strings of genes called *genomes*, where each gene occurs exactly once. The first biologically motivated edit operation that was studied is the *reversal* of a substring of a genome [@watterson1982; @Sankoff92]. When the relative direction of gene transcription is known, this problem is called [[<span style="font-variant:small-caps;">Signed Sorting By Reversals</span>]{}]{}([[<span style="font-variant:small-caps;">SSBR</span>]{}]{}), and is intimately linked to the *breakpoint graph* [@kececioglu1993exact; @bafna1993genome]. Roughly speaking, this multi-graph has a vertex for each gene extremity, and for every pair of adjacent gene extremities there is an edge. Thus, the edges for a single genome constitute a perfect matching. Say there is a gray genome and a black genome, then we have a gray matching and a black matching in the breakpoint graph. In this case the graph is 2-regular, and therefore decomposes into disjoint cycles of alternating color. A reversal on the black genome replaces two adjacencies. In this way, [[<span style="font-variant:small-caps;">SSBR</span>]{}]{}can be seen as a graph edit problem where the graph edit operation is a replacement of 2 edges, simulating the reversal of a substring. The [[<span style="font-variant:small-caps;">SSBR</span>]{}]{}problem can be solved in polynomial-time [@hannenhalli1999transforming].
When relative transcription directions are unknown for the genes, we have the [[<span style="font-variant:small-caps;">Sorting By Reversals</span>]{}]{}([[<span style="font-variant:small-caps;">SBR</span>]{}]{}) problem. In [[<span style="font-variant:small-caps;">SBR</span>]{}]{}, since the orientation of gene extremities are unknown, the vertices for the two extremities of each gene are merged into a single vertex in the breakpoint graph. The result is a 4-regular graph. While finding the alternating-color cycle decomposition of the 2-regular breakpoint graph is trivial, finding the same for a 4-regular graph is not: Caprara showed that [[<span style="font-variant:small-caps;">SBR</span>]{}]{}is as hard as finding a [[<span style="font-variant:small-caps;">Maximum Alternating Cycle Decomposition</span>]{}]{}([[<span style="font-variant:small-caps;">MACD</span>]{}]{}), and that [[<span style="font-variant:small-caps;">MACD</span>]{}]{}is NP-Hard [@caprara1997sorting].
Currently the most mathematically clean model for genome rearrangement is called the *Double Cut and Join* (DCJ) model [@DCJ; @Bergeron2006]. Genome extremities that are adjacent are paired, and transformations of these pairs occur by swapping elements of the pairs. DCJ is a generalization of the [[<span style="font-variant:small-caps;">SSBR</span>]{}]{}paradigm, since a DCJ operation on a genome can simulate a reversal. This implies that DCJ is a less restrictive graph edit distance model, where edge pair $\{(a,b), (c,d)\}$ is replaced by either $\{(a,c), (b,d)\}$ or $\{(a,d), (b,c)\}$ (in [[<span style="font-variant:small-caps;">SSBR</span>]{}]{}only one of the two replacements would be a valid reversal). The DCJ edit distance is inversely proportional to the number of cycles in the breakpoint graph.
While much of the work on genome rearrangement is on models where exactly one occurrence of each gene exists in each genome, content modifying operations like gene insertion, deletion, and duplication have been considered. Approaches to these problems are inextricably tied to finding cycle decompositions on breakpoint graphs [@el2000genome; @braga2011double; @shao2015exact; @shao2015comparing]. In the 2-break model, the [[<span style="font-variant:small-caps;">Maximum Alternating Cycle Decomposition</span>]{}]{}for each of these graphs implies the 2-break distance. Figure \[figBPGraph\] shows breakpoint graphs (see Section \[mlpsbreakpoint\] for a definition) for genomes with unequal gene content. Our results on 2-breaks, up to and including Section \[secPolyMCPS\], are directly applicable to the DCJ model, even in the presence of content modifying operations.
Distinguishing features of our model {#secDistFeatures}
------------------------------------
We have previously worked on DCJ problems with weight functions [@Swenson2016; @simonaitis_et_al:LIPIcs:2017:7660]. These papers address the 2-break model for 2-regular breakpoint graphs with a cost function on the edges instead of the vertices. The differences between the vertex-cost model from this paper, and edge-cost model of previous work are subtle. Indeed, the algorithmic results from Sections \[circlecost\],\[secPolyMCPS\], and \[mlpsbreakpoint\] are similar to those in [@Swenson2016], but bear the hallmark of this vertex colored model: simplicity. This simplicity facilitates the much more general nature of this work.
From a practical perspective, the vertex-cost model we present here sidesteps the weakest feature of the edge-cost model: the cost function. In the edge-cost model, each DCJ on a particular pair of adjacencies has two possible states. Which state is chosen determines the cost of subsequent DCJs on these adjacencies. In this way, the edge-cost function is state dependent.
For our vertex-cost model, the natural bijection between vertices representing the same extremity in the two genomes allows for a stateless cost function. This way, the weight of a DCJ can be computed as function of both genomes instead of just one. We also speculate that the computation of a general cost function under the vertex-cost model will be simpler. The only known polynomial time algorithm for the edge-cost model is restricted to a color based cost function.
k-breaks on a 2-edge-colored graph
==================================
Definitions
-----------
In our work a *graph* will be an Eulerian 2-edge-colored undirected multi-graph $G=(V,E^{b}\cup E^{g})$ with black and gray edges. We set $e(G)=|E^{b}|=|E^{g}|$.
For a graph $G$ and a vertex $v$ we set $\bar{d}(G,v)=d^{b}(G,v)-d^{g}(G,v)$ where $d^{b}$ and $d^{g}$ are black and gray degrees of $v$. $G$ is *Eulerian* if $\bar{d}(G,v)=0$ for every vertex $v$. A cycle is *alternating* if it is Eulerian. All the cycles in our work will be alternating unless specified otherwise. The *length* of a cycle is its number of black edges.
A *k-break* is a transformation of a graph $G$ into $G'$ that replaces $k$ black edges $(x_{1},x_{2}),\ldots, (x_{2k-1}, x_{2k})$ by $(x_{q_{1}},x_{q_{2}}),\ldots, (x_{q_{2k-1}}, x_{q_{2k}})$ while preserving the degree of all the vertices of $G$. In other words, $d^{b}(G,v)=d^{b}(G',v)$ for all $v$ and the multi-sets $\{x_{q_{1}}, \ldots, x_{q_{2k}}\}$ and $\{x_{1}, \ldots, x_{2k}\}$ are equal.
Since $G$ is Eulerian, it admits a decomposition into edge-disjoint alternating cycles. We denote the size of a [[<span style="font-variant:small-caps;">Maximum Alternating Cycle Decomposition</span>]{}]{}([[<span style="font-variant:small-caps;">MACD</span>]{}]{}) by $c(G)$ . If $c(G)=1$ we say that $G$ is a *simple cycle*. All the cycles in a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}are simple. A graph with equal multi-sets of black and gray edges will be called *terminal*. A *scenario* will be a sequence of k-breaks transforming a graph into a terminal graph.
Cycle decompositions for a k-break scenario {#decompositionOFscenario}
-------------------------------------------
To a k-break scenario we will associate a cycle decomposition such that all the edges replaced by a k-break belong to a single cycle. In Section \[generalmlps\] this will allow us to concentrate on the scenarios for the simple cycles instead of the general graphs.
We say that two edges are *$l$-related* according to a given scenario if among the first $l$ k-breaks of the scenario there is one replacing these two edges. $P_{l}$ will be a set of the equivalence classes of the transitive closure of this relation. We start by constructing $P_{l}$, which is a partition of the black edges of a graph, and proceed by showing how the gray edges can be incorporated to obtain a cycle decomposition of a graph.
Take a k-break scenario of length $m$ for a graph $G=G_{0}$. We denote the graph after $l\geq 0$ k-breaks by $G_{l}$. We label the $G_{0}$ black edges by the set $\{1,\ldots, e(G)\}$ and partition them into singletons $P_{0}=\{\{1\},\ldots,\{e(G)\}\}$. The $l$th k-break replaces $k$ black edges labeled $e_{1}, \ldots, e_{k}$ of $G_{l-1}$ by $k$ black edges that we arbitrarily label $e_{1},\ldots,e_{k}$. Let $Q_{i}$ be the subset of a partition $P_{l-1}$ including the edge $e_{i}$. We obtain a new partition $P_{l}$ from $P_{l-1}$ by merging all $Q_i$ into the set $\bigcup_{i=1}^{k} Q_{i}$. After applying all $m$ merges of a scenario we obtain a partition $P_{m}$ of the set $\{1,\ldots, e(G)\}$.
We label the gray edges of $G_{m}$ by $\{e(G)+1, \ldots, 2e(G)\}$, keeping the same labeling for all $G_{l}$. Since $G_{m}$ is terminal, its multi-sets of black and gray edges are equal. This implies a one-to-one mapping of labels between $\{1, \ldots, e(G)\}$ and $\{e(G)+1, \ldots, 2e(G)\}$ such that corresponding labels have the same endpoints in $G_{m}$. Using this mapping, we produce a partition on $\{1,\ldots, 2e(G)\}$ called $R_l$, that includes all black edge labels $e_i \in P_l$ along with the gray edge label that $e_i$ maps to. We will show that every subset of $R_{m}$ defines an Eulerian subgraph of $G$. stack For a subset $Q\subset\{1,\ldots, 2e(G)\}$ and $l\in\{0,\ldots,m\}$, we define $s(G_{l}, Q)$ to be the edge-induced subgraph of $G_{l}$ having only the edges of $G_{l}$ labeled by elements of $Q$. By construction, for $Q\in R_{m}$ and a vertex $v$ we have $\bar{d}(s(G_{m},Q),v)=0$. Lemma \[degrees\], proven in the Appendix, establishes the equality $\bar{d}(s(G_{0},Q),v)=\bar{d}(s(G_{m},Q),v)$, which means that $s(G,Q)$ is Eulerian. For a scenario $\rho$ $C(\rho) = R_{m}$ denotes the *cycle decomposition* of the scenario $\rho$.
\[degrees\] $\bar{d}(s(G_{0},Q),v)=\bar{d}(s(G_{l},Q),v)$ for a vertex $v$ and $Q\in R_{l}$ with $l\in\{0,\dots m\}$.
\[minlength2break\] The minimum length of a 2-break scenario $l(G)$ is equal to $e(G)-c(G)$.
We first show that for a 2-break scenario $\rho$ of length $m$ we have $|C(\rho)|\geq e(G)-m$. Recall that the size of $C(\rho)$ is equal to $|R_{m}|=|P_{m}|$ where $P_{m}$ is a partition of a set $\{1,\ldots, e(G)\}$ encountered in the construction of $C(\rho)$. The $l$th 2-break of $\rho$ replaces two edges $e_{1}$ and $e_{2}$ and merges the subsets $Q_{1}$ and $Q_{2}$ of $P_{l-1}$ (recall that $e_{1} \in Q_{1}$ and $e_{2}\in Q_{2}$) to obtain a partition $P_{l}$. By construction, $|P_{l}|\geq |P_{l-1}|-1$ as at most two subsets get merged. The size of $P_{0}$ is equal to $e(G)$, thus the size of $P_{m}$ is at least $e(G)-m$, meaning that $c(G)\geq e(G)-l(G)$.
On the other hand, for any cycle $c$ of length $l > 1$ there is a 2-break transforming $c$ into a union of length 1 and length $l-1$ cycles. In this way we obtain a scenario of length $l-1$ for $c$, and can transform every cycle of a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}of $G$ independently, obtaining a 2-break scenario of length $e(G)-c(G)$. Thus, $l(G)\leq e(G)-c(G)$.
$C(\rho)$ for a parsimonious 2-break scenario $\rho$ is a [[<span style="font-variant:small-caps;">Maximum Alternating Cycle Decomposition</span>]{}]{}of $G$.
[[<span style="font-variant:small-caps;">Minimum Cost Parsimonious Scenario</span>]{}]{} {#generalmlps}
========================================================================================
Consider a non-negative cost function $\varphi$ for the 2-breaks on $V$. By *2-breaks on* $V$ we mean the set of all the 2-breaks on the complete graph with vertices $V$. The cost of a scenario on a graph $G=(V,E^{b}\cup E^{g})$ is the sum of the costs of its 2-breaks. We provide an example of a cost function $\varphi$ in Section \[colorcost\]. The [[<span style="font-variant:small-caps;">Minimum Cost Parsimonious Scenario</span>]{}]{}() for a graph under cost function $\varphi$ is a minimum cost scenario among the scenarios for $G$ of minimum length. ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(G)$ denotes the cost of a for a graph $G$ and a cost function $\varphi$.
\[sumofcosts\] ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(G)$ is the minimum over $\big\{\sum_{c \in C} {\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_\varphi(c) ~\big|~ $C$~\text{is a {{\textsc{MACD}}\xspace}of}~$G$ \big\}$.
Take a $\rho$ for $G$ and $\varphi$. $C(\rho)$ is a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}of $G$ due to Theorem \[minlength2break\]. The subsequence of $\rho$ consisting of the 2-breaks replacing the edges of a cycle $c\in C(\rho)$ is a parsimonious scenario for $c$ that we name $\rho_{c}$. A 2-break sequence $\rho'$ obtained by performing $\rho_{c}$ one by one is a parsimonious scenario for $G$. The costs of $\rho$ and $\rho'$ are equal as they consist of the same 2-breaks performed in different order. This way we know that the cost of $\rho$ is smaller or equal to $\sum_{c\in C(\rho)} {\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(c)$. On the other hand, for a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}$C$ minimizing the sum $\sum_{c\in C} {\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(c)$ we can construct a parsimonious scenario of cost $\sum_{c\in C} {\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(c)$ transforming each cycle separately.
for a simple cycle {#simplecost}
------------------
A simple cycle $S$ might have a certain number $d(S)$ of vertices $v$ with $d^{b}(S,v)=d^{g}(S,v)=2$. It is easy to check that $d^{b}(G,v)=d^{g}(G,v)<3$ for any vertex. If $d(S)=0$, then we call $S$ a *circle*. See Figure \[simplecycles\] for an example of a simple cycle that is not a circle.
Take a simple cycle $S$ with $d(S)>0$, a cost function $\varphi$ for the 2-breaks on its vertices and a vertex $v_{0}$. For every Eulerian cycle $(v_{0},\ldots,v_{m-1},v_{0})$ of $S$ we construct a circle $(u_{0},\ldots,u_{m-1},u_{0})$ with a cost function $\varphi'$ defined as follows for $i,j,k,l\in\{0,\ldots,m-1\}$: $$\begin{aligned}
\varphi'((u_{i},u_{j}),(u_{k},u_{l})\rightarrow(u_{i},u_{k}),(u_{j},u_{l}))=\varphi((v_{i},v_{j}),(v_{k},v_{l})\rightarrow(v_{i},v_{k}),(v_{j},v_{l})).\end{aligned}$$ There are no more than $2^{d(S)}$ of such circles. In Figure \[simplecycles\] a simple cycle $S$ is given together with its two Eulerian circles. A simple cycle $S$ is recovered from its Eulerian circle by merging $d(S)/2$ pairs of vertices. In the Appendix we prove the following theorem:
\[costOFsimple\] The ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}$ cost of a simple cycle $S$ is equal to the minimum over all ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}$ costs of its Eulerian circles.
Given a subroutine computing ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}$ for a circle we can compute ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}$ for every simple cycle of $G$ using Theorem \[costOFsimple\]. Then ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(G)$ can be computed by choosing a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}of $G$ maximizing the sum of the costs of its cycles using Theorem \[sumofcosts\]. In Section \[colorcost\] we define a particular cost function $\varphi$ and provide a polynomial time algorithm computing ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}$ for a circle in Section \[secPolyMCPS\]. Then in Section \[mlpsbreakpoint\] we show how, using this algorithm as a subroutine, a minimum cost parsimonious DCJ scenario transforming genome $A$ into $B$ can be found in polynomial time.
A colored cost for 2-break scenarios {#colorcost}
====================================
In this section we partition $G$’s vertices into subsets of different colors so as to define a *colored cost function* $\varphi$ on 2-breaks: $\varphi((x_{1},x_{2}),(x_{3},x_{4})\rightarrow (x_{q_{1}},x_{q_{2}}),(x_{q_{3}},x_{q_{4}}))=0$ if $$\begin{aligned}
\big\{\{col(x_{1}),col(x_{2})\},\{col(x_{3}),col(x_{4})\}\big\}=\big\{\{col(x_{q_{1}}),col(x_{q_{2}})\},\{col(x_{q_{3}}),col(x_{q_{4}})\}\big\},\end{aligned}$$ and $\varphi$ is 1 otherwise. For example a 2-break $(x,y),(z,t)\rightarrow (x,z),(y,t)$ is of zero cost if $col(x)=col(t)$ or $col(y)=col(z)$, otherwise it is of cost 1. This cost function is convenient for use with spacial proximity constraints given by the packing of the chromosomes into the nucleus [@pulicani2017rearrangement]. See Figure \[0\] for an example.
In what follows, by a *graph* we will mean a graph together with a coloring $col$ of its vertices and by cost we will mean the cost $\varphi$ obtained using this coloring. We define a *color-merged graph* $J(G,col)$ obtained by merging the vertices of $G$ of the same color. For a 2-break $G\rightarrow G'$ a transformation $J(G,col)\rightarrow J(G',col)$ is also a 2-break. This means that a scenario $\rho$ for $G$ defines a scenario $\rho_{J}$ for $J(G,col)$. If the cost of $G\rightarrow G'$ is $0$, then $J(G,col)=J(G',col)$, and such a move can be omitted from $\rho_{J}$, leaving us with a scenario of length equal to the cost of $\rho$. This observation leads us to the following theorems proven in the Appendix.
\[mlsforgraph\] The minimum cost of a scenario for a graph $G$ is equal to $l(J(G,col))$.
\[mcsnphard\] Deciding for a circle $O$ and a bound $k$ whether there exists a scenario of cost at most $k$ for $O$ is NP-hard.
Cycle decomposition of a color-merged graph {#junctiondecomposition}
-------------------------------------------
Edges of $G$ and $J(G,col)$ can be labeled in such a way that for the edges $(a,b)$ and $(x,y)$ labeled $i$ in $G$ and $J(G,col)$ we have equality $(x,y)=(col(a),col(b))$. We say that such labelings *conform*. For a scenario $\rho$ of cost $w$ for $G$ we will construct a scenario $\rho_{J}$ for $J(G,col)$ of length $w$ that acts on the same edges as $\rho$. We will mimic the process of Section \[decompositionOFscenario\], taking more care this time when relabeling the edges during the scenario and mapping the black and gray edges at the end of the scenario.
We label the black edges of $G$ and $J(G,col)$ with $\{1,\ldots, e(G)\}$ and gray edges with $\{e(G)+1,\ldots, 2e(G)\}$. Given a labeling $L_{G}$ we obtain a conforming labeling $L_{J}$ by merging the vertices of $G$ having the same color while keeping the labels of the edges. We set $P_{0}=\big\{\{1\},\ldots,\{e(G)\}\big\}, G_{0}=G$ and $J_{0}=J(G,col)$. Let us fix a 2-break scenario $\rho$ of cost $w$ and length $m$ for $G$. We take the $l$th 2-break $(a,b),(c,d)\rightarrow (a,c),(b,d)$ of $\rho$ transforming $G_{l-1}$ into $G_{l}$ with $L_{G_{l-1}}(a,b)=i$ and $L_{G_{l-1}}(c,d)=j$.
If a 2-break is of cost 1, then we label the newly added edges $(a,c)$ and $(b,d)$ with $i$ and $j$ respectively to obtain $L_{G_{l}}$ and merge the subsets in $P_{l-1}$ containing $i$ and $j$ to obtain $P_{l}$. In $J(G_{l-1},col)$ we replace the edges labeled $i$ and $j$ by an edge $(col(a),col(c))$ labeled $i$ and $(col(b),col(d))$ labeled $j$ to obtain a labeling $L_{J_{l}}$ of $J(G_{l},col)$ that conforms to $L_{G_{l}}$. If a 2-break is of cost 0, then without loss of generality we can suppose that $col(a)=col(d)$. In this case we label the newly added edges $(b,d)$ and $(a,c)$ with $i$ and $j$ respectively to obtain $L_{G_{l}}$. $J(G_{l-1},col)=J(G_{l},col)$ and $L_{G_{l}}$ is chosen in such a way that $L_{J_{l-1}}$ and $L_{G_{l}}$ still conform, thus we keep $L_{J_{l}}=L_{J_{l-1}}$ and $P_{l}=P_{l-1}$.
At the end of the scenario we obtain a partition $P_{m}$ of $\{1,\ldots, e(G)\}$. Since $G_{m}$ is terminal, we can map the subsets $\{1,\ldots, e(G)\}$ and $\{e(G)+1,\ldots, 2e(G)\}$ one-to-one in a way that for a pair $(i,j)$ of mapped labels the black edge $i$ and gray edge $j$ have the same endpoints. For $j\in\{e(G)+1,\ldots, 2e(G)\}$ we include it into a subset of $P_{m}$ containing a label to which $j$ is mapped. This way a partition $R_{m}$ of a set $\{1,\ldots, 2e(G)\}$ is obtained. Only the 2-breaks of cost 1 of $\rho$ modify the colored-merged graph. This provides us with a 2-break scenario $\rho_{J}$ of length $w$ for $J(G,col)$. In addition to that $R_{m}$, when seen as a partition of the edges of $J(G,col)$, is exactly the cycle decomposition $C(\rho_{J})$ of the scenario $\rho_{J}$. In Section \[circlecost\] we study the structure of $C(\rho_{J})$ for a parsimonious scenario $\rho$.
for a circle {#circlecost}
------------
For a circle $O$ we take conforming labelings $L_{O}$ and $L_{J}$ of $O$ and $J(O,col)$. For a subset $S$ of edges of $J(O,col)$ (resp. $O$) we define a subset $O(S)$ (resp. $J(S)$) of edges of $O$ (resp. $J(O,col)$) labeled with the same labels. To a scenario $\rho$ for $O$ we have associated a scenario $\rho_{J}$ for $J(O,col)$ and its cycle decomposition $C(\rho_{J})$ in Section \[junctiondecomposition\].
Two disjoint subsets of edges $S_{1}$ and $S_{2}$ of a circle $O$ *cross* if there are edges $e_{1},e_{2}$ in $S_{1}$ and $f_{1}, f_{2}$ in $S_{2}$ such that a path in $O$ joining $e_{1}$ and $e_{2}$ contains exactly one of the edges $f_{1}$ or $f_{2}$. A cycle decomposition $C$ of $J(O,col)$ is said to be non-crossing if none of the subsets of $O(C)$ cross.
\[parsimoniousTOdecomposition\] $C(\rho_{J})$ is non-crossing for a parsimonious scenario $\rho$ for $O$.
If the theorem is false, then the set of the circles for which there exists a parsimonious scenario contradicting the theorem is non-empty. Let us take a circle $O^{*}$ in this set having the minimum number of edges and a scenario $\rho^{*}$ for $O^{*}$ such that $O(C(\rho^{*}_{J}))$ is crossing. $O^{*}$ has at least 2 black edges as otherwise $C(\rho^{*}_{J})$ contains a single subset.
Using Theorem \[minlength2break\] and the structure of a circle we get that a 2-break of a parsimonious scenario for a union of vertex-disjoint circles transforms one of its circles into a union of two vertex-disjoint circles. This means that the first 2-break of $\rho^{*}$ replaces edges $i$ and $j$ and transforms $O^{*}$ into a union of two smaller circles $\bar{O}$ and $\hat{O}$. The following 2-breaks of $\rho^{*}$ replace 2 edges with labels belonging to either $\bar{O}$ or $\hat{O}$, which provides us with the parsimonious scenarios $\bar{\rho}$ for $\bar{O}$ and $\hat{\rho}$ for $\hat{O}$. By the minimality of $O^{*}$ we get that $C(\bar{\rho}_{\bar{J}})$ and $C(\hat{\rho}_{\hat{J}})$ are non-crossing. $C(\rho^{*}_{J})$ can be easily obtained from $C(\bar{\rho}_{\bar{J}})$ and $C(\hat{\rho}_{\hat{J}})$ by taking their union and then merging the subsets of edges including $i$ and $j$ if the first 2-break of $\rho^{*}$ is of cost 1. Now it is easy to check that $C(\rho^{*}_{J})$ is non-crossing, a contradiction.
An arc is an alternating path joining two vertices of the same color, called *endpoints*, and having the same number of black and gray edges. Edges of an arc adjacent to its endpoints are called *ends* and one of them is black and another is gray.
A circle with $n$ vertices has $n$ arcs if all the vertices are of different colors and $n^2/2$ arcs if all the vertices share the same color. We say that two arcs do not *overlap* if they are edge-disjoint or one is included in another but none of their ends coincide. A set of pairwise non-overlapping arcs will be called an *independent* set of arcs. A [[<span style="font-variant:small-caps;">Maximum Independent Set of Arcs</span>]{}]{}([[<span style="font-variant:small-caps;">MISA</span>]{}]{}) of $O$ is an independent subset of arcs of $O$ of maximum cardinality. For a set of arcs $S$ a *maximal arc* is an arc that is not included in any other arc, while a *minimal arc* is an arc that does not include any other arc. In the Appendix Lemma \[decompositionTOmisa\] is proven, leading us to Theorem \[nestedscenario\].
\[decompositionTOmisa\] The size of a maximum non-crossing cycle decomposition of $J(O,col)$ is equal to the size of a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}of $O$.
\[nestedscenario\] The cost for a circle $O$ is equal to $e(O)-|I|$ for a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I$ of $O$.
We first show that there exists a parsimonious scenario for $O$ of cost at most $e(O)-|I|$. A parsimonious scenario for $O$ is of length $e(O)-1$, thus of cost at most $e(O)-1$, due to Theorem \[minlength2break\]. This means that if $|I|=1$, then inequality is trivial. Otherwise take a minimal arc $U$ in $I$. Perform a 2-break replacing the black end of $U$ and the black edge not belonging to $U$ adjacent to the gray end of $U$ and transforming $O$ into a union of two circles. One of these circles is of length equal to the length of $U$ and we call it $\hat{O}_{1}$. Another is such that its [[<span style="font-variant:small-caps;">MISA</span>]{}]{}is of size $|I|-1$ and we call it $O_{1}$. The cost of this 2-break is 0 as the endpoints of an arc have the same color. We iterate this procedure for $O_{1}$ until we end up with $|I|$ circles $O_{|I|-1}, \hat{O}_{1}, \ldots, \hat{O}_{|I|-1}$. A parsimonious scenario for this set of circles is of length $e(O)-|I|$ and thus of cost at most $e(O)-|I|$.
On the other hand, we take a scenario $\rho$ of cost $w$ for a circle $O$. A cycle decomposition $C(\rho_{J})$ of $J(O,col)$ is of size at least $e(O)-w$ using Lemma \[minlength2break\]. Using Theorem \[parsimoniousTOdecomposition\] we obtain that $C(\rho_{J})$ is non-crossing and using Lemma \[decompositionTOmisa\] we obtain an independent set of arcs of size $|C(\rho_{J})|$ and thus $w\geq e(O)-|C(J(\rho))|\geq e(O)-|I|$ where $I$ is a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}.
A polynomial time algorithm for on a circle {#secPolyMCPS}
-------------------------------------------
Fix a vertex $v$ of a circle $O$ having $n$ vertices. We say that an arc $U$ *crosses* $v$ if $v$ is in $U$ without being its endpoint. We replace $v$ by two vertices of the same color transforming $O$ into an alternating path $P$. [[<span style="font-variant:small-caps;">MISA</span>]{}]{}sizes of $P=(v_{0},\ldots,v_{l})$ and $O$ are equal due to Lemma \[arcTOinterval\], proven in the Appendix.
\[arcTOinterval\] There exists a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I'$ of $O$ such that there is no arc crossing $v$ in $I'$.
For $0\leq i\leq j\leq l$ ${{\textsc{MISA}}\xspace}(i,j)$ denotes the size of a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}of an alternating path $(v_{i},\ldots, v_{j})$, ${{\textsc{MISA}}\xspace}(i,i)=0$ for all $i$. We say that $i$ and $j$ are *compatible* if $v_{i}$ and $v_{j}$ are the endpoints of an arc. Take $I$ a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}of a path $(v_{i},\ldots,v_{j})$ for $j\geq i$. Vertex $v_{j}$ belongs to at most one arc in $I$ and this creates three possibilities. If $v_{j}$ does not belong to any arc in $I$, then $|I|={{\textsc{MISA}}\xspace}(i,j-1)$. If an arc with the endpoints $v_{i}$ and $v_{j}$ is in $I$, then $|I|={{\textsc{MISA}}\xspace}(i+1,j-1)+1$. If an arc with the endpoints $v_{k}$ and $v_{j}$ is in $I$ with $i<k<j$, then $|I|={{\textsc{MISA}}\xspace}(i,k)+{{\textsc{MISA}}\xspace}(k,j)$. This leads to the following recurrence: $${{\textsc{MISA}}\xspace}(i,j)=max
\begin{cases}
{{\textsc{MISA}}\xspace}(i,j-1),& \\
{{\textsc{MISA}}\xspace}(i+1,j-1)+1,&\text{if }i\text{ and }j\text{ compatible} \\
{{\textsc{MISA}}\xspace}(i,k)+{{\textsc{MISA}}\xspace}(k,j),&\text{for }k,i<k<j\text{ compatible with }j\text{,}
\end{cases}$$ which provides us with a dynamic program with time complexity $O(n^3)$. It is easy to modify the algorithm to give a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I$ of $O$. From $I$ it we can obtain a scenario for $O$ using Theorem \[nestedscenario\]. Since $I$ is of size $O(n)$ this can easily be done in $O(n^2)$ time.
We partition the vertices of $P$ into the subsets $S_{1},\ldots,S_{m}$ of pairwise compatible vertices and set $s_{i}=|S_{i}|$. One can show that our dynamic program computes ${{\textsc{MISA}}\xspace}(0,l)$ in $cn(s_{1}^{2}+\ldots+s_{m}^{2})$ steps for some constant $c$. If the subsets are of equal sizes then the number of steps is $cn^{3}/m$. This means that the best-case time-complexity of our program is $O(n^2)$.
DCJ scenarios for genomes {#mlpsbreakpoint}
=========================
A *genome* consists of *chromosomes* that are linear or circular orders of genes separated by potential *breakpoint* regions. In Figure \[0\] the tail of an arrow represents the *tail extremity*, and the head of an arrow represents the *head extremity* of a gene. We can represent a genome by a set of *adjacencies* between the gene extremities. In Figure \[0\] this set is $\big\{\{1t\}, \{1h, 2t\}, \{2h,3h\}, \{3t\}\big\}$ for genome $A$ and $\big\{\{1t\}, \{1h, 2h\}, \{2t,3h\}, \{3t\}\big\}$ for genome $B$. An *adjacency* is either an unordered pair of the extremities that are adjacent on a chromosome, called *internal* adjacency, or a single extremity adjacent to one of the two ends of a linear chromosome, called an *external* adjacency. In what follows we will suppose two genomes $A$ and $B$ that share the same genes, and our goal will be to transform $A$ into $B$ using a sequence of DCJs.
\[dcjgenome\] A DCJ cuts one or two breakpoint regions and joins the resulting ends of the chromosomes back in one of the four following ways: $\{a,b\},\{c,d\}\rightarrow\{a,c\},\{b,d\}$; $\{a,b\},\{c\}\rightarrow\{a,c\}$; $\{a,b\}\rightarrow\{a\},\{b\}$; $\{a\},\{b\}\rightarrow\{a,b\}$.
We partition the gene extremities into subsets of different colors. $col(a)$ denotes the color of a gene extremity $a$. The colored internal adjacency $\{a,b\}$ is $\{col(a), col(b)\}$ and the colored external adjacency $\{a\}$ is $\{col(a),\circ\}$, where $\circ$ does not coincide with any of the colors of the gene extremities. A DCJ $A\rightarrow A'$ is said to be of zero cost if the sets of colored adjacencies of $A$ and $A'$ are equal. It is of cost 1 otherwise. For example $\{a,b\},\{c\}\rightarrow\{a,c\},\{b\}$ is of cost 0 if $\big\{\{col(a),col(b)\},\{col(c),\circ\}\big\}=\big\{\{col(a),col(c)\},\{col(b),\circ\}\big\}$, that is if $col(b)=col(c)$. The cost of a DCJ scenario is the sum of the costs of its rearrangements.
In [@Bergeron2006], a linear time algorithm for finding a parsimonious DCJ scenario was proposed. The algorithm is based on the analysis of the connected components of the *adjacency graph*. Here, we use a slightly different structure associated to a genome pair $(A,B)$ called the breakpoint graph [@kececioglu1993exact; @bafna1993genome; @caprara1999sorting].
$G(A,B)=(V, E^{b}\cup E^{g})$ for genomes $A$ and $B$, sharing $n$ genes, is a 2-edge-colored Eulerian undirected multi-graph. $V$ consists of $2n$ gene extremities and an additional vertex $\circ$. For every internal adjacency $\{a,b\}\in A$ (resp. $\{a,b\}\in B$) there is a black (resp. gray) edge $(a,b)$ in $G(A,B)$ and for every external adjacency $\{a\}\in A$ (resp. $\{a\}\in B$) there is a black (resp. gray) edge $(a,\circ)$ in $G(A,B)$. We add additional black and gray loops $(\circ, \circ)$ to obtain $d^{b}(\circ)=d^{g}(\circ)=2n$. The breakpoint graph $G(A,B)$ for the genomes from Figure \[0\] is given in Figure \[1\].
\[dcjTObrekapointgraph\] For the DCJ scenarios transforming genome $A$ into $B$ the minimum length is $l(G(A,B))=e(G(A,B))-c(G(A,B))$, the minimum cost is $l(J)=e(J)-c(J)$ with $J=J(G(A,B),col)$ and the minimum cost of a parsimonious DCJ scenario is ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}(G(A,B)$.
is polynomial-time solvable for a breakpoint graph $G(A,B)$.
Take genomes $A$ and $B$ sharing $n$ genes. For all the vertices $v\neq\circ$ we have $d^{g}(G(A,B),v)=d^{b}(G(A,B),v)=1$. From this we obtain that for an edge belonging to a circle this is the only simple cycle in $G(A,B)$ including this edge. Thus a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}of $G(A,B)$ includes all of its circles. These set aside, we are left with $G(A,B)'$, which is a union of alternating paths starting and ending at $\circ$ and having the end edges of the same color. If this color is black we call a path $AA$, and $BB$ otherwise. Every simple cycle of $G(A,B)'$ is a union of a $BB$ path and a $AA$ path. We proceed by constructing a complete bipartite graph $H$ having $AA$ and $BB$ paths as vertices. An edge joining paths $a$ and $b$ is assigned the weight equal to the cost of $a\cup b$.
\[weightinggraph\] The weights of the edges of $H$ can be assigned in $O(n^4)$ time.
Lemma \[weightinggraph\] is proven in the Appendix. We proceed by computing a maximum weight matching for $H$ to obtain a [[<span style="font-variant:small-caps;">MACD</span>]{}]{}of $G(A,B)'$ minimizing cost, this can be done in $O(n^3)$ time using Hungarian algorithm. The cost of $G(A,B)$ is obtained by adding the costs of the circles and the cost of $G(A,B)'$. In this way we obtain a $O(n^4)$ time algorithm for computing the cost, which it can be easily modified to give a scenario.
Conclusions and further work
============================
Sorting by mathematical transpositions
--------------------------------------
Finding a parsimonious 2-break scenario on a circle is closely related to the problem of sorting a circular permutation by mathematical transpositions. For a permutation $\pi$ of $\{1,\ldots, n\}$ we define a digraph $D(\pi)$ with vertices $i$ and directed edges $(i,\pi(i))$. A transposition on $\pi$ defines a 2-break on $D(\pi)$ as illustrated in Figure \[-2\]. The problem of finding a minimum cost parsimonious scenario of transpositions for permutations with a cost function $\varphi$ defined for every pair of elements of $V$ was treated in [@Farnoud(Hassanzadeh):2012:SPC:2331863.2332291]. Such a $\varphi$ defines a natural cost function $\varphi'$ for the 2-breaks on $V$ such that $\varphi'(\{a,b\},\{c,d\}\rightarrow\{a,c\},\{b,d\})=min(\varphi(a,d),\varphi(b,c))$. The cost function used in Section \[colorcost\] is precisely of such a type. The paper described a polynomial-time algorithm for a minimum cost parsimonious scenario of transpositions for a general cost function $\varphi$. We speculate that this algorithm can be adapted to obtain a polynomial algorithm for ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}$ on a circle. Another problem of interest is finding a minimum cost scenario among the scenarios of length smaller than some fixed length. This would be an important step towards finding more realistic evolutionary scenarios, since the most likely scenario may not always be of minimum length.
Conclusions
-----------
The DCJ models on genes that are oriented, or unoriented [@Chen2013], have insertions and deletions [@Shao2012], or have segmental duplications [@shao2015comparing], are all intimately tied to the breakpoint graph. They all can be easily formulated in our setting of 2-edge-colored Eulerian multi-graphs and 2-break scenarios; all of our results about cycle decompositions and apply directly in these cases. We showed one example of how to use our work algorithmically, but we expect our framework to lead to further algorithmic results on general cost functions and general DCJ distances that consider unequal content.
Lemma \[degrees\]
-----------------
For a vertex $v$ and a subset $Q\in R_{l}$ with $l\in\{0,\dots m\}$ we have $$\begin{aligned}
\bar{d}(s(G_{0},Q),v)=\bar{d}(s(G_{l},Q),v)\end{aligned}$$
Equality is true for $l=0$. Suppose that equality is true for every $Q$ and $v$ with $l-1$ and proceed by induction on $l>0$. Fix a vertex $v$ and a subset $Q\in R_{l}$. The $l$th k-break replaces the edges labeled $e_{1},\ldots,e_{k}$. By construction, these edges belong to the same subset $Q'\in R_{l}$. There are two possibilities:
- ($Q'\neq Q$) In this case $Q\in R_{l-1}$ and $s(G_{l},Q)=s(G_{l-1}, Q)$, as the edges in $Q$ are unaffected by the $l$th k-break. Using the inductive hypothesis we obtain $$\bar{d}(s(G_{l},Q),v)=\bar{d}(s(G_{l-1},Q),v)=\bar{d}(s(G_{0},Q),v)$$
- ($Q=Q'$) In this case $Q'=\bigcup_{i=1}^{k} Q_{i}$ where $e_{i}\in Q_{i}\in R_{l-1}$. There may exist $i,j \in\{1,\ldots,k\}$ such that $Q_{i}=Q_{j}$, thus we select such $q_{1},\ldots,q_{r}\in\{1,\ldots,k\}$ so that $Q=Q'=\bigcup_{i=1}^{r} Q_{q_{i}}$ and all of these subsets are different. To begin with, $$\begin{aligned}
\bar{d}(s(G_{l},Q),v) = \bar{d}(s(G_{l-1},Q),v) = \sum_{i=1}^{r}\bar{d}(s(G_{l-1},Q_{q_{i}}),v).\end{aligned}$$ The first equality is due to the fact that a graph $s(G_{l},Q)$ is obtained from $s(G_{l-1},Q)$ by a k-break and k-break does not modify the degrees of the vertices in a graph. The second equality is guaranteed since no two $Q_{q_{1}},\ldots, Q_{q_{r}}$ intersect. Then $$\begin{aligned}
\sum_{i=1}^{r}\bar{d}(s(G_{l-1},Q_{q_{i}}),v) = \sum_{i=1}^{r}\bar{d}(s(G_{0},Q_{q_{i}}),v)=\bar{d}(s(G_{0},Q),v).\end{aligned}$$ With the first equality following from the inductive hypothesis and the latter once again since no two $Q_{q_{1}},\ldots, Q_{q_{r}}$ intersect.
As equality is preserved by a k-break, and true for $l=0$, we obtain the result by induction.
Theorem \[costOFsimple\]
------------------------
The ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}$ cost of a simple cycle $S$ is equal to the minimum over all ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}$ costs of its Eulerian circles.
Take a simple cycle $S$ with $d(S)=k$. If $k=0$, then $S$ is its own Eulerian circle. If $k>0$, then we choose a vertex $v$ satisfying $d^{b}(S,v)=d^{g}(S,v)=2$ and construct a graph with $v$ replaced by two vertices $v_{1}$ and $v_{2}$ satisfying $d^{b}(S,v_{i})=d^{g}(S,v_{i})=1$. We start with a copy of $S$ from which we remove $v$ with the adjacent edges and add two new vertices $v_{1}$ and $v_{2}$ to obtain $S'$. If there is a black loop $(v,v)$ in $S$ then we add a black edge $(v_{1},v_{2})$ to $S'$. Otherwise there are two black edges $(v,u_{1})$ and $(v,u_{2})$ in $S$ with possibly $u_{1}=u_{2}$. We add black edges $(v_{1},u_{1})$ and $(v_{2},u_{2})$ to $S'$. If there is a gray loop $(v,v)$ in $S$, then we add a gray edge $(v_{1},v_{2})$ to obtain a graph $S_{0}$. If there are two gray edges $(v,z_{1})$ and $(v,z_{2})$ in $S$, then we construct two graphs $S_{1}$ and $S_{2}$. $S_{1}$ is obtained by adding gray edges $(z_{1},v_{1})$ and $(z_{2},v_{2})$ to $S'$ and $S_{2}$ by adding gray edges $(z_{1},v_{2})$ and $(z_{2},v_{1})$ to $S'$. In Figure \[simplecycles\] the graphs $S_{1}$ and $S_{2}$ are given for a simple cycle $S$.
$S_{0}, S_{1}, S_{2}$ have the same set of vertices $V'$. We set $\bar{v}_{1}=\bar{v}_{2}=v$ and $\bar{u}=u$ for other vertices in $V'$. We define a cost function $\varphi'$ for the 2-breaks on $V'$ as follows: $$\begin{aligned}
\varphi'((x_{1},x_{2}),(x_{3},x_{4})\rightarrow(x_{1},x_{3}),(x_{2},x_{4}))=\varphi((\bar{x}_{1},\bar{x}_{2}),(\bar{x}_{3},\bar{x}_{4})\rightarrow(\bar{x}_{1},\bar{x}_{3}),(\bar{x}_{2},\bar{x}_{4})).\end{aligned}$$
From $S_{i}$ we obtain $S$ by merging the vertices $v_{1}$ and $v_{2}$. We will show that ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(S)$ of $S$ is equal to the ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{0})$ or the minimum of ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{1})$ and ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{2})$ depending on whether $S$ contains a gray loop $(v,v)$ or not. $c(S_{i})=1$ and $e(S_{i})=e(S)$, thus the lengths of their parsimonious scenarios are equal due to Theorem \[minlength2break\] .
For every 2-break transforming $S_{i}$ to $S_{i}'$ there is a unique 2-break on $S$ of the same cost transforming $S$ into a graph that we obtain from $S_{i}'$ by merging $v_{1}$ and $v_{2}$. If we take a scenario for $S_{i}$ we obtain a parsimonious scenario for $S$ of the same cost. This means that ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(S)$ is smaller or equal to ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{0})$ or the minimum of ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{1})$ and ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{2})$ depending on whether $S$ has a gray loop $(v,v)$ or not.
On the other hand for a 2-break transforming $S$ into $S'$ there exists a 2-break of the same cost transforming $S_{i}$ into such $S_{i}'$ that by merging its vertices $v_{1}$ and $v_{2}$ we obtain $S'$. A scenario for $S$ provides us with a sequence of 2-breaks $\rho$ on $S_{i}$ of the same cost transforming it into a graph $\hat{S_{i}}$ from which we obtain a terminal graph by merging $v_{1}$ and $v_{2}$.
The structure of $\hat{S_{i}}$ is fairly simple and the two possible cases can be checked by hand. If $S$ contains a gray loop $(v,v)$ then $\hat{S_{0}}$ must be already terminal as it is Eulerian and there is a single gray edge adjacent to $v_{0}$ or $v_{1}$. In this case a scenario for $S$ and $\varphi$ provides us with a parsimonious scenario of the same cost for $S_{0}$ and $\varphi'$.
Sets of black edges of $S_{1}$ and $S_{2}$ are equal. This means that a scenario for $S$ provides us with a single sequence $\rho$ of 2-breaks of the same cost transforming $S_{1}$ into $\hat{S_{1}}$ and $S_{2}$ into $\hat{S_{2}}$ such that by merging $v_{1}$ and $v_{2}$ we obtain terminal graphs. The sets of black edges of $\hat{S_{1}}$ and $\hat{S_{2}}$ stay equal. If $\hat{S_{1}}$ is already terminal, then we are done. Otherwise $\hat{S_{1}}$ is a union of a terminal graph and a cycle of length 2 containing gray edges $(z_{1},v_{1})$ and $(z_{2},v_{2})$ and black edges $(z_{2}, v_{1})$ and $(z_{1}, v_{2})$, however this means that $\hat{S_{2}}$ is terminal. Thus in this case ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(S)$ is equal to the minimum of ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{1})$ and ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi'}(S_{2})$.
Either $S_{i}$ are circles and we are done or we can proceed by choosing another vertex $v$. At the end we obtain Eulerian circles of $S$ and ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}_{\varphi}(S)$ is equal to the minimum of the ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}$ costs for these circles.
Theorem \[mlsforgraph\]
-----------------------
We start by proving an auxiliary lemma.
If $J(G,col)$ is terminal, then there exists a zero cost 2-break scenario for $G$.
We denote the maximum number of length 1 cycles in a cycle decomposition of $G$ by $c_{1}(G)$. If $c_{1}(G)=e(G)$, then $G$ is terminal and we are done. Otherwise we demonstrate how we can transform $G$ with a sequence of zero cost 2-breaks into $G'$ with $c_{1}(G')>c_{1}(G)$. When iterated this provides us with a zero cost scenario for $G$.
Take a cycle decomposition $C$ of $G$ having $c_{1}(G)$ length 1 cycles and remove these $c_{1}(G)$ cycles from $G$ obtaining $\bar{G}$. Take a black edge $(u,v)$ of $\bar{G}$. $J(\bar{G},col)$ is terminal, thus there exists a gray edge $(col(u),col(v))$ in $J(\bar{G},col)$. This means that there is a gray edge $(u',v')$ in $\bar{G}$ with $col(u)=col(u')$ and $col(v)=col(v')$. $G$ is Eulerian thus $\bar{G}$ is also Eulerian and there are gray edges $(p,u')$ and $(r, v')$ in $\bar{G}$.
- If $(u,v)$ equals one of these edges, let’s say $(p,u')$, then $col(p)=col(v)=col(v')$. This means that a 2-break $(p,u'),(r,v')\rightarrow (p,r),(u',v')$ transforming $\bar{G}$ into $\bar{G}'$ is of zero cost and creates a length 1 cycle.
- If $(u,v)$ does not coincide with $(p,u')$ or $(r, v')$, then a 2-break $(u,v),(p,u')\rightarrow (u,p),(v,u')$ is of zero cost as $col(u)=col(u')$. So is a 2-break $(v,u'),(v',r)\rightarrow (v,r),(u',v')$ as $col(v)=col(v')$.
These 2-breaks transforming $\bar{G}$ into $\bar{G}'$ are of zero cost and create a length 1 cycle in $\bar{G}'$. Once the length 1 cycles deleted from $G$ are reintroduced to $\bar{G}'$ we obtain a graph $G'$ with $c_{1}(G')>c_{1}(G)$ and a sequence of zero cost 2-breaks transforming $G$ into $G'$.
The minimum cost of a scenario for a graph $G$ is equal to $l(J(G,col))$.
For a 2-break $G\rightarrow G'$ a transformation $J(G,col)\rightarrow J(G',col)$ is also a 2-break. If the cost of $G\rightarrow G'$ is $0$, then $J(G,col)=J(G',col)$. This means that for a scenario of cost $w$ for $G$ there exists a scenario of length $w$ for $J(G,col)$ and thus $l(J(G,col))\leq w$. On the other hand, for every 2-break $J(G,col)\rightarrow J'$ a 2-break $G\rightarrow G'$ can be found such that $J(G',col)=J'$. For $J(G,col)$ scenario of length $l=l(J(G,col))$ we obtain a 2-break sequence of length $l$, thus of cost at most $l$, transforming $G$ into such $G'$ that $J(G',col)$ is terminal. Using the previously proven lemma we get a scenario for $G$ of cost at most $l$ establishing $l(J(G,col))\geq w$.
Theorem \[mcsnphard\]
---------------------
Deciding for a circle $O$ and a bound $k$ whether there exists a scenario of cost at most $k$ for $O$ is NP-hard.
The problem is clearly in NP. We reduce the decision version of a maximum cycle decomposition on simple Eulerian graphs, which is NP-hard [@Holyer81], to our problem. Without loss of generality, take an instance $G=(V,E)$ and a bound $k$, where $G$ is Eulerian and connected. Consider an Eulerian cycle $(u_{1},u_{2},\dots,u_{n},u_{1})$ of $G$ and construct a circle $O=(v_{1},\ldots,v_{2n},v_{1})$ with gray edges $(v_{2i-1},v_{2i})$, black edges $(v_{2i},v_{2i+1})$ and $col(v_{2i-1})=col(v_{2i})=u_{i}$ for $i\in\{1,\ldots, n\}$. The gray edges of $J(O,col)$ are loops and black edges of $J(O,col)$ are exactly the edges of $G$. There is a scenario of cost at most $e(G)-k$ for $O$ if and only if there exists an alternating cycle packing of $J(O,col)$ of size at least $k$ due to Theorem \[mlsforgraph\] and Theorem \[minlength2break\]. And the latter is true if and only if $G$ admits a maximum cycle decomposition of size least $k$.
Lemma \[decompositionTOmisa\]
-----------------------------
We start by proving an auxiliary lemma.
\[every\] Every edge of a circle $O$ is included in at least one arc of its [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I$.
Let us direct the edges of $O$ to obtain a directed cycle and suppose by contradiction that an edge $e$ entering a vertex of color $x$ does not belong to any arc in $I$. Without loss of generality we can suppose that the color of $e$ is black. The number of gray edges leaving the vertices of color $x$ is equal to the number of black edges entering the vertices of color $x$ in $O$ and in any arc of $O$. This is true due to the constraint that an arc must have the vertices of the same color as its endpoints and the edges of different colors as its ends. This means that there is a gray edge $f$ in $O$ not included in any arcs of $I$ leaving a vertex of color $x$. We define a new arc starting and ending at the vertices of color $x$ with edges $e$ and $f$ as its ends. By construction, this new arc does not overlap with any of the arcs in $I$, which contradicts the maximality of $I$.
Let us take a set $D$ of edge-disjoint non-crossing cycles of $J(O,col)$ and an independent set of arcs $Y$ of $O$. We call $(D,Y)$ a valid pair if the set of edges included in $Y$ and $O(D)$ partition the set of edges of $O$.
The size of a maximum non-crossing cycle decomposition of $J(O,col)$ is equal to the size of a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}of $O$.
For a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I$ of $O$ we have that $(\emptyset,I)$ is a valid pair due to the previously proven lemma. We start by showing how from a given valid pair $(D,Y)$ with a non-empty $Y$ we can obtain a valid pair $(D',Y')$ with $|D'|=|D|+1$ and $|Y'|=|Y|-1$. Take a maximal arc $U$ in $Y$. The set $c$ of edges of $U$ not belonging to any other arc in $Y$ is non-empty as it includes its ends. $J(V)$ is a cycle of $J(O,col)$ for every arc $V$ of $O$, thus we obtain that $J(c)$ is also a cycle of $J(O,col)$ and, by construction, it does not cross with the cycles in $D$. We include this cycle in $D$ and eliminate $U$ from $Y$ to obtain $(D',Y')$. Iterating this step we obtain a non-crossing cycle decomposition of $J(O,col)$ of size $|I|$.
Take a maximum non-crossing cycle decomposition $C$ of $J(O,col)$. $(C, \emptyset)$ is a valid pair. We proceed by showing how from a given valid pair $(D,Y)$ with non-empty $D$ we can obtain a valid pair $(D',Y')$ with $|D'|=|D|-1$ and $|Y'|=|Y|+1$. For every cycle $c$ in $D$ we define the minimum length path $P(c)$ in $O$ that contains every edge of $O(c)$. We choose such $c$ for which $P(c)$ is of minimum length. Since $P(c)$ is the minimum length path containing $O(c)$, we have that the ends of $P(c)$ belong to $O(c)$ and thus do not belong to any arc in $Y$. We remove from $P(c)$ any arcs of $Y$ that it might contain to obtain $P'$, a set of edges of $O$, and show that $P'$ consists entirely of edges of $O(c)$. Suppose by contradiction that there is an edge $e\in P'$ such that $e\notin O(c)$. As it does not belong to any arc in $Y$ and $(D,Y)$ is valid, there must be a cycle $c'\in C$ that contains $e$. However $O(c)$ and $O(c')$ can not cross and the endpoints of $P(c)$ belong to $O(c)$ and stay in $P'$, thus $O(c')$ must be properly included in $P'$ and thus in $P(c)$. However this means that $P(c')$ is also properly included in $P(c)$ and thus shorter in length which contradicts the minimality of $P(c)$. Thus we obtain that $P'$ consists entirely of edges of $O(c)$. Now we show that $P(c)$ is an arc non-overlapping with any arc in $Y$. We have already shown that $P(c)$ is a union of $O(c)$ where $c$ is a cycle in $J(O,col)$ and maybe some arcs from $Y$. This establishes that there is an equal number of gray and black edges in $P(c)$ and that the colors of its endpoints are the same, which means that $P(c)$ is an arc. By construction, the ends of this arc do not belong to any arc in $Y$ and thus $\{P(c)\}\cup Y$ is an independent set of arcs. We remove $c$ from $C$ and add $P(c)$ to $Y$ to obtain a valid pair $(D',Y')$. Iterating this step we obtain an independent set of arcs of $O$ of cardinality equal to the size of a maximum non-crossing cycle decomposition.
Lemma \[arcTOinterval\]
-----------------------
A chain $H$ in the set of arcs $I$ is a set of edge-disjoint arcs $U_{1},\ldots, U_{m}$ in $I$ with endpoints of $U_{i}$ being $u_{i}$ and $v_{i}$ such that all these endpoints are of the same color and $v_{i}=u_{i+1}$ for $i\in\{1,\ldots,m-1\}$. We say that $u_{0}$ and $v_{m}$ are the endpoints of $H$.
There exists a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I'$ of $O$ such that there is no arc crossing $v$ in $I'$.
If there is no [[<span style="font-variant:small-caps;">MISA</span>]{}]{}having an arc crossing $v$ then we are done. Otherwise take a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}$I$ of $O$ having at least one arc crossing $v$. All such arcs include the edges adjacent to $v$, thus there exists the single maximal arc $U$ in $I$ crossing $v$ that we remove from $I$. If after this $v$ does not belong to any arc in $I$ then we can include in $I$ an arc whose both endpoints are $v$ to obtain an independent set of arcs $I'$ with no arc crossing $v$. Otherwise we take a maximum length chain $H$ in $I$ including $v$. Due to the maximality of $U$ we obtain that $H$ is included in $U$ and does not include its ends, which means that the endpoints of $H$ do not coincide. We take an edge $e$ of $O$ adjacent to an endpoint of $H$ but not belonging to $H$ and show that $e$ does not belong to any arc in $I$. $e$ can not be an end of an arc as this arc would intersect with an arc in $H$ or could extend it. If $e$ belongs to an arc $V$ without being its end, then all of $H$ must belong to $V$, which contradicts the maximality of $H$. We obtain that an arc joining the endpoints of $H$ and not crossing $v$ can be included in $I$ giving an independent set of arcs $I'$ equal in size to $I$ and having the number of arcs crossing $v$ smaller than $I$. Iterating this process we obtain a [[<span style="font-variant:small-caps;">MISA</span>]{}]{}with no arcs crossing $v$.
Lemma \[dcjTObrekapointgraph\]
-------------------------------
For the DCJ scenarios transforming genome $A$ into $B$ the minimum length is $l(G(A,B))=e(G(A,B))-c(G(A,B))$, the minimum cost is $l(J)=e(J)-c(J)$ with $J=J(G(A,B),col)$ and the minimum cost of a DCJ scenario of minimum length is ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}(G(A,B)$.
$G(A,B)$ is constructed in such a way that for every DCJ $A\rightarrow A'$ the transformation $G(A,B)\rightarrow G(A',B)$ is a 2-break. Notably, a DCJ $\{a,b\}\rightarrow\{a\},\{b\}$ results in a transformation $(a,b),(\circ,\circ)\rightarrow (a,\circ),(b,\circ)$, as the construction of a breakpoint graph guarantees that there are enough black loops $(\circ,\circ)$ to realize such a 2-break. For any 2-break $G(A,B)\rightarrow G'$ with $G'\neq G(A,B)$ there exists a DCJ $A\rightarrow A'$ such that $G(A',B)=G'$. Since $G(B,B)$ is terminal, it follows that the minimum length of a scenario transforming $A$ into $B$ is $l(G(A,B))$.
Vertices of $G(A,B)$ are the gene extremities of $A$ and $B$ plus an additional vertex $\circ$. We color $\circ$ with a unique color. The rest of the vertices of $G(A,B)$ have the colors of their gene extremities. By construction, the cost of a DCJ $A\rightarrow A'$ is equal to the cost of a 2-break $G(A,B)\rightarrow G(A',B)$. On the other hand for a 2-break $G(A,B)\rightarrow G'$ there exists a DCJ $A\rightarrow A'$ of the same cost such that $G(A',B)=G'$. This means that the minimum cost of a scenario transforming $A$ into $B$ is equal to the minimum cost of a scenario for $G(A,B)$ and we conclude using Theorem \[mlsforgraph\] that this cost is $l(J)=e(J)-c(J)$ with $J=J(G(A,B),col)$.
Let us denote the minimum cost of a parsimonious DCJ scenario transforming $A$ into $B$ by $w_{min}$. In the two previous paragraphs we have seen that for a DCJ scenario of length $l$ and cost $w$ transforming $A$ into $B$ there exists a 2-break scenario on $G(A,B)$ of length $l$ and cost $w$. This means that ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}(G(A,B))\leq w_{min}$. On the other hand we have also seen that for a 2-break scenario on $G(A,B)$ of length $l$ and cost $w$ there exists a DCJ scenario of length $l$ and cost $w$ transforming $A$ into $B$ and this establishes the equality ${\hyperref[probMLPS]{{\textsc{MCPS}}}\xspace}(G(A,B))=w_{min}$.
Lemma \[weightinggraph\]
------------------------
The weights of the edges of $H$ can be assigned in $O(n^4)$ time.
$a_{1},\ldots,a_{x}$ denotes the sizes of $AA$ paths and $b_{1},\ldots,b_{y}$ denotes the sizes of $BB$ paths with $\sum_{i=0}^{x}a_{i}=|AA|$ and $\sum_{j=0}^{y}b_{j}=|BB|$. By construction, $d^{b}(G(A,B),\circ)=d^{g}(G(A,B),\circ)=2n$, meaning that $x,y\leq n$. cost of a union of two paths having $a_{i}$ and $b_{j}$ vertices can be computed in $2c(a_{i}+b_{j})^{3}$ steps as cost of a circle of size $n$ can be computed in $cn^3$ steps for some constant $c$ and we need to compute this cost for two Eulerian circles. We can compute cost for every pair of $AA$ and $BB$ path in the number of steps equal to $$\begin{aligned}
\sum_{i=0}^{x}\sum_{j=0}^{y}2c(a_{i}+b_{j})^{3}=&2c\sum_{i=0}^{x}\sum_{j=0}^{y} (a_{i}^3+3a_{i}^{2}b_{j}+3b_{j}^{2}a_{i}+b_{j}^3)\\
=&2c(y\sum_{i=0}^{x}a_{i}^{3}+x\sum_{j=0}^{y}b_{j}^{3}+3|BB|\sum_{i=0}^{x}a_{i}^{2}+3|AA|\sum_{j=0}^{x}b_{j}^{2})\end{aligned}$$ The terms $y,x,|AA|$ and $|BB|$ are clearly $O(n)$ and we obtain the worst-case time-complexity $O(n^4)$ for weighting the bipartite graph.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The QCDMAPT program package facilitates computations in the framework of dispersive approach to Quantum Chromodynamics. The QCDMAPTF version of this package enables one to perform such computations with Fortran, whereas the previous version was developed for use with Maple system. The QCDMAPTF package possesses the same basic features as its previous version. Namely, it embodies the calculated explicit expressions for relevant spectral functions up to the four–loop level and the subroutines for necessary integrals.'
address:
- 'BLTPh, Joint Institute for Nuclear Research, Dubna, 141980, Russian Federation'
- 'ISAC–CNR, I–40129, Bologna, Italy'
author:
- 'A.V. Nesterenko'
- 'C. Simolo'
title: 'QCDMAPTF: Fortran version of QCDMAPT package'
---
[**NEW VERSION PROGRAM SUMMARY**]{}
2.5mm
[*Manuscript Title:*]{} QCDMAPTF: Fortran version of QCDMAPT package\
[*Authors:*]{} A.V. Nesterenko and C. Simolo\
[*Program Title:*]{} QCDMAPTF\
[*Journal Reference:*]{} Comput. Phys. Comm. **182** (2011) 2303–2304\
[*Catalogue identifier:*]{} AEGPv20\
[*Programming language:*]{} Fortran 77 and higher\
[*Computer:*]{} Any which supports Fortran 77\
[*Operating system:*]{} Any which supports Fortran 77\
[*Keywords:*]{} Nonperturbative QCD; Dispersion relations\
[*Classification:*]{} 11.1, 11.5, 11.6\
[*External routines/libraries:*]{} MATHLIB routine RADAPT (D102) from CERNLIB Program Library [@MATHLIB].\
[*Catalogue identifier of previous version:*]{} AEGPv10\
[*Journal reference of previous version:*]{} Comput. Phys. Comm. **181** (2010) 1769–1775\
[*Does the new version supersede the previous version?:*]{} No\
[*Nature of problem:*]{} A central object of the dispersive (or “analytic”) approach to Quantum Chromodynamics [@APT; @MAPT] is the so–called spectral function, which can be calculated by making use of the strong running coupling. At the one–loop level the latter has a quite simple form and the relevant spectral function can easily be calculated. However, at the higher loop levels the strong running coupling has a rather cumbersome structure. Here, the explicit calculation of corresponding spectral functions represents a somewhat complicated task (see Sect. 3 and App. B of Ref. [@QCDMAPT]), whereas their numerical evaluation requires a lot of computational resources and essentially slows down the overall computation process.\
[*Solution method:*]{} The developed package includes the calculated explicit expressions for relevant spectral functions up to the four–loop level and the subroutines for necessary integrals.\
[*Reasons for the new version:*]{} The previous version of the package (Ref. [@QCDMAPT]) was developed for use with Maple system. The new version is developed for Fortran programming language.\
[*Summary of revisions:*]{} The QCDMAPTF package consists of the main program (QCDMAPTF.f) and two samples of the file containing the values of input parameters (QCDMAPTF.i1 and QCDMAPTF.i2). The main program includes the definitions of relevant spectral functions and subroutines for necessary integrals. The main program also provides an example of computation of the values of (M)APT spacelike/timelike expansion functions for the specified set of input parameters and (as an option) generates the output data files with values of these functions over the given kinematic intervals.\
[*Additional comments:*]{} For the proper functioning of QCDMAPTF package, the “MATHLIB” CERNLIB library [@MATHLIB] has to be installed.\
[*Running time:*]{} The running time of the main program with sample set of input parameters specified in the file QCDMAPTF.i2 is about a minute (depends on CPU).\
[0]{}
Subroutine D102 of the “MATHLIB” CERNLIB library, URL addresses:\
http://cernlib.web.cern.ch/cernlib/mathlib.html\
http://wwwasdoc.web.cern.ch/wwwasdoc/shortwrupsdir/d102/top.html D.V. Shirkov and I.L. Solovtsov, Phys. Rev. Lett. **79** (1997) 1209;\
K.A. Milton and I.L. Solovtsov, Phys. Rev. D **55** (1997) 5295;\
K.A. Milton and I.L. Solovtsov, Phys. Rev. D **59** (1999) 107701;\
I.L. Solovtsov and D.V. Shirkov, Theor. Math. Phys. **120** (1999) 1220;\
D.V. Shirkov and I.L. Solovtsov, Theor. Math. Phys. **150** (2007) 132. A.V. Nesterenko, Phys. Rev. D **62** (2000) 094028;\
A.V. Nesterenko, Phys. Rev. D **64** (2001) 116009;\
A.V. Nesterenko, Int. J. Mod. Phys. A **18** (2003) 5475;\
A.V. Nesterenko, Nucl. Phys. B (Proc. Suppl.) **133** (2004) 59;\
A.V. Nesterenko and J. Papavassiliou, J. Phys. G **32** (2006) 1025;\
A.V. Nesterenko, Nucl. Phys. B (Proc. Suppl.) **186** (2009) 207;\
A.V. Nesterenko, arXiv:1106.4006 \[hep-ph\]. A.V. Nesterenko and C. Simolo, Comput. Phys. Commun. **181** (2010) 1769.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
-
bibliography:
- 'refs.bib'
title: 'Score-based Tests for Explaining Upper-Level Heterogeneity in Linear Mixed Models'
---
=1
Introduction
============
The estimation and testing of cross-level interactions in linear mixed models (LMMs) is often complicated by the multiple variance terms in the model. For example, when a cross-level interaction (such as a treatment-subgroup interaction) exists, it may not be detected due to heterogeneity in a random effect variance or in a residual variance [@fok15]. If this heterogeneity is unaccounted for, it can lead to biased standard error estimates [@lec14; @ver96] and incorrect significance tests [@kwo07]. Along with treatment-subgroup interactions in cross-sectional studies, heterogeneous (co-)variances often occur in longitudinal studies, where the heterogeneity is observed across individuals. For example, recent ecological momentary assessment (EMA) studies have shown that substantial heterogeneity exists across individuals in the variance and covariance of emotional states over time [@roc09; @tho12; @kni13; @ebn15]. A similar phenomenon occurs in education research, where it is of interest to model student achievement over time [@loc07]. This heterogeneity is often due to unobserved covariates or to confounding with other covariates ; issues that are inevitable given the complicated observation scenario [@loc07]. Hence, it is important to develop accessible methods to detect heterogeneity in the multiple variance parameters observed in mixed models.
Several methods can be used to test for heterogeneity in specific situations, including the likelihood ratio (LR) test and the Kolmogorov-Smirnov test [@ver96; @mcl88; @ste74]. The LR test can be used when heterogeneity can be explained by an observed, unordered categorical variable. However, this test can be cumbersome when the variable has many categories, suboptimal when the variable has ordered categories, and not applicable when the variable is continuous instead of categorical. In contrast, the Kolmogorov-Smirnov test utilizes a mixture model framework, and checks the correct number of mixture components in an omnibus goodness-of-fit test. However, as shown by , this approach is subject to low power, even decrease to zero when residual variance is large.
In this paper, we aim to extend a family of score-based tests to linear mixed models, focusing on distinguishing cross-level interactions from heterogeneous residuals. We also discuss graphical methods associated with the tests, which can be helpful for identifying clusters that exhibit similar variance estimates. In the following section, we provide a brief overview of the score-based tests’ generalizations to linear mixed models. Next, we report on the results of a simulation to examine the tests’ abilities in the context of linear mixed models. Finally, we provide an empirical example and discuss the tests’ future generalizations.
Linear Mixed Model
==================
The linear mixed model (LMM) can be expressed in both conditional and marginal forms. The former facilitates theoretical understanding, and the latter simplifies the computational expression. We will detail these two expressions in the following sections, focusing on a two-level model where individual observations are nested within a series of clusters.
Conditional Expression
----------------------
The conditional version of the LMM can be written as $$\begin{aligned}
\label{eq:lmmcond}
\bm y_j |\bm b_j &\sim& N(\bm X_j \bm \beta+\bm Z_j\bm b_j, \bm R_j)\\
\label{eq:lmmran}
\bm b_j &\sim& N(\bm 0, \bm D)\\
\label{eq:lmmres}
\bm R_j &=& \sigma_{r}^2\bm I_{n_j},\end{aligned}$$ where $\bm y_j$ is the observed data vector for the $j$th cluster, $j=1,\ldots,J$ (so that the level 1 sample size is given as $\sum_{j=1}^{J} n_j$); $\bm X_j$ is an $n_j \times p$ matrix of fixed covariates; $\bm \beta$ is the fixed effect vector of length $p$; $\bm Z_j$ is an $n_j \times q$ design matrix of random effects; and $\bm b_j$ is the random effect vector of length $q$.
The vector $\bm b_j$ is assumed to follow a normal distribution with mean $\bm 0$ and covariance matrix $\bm D$, where $\bm D$ is a matrix composed of variances/covariances for random effect parameters. The residual covariance matrix, $\bm{R}_j$, is the product of the residual variance $\sigma_{r}^2$ and an identity matrix of dimension $n_j$. Later, the matrix $\bm R$ will include residuals across all clusters, so the identity matrix is of dimension $\sum_{j=1}^{J} n_j$.
Using the notation above, the following notation is used to represent data and parameters across all clusters in the data. $$\begin{aligned}
\label{eq:redefine1}
\bm y &=& \{\bm y_1, \bm y_2, \ldots, \bm y_j, \ldots, \bm y_J \}\\
\label{eq:redefine2}
\bm X &=& \{\bm X_1, \bm X_2, \ldots, \bm X_j, \ldots, \bm X_J \}\\
\label{eq:redefine3}
\bm Z &=& \{\bm Z_1, \bm Z_2, \ldots, \bm Z_j, \ldots, \bm Z_J \}\\
\label{eq:redefine4}
\bm b &=& \{\bm b_1, \bm b_2, \ldots, \bm b_j, \ldots, \bm b_J \}\\
\bm G &=& \bm D \otimes \bm I_{J},\end{aligned}$$ where $\otimes$ is the Kronecker product.
Finally, we define $\bm \sigma^2$ to be a vector of length $K$, containing all variance/covariance parameters (including those of the random effects and the residual). This implies that the matrix $\bm D$ has $(K-1)$ unique elements. For example, in a model with two random effects that are allowed to covary, $\bm \sigma^{2}$ is a vector of length 4 (i.e., $K = 4$). The first three elements correspond to the unique entries of $\bm D$, which are commonly expressed as $\sigma_0^2$, $\sigma_{01}$, and $\sigma_1^2$. The last component is then the residual variance $\sigma_r^2$.
Marginal Expression
-------------------
Based on Equations \[eq:lmmcond\], \[eq:lmmran\], and \[eq:lmmres\], the marginal distribution of the LMM is $$\label{eq:marginml}
\bm y_j \sim N(\bm X_j \bm \beta, \bm V_j),$$ where $$\label{eq:marginv}
\bm V_j = \bm Z_j \bm D \bm Z_j^{\top} + \bm R_j.$$ By using the combined notation from Equation to Equation , we can further define $\bm V$ as $$\begin{aligned}
\label{eq:marginvtotal}
\bm V &=& \bm V_{j} \otimes \bm I_{J}\\
&=& \bm Z \bm G \bm Z^{\top} + \bm R.\end{aligned}$$ Thus, Equation can be rewritten as: $$\label{eq:marginml2}
\bm y \sim N(\bm X \bm \beta, \bm V),$$ From Equation , we can perceive the LMM as a regular linear model with correlated residual variance $\bm V$. From this perspective one can easily deduce that heterogeneity in $\bm V$ has little impact on the estimate of $\bm \beta$, but can have a large impact on the significance test of $\bm \beta$ [@bates2015].
Problems Stemming from Heterogeneity
====================================
We now illustrate implications of heterogeneity via both theoretical results and simulation.
Theoretical Demonstration
-------------------------
The variance-covariance matrix w.r.t. the fixed parameter corresponds to the inverse of the model’s Fisher information, the relevant part of which can be expressed as : $$\begin{aligned}
\label{eq:vfix}
\text{V}_{\beta} &=& (\bm X \bm V^{-1} \bm X^{T})^{-1}\\
&=& (\bm X (\bm Z \bm G \bm Z^{T} + \bm R)^{-1}\bm X)^{-1}\\
&=& \bm X^{-1}(\bm Z \bm G \bm Z^{T} + \bm R)(\bm X^{T})^{-1}.\end{aligned}$$ The standard error of fixed parameter, $\text{SE}_{\beta}$, is then the square root of the diagonal elements of $\text{V}_{\beta}$. This shows that $\bm V$ directly contributes to the fixed parameters’ standard errors, which in turn influences the fixed parameters’ test statistics. With the under/over-estimates of $SE_{\beta}$, the *t*-statistic will be larger/smaller than it should be. Generally one can expect that the increasing of $\bm V$ results in Type II error whereas decreasing of $\bm V$ leads to Type I error. In practice, the former happens more often. conducted a series of Monte Carlo simulations and found underspecification and misspecification of $\bm V$ result in overestimation of $\text{SE}_{\beta}$, which lead to lower statistical power in significance tests of the fixed parameters. Although their simulations only examined main effects, one can expect similar results for interaction effects. We illustrate this issue in the next section.
Data Demonstration
------------------
In this section, we specifically illustrate how the change (increase) in $\bm V$ could impact the significance of fixed parameters by using artificial data similar to the *sleepstudy* data [@belenky03] included in *lme4*. This dataset includes $18$ subjects participating in a sleep deprivation study, where each subject’s reaction time (RT) [^1] was monitored for $10$ consecutive days. The reaction times are nested by subject and continuous in measurement. Then we fit a model with day of measurement (“*Days*”) as the covariate, including random intercept and slope effects that are allowed to covary. This leads to a model whose free parameters include: the fixed intercept and slope $\beta_0$ and $\beta_1$; the random variance and covariances $\sigma_0^2$, $\sigma_1^2$, and $\sigma_{01}$; and the residual variance $\sigma_r^2$. To illustrate the impact of heterogeneity on cross-level interactions, we also simulate an ordinal variable with four levels loosely called *Cognitive Ability* (CA), with its own main effect coefficient as $\beta_2$ and its interaction effect *Cognitive Ability* (CA) $\times$*Days* coefficient as $\beta_3$. In the simulation, we focus on the significance test results of $\beta_1$ and $\beta_3$. The true values were set to be $10.47$ and $6.27$, respectively, with both far different from $0$. The full model is displayed from Equation to Equation . $$\begin{aligned}
\label{eq:prob1}
\text{RT}_j|\text{Subject}_j &\sim& N(\beta_0 + \beta_1\text{Days}
+ \beta_2\text{CA}
+ \beta_3\text{Days} \times \text{CA}, \bm R_j)\\
\label{eq:prob2}
\text{Subject}_j &\sim& N(\bm 0, \left[ {\begin{array}{cc}
\sigma_0^2 & \sigma_{01}\\
\sigma_{01} & \sigma_{1}^2\\
\end{array} } \right]) \\
\label{eq:prob3}
\bm R_j &=& \sigma_r^2\bm I_{10}\end{aligned}$$
From Equation , it can be observed that changes in $\bm V$ can come from either $\bm G$, which is composed of between-subjects variance parameters $\sigma_0, \sigma_1$ and $\sigma_{01}$, or the residual variance $\sigma_r^2$. We generated data so that $\bm V$ changed with each of these four parameters, including the between subjects intercept variance $\sigma_{0}^2$, slope variance $\sigma_{1}^2$, covariance $\sigma_{01}$ and residual variance $\sigma_{r}^2$, along with different sample sizes as small ($n = 120$), medium ($n = 480$) and large ($n = 960$). Changes in these variance parameters began at cognitive ability level 2 and were consistent thereafter. Participants below cognitive ability level 2 deviated from participants at or above level 2 by $d$times the parameters’ asymptotic variance, with $d = 0, 1, 2, 3, 4$.
The magnitude of change is reflected in $d$. When $d$ is 0, it represents homogeneity in the corresponding parameter, which serves as the baseline; when $d$ is greater than 0, it represents heterogeneity in $\bm V$ (increasing with Cognitive Ability in this example), with larger $d$ indicating more severe heterogeneity. One example of data with and without heterogeneity is displayed in Figure \[fig:probdis\]. In the left panel, data were generated without heterogeneity in random slope ($d = 0$); whereas in the right panel data were generated with heterogeneity in random slope as large as $d = 4$. Within each panel, 12 different colored lines represent 12 subjects’ response time trajectory during 10 days, composing the small sample size as 120. It can be observed heterogeneity results in more varying individual trajectories. To formally examine the impact of heterogeneity, we computed the percentage of significant fixed parameters ($\beta_{1}$ and $\beta_3$) among $1,000$ replications in each condition.
![[]{data-label="fig:probdis"}](lmm-probdis)
The full simulation results for $\beta_{1}$ and $\beta_3$ are demonstrated in Figure \[fig:probres\], with the panel titles first indicating the tested parameter and then indicating the heterogeneous parameter, and the y-axis representing power (using $\alpha = 0.05$). In general, when sample size is medium or large, increasing heterogeneity in the slope variance $\sigma_1^2$ or covariance $\sigma_{01}$ reduces power for both the main effect and interaction effect. Heterogeneity in the residual variance or intercept variance does not impact power for $\beta_1$ or $\beta_3$, because they can be compensated for during estimation [@kwo07]. That is to say, when the intercept variance (or residual variance) increases, the residual variance (or intercept variance) estimation will decrease to compensate for the change, leading to the diagonal of $\bm V$ being unchanged. This compensation effect exists because the intercept covariate in the random effect design matrix ($\bm Z$) is all $1$, which results in the same contribution to the diagonal of $\bm V$ as the residual variance matrix.
When sample size is small ($n = 120$), power is generally lower in all scenarios. In addition, greater heterogeneity in the residual variance also leads to lower power, which might be due to the fact that heterogeneity combined with small sample size is more likely to result in unstable variance/covariance estimates, or even convergence issues.
Overall, however, failing to account for the upward changes in $\bm V$ would generally result in Type II error. Although it is important to systematically monitor heterogeneity in variance components, it is also plausible that a fixed parameter indeed changes according to another variable (e.g., that an interaction exists). Ideally, there would exist a statistical test that can differentiate between these two kinds of changes. In the next section, we will introduce a score-based family of statistical tests that can fulfill this need.
![[]{data-label="fig:probres"}](lmm-probres1)
Score-based Tests
=================
In this section, we will introduce the score-based test in the framework of LMM. This introduction draws on LMM results described by , and is related to tests described by, e.g., , , and .
Scores
------
Based on the marginal model expression shown in Equation , the log likelihood of the LMM can be expressed as:
$$\label{eq:obj}
\ell(\bm \sigma^2, \bm \beta; \bm y) = -\frac{n}{2}\log(2\pi) -
\frac{1}{2}\log(|\bm V|) - \frac{1}{2}
(\bm y - \bm X \bm \beta)^{\top}\bm V^{-1} (\bm y- \bm X \bm \beta).$$
Scores, denoted $s_i()$ in this paper, are based on the first partial derivatives of $\ell$ w.r.t. $\bm \xi = (\bm \sigma^2\ \bm \beta)^\top$. The scores involve these partial derivatives evaluated for each case $i$, and they can be viewed as a residual: values close to 0 imply that the model provides a good fit to case $i$, and values far from 0 imply the opposite.
The model gradient is equal to the sum of scores across all individuals and clusters: $$\begin{aligned}
\label{eq:firstder}
\ell^{'}(\bm \xi; \bm y) &=& \sum_{j=1}^{J}
\frac{\partial \ell(\bm \xi; \bm y_{j})}
{\partial \bm \xi} = \sum_{j=1}^{J}
\sum_{j \in c_j} s_i(\bm \xi; y_i)
$$ where $\frac{\partial \ell(\bm \xi; \bm y_{j})}{\partial \bm \xi}$ represents the first derivative within cluster $c_j$, which can be expressed as the sum of the casewise score $s_i()$ belonging to cluster $j$. For LMMs, the function $s_i()$ w.r.t. $\bm \sigma^{2}$ and $\bm \beta$ can be expressed as the $i$th component of the vectors [@wanmer16; @mcc01; @stroup12]:
$$\begin{aligned}
\label{eq:scoresigma}
s(\sigma_k^2; \bm y) &= -\frac{1}{2}
\text{diag} \left [\bm V^{-1} \frac{\partial \bm V}
{\partial \sigma_k^2}\right ] +
\left \{\frac{1}{2}(\bm y-\bm X \bm \beta)^{\top} \bm V^{-1} \left
(\frac{\partial \bm V}{\partial \sigma_k^2}\right ) \bm V^{-1}\right\}^{T}
\circ (\bm y- \bm X \bm \beta) \\
\label{eq:scoresbeta}
s(\bm \beta; \bm y) &= \left\{\bm X^{\top} \bm V^{-1} \right\}^{T} \circ (\bm y-\bm X
\bm \beta),\end{aligned}$$
where $\circ$ represents the Hadamard product (component-wise multiplication).
These equations provide scores for each observation $i$, and we can construct the clusterwise scores by summing scores within each cluster. In situations with one grouping (clustering) variable, the clusterwise scores can be obtained from a fitted model via R package [@merDeriv].
Statistics
----------
As applied to the LMMs considered here, score-based tests can be used to study heterogeneity that is potentially explained by an auxiliary variable $T$; for example, in the data demonstration considered earlier, the auxiliary variable could have been Cognitive Ability. Because the scores can be viewed as a type of residual, the score-based tests basically help us judge whether the residual magnitudes are associated with $T$. Because we have unique scores for each model parameter, we can also obtain information about where heterogeneity occurs.
Statistically, the tests considered here can be viewed as a generalization of the Lagrange multiplier test. The tests are based on a cumulative sum of scores, where the order of accumulation is determined by $T$. If there is no heterogeneity explained by $T$, then this cumulative sum should fluctuate around zero. Otherwise, the cumulative sum would systematically diverge from zero.
To formalize these ideas, we first define the (scaled) cumulative sum of the ordered scores. This can be written as $$\label{eq:cumscore}
{\bm B}(t; \hat {\bm \theta}) ~=~ \hat {\bm I}^{-1/2} J^{-1/2}
\sum_{j = 1}^{\lfloor J \cdot t \rfloor} {\bm s}(\hat {\bm \theta}; x_{(j)})
\qquad (0 \le t \le 1)$$ where $\hat{\bm I}$ is an estimate of the information matrix, $\lfloor
jt \rfloor$ is the integer part of $jt$ (i.e., a floor operator), and $x_{(j)}$ reflects the cluster with the $j$-th smallest value of the auxiliary variable $T$. While the above equation is written in general form, we can restrict the value of $t$ in finite samples to the set $\{0,
1/J, 2/J, 3/J, \ldots, J/J\}$. We focus on how the cumulative sum fluctuates as more clusters’ scores are added to it, e.g., starting with the person of lowest cognitive ability and ending with the person of highest cognitive ability. The summation is premultiplied by an estimate of the inverse square root of the information matrix, which serves to decorrelate the fluctuation processes associated with model parameters. For LMMs, $\bm \hat{I}$ can be written as expected information matrix :
$$\bm \hat{I} = \left[\begin{array}{ccc|ccc}
&&\\
& \bm X^{\top} \bm V^{-1} \bm X &&& \bm 0 &\\
&&\\
\hline
&&\\
& \bm 0 &&& \bm \left (\frac{1}{2}\right) \text{tr}
\left [ \bm V^{-1} \left (\frac{\partial \bm V}
{\partial \sigma_{k1}^2}\right)
\bm V^{-1} \left (\frac{\partial \bm V}
{\partial \sigma_{k2}^2}\right)\right ]&\\
&&\\
\end{array}\right].$$
Score-based tests statistics
----------------------------
To obtain an official test statistic, we must summarize the behavior of the cumulative sum in a scalar. Multiple summaries are available, leading to multiple tests of the same hypothesis. For example, one could take the absolute maximum that the cumulative sum attains for any parameter of interest, resulting in a [*double max*]{} statistic (the maximum is taken across parameters and clusters entering into the cumulative sum). Alternatively, one could sum the (squared) cumulative sum across parameters of interest and take the maximum or the average across clusters, resulting in a [*maximum Lagrange multiplier*]{} statistic and *Cramér-von Mises* statistic, respectively . These statistics are given by $$\begin{aligned}
\label{eq:dmax}
\mathit{DM} & = & \max_{j = 1,\dots, J} \max_{k = 1, \dots, K} | {\bm B}(\hat {\bm \theta})_{jk} | \\
\label{eq:cvm}
\mathit{CvM} & = & J^{-1} \sum_{j = 1,\dots, J} \sum_{k = 1, \dots, K} {\bm B}(\hat {\bm \theta})_{jk}^2, \\
\label{eq:maxlm}
\max \mathit{LM} & = & \max_{j = \underline{j}, \dots, \overline{\jmath}} ~
\left\{ \frac{j}{J} \left( 1 - \frac{j}{J} \right) \right\}^{-1}
\sum_{k = 1, \dots, K} {\bm B}(\hat {\bm \theta})_{jk}^2.\end{aligned}$$
For an ordinal auxiliary variable $T$ with $m$ levels, we can modify the statistics above so that the maximum is only considered after all clusters at the same level of $T$ have entered the summation. This leads to test statistics that are especially sensitive to heterogeneity that is monotonic with $T$ [@MerFanZei]. Formally, we define $t_{L}$ $(L=1,\ldots,m-1)$ to be the empirical, cumulative proportions of clusters observed at the first $m-1$ levels of $T$. The modified statistics are then given by $$\begin{aligned}
\label{eq:wdm}
\mathit{WDM}_o & = & \max_{j \in \{j_1, \dots, j_{m-1} \}} ~ \left\{ \frac{j}{J} \left( 1 - \frac{j}{J} \right) \right\}^{-1/2}
\max_{k = 1, \dots, K} | {\bm B}(\hat {\bm \theta})_{jk} |,\\
\label{eq:maxlmo}
\max \mathit{LM}_o & = & \max_{j \in \{j_1, \dots, j_{m-1} \}} ~
\left\{ \frac{j}{J} \left( 1 - \frac{j}{J} \right) \right\}^{-1}
\sum_{k = 1, \dots, K} {\bm B}(\hat {\bm \theta})_{jk}^2,\end{aligned}$$ where $j_{L}=\lfloor n \cdot t_{L} \rfloor$ $(L=1,\ldots,m-1)$.
Finally, if the auxiliary variable $T$ is only nominal/categorical, the cumulative sums of scores can be used to obtain a Lagrange multiplier statistic. This test statistic can be formally written as $$\label{eq:lmuo}
\mathit{LM}_{uo} = \sum_{L = 1, \dots, m} \sum_{k = 1, \dots, K}
\left( {\bm B}(\hat {\bm \theta})_{j_L k} - {\bm B}(\hat {\bm \theta})_{j_{L - 1}k} \right)^2,$$ where ${\bm B}(\hat {\bm \theta})_{j_{0}k} = 0$ for all $k$. This statistic is asymptotically equivalent to the usual, likelihood ratio test statistic, and it is advantageous over the likelihood ratio test because it requires estimation of only one model (the restricted model). We make use of this advantage in the simulations, described later.
In the following section, we apply these tests to a linear mixed model with one grouping variable, studying the tests’ ability to distinguish between heterogeneity and interactions.
Simulation
==========
The goal of the simulation is to examine score-based tests’ ability to differentiate between changes in fixed effect parameters (i.e., interaction effects) and changes in variance parameters (i.e., heterogeneity). For ease of description, we frame the data-generating model as being based on a longitudinal depression intervention study administered to participants with different levels of cognitive ability (here, we assume that $m = 4$, i.e., that there are four ordered levels of cognitive ability). Each participant’s depression magnitude is measured once per month during a 10 month period. Thus $10$ measurements are nested within each participant, which comprises a typical application for LMMs. It is plausible that the months needed to change the magnitude of depression is dependent on subjects’ cognitive ability. If so, there exists an interaction between time and cognitive ability. However, it is also possible that patient with higher cognitive ability has larger intercept variance ($\sigma_0^2$) or residual variance ($\sigma_r^2$). In addition, the interaction and heterogeneity might occur simultaneously. Since both interaction and heterogeneity can be viewed as parameter instability w.r.t. an auxiliary variable, we aim to examine the extent to which the score-based tests could attribute the parameter instability to the truly changing parameter(s) in an LMM framework.
Method
------
Data were generated from an LMM. The predictor is time, with its associated coefficient as $\beta_1$, and $\beta_0$ serves as the fixed intercept, which completes the fixed parameters in the model. For the random effect, we have two covarying random effect as intercept and slope as well, with the variance and covariance as $\sigma_0^2,
\sigma_{01}, \sigma_1^2$ respectively. The variances not captured by the random effects’ variance covariance is modeled in residual variance $\sigma_r^2$. The true parameter change occur in one of the four places: fixed parameter for time coefficient $\beta_1$, random intercept variance $\sigma_0^2$, residual variance $\sigma_r^2$ or simultaneously $\beta_1$ and $\sigma_r^2$. The fitted models matched the data generating model, and parameter estimates were obtained by marginal maximum likelihood. Parameter changes were tested in each of the $6$ estimated parameters, respectively.
Power and Type I error were examined across three sample sizes (n = 120, 480, 960), testing parameters and 5 magnitudes of parameter change. The parameter change point and changing magnitude is manipulated in the same way as an artificial example.
For each combination of sample size ($n$) $\times$ testing parameter $\times$ violation magnitude ($d$), $1,000$ data sets were generated and tested. Two ordinal statistics ($\text{maxLM}_o, \text{WDM}_o$) and one categorical statistic were examined ($\text{LM}_{uo}$) [@MerFanZei; @WanMerZei14]. The categorical statistic is asymptotically equivalent to the usual likelihood ratio test. Thus, this statistic provides information about the relative performance of the ordinal statistics vs. the LRT.
Results
-------
Full simulation results are presented in Figure \[fig:sim11res\] to Figure \[fig:sim14res\]. In each graph, the x-axis represents the violation magnitude and the y-axis represents power. Figure \[fig:sim11res\] demonstrates power curves as a function of violation magnitude in $\beta_1$, with sample size changing across rows, the tested parameters changing across columns, and lines reflecting different test statistics. Figure \[fig:sim12res\] and Figure \[fig:sim13res\] display similar power curves when the true changing parameter is $\sigma_0$ and $\sigma_r^2$, respectively. Figure \[fig:sim14res\] shows the power curves when there exist two changing parameters, $\beta_1$ and $\sigma_r^2$.
From these figures, one can generally observe that the score-based statistics could isolate the truly-changing parameter, with non-zero power curves for changing parameter(s), whereas near-zero power curves for non-changing parameters. For example, in Figure \[fig:sim11res\], the power curves for $\beta_1$ get larger with increasing violation magnitude $d$ and sample size (across rows); by contrast, all the power curves for the other five non-changing parameters (across columns) stay near zero.
Within each non-zero power curve panel of Figure \[fig:sim11res\] to Figure \[fig:sim14res\], the two ordinal statistics, $\text{maxLM}_o$ and $\text{WDM}_o$, exhibit higher (when testing fixed parameter or random intercept variance) or similar (when testing residual variance) power compared with categorical statistic $\text{LM}_{uo}$. This is partially consistent with the results demonstrated in , where ordinal statistics are shown to be more sensitive to monotonic parameter changes. The residual variance results might be due to a ceiling effect, where all three power curves quickly increase to 1. In conditions with only one changing parameter (Figure \[fig:sim11res\] to Figure \[fig:sim13res\]), $\text{maxLM}_o$ and $\text{WDM}_o$ are mathematically equivalent [@MerFanZei]. In conditions with two changing parameters (Figure \[fig:sim14res\]), $\text{maxLM}_o$ and $\text{WDM}_o$ still demonstrate similar power curves. The advantages of $\text{WDM}_o$ are only apparent when testing many (more than two) parameters at a time [@MerFanZei; @WanMerZei14].
Comparing the non-zero power curves across these four figures, it shows the score-based tests have somewhat higher power to detect residual variance change when sample size is medium or large, followed by fixed parameter change and random intercept variance parameter change. This phenomenon is most apparent by comparing Figure \[fig:sim12res\] and Figure \[fig:sim14res\], with the power curve for the residual variance and fixed parameter approaching 1 in conditions with medium or large sample sizes, while the power curve for the random intercept variance no greater than 0.8 for all values of sample size and $d$. In summary, we found that the score-based tests can attribute heterogeneity to the truly problematic parameter(s) in LMM context. Additionally, the tests were more sensitive to changes in variance parameters, as compared to fixed effect parameters. In the next section, we will fit real data set to illustrate the potential usage of the score-based tests in an empirical study. The general approach is to fit a LMM of interest, then obtain the score-based test statistics associated with each parameter w.r.t. an auxiliary variable in level 2 (e.g. person’s cognitive ability). If the variance (either random effect or residual) component is detected to have parameter instability, it indicates heterogeneity present in the data; if the fixed parameter demonstrates instability, then we can claim interaction between the covariate and the auxiliary variable.
![[]{data-label="fig:sim11res"}](lmm-sim11res)
![[]{data-label="fig:sim12res"}](lmm-sim12res)
![[]{data-label="fig:sim13res"}](lmm-sim13res)
![[]{data-label="fig:sim14res"}](lmm-sim14res)
Application
===========
We illustrate the tests’ applications using data from the 1982 “High School and Beyond” survey funded by National Center for Education Statistics (NCES), which is available in package . The aim of the current analyses is to determine how students’ math achievement scores are related to their family socioeconomic status.
In the dataset, 7185 U.S. high-school students from 160 schools completed a math achievement test, with the students’ socioeconomic status (SES) as a level 1 predictor. For the ease of parameter interpretation, we center SES around its school average, yielding a variable denoted cSES.
It is plausible that the relationship between cSES and math achievement differs across schools with different mean SES (denoted as meanSES). If so, then the model should include an interaction term between meanSES and cSES. However, as shown in the Problem Demonstration section, the significance test for the interaction might be impacted by variance/covariance heterogeneity in random effects. Thus, we use the score-based tests to distinguish between the cross-level interaction and variance heterogeneity. To illustrate the ordinal score-based test statistics, we recode school-level meanSES into five ordered categories.
Results
-------
The fitted LMM includes 6 parameters, which are $\beta_0$ $\beta_1$, $\sigma_0^2$ $\sigma_{01}$, $\sigma_1^2$ and $\sigma_r^2$. Applying the score-based tests to each individual parameter, we obtain the results in Figure \[fig:ordres\]. In that figure, the first column displays the fluctuation process associated with $\max \mathit{LM}_o$, and the second column displays the fluctuation process associated with $\mathit{WDM}_o$. Within each panel, the horizontal line represents the 5% critical value. If the solid line crosses the critical value, then there is evidence that the corresponding parameter fluctuates across meanSES. Since the final level’s statistics always equal zero, they are not displayed.
In Figure \[fig:ordres\], it is observed that the $\beta_0$, $\beta_1$, $\sigma_0^2$ and $\sigma_{1}^2$ demonstrate parameter instability, whereas $\sigma_{01}$ and $\sigma_r^2$ do not. The instability of $\beta_0$ indicates that there exists a main effect of meanSES, and the instability of $\beta_1$ implies that there exists a cross-level interaction effect between meanSES and cSES. In addition, the random intercept and the random slope demonstrate instability. As discussed earlier, this heterogeneity in random variance/covariance, could “mask” the significance of interaction term. In this specific example, fitting the model with the cross-level interaction included in the model leads to a non-significant interaction due to the heterogeneity in the random effects’ variances. On the other hand, the score-based test could differentiate these changes by fitting a model with only one main effect (cSES), using meanSES as an auxiliary variable.
Beyond the formal score-based tests, Figure \[fig:ordres\] provides information about levels of meanSES where parameters differ from one another; this can be discerned from levels where the solid line crosses the dashed horizontal line. Thus, the intercept parameter changes w.r.t. each of the four levels of meanSES (all points are above the line); the effect of SES ($\beta_1$) at meanSES level 1 differs from the other levels; random intercept $\sigma_0^2$ has two changing points: one is level 1 and the other is level 4; and the random slope $\sigma_1^2$ differs between level 1 and the remaining levels. Taken together, results imply that associations between students’ SES and math achievement function differently in schools with lowest meanSES, as compared to schools with higher meanSES.
Given the changing point information provided by the empirical plot, we can create a dummy variable by separating students at level 1 (coded as 1) or not (coded as 0). The dummy variable replace the previous position of meanSES. The interaction between cSES and dummy variable is 1.09 with $p < 0.05$, indicating significant interaction. Informed by this optimal cutting point, we can detect the previous “missing” interaction effect. Alternatively, we can fit the model separately for students at schools with meanSES at level 1 and students at other schools. The former model results in $\beta_1$ (coefficient of SES) as 1.21, whereas the latter model has $\beta_1$ as 2.31. These results indicate that students cSES has stronger relationship with math achievement in schools with higher meanSES. In summary, score-based tests provide a statistical tool to closely examine an LMM’s parameter estimates with respect to an auxiliary, level-2 variable. The examination of variance components (random effect (co)variance and residual variance) provide tests of heterogeneity. Additionally, the fluctuation plots can be used to interpret the nature of heterogeneity or interactions, without arbitrary median splits or use of subsamples with small numbers of observations.
![Empirical fluctuation processes of the $\max \mathit{LM}_o$ statistic (first column) and $\mathit{WDM}_o$ (second column) for $\beta_0$ (first row), $\beta_1$ (second row), $\sigma_0^2$ (third row) and $\sigma_r^2$ (fourth row), using M1 model.[]{data-label="fig:ordres"}](lmm-sctest-ord-graphics)
Discussion
==========
In this paper, we extended a family of score-based tests to linear mixed models, focusing on models with one grouping variable. We found that the tests can isolate specific parameters that exhibit instability, which avoids spurious cross-level interaction effects in the presence of heterogeneity. They also provide specific information about groupings of the auxiliary variable whose parameter values differ. The tests developed in this paper can currently only be carried out on an auxiliary variable measured at the model’s upper level (level 2), a restriction that leads to the future directions described below.
Grouping with multiple variables
--------------------------------
The auxiliary variable is specifically required to be at the upper level because the tests described here require that the scores be independent. This independence assumption challenges models with at least two variables defining clusters, such as models with (partially) crossed random designs, or models with multilevel nested designs [@bates10 Ch.2]. In these cases, we cannot simply sum scores within a cluster to obtain independent, clusterwise scores, because observations in different clusters on the first grouping variable may be in the same cluster on the second grouping variable. A related issue occurs when the auxiliary variable is at the lowest (first) level of the model: scores at the lowest level are not independent, so the tests described here cannot be immediately used to test parameters with respect to a level-one variable.
A natural approach to deal with the issue of dependent scores is to find a heteroskedasticity and autocorrelation consistent (HAC) covariance matrix estimator. The traditional Hessian matrix only accounts for the correlations among score columns, whereas the HAC estimator is a robust Hessian estimator that can serve the purpose of de-correlating the scores under a generalized linear model framework. Several methods have been proposed here, including kernel HAC estimators with automatic bandwidth selection and weighted empirical adaptive variance estimators [@and93]. With regard to the model with multiple grouping variables, recently implement sandwich approach to obtain robust variance covariance. It might be possible to deploy these methods in the linear mixed model context.
Model Estimation
----------------
Along with independence issues, general model estimation issues may influence the score-based tests’ accuracies. For example, in the relatively-common case where a parameter estimate lies on the boundary (e.g., a correlation between random effects near $\pm1$ or a variance approaching to zero), then it may be impossible to carry out the proposed tests due to the non-positive definite structure of the model information matrix. Additionally, model misspecification can also influence the tests. examined the tests’ performance in the factor analysis framework, and they found that unmodeled parameters’ instability would lead the tests to identify instability in related model parameters. Thus, it is important to carefully consider the covariates and random effects that enter into the model.
Summary
-------
In this paper, we generalized a family of score-based tests to two-level linear mixed models, which allow researchers to test whether model parameters fluctuate with an unmodeled level two variable. We found that the tests could successfully decouple cross-level interactions from variance heterogeneity, whereas heterogeneity could cause the traditional significance test of a cross-level interaction to exhibit a Type II error. Along with providing information about parameter stability across all estimated LMM parameters, the tests provide additional information about heterogeneous subgroups when parameter instability is detected. Thus, applied researchers in psychology and education can use the tests to examine potential cross-level interactions while ruling out possible spurious results due to heterogeneity.
Computational Details {#computational-details .unnumbered}
=====================
All results were obtained using the system for statistical computing [@R11], version 3.4.4, employing the add-on package 1.1-13 [@lme4] for fitting of the linear mixed models and 1.5-1 [@ZeiLei02; @Zei06] for evaluating the parameter instability tests. and both packages are freely available under the General Public License from the Comprehensive Archive Network at <http://CRAN.R-project.org/>. code for replication of our results is available at <http://semtools.R-Forge.R-project.org/>.
[^1]: strictly speaking, the response variable should be log(RT) to have a meaningful infinity support.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Time-correlated noise is a significant source of uncertainty when modeling exoplanet light-curve data. A correct assessment of correlated noise is fundamental to determine the true statistical significance of our findings. Here we review three of the most widely used correlated-noise estimators in the exoplanet field, the time-averaging, residual-permutation, and wavelet-likelihood methods. We argue that the residual-permutation method is unsound in estimating the uncertainty of parameter estimates. We thus recommend to refrain from this method altogether. We characterize the behavior of the time averaging’s rms-vs.-bin-size curves at bin sizes similar to the total observation duration, which may lead to underestimated uncertainties. For the wavelet-likelihood method, we note errors in the published equations and provide a list of corrections. We further assess the performance of these techniques by injecting and retrieving eclipse signals into synthetic and real Spitzer light curves, analyzing the results in terms of the relative-accuracy and coverage-fraction statistics. Both the time-averaging and wavelet-likelihood methods significantly improve the estimate of the eclipse depth over a white-noise analysis (a Markov-chain Monte Carlo exploration assuming uncorrelated noise). However, the corrections are not perfect, when retrieving the eclipse depth from Spitzer datasets, these methods covered the true (injected) depth within the 68% credible region in only ${\ifmmode\oldsim\else{$\oldsim$}\fi}$45–65% of the trials. Lastly, we present our open-source model-fitting tool, Multi-Core Markov-Chain Monte Carlo ([MC$^3$]{}). This package uses Bayesian statistics to estimate the best-fitting values and the credible regions for the parameters for a (user-provided) model. [MC$^3$]{} is a Python/C code, available at <https://github.com/pcubillos/MCcubed>.'
author:
- 'Patricio Cubillos, Joseph Harrington, Thomas J. Loredo, Nate B. Lust, Jasmina Blecic, and Madison Stemm'
title: 'On Correlated-Noise Analyses Applied To Exoplanet Light Curves'
---
INTRODUCTION {#sec:introduction}
============
Whether one’s goal is the detection or the characterization of exoplanets through transit or eclipse observations, the large contrast between the stellar and planetary emission (e.g., about a thousand times in the infrared for a hot Jupiter around an FGK dwarf star) make the data analysis an intrinsically challenging task. For example, for the Spitzer Space Telescope, most planetary signals [e.g., @StevensonEtal2010natGJ436b; @DemoryEtal2012apjl55cnceEclipse] lie below the instrument’s design criteria for photometric stability [@FazioEtal2004apjsIRAC]. Extracting planetary signals at this precision requires meticulous data reduction. Despite our best attempts to account for all known systematics, time-correlated residuals (or red noise) between the data and models often remain. These systematics may originate from instrumental or astrophysical sources, for example: stellar flux variations from flares or granulation; imperfect flat fielding; or telluric variations from changing weather conditions, differential extinction, or imperfect telescope systematics corrections from changing telescope pointing. Many authors have acknowledged correlated noise as an important source of noise in time-series datasets [e.g., @PontEtal2006mnrasRednoise; @WinnEtal2007ajHATP1b; @AgolEtal2010apjHD189; @CubillosEtal2013apjWASP8b].
Correlated noise affects both the accuracy and the precision of estimates of the model parameters. The typical statistical analyses neglect the correlation between data points (e.g., likelihood functions such as $\chi\sp{2}$, based on uncorrelated noise). Hence, their estimated best-fitting values may be biased, whereas their credible regions (Appendix \[sec:CredRegion\]) can be incorrect. This paper reports our study of three common correlated-noise estimators found in the exoplanet literature; specifically, in analyses of transit and eclipse exoplanet light curves. First, the time-averaging method [@PontEtal2006mnrasRednoise; @WinnEtal2007ajHATP1b] compares the standard deviation of the data to the (expected) uncorrelated-noise standard deviation, scaling the uncertainties accordingly. Next, the residual-permutation (or “prayer bead”) method [@BouchyEtal2005CorrNoise] uses a data-shifting (bootstrap-like) algorithm that preserves the structure of the residuals. Lastly, the method of @CarterWinn2009apjWavelets calculates the likelihood function in a wavelet basis, where the correlation between the wavelet coefficients is negligible. Qualitatively speaking, these methods do return larger parameter uncertainties for stronger correlated noise. However, besides @CarterWinn2009apjWavelets, there are few efforts to validate their quantitative accuracy. We have implemented these methods, testing them with real and synthetic exoplanet eclipse data.
Although recently there has been development of additional sophisticated methods to model exoplanet light curves [e.g., @GibsonEtal2012mnrasGaussProc; @JordanEtal2013apjWASP6bPCA; @Morello2015apjPixelICA; @EvansEtal2015mnrasHD209bGaussianProcesses], we explicitly exclude them from this study because the large number of runs and dataset sizes would require unfeasible amounts of computing power. For example, Gaussian processes can become computationally prohibitive for datasets larger than ${\ifmmode\oldsim\else{$\oldsim$}\fi}1000$ data points [@Gibson2014mnrasModelSelecGP].
With a focus on atmospheric characterization, we concentrate on estimating the eclipse depths from Spitzer exoplanet light curves, since they represent the largest and best-quality sample of exoplanet data beyond 2 [[[m]{}]{}]{}. Spitzer data are affected by two well-known systematics: time-varying sensitivity (ramp) and intra-pixel sensitivity variations [@KnutsonEtal2009apjHD149026bphase; @CharbonneauEtal2005apjTrES1]. Although several models have been proposed to correct for these systematics [@HarringtonEtal2007natHD149026b; @KnutsonEtal2008apjHD209; @BallardEtal2010paspIntraPixel; @AgolEtal2010apjHD189; @StevensonEtal2012apjHD149026b; @LewisEtal2013apjHAT2bPhase; @DemingEtal2014IntraPixelModel], the corrections are not always perfect, and thus many light-curve fits exhibit time-correlated residuals.
In Section \[sec:impact\], we discuss the impact of correlated noise on determining model-parameter uncertainties. In Section \[sec:estimators\], we review the most commonly used correlated-noise estimators in exoplanet analyses. In Section \[sec:simulations\], we test and compare the correlated-noise methods by retrieving synthetic eclipse curves that were injected into synthetic and real light-curve data. In Section \[sec:mc3\] we present our open-source package, Multi-Core Markov-Chain Monte Carlo (MC), to calculate the model-parameters’ credible regions. Finally, in Section \[sec:conclusions\] we present our conclusions.
THE IMPACT OF CORRELATED NOISE {#sec:impact}
==============================
A central ingredient in both frequentist and Bayesian parametric modeling is the sampling distribution for the data: the joint probability density function (PDF) for the data values, as a function of the model parameters. In our time series setting, we denote the data by ${\vec{y}}= (y_1,\ldots,y_n)$, with $y_i$ denoting the value of a measurement at time $t_i$. Similarly, we denote the model predictions by ${\vec{f}}= (f_1,\ldots,f_n)$, with $f_i = f(t_i;{\theta})$ for a model function with parameters ${\theta}$. Note that the predictions are functions of the parameters, $f_i({\theta})$, but we often suppress the parameter dependence for convenience. The sampling distribution is the $n$-dimensional joint PDF, $p({\vec{y}}|{\theta})$.
When ${\vec{y}}$ is fixed to an actually observed data vector, the sampling distribution as a function of the model parameters is called the likelihood function, ${\mathcal{L}}({\theta})$. Bayesian methods quantify uncertainty in the parameters via the dependence of ${\mathcal{L}}({\theta})$ on the parameters; Bayes’s theorem and the law of total probability convert this dependence into posterior probabilities for statements about the parameters. Frequentist methods quantify uncertainty by first defining statistics (functions of ${\vec{y}}$) that produce point estimates or intervals in the parameter space (perhaps using the ${\theta}$ dependence of the likelihood function), and then using the sampling distribution to quantify the variability of the statistics across ensembles of hypothetical data vectors. The variability in the sample space then is mapped into uncertainty quantifications in the parameter space (e.g., bias of a point estimate, or coverage of a confidence interval).
Commonly, the data are modeled as the sum of the predictions and independent, zero-mean, normally-distributed noise, $$y_i = f_i({\theta}) + {\epsilon}_i,
\label{y-fe}$$ with independent noise probabilities $$p({\epsilon}_i) = \frac{1}{\sigma_i\sqrt{2\pi}}\exp\left[-\frac{{\epsilon}_i^2}{2\sigma_i^2}\right],
\label{e-norm}$$ where $\sigma_i$ is the standard deviation for the noise contribution in measurement $i$. In this scenario, the sampling distribution factors, $$\begin{aligned}
p({\vec{y}}|{\theta})
& = \prod_i p(y_i|{\theta}) \nonumber\\
& =\frac{1}{{(2\pi)}^{n/2}}
\left(\prod_i\sigma_i\right)^{-1}
\exp\left[-\frac{1}{2}\sum\sb{i} \frac{{\epsilon}\sb{i}({\theta})\sp{2}}{\sigma\sb{i}\sp{2}}\right],
\label{eq:samp-ind}\end{aligned}$$ where the sum inside the exponential is the familiar “$\chi^2$” sum of squared, standardized residuals. Parameter estimation based on maximizing Eq. (\[eq:samp-ind\]) with respect to $\theta$ is called *weighted least squares* (WLS) regression when $f_i(\theta)$ is linear with respect to the parameters. If in addition the $\sigma_i$ values are all the same (homoskedastic), the approach is called *ordinary least squares* (OLS).
In the general case, correlated Gaussian noise has a multivariate normal PDF with a non-diagonal precision matrix $K$, $$p({\epsilon}) = \frac{\vert K\vert^{1/2}}{(2\pi)^{n/2}}
\exp\left[-\frac{1}{2}\sum_{ij} {\epsilon}_i K_{ij} {\epsilon}_j\right],
\label{eq:corrPDF}$$ where $\vert K\vert$ denotes the determinant of the precision matrix. The derivation to the specific case of uncorrelated Gaussian noise is trivial. As a consequence of independence, $K$ becomes a diagonal matrix with components $$\begin{aligned}
K_{ij} = \frac{1}{\sigma_i\sigma_j} \delta_{ij}.\end{aligned}$$ Therefore, the sum in Eq. (\[eq:corrPDF\]) collapses to the single sum in Eq. (\[eq:samp-ind\]).
Parameter estimation based on maximizing this likelihood function with respect to $\theta$ is called *generalized least squares* (GLS) regression when $f_i(\theta)$ is linear with respect to the parameters. When the model has nonlinear dependence on any of the parameters, maximum likelihood estimate is the *nonlinear least squares* estimate, regardless of whether the errors are correlated or homoskedastic.
Methods currently used to handle correlated noise fall into two broad classes. Methods like the @CarterWinn2009apjWavelets wavelet approach estimate a correlation matrix, and produce GLS estimates. In contrast, the time-averaging and residual-permutation approaches rely on WLS for estimation, but devise rules for inflating uncertainties to account for the ignored correlations. It is well known that non-linear least squares estimates are statistically consistent; that is, asymptotically (as $n\rightarrow\infty$), estimates converge to the true parameter values [@Wu1981NonlinearLSE]. Roughly speaking, although correlation complicates the way information accumulates across samples, infinite sample size ameliorates the complications. However, since WLS does not account for correlations, the quality of estimates can be significantly compromised with finite sample size.
In Appendix \[sec:OLSGLSexamples\] we describe two simple example calculations comparing WLS and GLS that provide some insight into the costs of ignoring noise correlation. The first example treats estimation of the amplitude of a *constant signal* in the presence of autoregressive noise, a simple and analytically tractable example of correlated noise. When noise is independent, with standard deviation $s$, the uncertainty in an estimate of a constant signal level is $s/\sqrt{n}$, the familiar “root-$n$” law. The WLS estimate has just this behavior. In contrast, when the noise is known to be positively correlated between adjacent samples, the uncertainty in the GLS estimate decreases more slowly than $1/\sqrt{n}$. This kind of example motivates approaches like time averaging that attempt to account for correlation merely by inflating uncertainties.
The second example replaces the constant signal with an eclipse-like *dip signal*. The dip location and width are presumed known; the background level and dip depth are to be estimated. In this case, a simple simulation study shows that noise correlation does not merely inflate uncertainties. It can also corrupt parameter estimates, with WLS estimates potentially taking values far away from the optimal estimates that account for noise correlations. This occurs when parameters of interest pertain to temporally localized structure in the model, for which noise correlations can significantly change the data projections needed for accurate inference.
Together, these examples show that methods that seek to account for correlations only by inflating parameter uncertainties are at best suboptimal (producing larger estimation errors than could be achieved with a good correlated noise model), and can sometimes be significantly misleading.
COMPUTING PARAMETER UNCERTAINTIES {#sec:estimators}
=================================
Markov-chain Monte Carlo
------------------------
In the Bayesian framework, a credible region for the parameters of a model, $\mathcal{M}$, can be computed via the Markov-chain Monte Carlo (MCMC) algorithm. The MCMC method generates a large number of random samples from the parameter space with a probability density proportional to the posterior probability distribution: $$p(\theta|{\vec{y}}, \mathcal{M}) \propto p(\theta|\mathcal{M})
p({\vec{y}}|\theta, \mathcal{M}),$$ where $p(\theta|\mathcal{M})$ is the prior probability distribution. A marginal highest-posterior-density (HPD) credible region for each parameter is then obtained from the interval that contains a certain fraction of the highest posterior density (typically 68%, 95%, or 99%) of the marginalized posterior (see Appendix \[sec:CredRegion\]). For example, when the posterior follows a normal distribution, the 68.3% marginal credible interval corresponds to the interval contained within one standard deviation from the mean.
Inference based on the likelihood function of Eq. (\[eq:samp-ind\]) works well when the noise contributions are independent and normally distributed; however, it does not account for time-correlated noise. Alternatively, an inference that uses the full covariance matrix, as in of Eq. (\[eq:corrPDF\]), should account for correlated noise, although its calculation often becomes computationally prohibitive.
Time Averaging {#sec:timeavg}
--------------
@PontEtal2006mnrasRednoise developed a method to compute the uncertainty of a transit or eclipse-depth estimation using the light-curve data points themselves. They considered the noise as the sum in quadrature of two components, a purely white (uncorrelated) source (characterized by a standard deviation per data point $\sigma\sb{w}$), and a purely time-correlated source (characterized by $\sigma\sb{r}$). @PontEtal2006mnrasRednoise assumed the white-noise component to scale as $\sigma\sb{w}/\sqrt{n}$, with $n$ the number of data points in the transit; whereas the time-correlated standard deviation, $\sigma\sb{r}$, to be independent of the number of data points. Then for any given signal, the uncertainty of a measurement should scale as: $$\sigma\sb{d} = \sqrt{\frac{\sigma\sb{w}\sp{2}}{n} + \sigma\sb{r}\sp{2}}.
\label{eq:deptherror}$$
For small $n$, $\sigma\sb{d}$ may be dominated by $\sigma\sb{w}/\sqrt{n}$, whereas as $n$ increases, $\sigma\sb{d}$ approaches $\sigma\sb{r}$. The time-averaging method uses this fact to estimate the contribution from the correlated noise. Note, however, that this hypothesized behavior is not typical of stationary correlated noise models exhibiting long-range dependence, which instead have $\sigma\sb{d}$ decreasing at a rate slower than $1/\sqrt{n}$, but still monotonically decreasing to zero [@BeranEtal2013LongMemoryProcesses].
We implement the time-averaging procedure as described by @WinnEtal2007ajHATP1b. First, we calculate the residuals between the data points and the best-fitting model. Then, we group the residuals in time-ordered, non-overlapping bins of $N$ elements each, and calculate their mean values. Lastly, we calculate the standard deviation (or root mean squared, rms) of the binned residuals, rms$\sb{N}$. We repeat the process for a range of bin sizes from one to half the data size. The uncertainty of rms$\sb{N}$ is approximately $\sigma\sb{\rm rms}=$${\rm rms}\sb{N}/\sqrt{2M}$ (see Appendix \[sec:StdUncert\]).
Now, let $\sigma\sb{1}$ be the rms value of the non-binned residuals (presumed to be dominated by white noise). In the absence of correlated noise, the expected rms for the set of $M$ bins, each containing $N$ points, is given by the extrapolation of $\sigma\sb{1}$ [@WinnEtal2008apjXO3bRedNoise]: $$\sigma\sb{N} = \frac{\sigma\sb{1}}{\sqrt{N}} \sqrt{\frac{M}{M-1}}.
\label{eq:rmsvsbin}$$
![Binned residuals rms vs. bin size (black curve with gray error bars) of WASP-8b Spitzer eclipse at 3.6 [[[m]{}]{}]{} (PI J. Harrington, Program ID 60003, see [@CubillosEtal2013apjWASP8b]). The red curve corresponds to the expected rms for white noise (Equation \[eq:rmsvsbin\]). The saw-tooth look of the curve arises from the discreet change in $M$, which becomes more significant as $N$ increases. The vertical dashed lines mark the duration of ingress/egress (left) and eclipse (right). The gray vertical error bars denote the $1\sigma$ uncertainty of the rms residuals (${\rm rms}\sb{N}/\sqrt{2M}$).[]{data-label="fig:wasp8bRMS"}](WASP8b_RMS-ch1_2014-05-27.ps){width="\linewidth"}
The rms$\sb{N}$ and $\sigma\sb{N}$ curves are analogous to $\sigma\sb{d}$ and $\sigma\sb{w}$, respectively. The time-averaging correction inflates the data uncertainties multiplying them by the ratio $\beta\sb{N} = {\rm rms}\sb{N} / \sigma\sb{N}$ if $\beta\sb{N}$ is statistically larger than one (i.e., by more than $1\sigma\sb\beta = \sigma\sb{\rm rms}/\sigma\sb{N}$). Finally, one runs a white-noise MCMC analysis, i.e., with Eq. (\[eq:samp-ind\]), using the inflated data errors. One typically visualizes both curves in an rms vs. bin size plot (Figure \[fig:wasp8bRMS\]).
### Behavior at Large Bin Sizes {#sec:TAzerotest}
One has to be aware that the binned-rms uncertainty $\sigma\sb{\rm
rms}$ is an asymptotic approximation. In the large bin-size regime (equivalently, small $M$) that approximation is not justified. The reason is that the marginalized posterior distribution for ${\rm
rms}\sb{N}$ (Eq. \[eq:fSigma\]), which has the form of an inverse-gamma distribution, becomes increasingly skewed as $M$ decreases.
By comparing the 68%-credible-region error bars of the inverse-gamma formula with the asymptotic approximation, we find that the latter moderately overestimates the lower error bar by 5%–10% for $M<200$. For the upper error bar the asymptotic approximation underestimates the error by 5% to 450% between $M=200$ and $M=2$. If one does not consider the skewed error bars, it may seem that the rms curves deviate below the $\sigma\sb{N}$ curve at large bin sizes [e.g., @StevensonEtal2012apjHD149026b; @CubillosEtal2013apjWASP8b; @BlecicEtal2013apjWASP14b]. However, this deviation is not statistically significant when one computes the uncertainties with the correct posterior distribution.
Since the typical transit (or eclipse) observation does not last much longer than the duration of the transit itself (usually one–two hours of out-of-transit before and after), it is important to consider the asymmetric rms error bars to properly account for correlated noise. Furthermore, since the signal-to-noise ratio of rms$_N$ decreases proportionally to $\sqrt{M}$ as $N$ increases, one ideally wants the longest-possible observation duration to improve the constraint on $\beta\sb{N}$ at the desired timescale of the event.
Residual Permutation {#sec:prayer}
--------------------
Residual permutation (also called the prayer bead method) is inspired by nonparametric bootstrapping methods from frequentist statistics. Nonparametric bootstrap methods directly use the sampled data (typically via resampling) to generate a distribution that approximates the sampling distribution, $p({\vec{y}}| \theta^*)$, for the true parameter values, $\theta^*$. Nonparametric bootstrapping typically relies on independent resampling of the data or residuals (possibly re-scaled), with replacement [e.g., @DavisonHinkley1997BootstrapMethods; @RuppertMatteson2015StatisticsForFinancialEngineering].
The motivating idea of the residual-permutation approach is to shift the data while preserving the time ordering and, thus, preserving the correlation structure. While the structure is indeed preserved, the residual-permutation method does not resample with replacement, a crucial requrement for nonparametric bootstrapping to work, i.e., to produce independent replicated datasets (a resample of the entire time-series observation). When there is correlated noise, the shifted datasets do not correspond to an independent resampling from any distribution, and thus do not exhibit the variability necessary for correct uncertainty quantification (e.g., computing confidence levels or estimator bias).
In the exoplanet field, the residual-permutation technique has been repeatedly used to estimate parameter uncertainties. However, the name of the technique has been loosely used to describe similar, but not equivalent procedures over the past decade. @BouchyEtal2005CorrNoise, @GillonEtal2007aaGJ436bspitz, and @Southworth2008HomogeneousStudyI all describe different methods, when referring to residual permutations. Some authors reference @JenkinsEtal2002apjDetectionConfidence, who actually use a “segmented bootstrap”, applying the method for detection instead of parameter estimation. Furthermore, several authors have wrongly attributed the method to @MoutouEtal2004aaOgle132b. Thus, there is a visible lack of rigor in the use of this method.
Currently, the most widely-used version of residual permutation is the one described by @Southworth2008HomogeneousStudyI or @WinnEtal2008apjXO3bRedNoise. This implementation computes the residuals between the light curve and the best-fitting model, cyclically shifts the residuals (preserving the point-to-point structure and thus the “redness” of the noise) by a given number of data points, adds the residuals back to the model, and finds a new set of best-fitting parameters. Usually, either one repeats the shift–fit process for a large number of iterations with random shifts, or one sequentially shifts the residuals by one data point at a time, fitting all possible shifts. Each parameter uncertainty is then given by the respective standard deviation of the distribution of best-fitting values. As already noted, this does not correspond to a sound resampling procedure, thus we will not consider residual permutations for the subsequent analyses.
There is a significant literature on generalizing the independent and identically distributed nonparametric bootstrap idea to address time series problems with correlated noise; this is a topic of ongoing research. One widely used approach is the block bootstrap. Presuming the investigator knows or can estimate a longest scale for correlations, $\Delta t$, the data are divided into blocks of length greater than $\Delta t$, and bootstrap resampling is done by drawing blocks of data at random to build a replication. A particular block rigidly preserves the time ordering of a subset of the data; in replications, it will appear shifted in time by various amounts. This behavior resembles the behavior of the prayer bead method. But block resampling produces greater variability than shifting the entire data vector, and by sampling with replacement, it produces ensembles that approximate independent draws from a (dependent) sampling distribution. The “segmented bootstrap” devised by @JenkinsEtal2002apjDetectionConfidence for analysis of ground-based transit photometry is similar to the block bootstrap. The block bootstrap only works if the correlation scale is significantly shorter than the span of the data, which will often not be true for Spitzer exoplanet eclipse data, so we do not consider it further here. Further details about the block bootstrap and other methods for resampling dependent data may be found in @Lahiri2003.
Wavelet Analysis {#sec:wavelet}
----------------
@CarterWinn2009apjWavelets introduced to the exoplanet field a technique where the time-correlated noise is modeled using wavelet transforms [@DericheTewfik1993tspLikeFractalNoise; @Wornell1993ieeeWavelet; @WornellOppenheim1992ieeeFractalSignalEstimation; @WornellOppenheim1992ieeeWaveletFractalModulation]. This method projects the time series residuals into an orthonormal wavelet basis, where the off-diagonal terms of the covariance matrix become negligible, thus simplifying the likelihood function calculation. Furthermore, they assumed noise that has a power spectral density with frequency $f$, varying as $1/f\sp{\gamma}$. They parameterized the noise with three parameters, $\gamma$, $\sigma\sb{\omega}$, and $\sigma\sb{r}$, as described in Equations (41)–(43) of @CarterWinn2009apjWavelets.
A thorough review of wavelets is beyond the scope of this work; see @Mallat2008WaveletTour and @Wornell1996Wavelet for more comprehensive discussions. Briefly, a wavelet transform projects a time-series signal onto a basis of functions that are dilations and translations of a compact parent (“wavelet”) function. The resulting transform has two dimensions, scale and location (in time). The discrete wavelet transform (DWT) consists of the hierarchical application over $M$ dilation scales of an orthonormal wavelet transform on a discrete time-series signal. For a signal consisting of $N=N\sb{0}2\sp{M}$ uniformly-spaced samples (with $N\sb{0}$ integer), and a wavelet function with $2N\sb{0}$ coefficients, the DWT produces $N\sb{0}$ scaling coefficients and $N\sb{0}2\sp{m-1}$ wavelet coefficients at each scale $m$, totaling $N\sb{0}(2\sp{M}-1)$ wavelet coefficients.
@CarterWinn2009apjWavelets recommend the fourth-order Daubechies wavelet basis [@Daubechies1988wavelets] for modeling time-series correlated noise, which we adopt in the current work. This is a basis well localized in time and frequency, well suited for $1/f\sp{\gamma}$ noise [@Wornell1996Wavelet]. @CarterWinn2009apjWavelets found that correlations between the wavelet and scaling coefficients decays faster for the Daubechies basis than the Haar basis, producing negligible covariances. Another advantage is that since the Daubechies basis is well localized in time, it reduces artifacts arising from the assumption of a periodic boundary condition by the wavelet transform.
### Wavelet-based Likelihood {#sec:waveletlike}
The likelihood function in the wavelet analysis is calculated in the following way. Let $\epsilon(t)$ be the fitting residuals of a time-series signal. Considering $\epsilon(t)$ as the contribution of a time-correlated ($\gamma \neq 0$) and an uncorrelated ($\gamma = 0$) component: $$\epsilon(t) = \epsilon\sb{\gamma}(t) + \epsilon\sb{0}(t),$$ this method calculates the DWT of $\epsilon(t)$ to produce the wavelet, $r\sb{n}\sp{m}$, and scaling, $\bar{r}\sb{n}\sp{1}$, coefficients of the signal. The variances of these coefficients are computed, respectively, as: $$\begin{aligned}
\label{eq:varwavelet}
\sigma\sb{W}\sp{2} & = & \sigma\sb{r}\sp{2}2\sp{-\gamma m} +
\sigma\sb{\omega}\sp{2} \\
\label{eq:varscale}
\sigma\sb{S}\sp{2} & = & \sigma\sb{r}\sp{2}2\sp{-\gamma} g(\gamma) +
\sigma\sb{\omega}\sp{2},\end{aligned}$$ where $\sigma\sb{\omega}$ and $\sigma\sb{r}$ parameterize the standard deviation of the uncorrelated and the correlated-noise signals, respectively, and $g(\gamma)=1/(2^{1-\gamma}-1)$ for $\gamma\ne1$ [following derivations from, e.g., @FadiliBullmore2002; @Wornell1993ieeeWavelet] and $g(\gamma) =
1/2\ln2$ for $\gamma=1$ [@CarterWinn2009apjWavelets] (see Appendix \[sec:WaveletVariance\]). Therefore, the wavelet-based likelihood function is given by $$\begin{aligned}
\mathcal{L}({\bf x}, \sigma\sb{\omega}, \sigma\sb{r})
& = & \left\{ \prod\sb{m=1}\sp{M} \prod\sb{n=1}\sp{N\sb{0}2\sp{m-1}}
\frac{1}{\sqrt{2\pi\sigma\sb{W}\sp{2}}}
\exp\left[{-\frac{(r\sb{n}\sp{m})\sp{2}}{2\sigma\sb{W}\sp{2}}} \right] \right\} \times \nonumber \\
& & \left\{ \prod\sb{n=1}\sp{n\sb{0}}
\frac{1}{\sqrt{2\pi\sigma\sb{S}\sp{2}}}
\exp\left[{-\frac{(\bar{r}\sb{n}\sp{1})\sp{2}}{2\sigma\sb{S}\sp{2}}} \right] \right\}.
\label{eq:wavelike}\end{aligned}$$
Equation (\[eq:wavelike\]) allows one to fit a model, sample its parameter’s posterior distribution, and determine the credible regions, while taking into account the effects of time-correlated noise.
During our review and implementation of the wavelet-likelihood technique from @CarterWinn2009apjWavelets, we found a few oversights in their equations and code (available in the Astronomical Source Code Library, ASCL[^1]). See details in Appendix \[sec:errata\].
{width="\linewidth"}
CORRELATED-NOISE TESTS FOR EXOPLANET ECLIPSE DATA {#sec:simulations}
=================================================
To assess the performance of the correlated-noise estimators described in Section \[sec:estimators\], we carried out injection–retrieval eclipse simulations. We focus on estimating the secondary-eclipse depth in a light curve observation, creating synthetic light curves that resemble Spitzer InfraRed Array Camera (IRAC) observations in terms of the signal-to-noise ratio (S/N), known systematics, cadence, observation duration, and eclipse shape.
In our first experiment we test the estimators’ performances when the time-correlated noise is described by a stochastic wavelet signal with a $1/f$ power spectral density [similar to the experiment of @CarterWinn2009apjWavelets]. We test the case when the observation time span is similar to the eclipse-event duration (Section \[sec:synthetic\], typical of real Spitzer secondary eclipse observations) and for the hypothetical case when the time span lasted an order of magnitude longer than the eclipse event (Section \[sec:synthetic2\]). In a second experiment (Section \[sec:semisynth3\]) we test the estimators on a more realistic case by injecting a synthetic eclipse signal into Spitzer phase-curve datasets.
Synthetic-noise Simulation {#sec:synthetic}
--------------------------
In this simulation we generate synthetic light curves by combining a @MandelAgol2002ApJtransits eclipse model, a linear ramp model, and a signal with both correlated and uncorrelated noise. The light-curve parameters closely follow those of a Spitzer observation of the WASP-12 system (Table \[table:parameters\]). The signal consists of 1700 data points, with a cadence of ${\ifmmode\oldsim\else{$\oldsim$}\fi}$12 seconds between data points, spanning an orbital-phase range from 0.39 to 0.63, about twice the eclipse duration.
Parameter Value
------------------------------------- -----------------
Eclipse depth (counts) 98.1
Eclipse duration (phase) 0.1119
Eclipse mid point (phase) 0.5015
Eclipse ingress/egress time (phase) 0.013
Ramp slope (counts/phase) 0.006
System flux (counts) 25815
$\sigma\sb{\omega}$ (counts) 64.5
$\sigma\sb{r}$ (counts) 0, 230, and 459
: \[table:parameters\] Synthetic light curve parameters
We created three sets of 5000 light-curve realizations each. For each realization, we generate a zero-mean random normal distribution, which we add to the light curve as the uncorrelated noise signal. We adjust the variance of this signal ($\sigma\sb{\omega}\sp{2}$) to yield an eclipse-depth signal-to-noise ratio of 30. Additionally, we generate purely-correlated $1/f$ signals ($\sigma\sb{\omega}=0$) using a Gaussian random number generator to produce wavelet coefficients with variances given by Equations (\[eq:varwavelet\]) and (\[eq:varscale\]). Then, we apply the inverse DWT to transform the signal from the wavelet basis to the time domain. Following the notation of @CarterWinn2009apjWavelets, we denote by $\alpha$ the ratio between the rms of the uncorrelated and correlated noise signals.
We constructed the signals in each of the three sets to have a pure uncorrelated noise, a weak time-correlated signal, and a strong correlated signal ($\alpha=0.0$, $0.25$, and $0.5$, respectively). Figure \[fig:synthetic\] shows two synthetic light curves for $\alpha=0.25$ and $0.5$. Note that our designations of “weak” and “strong” are, to some extent, arbitrary. We selected these limits based on our experience and tests: for $\alpha \lesssim 0.20$, the time-correlated signal becomes negligible compared to the uncorrelated-noise signal, whereas values of $\alpha {\ifmmode\oldsim\else{$\oldsim$}\fi}0.5$ are on the level of what we have observed in some cases [e.g., WASP-8b, @CubillosEtal2013apjWASP8b].
For each realization, we compute the parameter posteriors using the methods described in Section \[sec:estimators\], excluding residual permutation, which we deem to be unsound. Our model-fitting routines only fix the eclipse ingress/egress-time parameter (usually poorly-constrained by eclipse data), leaving free the system flux, eclipse depth, eclipse midpoint, eclipse duration, and ramp slope. First, we carry out a “white analysis” (i.e., ignoring the time-correlation between data points) by using Equation (\[eq:samp-ind\]) to compute the model-parameter best-fitting values (using the Levenberg-Marquardt algorithm) and their posterior distributions (using a MCMC).
![[**Top:**]{} Normalized distribution of $\beta$ for the $\alpha=0.25$ set. The histograms represent $\beta$ measured at the ingress-time and eclipse-duration timescales, and at the maximum value of $\beta$. The vertical dashed line marks $\beta=1$. The colored shaded areas denote the corresponding fraction of trials that were more than $1\sigma\sb{\beta}$ greater than one. The distributions for the other two sets ($\alpha=0$ and 0.5) were similar. [**Bottom:**]{} Normalized distribution of the bin sizes for $\beta\sb{\rm max}$. The vertical dashed lines indicate the ingress time and eclipse duration.[]{data-label="fig:beta"}](beta_hists_2016-09-16.ps){width="\linewidth"}
Next, we use the best-fit results to calculate the time-averaging rms-vs.-bin size curves. We retrieve the $\beta$ factor at three timescales: at the ingress time, at the eclipse duration, and at the time of maximum $\beta$ ($\beta\sb{\rm max}$, Figure \[fig:beta\]). In accordance with the discussion in Section \[sec:TAzerotest\], most $\beta$ values at the eclipse-duration timescale (similar to the total observation duration) were not significant. Thus, we adopted $\beta\sb{\rm max}$ as the scaling factor to calculate the time-averaging method uncertainties. Finally, we apply the wavelet-based likelihood method in an MCMC guided by Equation (\[eq:wavelike\]), simultaneously fitting the noise parameters ($\sigma\sb{\omega}$ and $\sigma\sb{r}$) and the model parameters, while keeping $\gamma$ fixed at 1. We find that a non-informative logarithmic prior on $\sigma\sb{r}$ handle the case with no correlated noise better than a flat prior. A log-flat prior is a scale-invariant prior that has an equal probability per order of magnitude. This is a more convenient prior when the parameter may range over several orders of magnitude [@Gregory2005BayesianBook]. The only requirement is that the parameter value must be positive.
### Results {#sec:simresults}
To assess the quality of the inferences from the time averaging and wavelet likelihood methods, we performed calibration tests, i.e., tests of the repeated-sampling (frequentist) performance of the inferences, when applied across an ensemble of simulated datasets.
The first test computes a measure of the *relative* accuracy of the eclipse depth estimates (i.e., accuracy relative to the reported uncertainty), also known as “number-of-sigma” statistic as described by @CarterWinn2009apjWavelets. The simulated datasets are large enough that the marginal posterior PDFs for the eclipse depth are typically nearly normal. This motivates measuring relative accuracy by the number of posterior standard deviations between the best-fit value and the true value used for a simulation, $$\begin{aligned}
{\cal N}\sb{p} = \frac{\hat p - p}{\sigma\sb{p}},
\label{eq:sigmastats}\end{aligned}$$ where $p$ is the true eclipse depth used for the simulation, $\hat p$ is the best-fit estimate, and $\sigma_p$ is the standard deviation of the marginal posterior for $p$. If the marginal posterior were normal, and if $\sigma_p$ were constant across the simulations, we would expect ${\cal N}_p$ to have normal standard deviation with mean $\langle{\cal N}_p\rangle$ zero, and standard deviation $\sigma\sb{\cal N}$ unity. In principle, a departure of $\langle{\cal
N}_p\rangle$ from zero would suggest a lack of accuracy of an analysis, whereas a departure of $\sigma\sb{\cal N}$ from unity would suggest an under- or overestimated precision ($\sigma\sb{\cal N}>1$ or $\sigma\sb{\cal N}<1$, respectively). In general, however, neither condition rigorously holds; the posteriors are slightly non-normal, and the value of $\sigma_p$ varies a bit. As a result, we do not expect ${\cal N}_p$ to have a standard normal distribution exactly (even for many simulations). Nevertheless, the mean and shape of the ${\cal N}_p$ distribution can reveal significant calibration failures of inferences.
The second test examines the conditional coverage of marginal credible regions for the eclipse depth. For a set of simulations with parameters $\theta$, we compute $C_Q(\theta)$, the coverage of marginal credible regions for the eclipse depth that were computed to contain a fraction $Q$ of the posterior probability. That is, we fix a size for the eclipse depth credible region to be tested (say, $Q=0.683$ for a conventional “1$\sigma$” region), and we compute the fraction of times the credible region contains the true eclipse depth value.
Some caution is required in interpreting results of conditional coverage tests. Bayesian credible regions will not in general be perfectly calibrated for fixed $\theta$. Rather, Bayesian methods produce regions with exact *average* coverage, i.e., with $\langle C_Q(\theta)\rangle = Q$ if one averages over the prior. The bottom line of these considerations is that, for a set of simulations with a fixed set of true parameter values, we do not expect $C_Q(\theta)$ to equal $Q$ exactly. But large departures from $Q$ likely indicate problems with an inference procedure. It is possible, at least in principle, to more thoroughly verify calibration of Bayesian MCMC algorithms [e.g., to average over $\theta$, or to consider all possible sizes of credible regions; see @Cook2006jcgsBayesianValidation], but we focus on simpler conditional tests here. For the coverage calculations reported below, we used the 68.3% marginal highest posterior density credible region.
![Eclipse-depth histogram of the relative-accuracy statistic for synthetic-data simulations. The top, middle, and bottom panels show the results for the white-MCMC, wavelet-likelihood, and time-averaging methods, respectively. The color code denotes the sample (see legend). The background gray contour denotes a standard normal distribution for comparison. All histograms are normalized such that the integral of each curve adds to one.[]{data-label="fig:syntheticResults"}](N-hist_synthetic_2016-08-30.ps){width="\linewidth"}
Figure \[fig:syntheticResults\] shows the ${\cal N}\sb{\rm p}$ distributions for each method and dataset. Table \[table:results\] shows the mean and standard deviation of ${\cal N}\sb{\rm p}$ and coverage fraction $C\sb{Q=0.68}$ (i.e., the fraction of trials where the 68% HPD covers the true depth value). The white analysis of the uncorrelated-noise set serves as a control sample. As expected, the ${\cal N}_p$ distribution for this case shows a negligible deviation from zero, a standard deviation close to one, and a 68%-HPD coverage fraction of ${\ifmmode\oldsim\else{$\oldsim$}\fi}$68%. The correlated-noise runs reveal the failure of the white analysis to account for correlated noise; as the correlated-noise component increases, the mean and standard deviation of ${\cal N}_p$ increase (suggesting a decrease of accuracy and underestimated uncertainties), which is well correlated with the lower coverage fraction.
The time-averaging method seems to improve the precision for the correlated-noise runs with respect to those of the white analysis (less underestimated uncertainties), as shown by the smaller $\sigma\sb{\cal N}$. Accordingly, the coverage fractions closer to 68% indicate an improvement in the parameter estimation. Note that the time-averaging method does not affect the accuracy (with respect to the white analysis). Since all the data uncertainties are inflated by a common scaling factor, the best-fitting solution does not change. Therefore, the variation in $\langle{\cal N}_p\rangle$ is a consequence of the variation in the precision and any possible underlying correlation between precision and accuracy for the sample.
The wavelet method also seems to improve the parameter estimation of the correlated samples with respect to the white analysis. However, this time the method seems to overestimate the uncertainties, as shown by the coverage fractions greater than 68% and the values of $\sigma\sb{\cal N}$ smaller than one. The values of $\langle{\cal
N}_p\rangle$ lie at the same level as those of the white analysis.
Overall, both the time-averaging and wavelet methods improve the eclipse-depth estimation over a white analysis. For a sample size of 3000 trials, the coverage uncertainty is roughly 2% (from a root-$N$ estimate). Since we carried out only a conditional study, there is an additional error budget to consider. Only a substantial mismatch between coverage and credible region size would be evidence that there is a problem. The results of the wavelet analysis may be evidence of a real coverage mismatch, but it is not at a level that would be surprising for conditional [[*vs.*]{}]{} average coverage. If real, this emphasizes the challenges of retrieving reliable parameter estimates from light curves affected by correlated noise, considering that we generated the synthetic signal with wavelet function.
[@lccc]{} Estimation method & $\langle{\cal N}_p\rangle$ & $\sigma\sb{\cal N}$ & $C\sb{0.68}$\
\
${\alpha=0.00}$ & 0.048 & 1.008 & 0.67\
${\alpha=0.25}$ & 0.091 & 2.230 & 0.33\
${\alpha=0.50}$ & 0.420 & 3.826 & 0.21\
[**Time Averaging**]{}\
${\alpha=0.00}$ & 0.044 & 0.954 & 0.71\
${\alpha=0.25}$ & 0.062 & 1.418 & 0.52\
${\alpha=0.50}$ & 0.177 & 1.581 & 0.48\
[**Wavelet**]{}\
${\alpha=0.00}$ & 0.056 & 1.128 & 0.63\
${\alpha=0.25}$ & 0.111 & 0.897 & 0.74\
${\alpha=0.50}$ & 0.408 & 0.846 & 0.78\
Synthetic-noise for Long-duration Simulation {#sec:synthetic2}
--------------------------------------------
Here we describe tests of the time-averaging method for datasets long enough such that the eclipse duration lies at timescales where the asymptotic approximation is still valid. To do so, we replicate the previous simulation (synthetic transit, white noise, and 1/$f$ noise signals) for an observation lasting ${\ifmmode\oldsim\else{$\oldsim$}\fi}20$ times the eclipse duration (akin to a phase-curve observation). We generate the light curve with the same eclipse configuration and system flux as in Section \[sec:synthetic2\], keeping the cadence (17,000 data points total) and the value of $\sigma\sb{\omega}$ at 64.5 counts. To conserve the noise rms ratios at $\alpha=0.25$ and 0.5, we set $\sigma\sb{r}=774$ and 1549 counts, respectively.
In this case, we find that the time-averaging $\beta$ scaling factors accurately inflate the data uncertainties to account for the time-correlated noise (Fig. \[fig:longrun\]).
![Time-averaging eclipse-depth histogram of the relative-accuracy statistic for the long-duration (20 times the eclipse duration) synthetic-data simulations. The color code denotes the sample (see legend). The background gray contour denotes a standard normal distribution for comparison. All histograms are normalized such that the integral of each curve adds to one.[]{data-label="fig:longrun"}](N-hist_longrun_2016-06-13.ps){width="\linewidth"}
Simulation with Spitzer-IRAC Noise {#sec:semisynth3}
----------------------------------
In this section we describe tests of the correlated-noise estimators for real exoplanet signals from the Spitzer IRAC instrument, which is more compelling than the previous test with synthetic data. We select two published phase-curve observations that are affected by correlated noise, a 4.5 [[[m]{}]{}]{} HD 209458b [@ZellemEtal2014apjHD209Phase] and a 3.6 [[[m]{}]{}]{} WASP-14b [@WongEtal2015apjWASP14bSpitzerPhaseCurve] dataset. The strength of the correlated noise in these two datasets is markedly different. The WASP-14b not only presents higher levels of correlated noise (as reported by the time-averaging curves), but also presents sporadic short-duration flux anomalies (dips) along the observation. Thus, these two datasets allow us to test the correlated-noise estimators under true instrumental noise as detected by the telescope, under two different correlated-noise regimes. We specifically selected phase-curve observations to remove the astrophysical signals and trace the telescope systematics to the best of our knowledge.
We processed the Spitzer BCD data to obtain raw light curves using the Photometry for Orbits, Eclipses, and Transits (POET) pipeline [@StevensonEtal2010natGJ436b; @StevensonEtal2012apjHD149026b; @StevensonEtal2012apjGJ436c; @CampoEtal2011apjWASP12b; @NymeyerEtal2011apjWASP18b; @CubillosEtal2013apjWASP8b; @CubillosEtal2014apjTrES1]. The POET pipeline involves bad-pixel masking (sigma rejection), 2-dimensional Gaussian fitting to determine the target location, and interpolated aperture photometry to obtain raw light curves [for details see, e.g., @CubillosEtal2014apjTrES1]. We remove the first couple hours of observation to avoid the time-dependent systematic.
We model the light curves using @MandelAgol2002ApJtransits eclipse and transit models, a BLISS map model [to account for the intrapixel effect, @StevensonEtal2012apjHD149026b], and and a sinusoidal function [for the phase-curve variation, following @ZellemEtal2014apjHD209Phase]: $$F(t) = 1 + c\sb{0} + c\sb{1} \cos(2\pi t) + c\sb{2} \sin(2\pi t),$$ where $c\sb{0}$, $c\sb{1}$, and $c\sb{2}$ are the model fitting parameters, and $t$ is the time of the observation (measured in orbital phase). To avoid degeneracy with the other fitting parameters, we constrain $c\sb{0}$ by requiring $F(t\sb{0}) = 1$, with $t\sb{0}$ the eclipse midpoint time. We finally remove all astrophysical variations from the signal by dividing out the sinusoidal model and trimming the HD 209458b phase curves to the span between the eclipse and transit (38.9 h long) and the WASP-14b phase curve between the transit and eclipse (22.6 h long). The resulting light curves consist of flat curves containing only the intra-pixel systematic variation and noise. These curves are our baseline to create the synthetic transit observations.
To construct the trial samples, we inject an eclipse curve at random uniformly distributed times into the baseline, generating 3000 realizations for each dataset. We adopt eclipse parameters (duration, depth, ingress, and egress) similar to the observed values for each dataset [@ZellemEtal2014apjHD209Phase; @WongEtal2015apjWASP14bSpitzerPhaseCurve].
We analyze the data and outputs in the same manner as in Section \[sec:simresults\]. Our fitting model includes an eclipse and a BLISS-map model. In practical terms, we found that many times the MCMC for the wavelet method failed to converge or failed to fit well the entire light curve. This may be result of the wavelet noise model attempting to overfit the transit curve, or because the wavelet cannot appropriately model correlated-noise structure. Thus, for this analysis we trim the dataset to a window of 2.5 times the eclipse duration, centered at the input midpoint.
![Eclipse-depth histogram of the relative-accuracy statistic for real-noise simulations. The top and bottom panels show the results for the HD 209458b and WASP-14b datasets, respectively. The color code denotes the parameter-estimation method (see legend). The background gray contour denotes a standard normal distribution for comparison. All histograms are normalized such that the integral of each curve adds to one.[]{data-label="fig:HD209simulation"}](N-hist_real-data_2016-06-16.ps){width="\linewidth"}
[@lccc]{} Estimation method & $\langle{\cal N}_p\rangle$ & $\sigma\sb{\cal N}$ & $C\sb{0.68}$\
\
White MCMC & 0.726 & 2.772 & 0.20\
Time Averaging & 0.444 & 1.553 & 0.46\
Wavelet & 0.428 & 1.176 & 0.54\
[**3.6 [[[m]{}]{}]{} WASP-14b**]{}\
White MCMC & 0.720 & 5.901 & 0.15\
Time Averaging & 0.202 & 1.682 & 0.64\
Wavelet & 1.286 & 1.766 & 0.54\
Figure \[fig:HD209simulation\] shows the resulting ${\cal N}\sb{\rm
p}$ histograms for each method and dataset. The ${\cal N}\sb{\rm p}$ histograms for the WASP-14b dataset are noticeably more irregular (non-Gaussian) than the histograms for the HD 2019458b dataset. This may be result of the stronger correlated-noise systematics. This clearly complicates the interpretation of the ${\cal N}\sb{\rm p}$ statistics. Thus, we rely mostly on the coverage-fraction statistic, which is not affected by these nuances.
Table \[table:HD209results\] presents the ${\cal N}\sb{\rm p}$ statistic and coverage-fraction results. Again, in both cases, both the time-averaging and wavelet methods improve the parameter estimation with respect to the white analysis (coverage fractions closer to 68%). However, none of the analyses completely correct the eclipse depth estimation. The low values of the coverage fractions hint towards an underestimation of the uncertainties, a lack of accuracy, or a combination of both. The irregular shape of the ${\cal
N}\sb{\rm p}$ histograms, particularly for the WASP-14b dataset, suggests that there are correlations between the accuracy and precision for the trials, which would be expected given the stronger correlated-noise component in the data.
MULTI-CORE MARKOV-CHAIN MONTE CARLO (MC) CODE {#sec:mc3}
=============================================
We implemented and made available all of the discussed statistical methods into the open-source Python package Multi-Core Markov-Chain Monte Carlo ([[MC]{}, https://github.com/pcubillos/MCcubed](https://github.com/pcubillos/MCcubed)). Unlike other exoplanet model-fitting tools that are tailored to specific tasks, [[MC]{}]{} allows the user to define the modeling function and, thus, it is a general-purpose statistical package. We developed the main bulk of the code in Python, with several extensions written in C, combining simplicity and high performance. The code runs in multiple parallel processors (through the built-in `multiprocessing` Python package). [[MC]{}]{} provides statistically-robust model optimization (via Levenberg-Marquardt minimization) and credible-region estimation (via MCMC sampling) routines.
The MCMC random sampling is done via the Metropolis Random Walk (MRW, using multivariate Gaussian proposals), the Differential-Evolution Markov-chain Monte Carlo [DEMC, @Braak2006DifferentialEvolution], or the Snooker-updater DEMC algorithms [@Braak2008SnookerDEMC]. While the proposal step sizes of the MRW are predetermined by the user and have to be manually adjusted before each run, the DEMC algorithms automatically adjust the scale and orientation of the proposal distribution. To do so, DEMC runs several chains in parallel, computing the proposed jump for a given chain from the difference between the parameter states of two other randomly selected chains. As the chains converge toward the posterior distribution, the proposal jumps will be mainly oriented along the desired distribution and will have adequate scales. Therefore, DEMC improves the MCMC efficiency in two ways: (1) it increases the acceptance rate to optimal levels [$\gtrsim
25$%, @RobertsEtal1997] by better sampling the parameter space, and (2) it eliminates the heuristic need for the user to adjust the proposal jump scales.
The Metropolis-Hastings acceptance rule implements both the ordinary likelihood function (Eq. \[eq:samp-ind\]) and the wavelet-based likelihood (Eq. \[eq:wavelike\]) using the fourth-order Daubechies wavelet. The priors can be bounded or unbounded uniform, log-scale uniform, or Gaussian. To assess that the MCMC is working properly, the code performs a chain-converge test using the @GelmanRubin1992 statistics. The code also produces several plots to help visualize the results: trace, rms-vs.-bin-size, marginal-posterior, and pairwise-posterior plots can indicate non-convergence, multi-modal posteriors, parameter correlations, correlated noise, or incorrect priors. At the end of the MCMC run the code returns the sampled posterior distribution of the parameters, their best-fitting values, their 68% HPD credible region, and the acceptance rate of the MCMC. The majority of the routines of this module derive from our POET pipeline and, thus, have been thoroughly tested for years.
The core structure of [[MC]{}]{} consists of a central hub, which drives the MCMC exploration, and the workers, which evaluate the model for the given free parameters. The hub and the worker processes are connected through shared memory. [[MC]{}]{} assigns one CPU to each worker (i.e., one for each chain). Each cycle (iteration) of the MCMC comprises the following steps: (1) generate the proposal state (the set of free parameters) for each chain, (2) evaluate the model for the proposed state, and (3) compute the Metropolis ratio and accept/reject the proposal state for each chain.
The [[MC]{}]{} code runs from both the shell prompt and the Python interactive interpreter, and is available for Python 2.7 and Python 3. The user can configure the MCMC run either through a configuration file, command line arguments (prompt), and/or function arguments (Python interpreter). The minimum required inputs are the modeling function, the data being fitted, and starting estimate values for the free parameters. As optional arguments, the user can supply the data uncertainties, priors, and any extra arguments of the modeling function (in a manner much like the `scipy.optimize.leastq` routine). Additionally, the package allows the user to configure multiple features of the MCMC, e.g.: number of chains, number of iterations, burn-in length, thinning factor, etc. The repository of the code includes a user manual and guided examples.
CONCLUSIONS {#sec:conclusions}
===========
Time-correlated noise is an important source of uncertainty for faint signals such as exoplanet light curves. Unless all systematics of the data are well understood, the correlated noise must be taken into account to obtain a reliable parameter estimation. We have reviewed three of the most widely used methods to assess time-correlated noise in exoplanet time-series data: time averaging, residual permutation, and wavelet-based likelihood, expanding the limited literature of tests to assess the quantitative results of these techniques. We focused specifically on the case of Spitzer secondary-eclipse time-series data.
We characterized the behavior of the time-averaging $\beta$ correction factor at large bin sizes (the typical case for a transit observation). In this regime the assumed uncertainty of the rms curve is no longer valid, since the posterior adopts the form of a skewed inverse-gamma distribution. We also found the residual-permutation method unsound as a tool for quantifying uncertainty in parameter estimates, because it does not produce ensembles that mimic the behavior of independent draws from a probability distribution. The method is not supported by a consistent statistical basis. Finally, for the wavelet-likelihood method we detected and corrected errors in the published equations @CarterWinn2009apjWavelets and code (Appendix \[sec:errata\]).
To quantitatively test the performance of these methods, we carried out injection–retrieval simulations on synthetic eclipse light curves, creating a large number of trials for each simulation. We analyzed the results by (1) comparing the expectation and standard deviation of the relative accuracy against a normal distribution [following @CarterWinn2009apjWavelets] and (2) computing the 68% coverage fraction (the fraction of trials where the 68% credible region contained the injected parameter). A correct Bayesian calculation would guarantee matching of the credible region probability and the average coverage (over the parameter space). Note that our simulations used a single true parameter value. A precise test of average coverage would require a substantial amount of computing time. Thus, our conditional coverage tests would indicate a problem with a method only if the coverage differed substantially from the credible level.
In our first simulation, we tested the case when the time-correlated noise has a power spectral density of the form $1/f$. Both the time-averaging and the wavelet-likelihood methods improved the eclipse-depth estimations over a white MCMC analysis. In this simulation the wavelet analysis is expected to perform well, since the wavelet precisely assumes a noise component with a $1/f$ power spectral density. We found small differences between the estimated conditional coverage and the credible levels. These diffrences are consistent with expectations, given both the limited precision from the size of the simulations, and the conditional nature of the tests. We also note that the performance of the time-averaging correction can be further improved if the total observation time is much longer than the eclipse duration (as in a phase-curve observation). This arises from the lower signal to noise of the ${\rm rms}\sb{N}$ curve at large bin sizes.
In a further simulation we generated eclipse light-curve samples by injecting an eclipse signal into real Spitzer 3.6 [[m]{}]{} and 4.5 [[m]{}]{} IRAC time-series datasets, two sets with notoriously different correlated-noise signals. This experiment allowed us to assess the performance of the time-correlated estimators without assuming a specific shape of the time-correlated signal. Both the time-averaging and the wavelet-likelihood methods significantly improved the uncertainty estimations compared to a white MCMC analysis, raising the coverage fraction from 15%–20% to ${\ifmmode\oldsim\else{$\oldsim$}\fi}$45–65%. However, they are not perfect, as the coverage fractions are still lower than the expected 68%, suggesting a lack of accuracy and (or) underestimated uncertainties.
In conclusion, it is always better to try to determine the best possible model for the systematics than simply inflating the parameter uncertainties (as in time averaging). However, these sub-optimal noise estimators are better than ignoring time-correlated noise. Luckily, the continuous development of advanced data analysis techniques like Gaussian Processes, Independent Component Analysis, or kernel regression decorrelation [see e.g., @IngallsEtal2016ajRepeatability] will help to improve the best practices required to extract exoplanet data.
Lastly, we presented the open-source Python package Multi-Core Markov-Chain Monte Carlo, available at <https://github.com/pcubillos/MCcubed>. [[MC]{}]{} implements all of the statistical routines described in this paper, allowing the user to estimate best-fitting model parameters and their credible region, while letting the user provide the modeling function. By releasing our code to the community, we hope not only to provide access to the routines discussed here, but also to encourage researchers to consider open development and cross-validation of the software tools used in the field.
We thank Rebekah Dawson for useful conversations. We thank the referee for comments that significantly improved the quality of the paper. We thank contributors to AstroPy [@Astropy2013aaAstroPy], SciPy, Matplotlib [@Hunter2007ieeeMatplotlib], the Python Programming Language, and the free and open-source community. PC was partly supported by the Fulbright Program for Foreign Students. JB was partly supported by the NASA Earth and Space Science Fellowship Program, grant NNX12AL83H. Part of this work is based on observations made with the Spitzer Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. Support for this work was provided by NASA through an award issued by JPL/Caltech and through the NASA Science Mission Directorate’s Astrophysics Data Analysis Program, grant NNX13AF38G, and its Planetary Atmospheres Program, grant NNX12AI69G.
[59]{} natexlab\#1[\#1]{}
, E., [Cowan]{}, N. B., [Knutson]{}, H. A., [Deming]{}, D., [Steffen]{}, J. H., [Henry]{}, G. W., & [Charbonneau]{}, D. 2010, , 721, 1861, [[ADS](http://adsabs.harvard.edu/abs/2010ApJ...721.1861A)]{}, [[1007.4378](http://arxiv.org/abs/1007.4378)]{}
2013, , 558, A33, [[h]({}{ADS})ttp://adsabs.harvard.edu/abs/2013A%26A...558A..33A]{}, [[1307.6212](http://arxiv.org/abs/1307.6212)]{}
, S. [et al.]{} 2010, , 122, 1341, [[ADS](http://adsabs.harvard.edu/abs/2010PASP..122.1341B)]{}, [[1009.0755](http://arxiv.org/abs/1009.0755)]{}
, J., [Feng]{}, Y., [Ghosh]{}, S., & [Kulik]{}, R. 2013, Long-memory processes (Springer, Heidelberg), xviii+884, [[ISBN: 978-3-642-35511-0; 978-3-642-35512-7](http://cosmologist.info/ISBN/978-3-642-35511-0;
978-3-642-35512-7)]{}, probabilistic properties and statistical methods
, J. [et al.]{} 2013, , 779, 5, [[ADS](http://adsabs.harvard.edu/abs/2013ApJ...779....5B)]{}, [[1111.2363](http://arxiv.org/abs/1111.2363)]{}
, F., [Pont]{}, F., [Melo]{}, C., [Santos]{}, N. C., [Mayor]{}, M., [Queloz]{}, D., & [Udry]{}, S. 2005, , 431, 1105, [[h]({}{ADS})ttp://adsabs.harvard.edu/abs/2005A%26A...431.1105B]{}, [[arXiv:astro-ph/0410346](http://arxiv.org/abs/arXiv:astro-ph/0410346)]{}
, C. J. [et al.]{} 2011, , 727, 125, [[ADS](http://adsabs.harvard.edu/abs/2011ApJ...727..125C)]{}, [[arXiv:1003.2763](http://arxiv.org/abs/arXiv:1003.2763)]{}
, J. A., & [Winn]{}, J. N. 2009, , 704, 51, [[ADS](http://adsabs.harvard.edu/abs/2009ApJ...704...51C)]{}, [[0909.0747](http://arxiv.org/abs/0909.0747)]{}
, D. [et al.]{} 2005, , 626, 523, [[ADS](http://adsabs.harvard.edu/abs/2005ApJ...626..523C)]{}, [[arXiv:astro-ph/0503457](http://arxiv.org/abs/arXiv:astro-ph/0503457)]{}
, S. R., [Gelman]{}, A., & [Rubin]{}, D. B. 2006, Journal of Computational and Graphical Statistics, 15, 675
, P., [Harrington]{}, J., [Madhusudhan]{}, N., [Foster]{}, A. S. D., [Lust]{}, N. B., [Hardy]{}, R. A., & [Bowman]{}, M. O. 2014, , 797, 42, [[ADS](http://adsabs.harvard.edu/abs/2014ApJ...797...42C)]{}, [[1411.3093](http://arxiv.org/abs/1411.3093)]{}
, P. [et al.]{} 2013, , 768, 42, [[ADS](http://adsabs.harvard.edu/abs/2013ApJ...768...42C)]{}, [[1303.5468](http://arxiv.org/abs/1303.5468)]{}
Daubechies, I. 1988, Communications on Pure and Applied Mathematics, 41, 909
, A. C., & [Hinkley]{}, D. V. 1997, Cambridge Series in Statistical and Probabilistic Mathematics, Vol. 1, Bootstrap methods and their application (Cambridge University Press, Cambridge), x+582, [[ISBN: 0-521-57391-2](http://cosmologist.info/ISBN/0-521-57391-2)]{}, with 1 IBM-PC floppy disk (3.5 inch; HD)
, D. [et al.]{} 2014, In prep.
, B.-O., [Gillon]{}, M., [Seager]{}, S., [Benneke]{}, B., [Deming]{}, D., & [Jackson]{}, B. 2012, , 751, L28, [[ADS](http://adsabs.harvard.edu/abs/2012ApJ...751L..28D)]{}, [[1205.1766](http://arxiv.org/abs/1205.1766)]{}
Deriche, M., & Tewfik, A. 1993, Trans. Sig. Proc., 41, 2977
, T. M., [Aigrain]{}, S., [Gibson]{}, N., [Barstow]{}, J. K., [Amundsen]{}, D. S., [Tremblin]{}, P., & [Mourier]{}, P. 2015, , 451, 680, [[ADS](http://adsabs.harvard.edu/abs/2015MNRAS.451..680E)]{}, [[1504.05942](http://arxiv.org/abs/1504.05942)]{}
Fadili, M., & Bullmore, E. 2002, NeuroImage, 15, 217
, G. G. [et al.]{} 2004, , 154, 10, [[ADS](http://adsabs.harvard.edu/cgi-bin/nph-bib_query?bibcode=2004ApJS..154...10F&db_key=AST)]{}
, A., & [Rubin]{}, D. B. 1992, Statistical Science, 7, 457
, N. P. 2014, , 445, 3401, [[ADS](http://adsabs.harvard.edu/abs/2014MNRAS.445.3401G)]{}, [[1409.5668](http://arxiv.org/abs/1409.5668)]{}
, N. P., [Aigrain]{}, S., [Roberts]{}, S., [Evans]{}, T. M., [Osborne]{}, M., & [Pont]{}, F. 2012, , 419, 2683, [[ADS](http://adsabs.harvard.edu/abs/2012MNRAS.419.2683G)]{}, [[1109.3251](http://arxiv.org/abs/1109.3251)]{}
, M. [et al.]{} 2007, , 471, L51, [[h]({}{ADS})ttp://adsabs.harvard.edu/abs/2007A%26A...471L..51G]{}, [[arXiv:0707.2261](http://arxiv.org/abs/arXiv:0707.2261)]{}
, P. 2005, [Bayesian]{} Logical Data Analysis for the Physical Sciences (New York, NY, USA: Cambridge University Press), [[ISBN: 052184150X](http://cosmologist.info/ISBN/052184150X)]{}
, J., [Luszcz]{}, S., [Seager]{}, S., [Deming]{}, D., & [Richardson]{}, L. J. 2007, , 447, 691, [[ADS](http://adsabs.harvard.edu/abs/2007Natur.447..691H)]{}
Hunter, J. D. 2007, Computing In Science & Engineering, 9, 90
, J. G. [et al.]{} 2016, ArXiv e-prints, [[ADS](http://adsabs.harvard.edu/abs/2016arXiv160105101I)]{}, [[1601.05101](http://arxiv.org/abs/1601.05101)]{}
, J. M., [Caldwell]{}, D. A., & [Borucki]{}, W. J. 2002, , 564, 495, [[ADS](http://adsabs.harvard.edu/abs/2002ApJ...564..495J)]{}
, A. [et al.]{} 2013, , 778, 184, [[ADS](http://adsabs.harvard.edu/abs/2013ApJ...778..184J)]{}, [[1310.6048](http://arxiv.org/abs/1310.6048)]{}
, H. A., [Charbonneau]{}, D., [Allen]{}, L. E., [Burrows]{}, A., & [Megeath]{}, S. T. 2008, , 673, 526, [[ADS](http://adsabs.harvard.edu/abs/2008ApJ...673..526K)]{}, [[arXiv:0709.3984](http://arxiv.org/abs/arXiv:0709.3984)]{}
, H. A., [Charbonneau]{}, D., [Cowan]{}, N. B., [Fortney]{}, J. J., [Showman]{}, A. P., [Agol]{}, E., & [Henry]{}, G. W. 2009, , 703, 769, [[ADS](http://adsabs.harvard.edu/abs/2009ApJ...703..769K)]{}, [[arXiv:0908.1977](http://arxiv.org/abs/arXiv:0908.1977)]{}
Lahiri, S. N. 2003, Resampling methods for dependent data, Springer Series in Statistics (Springer-Verlag, New York), xiv+374, [[ISBN: 0-387-00928-0](http://cosmologist.info/ISBN/0-387-00928-0)]{}
, N. K. [et al.]{} 2013, , 766, 95, [[ADS](http://adsabs.harvard.edu/abs/2013ApJ...766...95L)]{}, [[1302.5084](http://arxiv.org/abs/1302.5084)]{}
, S. 2008, A Wavelet Tour of Signal Processing, Third Edition: The Sparse Way, 3rd edn. (Academic Press), [[ISBN: 0123743702, 9780123743701](http://cosmologist.info/ISBN/0123743702, 9780123743701)]{}
, K., & [Agol]{}, E. 2002, , 580, L171, [[ADS](http://adsabs.harvard.edu/abs/2002ApJ...580L.171M)]{}, [[arXiv:astro-ph/0210099](http://arxiv.org/abs/arXiv:astro-ph/0210099)]{}
, G. 2015, , 808, 56, [[ADS](http://adsabs.harvard.edu/abs/2015ApJ...808...56M)]{}, [[1503.05309](http://arxiv.org/abs/1503.05309)]{}
, C., [Pont]{}, F., [Bouchy]{}, F., & [Mayor]{}, M. 2004, , 424, L31, [[h]({}{ADS})ttp://adsabs.harvard.edu/abs/2004A%26A...424L..31M]{}, [[arXiv:astro-ph/0407635](http://arxiv.org/abs/arXiv:astro-ph/0407635)]{}
, S. [et al.]{} 2011, , 742, 35, [[ADS](http://adsabs.harvard.edu/abs/2011ApJ...742...35N)]{}, [[1005.1017](http://arxiv.org/abs/1005.1017)]{}
, F., [Zucker]{}, S., & [Queloz]{}, D. 2006, , 373, 231, [[ADS](http://adsabs.harvard.edu/abs/2006MNRAS.373..231P)]{}, [[arXiv:astro-ph/0608597](http://arxiv.org/abs/arXiv:astro-ph/0608597)]{}
, G., [Gelman]{}, A., & [Gilks]{}, W. 1997, Annals of Applied Probability, 7, 110
, D., & [Matteson]{}, D. S. 2015, Statistics and data analysis for financial engineering—with [R]{} examples, 2nd edn., Springer Texts in Statistics (Springer, New York), xxvi+719, [[ISBN: 978-1-4939-2613-8; 978-1-4939-2614-5](http://cosmologist.info/ISBN/978-1-4939-2613-8;
978-1-4939-2614-5)]{}
, D., & [Skilling]{}, J. 2006, [Data Analysis: A Bayesian Tutorial]{}, 2nd edn. (Oxford University Press, USA), [[ISBN: 0198568320](http://cosmologist.info/ISBN/0198568320)]{}
, J. 2008, , 386, 1644, [[ADS](http://adsabs.harvard.edu/abs/2008MNRAS.386.1644S)]{}, [[0802.3764](http://arxiv.org/abs/0802.3764)]{}
, K. [et al.]{} 2010, , 464, 1161, [[ADS](http://adsabs.harvard.edu/abs/2010Natur.464.1161S)]{}, [[arXiv:1010.4591](http://arxiv.org/abs/arXiv:1010.4591)]{}
, K. B. [et al.]{} 2012, , 754, 136, [[ADS](http://adsabs.harvard.edu/abs/2012ApJ...754..136S)]{}, [[1108.2057](http://arxiv.org/abs/1108.2057)]{}
—. 2012, , 755, 9, [[ADS](http://adsabs.harvard.edu/abs/2012ApJ...755....9S)]{}, [[1207.4245](http://arxiv.org/abs/1207.4245)]{}
, C. 2006, Statistics and Computing, 16, 239
, C. J. F., & [Vrugt]{}, J. A. 2008, Statistics and Computing, 18, 435
, J. N. [et al.]{} 2007, , 134, 1707, [[ADS](http://adsabs.harvard.edu/abs/2007AJ....134.1707W)]{}, [[0707.1908](http://arxiv.org/abs/0707.1908)]{}
—. 2008, , 683, 1076, [[ADS](http://adsabs.harvard.edu/abs/2008ApJ...683.1076W)]{}, [[0804.4475](http://arxiv.org/abs/0804.4475)]{}
, I. [et al.]{} 2015, , 811, 122, [[ADS](http://adsabs.harvard.edu/abs/2015ApJ...811..122W)]{}, [[1505.03158](http://arxiv.org/abs/1505.03158)]{}
, G., & [Oppenheim]{}, A. 1992, Signal Processing, IEEE Transactions on, 40, 611
—. 1992, Information Theory, IEEE Transactions on, 38, 785
, G. W. 1993, Proceedings of the IEEE, 81, 1428
—. 1996, Signal Processing with Fractals: A Wavelet-based Approach (Upper Saddle River, NJ: Prentice Hall PTR), [[ISBN: 013120999X](http://cosmologist.info/ISBN/013120999X)]{}
, C.-F. 1981, The Annals of Statistics, 9, 501
, R. T. [et al.]{} 2014, , 790, 53, [[ADS](http://adsabs.harvard.edu/abs/2014ApJ...790...53Z)]{}, [[1405.5923](http://arxiv.org/abs/1405.5923)]{}
, A. 1971, An introduction to [Bayesian]{} inference in econometrics (New York: J. Wiley), [[ISBN: 978-0-471-98165-7](http://cosmologist.info/ISBN/978-0-471-98165-7)]{}
BAYESIAN CREDIBLE REGION {#sec:CredRegion}
========================
In the Bayesian context, given the posterior probability density, $p(\theta|{\vec{y}})$, of a parameter, $\theta$, given the dataset, ${\vec{y}}$, the highest posterior density region (or credible region), $R$, is defined by $$C = \int\sb{R} {\rm d}\theta\;p(\theta|{\vec{y}})$$ where C is the probability contained in the credible region. The region $R$ is selected such that the posterior probability of any point inside $R$ is larger than that of any point outside.
In practice, to calculate the credible region, one constructs a histogram of the sampled posterior distribution (normalized such that the sum equals one) and sorts the bins in descending order. Then one sequentially adds the values of $p$ until reaching $C$. The credible-region boundaries are given by the smallest and largest values of $\theta$ for the samples considered in the sum, if the region is contiguous.
WLS [[*vs.*]{}]{} GLS EXAMPLES {#sec:OLSGLSexamples}
==============================
To gain insight into the difference between [*weighted least squares*]{} estimates (WLS, those considering Eq. (\[eq:samp-ind\])) and [*generalized least squares*]{} estimates (GLS, those considering Eq. (\[eq:corrPDF\])), we consider examples with a simple correlated noise model: AR(1) autoregressive noise, for regularly sampled data. Our treatment adapts analyses by @Zellner1971BayesianInference and @SiviaSkilling2006Bayesian on related models. In this model, the conditional expectation (regression) of the noise for sample $i$ is proportional to the previous noise value; the actual value of the noise is the sum of this expectation and a new *innovation* contribution, $\nu_i$: $${\epsilon}_i = \phi {\epsilon}_{i-1} + \nu_i,
\label{eps-AR1}$$ where $\phi$ is the autoregression parameter. The innovations are independent, with zero-mean normal PDFs with standard deviation $s$. The overall model for $f({\theta})$ is a *hidden Markov model* (HMM): “Markov” indicating that the prediction for the noise at time $t_i$ depends only on the noise at the previous time, and not on the whole noise history; and “hidden” because ${\epsilon}_i$ is not directly observed (as it would be in a standard AR(1) model), rather, $y_i$ is observed, mixing uncertain model and noise contributions.
The AR(1) model enables recursive construction of the joint distribution for the noise. The model specifies independent normal PDFs for the $\nu_i$ terms, so the goal is to express the ${\epsilon}_i$ values entirely in terms of $\nu_i$ values. The probability for the first noise sample, ${\epsilon}_1$, is slightly complicated by the fact that it depends on innovations at times before there is data. However, ${\epsilon}_i$ is a linear sum of terms that are each zero-mean normal, so it must itself have a normal PDF, with variance given by the sum of the variances of its contributions. Writing ${\epsilon}_{i-1} = \nu_{i-1}
+ \phi_{i-1}$, and recursing, we find $${\epsilon}_1 = \sum_{j=0}^{\infty} \phi^j\nu_{1-j}.
\label{eps1-AR1}$$ The standard deviation of each term is $\phi^j s$, so the sum of the variances is $$\begin{aligned}
\sigma_{{\epsilon}}^2 = s^2\sum_{j=0}^\infty \phi^{2j} = \frac{s^2}{1 - \phi^2},
\label{sig-eps}\end{aligned}$$ provided that $|\phi| < 1$. The marginal PDF for ${\epsilon}_i$ at any time is a zero-mean normal with this variance; the noise time series is thus *stationary* (with the same marginal distribution at each time).
We can write the joint PDF for all noise values in terms of factors that condition on the previous history: $$p({\vec{\epsilon}}) = p({\epsilon}_1)\, p({\epsilon}_2|{\epsilon}_1)\, p({\epsilon}_3|{\epsilon}_{1:2}) \cdots p({\epsilon}_n|{\epsilon}_{1:n-1}),
\label{eps-fac}$$ with ${\epsilon}_{i:j} = ({\epsilon}_i,\ldots,{\epsilon}_j)$. Given the Markov property of the AR(1) model, the joint PDF simplifies to $$p({\vec{\epsilon}}) = p({\epsilon}_1) \prod_{i=2}^n p({\epsilon}_i|{\epsilon}_{i-1}).
\label{evec-AR1}$$ Equation implies that $p({\epsilon}_i|{\epsilon}_{i-1})$ is the probability that $\nu_i = {\epsilon}_i - \phi{\epsilon}_{i-1}$. The factors appearing in Eq. (\[evec-AR1\]) are thus $$p({\epsilon}_1) = \frac{1-\phi^2}{s\sqrt{2\pi}} e^{-{\epsilon}_1^2/2s^2}
\label{eps1-pdf}$$ and $$p({\epsilon}_i|{\epsilon}_{i-1}) = \frac{1}{s\sqrt{2\pi}}
\exp\left[-\frac{1}{2s^2}({\epsilon}_i - \phi{\epsilon}_{i-1})^2\right].
\label{deps-pdf}$$
The observation equation, Eq. (\[y-fe\]), indicates that the probability for the data, ${\vec{y}}$, is the probability that the noise values take on the values ${\epsilon}_i = y_i - f_i({\theta})$. Let $r_i({\theta}) \equiv
y_i - f_i({\theta})$ denote the residuals from adopting the model with parameters ${\theta}$. Then the PDF for the data can be written $$p({\vec{y}}|{\theta}) = \frac{(1-\phi^2)^{1/2}}{s^n (2\pi)^{n/2}} e^{-Q(\theta)/2s^2},
\label{AR1-pdf-Q}$$ with $$\begin{aligned}
Q({\theta})
&= (1-\phi^2) r_1^2 + \sum_{i=2}^n (r_i - \phi r_{i-1})^2 \nonumber\\
&= \sum_{i=1}^n r_i^2 + \phi^2 \sum_{i=2}^{n-1} r_i^2
- 2\phi \sum_{i=2}^n r_i r_{i-1}.
\label{Q-resid}\end{aligned}$$ The first term—the sum of squared residuals—is just the “$\chi^2$” term that appears in WLS (see Eq. (\[eq:samp-ind\])). When $\phi \ne 0$, AR(1) noise correlations introduce new contributions to $Q({\theta})$, including a term resembling a lag-1 autocorrelation. These terms correspond to changes in the model basis projections entailed by the correlations in a GLS analysis, versus a WLS analysis.
Example: Constant Signal
------------------------
The simple case of a constant signal model of unknown amplitude, $f(t;\mu) = \mu$, is analytically tractable and is illuminating. Substituting $r_i = y_i - \mu$ and minimizing $Q(\mu)$ leads to the GLS estimate $$\hat{\mu} = \frac{w_n \bar{y} + w_2 (y_1+y_n)/2}{w_n + w_2},
\label{mu-hat}$$ where $\bar{y}$ is the sample mean, $\bar{y} \equiv (1/n)\sum_i y_i$, and we have defined weights $w_n = n(1-\phi)$ and $w_2 = 2\phi$. When $\phi=0$ (independent noise), $\hat{\mu}$ is just the sample mean. Otherwise, $\hat{\mu}$ is a weighted average of the full sample mean, and the average of the first and last (i.e., the most widely separated) samples. As $\phi$ approaches unity (strongly positively correlated noise), GLS instructs us to just average the most widely separated samples. In contrast, the WLS estimate is always the full sample mean. The WLS and GLS estimates thus will differ, not just in the uncertainties they assign to the mean, but also in the actual values of the estimates.
$Q(\mu)$ is quadratic in $\mu$, so the likelihood function is a Gaussian function in $\mu$. The reciprocal of the second derivative of $Q(\mu)/s^2$ at $\hat{\mu}$ gives the squared standard deviation of this Gaussian, $$\sigma_\mu^2 = \frac{s^2}{n(1-\phi)^2 - 2\phi(1-\phi)}.
\label{sig-mu}$$ When $\phi=0$, we have $\sigma_\mu = s/\sqrt{n}$, the familiar “root-$n$” result. As $\phi$ approaches unity, the denominator decreases toward zero, and the uncertainty in $\mu$ grows. Roughly speaking, growing positive correlation decreases the effective sample size, inflating uncertainties. This motivates approaches like time averaging that attempt to account for correlation merely by inflating uncertainties. But such approaches do not account for the effect of correlations on the actual value of a finite-sample estimate.
Example: Constant Baseline with One Dip
---------------------------------------
The effect of correlations on parameter estimates depends on the extent to which the correlations may mimic or distort the projections of the data onto the model components. When a model has components that vary slowly with respect to the correlation scale, the main effect of correlations is to change the effective sample size. But when a model has temporally localized components, correlations can significantly affect, not just the uncertainty scale, but also the best-fit parameter values.
To illustrate this, we used simulated AR(1) noise and the GLS likelihood function to model data generated from a baseline signal of amplitude $a$, with a localized dip of depth $\delta$. We took the dip location and width to be known. For the illustration we report here, we simulated 51 observations with true parameter values ${\theta}= (a,\delta) = (0, 2)$, with the dip spanning 10 samples in the middle of the time series. The noise was generated with an innovation standard deviation $s=1$, and $\phi=0.8$, producing data with autocorrelation time scales ${\ifmmode\oldsim\else{$\oldsim$}\fi}5$. Figure \[fig:series\] displays examples of the simulated data and WLS and GLS best-fit function estimates. It is visually apparent that the WLS and GLS estimates sometimes differ.
![Eight sampled time series from the baseline-dip model with AR(1) noise. Curves of matching color connect the simulated data (solid), the GLS best-fit points (dashed), and the WLS best-fit points (dotted). The solid black curve shows the true (noiseless) function.[]{data-label="fig:series"}](phi_p8-bad-OLS-DipSeries.eps){width="0.89\linewidth"}
Figure \[fig:contours\] shows contours of the posterior PDFs for $(a,\delta)$ from a representative simulation where the WLS and GLS estimates differ; here the WLS best-fit estimate is just outside of the 98% GLS credible region. Even when the WLS and GLS best-fit estimates did not differ too dramatically in, the WLS likelihood function not only has an incorrect uncertainty scale (which one might hope to fix via inflation), but does not correctly capture the shape of the PDF (i.e., the correlation between $a$ and $\delta$ estimates).
![ Contours of the posterior PDF for $(a,\delta)$, from GLS (larger, thicker contours) and WLS (smaller, thinner contours) analyses. From inside to outside, the contours bound highest posterior density regions with 50%, 75%, 90%, 95%, and 99% of the posterior probability. Dots indicate the modes. Crosshairs indicate the true parameter values.[]{data-label="fig:contours"}](phi_p8-bad-OLS-DipContours.ps){width="\columnwidth"}
The main message of these examples is that noise correlation not only can inflate uncertainties; it can also corrupt parameter estimates, particularly when parameters of interest pertain to temporally localized structure in the model, for which noise correlations can significantly change the data projections needed for accurate inference. Methods that seek to account for correlations only by inflating parameter uncertainties are at best suboptimal (producing larger estimation errors than could be achieved with a good correlated noise model), and can sometimes be significantly misleading.
STANDARD-DEVIATION UNCERTAINTY {#sec:StdUncert}
==============================
The uncertainty of a parameter estimate in a problem with a fixed model dimension (number of parameters) and growing sample size typically decreases asymptotically at the [$\sqrt{M}$]{} rate. That is, for estimating a Gaussian mean from [$M$]{} samples with the standard deviation $\sigma$, which is known, the uncertainty is [$\sigma/\sqrt{M}$]{}. However, this result says nothing about the actual [*size*]{} of the uncertainty at any particular sample size. When $\sigma$ is unknown, it becomes the target of estimation, instead of (or in addition to) the mean. Here, we elaborate on the derivation of the uncertainty for the standard deviation of a Gaussian. The derivation uses the Laplace approximation for a normal standard deviation and its uncertainty, i.e., it finds a Gaussian distribution with a peak and curvature matching the marginal probability density function.
Given a normal distribution of values with unknown mean $\mu$ and standard deviation $\sigma$, let $b\sb{i}$ be the means for a sample of $M$ groups of samples (“bins”) drawn from this distribution. The sample mean, $\bar{b}$, and the sample variance, $s\sp{2}$, are defined as usual: $$\label{Aeq:meanstd}
\bar{b} = \frac{1}{M}\sum\sb{i} b\sb{i}, \qquad
s\sp{2} = \frac{1}{M}\sum\sb{i} (b\sb{i} - \bar{b})\sp{2}.$$
If the residual $r\sb{i} = b\sb{i} - \bar{b}$ and $r\sp{2} = \sum\sb{i} r\sb{i}\sp{2}$, then the sample variance becomes $s\sp{2} = r\sp{2}/M$.
The likelihood function for our normal distribution with ($\mu$, $\sigma$) is: $$\begin{aligned}
\label{Aeq:likelihood}
\cal{L}(\mu, \sigma) & = & \prod\sb{i} p\,(b\sb{i}|\mu, \sigma) \nonumber \\
& = & \prod\sb{i} \frac{1}{\sigma\sqrt{2\pi}}
\exp\left[-\frac{(b\sb{i}-\mu)\sp{2}}{2\sigma\sp{2}}\right],\end{aligned}$$
so the likelihood can be written is terms of $\bar{b}$ and $r$ as: $$\begin{aligned}
\label{Aeq:likelihood2}
\cal{L}(\mu,\sigma)
& = & \frac{1}{\sigma\sp{M} (2\pi)\sp{M/2}}
\exp\left(-\frac{r\sp{2}}{2\sigma\sp{2}} \right)
\exp\left(-\frac{M(\mu-\bar{b})\sp{2}}{2\sigma\sp{2}}\right).\end{aligned}$$
To estimate $\mu$ and $\sigma$, we will adopt a flat prior for $\mu$ and a log-flat prior for $\sigma$, corresponding to $p(\sigma) \propto
1/\sigma$. Then, the joint posterior probability $p(\mu,\sigma|D)$ for $\mu$ and $\sigma$, given the data $D$ is:
$$\label{eq:posterior1}
p(\mu,\sigma|D) \propto p(\sigma) \times \cal{L}(\mu,\sigma),$$
with $p(\sigma)$ the prior probability on $\sigma$.
$$\begin{aligned}
\label{eq:posterior2}
p(\mu,\sigma|D)
& \propto & \frac{1}{\sigma\sp{M+1}}
\exp\left(-\frac{r\sp{2}}{2\sigma\sp{2}}\right)
\exp\left(-\frac{M(\mu-\bar{b})\sp{2}}{2\sigma\sp{2}}\right).\end{aligned}$$
Calculate the marginal posterior density for $\sigma$ by integrating over $\mu$: $$\begin{aligned}
\label{eq:margsigma}
p(\sigma|D)
\propto \int \frac{{\rm d}\mu}{\sigma\sp{M+1}}
\exp\left( -\frac{r\sp{2}}{2\sigma\sp{2}} \right)
\exp\left( -\frac{M(\mu-\bar{b})\sp{2}}{2\sigma\sp{2}}\right)\end{aligned}$$
The $\mu$ dependence is in the last exponential factor, a Gaussian that integrates to $\sigma\sqrt{2\pi}$. We denote the result as $f(\sigma)$: $$\begin{aligned}
\label{eq:margsigma2}
\label{eq:fSigma}
p(\sigma|D) \propto \frac{1}{\sigma\sp{M}}
\exp\left( -\frac{r\sp{2}}{2\sigma\sp{2}} \right)
= f(\sigma).\end{aligned}$$
We estimate $\sigma$ with its mode, $\hat\sigma$, which maximizes $f(\sigma)$. The first derivative of $f(\sigma)$ is: $$f'(\sigma) = f(\sigma)\left(\frac{r\sp{2}}{\sigma\sp{3}} -
\frac{M}{\sigma} \right),
\label{eq:Tom15}$$ so that setting $f'(\hat\sigma) = 0$ gives $\hat\sigma = r/\sqrt{M} = s$, as one might expect.
For a simple estimate of the uncertainty, let’s consider a Gaussian approximation with mean at $\hat\sigma$. The curvature (second derivative) of $\sigma$ at $\hat\sigma$: $$\label{eq:Tom16}
f''(\sigma) =
f'(\sigma) \left(\frac{r\sp{2}}{\sigma\sp{3}} - \frac{M}{\sigma}\right) +
f(\sigma) \left(\frac{M}{\sigma\sp{2}} - \frac{3r\sp{2}}{\sigma\sp{4}}\right),$$ determines the standard deviation. When $f(x)$ is of the form of a normal distribution with mean $m$ and standard deviation $w$, it is easy to show that $f''(m) = -f(m)/w\sp{2}$. So, if $\delta$ is the standard deviation for $\sigma$, in the normal approximation, that matches the curvature at the peak, we have $\delta\sp{2} \approx - f(\hat\sigma)/f''(\hat\sigma)$. Evaluating Equation (\[eq:Tom16\]) at $\hat\sigma$, the first term vanishes (since $f'(\hat\sigma)=0$), and the remaining term gives an approximate standard deviation of: $$\begin{aligned}
\label{eq:Tom17}
\delta \approx \frac{r}{M\sqrt{2}} = \frac{s}{\sqrt{2M}}.
\label{delta-approx}\end{aligned}$$
So, the mean and standard deviation sum for $\sigma$ for large $M$ is: $$\begin{aligned}
\label{eq:Tom18}
\sigma = s {\ifmmode\oldpm\else{$\oldpm$}\fi}\frac{s}{\sqrt{2M}}.
\label{sigma-est}\end{aligned}$$
Wavelet-likelihood Errata {#sec:errata}
=========================
This section reports three erratas found both in the published article of @CarterWinn2009apjWavelets and its associated ASCL code.
First, in the Likelihood equation, Eqs. (32) and (41) of @CarterWinn2009apjWavelets, the index for the scale, $m$, should start from 1 instead of 2. In this case the ASCL code has the correct value.
Next, the variance of the scaling coefficient in the ASCL code for $\gamma=1$, equation (34) of the paper, is missing the factor $2\sp{-\gamma}=2\sp{-1}$. The corrected equation should read: $$\sigma\sb{S}\sp{2} = \frac{\sigma\sb{r}\sp{2}}{4\ln{2}} +
\sigma\sb{\omega}\sp{2}.$$
The expression for the scaling coefficient for $\gamma\ne1$ also seems to be wrong in the ASCL code. We can compute the variance of the wavelet coefficients following equation (37) of @Wornell1993ieeeWavelet: $$\langle \epsilon^{m}_{n}\epsilon^{m}_{n} \rangle = \frac{2^{-m}}{2\pi}\int^{\infty}_{-\infty}
\frac{\sigma^{2}_{x}}{|\omega|^\gamma}|\Psi(2^{-m}\omega)|^{2}{\rm d}\omega.$$ Assuming an ideal bandpass —i.e., eq. (3) of @Wornell1993ieeeWavelet— and with a change of variable, $u =
2^{-m}\omega$, we reproduce eq. (24) of @CarterWinn2009apjWavelets: $$\begin{aligned}
\langle \epsilon^{m}_{n}\epsilon^{m}_{n} \rangle & = &
2^{-\gamma m} \frac{1}{2\pi}\int^{\infty}_{-\infty}
\frac{\sigma^{2}_{x}}{|u|^\gamma}|\Psi(u)|^{2}{\rm d}u. \\
& = & 2^{-\gamma m} \frac{2}{2\pi}\int^{2\pi}_{\pi}
\frac{\sigma^{2}_{x}}{u^\gamma}{\rm d}u. \\
& = & 2^{-\gamma m} \frac{\sigma^{2}_{x}}{\pi^\gamma}
\frac{[2^{1-\gamma}-1]}{1-\gamma}
\equiv 2^{-\gamma m} \sigma_r^2, \end{aligned}$$ Analogously, the variance for the scaling coefficient assuming an ideal bandpass —i.e., eq. (10) of @Wornell1993ieeeWavelet: $$\begin{aligned}
\langle {\bar\epsilon}^{m}_{n}{\bar\epsilon}^{m}_{n} \rangle & = &
2^{-\gamma m} \frac{1}{2\pi}\int^{\infty}_{-\infty}
\frac{\sigma^{2}_{x}}{|\omega|^\gamma}|\Phi(\omega)|^{2}{\rm d}\omega\\
& = & 2^{-\gamma m} \frac{2}{2\pi}\int^{\pi}_{0}
\frac{\sigma^{2}_{x}}{\omega^\gamma}{\rm d}\omega \\
& = & 2^{-\gamma m} \frac{\sigma^{2}_{x}}{\pi^\gamma} \frac{1}{1-\gamma}
\equiv 2^{-\gamma m} \sigma^{2}_{r} \frac{1}{2^{1-\gamma}-1}. \label{eq:vars2}\end{aligned}$$
This indicates that $g(\gamma) = 1/(2^{1-\gamma}-1)$ for $\gamma\ne1$, the inverse of the value given in the ASCL code from @CarterWinn2009apjWavelets. The same result can be derived from equations (16) and (17) of @FadiliBullmore2002. These derivations of $\langle \epsilon^{m}_{n}\epsilon^{m}_{n} \rangle$ and $\langle {\bar\epsilon}^{m}_{n}{\bar\epsilon}^{m}_{n} \rangle$ are not valid for $\gamma=1$; in fact, Eq. (\[eq:vars2\]) diverges to $+\infty$ from the left and to $-\infty$ from the right as we approach $\gamma=1$. Then, how can one get to $g(\gamma=1)=1/2\ln2$?.
Lastly, Section 4.1 of @CarterWinn2009apjWavelets mentions that they used a dataset of 1024 elements, and that their DWT produced 1023 wavelet coefficients and 1 scaling coefficient (implying $N\sb{0}=1$). This is inconsistent with the wavelet used (a fourth-order Daubechies wavelet), for which $N\sb{0}=2$. This wavelet’s DWT returns 2 scaling coefficients and 1022 wavelet coefficients (for the given dataset). The ASCL code is also suited to perform a likelihood calculation assuming $N\sb{0}=1$, resulting in each likelihood term having an $m$ value offset by 1.
[^1]: <http://asterisk.apod.com/viewtopic.php?f=35&t=21675>
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A recurrent neural net is described that learns a set of patterns $\{{\mbox{\boldmath $\xi $}}^{\mu}\}$ in the presence of noise. The learning rule is of a Hebbian type, and, if noise would be absent during the learning process, the resulting final values of the weights $w_{ij}$ would correspond to the pseudo-inverse solution of the fixed point equation in question. For a non-vanishing noise parameter, an explicit expression for the expectation value of the weights is obtained. This result turns out to be unequal to the pseudo-inverse solution. Furthermore, the stability properties of the system are discussed.'
address: 'Institute for Theoretical Physics, University of Amsterdam, Valckenierstraat 65, 1018 XE Amsterdam, The Netherlands'
author:
- W A van Leeuwen and B Wemmenhove
title: 'Learning by a neural net in a noisy environment – The pseudo-inverse solution revisited'
---
Introduction and summary
========================
In principle, it is our purpose to study learning in a neural net as it occurs in nature. The theory of recurrent neural nets [@MRS95] provides us with a model of content-addressable memory as it might be realized, to some extent, in the brain. Learning, in such a model, corresponds to adjusting the synaptic matrix $w_{ij}$ in such a way that $p$ memorized patterns ${\mbox{\boldmath $\xi $}}^{\mu}$, $(\mu=1, \ldots, p)$, become fixed points of the neuron state dynamics. This can be achieved in a recurrent neural net by sequentially clamping its neurons to a well-defined and unique set of patterns, and adjusting the weights of the connections according to some Hebbian learning rule. However, in reality, a neural net cannot be clamped to a fixed set of ideal patterns. A more realistic assumption would be that the clamping of the net to a pattern always is more or less distorted. Consider, for instance, the visual system as a system in which the clamping is imposed by input from the retina. Since neurons are noisy objects, which once in a while fire spontaneously, an internal representation of a stimulus in the brain will hardly ever be identical to the representation corresponding to a previous stimulus.
We therefore introduce noise to the set of patterns, thus making the set less well-defined and less unique. A network state array of a net of $N$ neurons is denoted by $${\mbox{\boldmath $x $}} := (x_1, \ldots, x_N) \label{vecx}$$ where $x_i=1$ if neuron $i$ is active and $x_i=0$ if it is non-active. At every learning step $n$, ${\mbox{\boldmath $x $}}$ will be similar to one of the $p$ given patterns ${\mbox{\boldmath $\xi $}}^{1}, \ldots, {\mbox{\boldmath $\xi $}}^p$, but it has nonzero probability, for each neuron $i$, $(i=1, \ldots, N)$, of deviating from it. At each learning step $n$, synaptic connections $w_{ij}$ will adapt themselves, according to a Hebbian learning rule which is a function of the weights $w_{ik}$ and of the (binary) neuron states $x_k$, $(k=1, \ldots, N)$. For the class of learning rules that we will use in our model, the case of noiseless learning has been studied in detail ([@PGD85], [@DieOp], [@Hee99]). At every learning step $n$ a pattern ${\mbox{\boldmath $\xi $}}^{\mu}$, $(\mu=1, \ldots p)$ is chosen. If, for each $n$, we put ${\mbox{\boldmath $x $}} = {\mbox{\boldmath $\xi $}}^{\mu}$, the resulting weights $w_{ij}$ for $n\rightarrow \infty$ are known to coincide with the pseudo-inverse solution. The pseudo-inverse solution is a particular solution of the under-determined set of $pN$ equations $$\displaystyle{\sum_{j=1}^N} w_{ij} \xi_j^{\mu} -\theta_i
= \kappa(2\xi_i^{\mu}-1)
\label{PIequations}$$ for the $N(N-1)$ unknowns $w_{ij}$ $(i,j=1,\ldots , N; i \neq j)$, $p<N$. Here, $\kappa$ is a positive number, and $\theta_i$ a constant. It is easily verified that these equations guarantee that the so-called stability coefficients $$\gamma_i({\mbox{\boldmath $\xi $}}^{\mu}) = (2\xi_i^{\mu}-1)(\displaystyle{\sum_{j=1}^N}
w_{ij} \xi_j^{\mu} -\theta_i)
\label{stabcoef}$$ are positive for all patterns ${\mbox{\boldmath $\xi $}}^{\mu}$. In our description, however, the network state during learning is determined by a probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$, centered around the patterns ${\mbox{\boldmath $\xi $}}^{\mu}$. By means of the Master equation derived in section \[mastersec\], we will arrive at the following equation for the expectation value of the weights in the limit of $n\rightarrow \infty$: $$\frac{1}{p}\displaystyle{\sum_{\mu=1}^p \sum_{{\mbox{\boldmath $x $}} \in \Omega}}
p^{\mu}({\mbox{\boldmath $x $}})
[\kappa(2x_i -1) - \bigl(\displaystyle{\sum_{k=1}^N}
\langle w_{ik} \rangle_{\infty} x_k -\theta_i\bigr)]x_j = 0
\label{Ourequation}$$ where $\Omega$ is the collection of all possible $2^N$ arrays ${\mbox{\boldmath $x $}}$. In contrast to the equations (\[PIequations\]), this is a completely determined set of equations, of which the solution is essentially different from the pseudo-inverse solution of (\[PIequations\]). It will turn out to exist only if the variance of the probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$ is non-zero, *i.e.*, in the presence of noise. This is the subject of section \[solutionsec\].
In section \[intermediatesec\] we will study numerically the weights $\langle w_{ij} \rangle_n$ as a function of $n$.
In section \[stabilitysection\] we will show that, on the average, this solution does yield stability coefficients close to $\kappa$ if the ‘noise parameter’, $b$, which will be introduced in the probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$, is small enough, which shows that the solution found is stable indeed.
Finally, in section \[basinssection\], we study the size of the basins of attraction around our new solution. It is found that the sizes are larger than those around the pseudo-inverse solution, a result that is in perfect agreement with earlier observations [@GSW89; @WS90a; @WS90b; @WS93] that learning with noise enlarges the basins of attraction. In these earlier studies, however, no analytical expression for the average values of the weights has been given.
Derivation of the Master Equation for a linear learning rule {#mastersec}
============================================================
We consider a recurrent net of $N$ binary neurons. The strengths of the synaptic connection between the post-synaptic neuron $j$ and the pre-synaptic neuron $i$ will be denoted by $w_{ij}$. The neurons $i$ $(i=1,\ldots, N)$ can take the values $x_i=0$ or $x_i=1$, corresponding to the non-active and active state, respectively. It is useful to associate with each neuron $i$ a set $V_i$, defined as the collection of neuron indices $j$ with which neuron $i$ has an adaptable, *i.e.*, a non-zero, non-constant afferent synaptic connection. In other words, for all $j \in V_i$, there is an axon going from neuron $j$ to a dendrite of neuron $i$, and the corresponding weight $w_{ij}$ is adaptable in a learning process. The collection of neurons $j$ with which $i$ has no connection, or a non-changing synaptic connection, will be denoted as the complementary set, $V_i^C$. We suppose that the synaptic strengths $w_{ij}$ between the neurons are changed in steps according to a rule of the general form $$\begin{aligned}
w_{ij}' = w_{ij} + \Delta w_{ij} \qquad& j \in V_i \\
w_{ij}' = w_{ij} & j \in V_i^C
\label{leerregel}\end{aligned}$$ where $\Delta w_{ij}$ is a function of the states $x_k$ of all $N$ neurons of the net and all afferent synaptic weights $w_{ik}$ ($i$ fixed, $k=1,2,\ldots,N)$. In general, the functions $\Delta w_{ij}$ will be linear in all $x_k$, since $x_k^2 = x_k$ (recall that $x_k$ equals $0$ or $1$), but non-linear in the weights $w_{ik}$. In this article we suppose, however, that the $\Delta w_{ij}$ do depend linearly on the weights $w_{ik}$. Hence, in this article, $$\Delta w_{ij} = \Delta w_{ij} ({\mbox{\boldmath $x $}},{\mbox{\boldmath $w $}}_i)$$ is a linear function in all $x_k$ and all $w_{ik}$ ($k=1,2,\ldots N$). We abbreviated $${\mbox{\boldmath $w $}}_i := (w_{i1}, \ldots, w_{iN}) \label{vecwi}$$ It is unrealistic to describe a biological neural net as a deterministic system, since there are many unknown parameters that influence its development in time. We therefore choose a probabilistic description. We suppose that the neuron states $x_i$ are mutually independent stochastic variables, *i.e.*, the probability that neuron $i$ has value $x_i$ is given by a probability distribution $p_i(x_i)$ which is independent of $j$ $(j\neq i)$. Since the changes $\Delta w_{ij}$ of the weights $w_{ij}$ are functions of the stochastic variables $x_i$ $(i=1, 2, \ldots N)$, and a function of a stochastic variable is a stochastic variable, the changes $\Delta w_{ij}$, and, hence, the $w_{ij}$ themselves, are stochastic variables.
Now let $T_{ij}(w_{ij}'|w_{ij},\{ w_{ik} \}_{k\neq j})$ be the probability that, due to a learning step, a transition takes place from the value $w_{ij}$ to the value $w_{ij}' = w_{ij} + \Delta w_{ij}$, for a given set $\{ w_{ik} \}_{k\neq j}$. Then we have $$T_{ij}(w_{ij}'|w_{ij},\{w_{ik}\}_{k\neq j}) =
\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p({\mbox{\boldmath $x $}})
\delta(w_{ij}'-w_{ij}- \Delta w_{ij}({\mbox{\boldmath $x $}}, {\mbox{\boldmath $w $}}_i))
\label{transition}$$ where $\Omega$ is the collection of all $2^N$ possible states of the neural net $(x_i=0,1;i=1, \ldots, N)$ and $p({\mbox{\boldmath $x $}})$ is the probability of occurrence of the network state ${\mbox{\boldmath $x $}}$, which we suppose to be independent of the variables $w_{ij}$. \[The relation between $p({\mbox{\boldmath $x $}})$ and $p_i(x_i)$ is left unspecified at this stage of the reasoning; compare, however, (\[pxvoorp\]) and (\[pmufactor\]) below.\] We have $$\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p({\mbox{\boldmath $x $}}) = 1
\label{somp(x)}$$ The delta-function in (\[transition\]) guarantees that only transitions take place which obey the learning rule (\[leerregel\]). Using (\[somp(x)\]) we find from (\[transition\]), that $$\int T_{ij}(w_{ij}'|w_{ij},\{w_{ik}\}_{k\neq j})dw_{ij}' = 1
\label{normTij}$$ Let $P_{ij}(w_{ij},n)$ be the probability of occurrence of the variable $w_{ij}$ at a time step $n$ $(n=0,1,2,\ldots)$. Then $P_{ij}$ and $T_{ij}$ are related according to $$P_{ij}(w_{ij}, n+1) = \int \ldots \int T_{ij}(w_{ij}|\{w_{ik}'\})
\displaystyle{\prod_{k=1}^N}[P_{ik}(w_{ik}',n) dw_{ik}']
\label{Pij}$$ Demanding that the probability $P_{ij}$ is normalized initially, $$\int P_{ij}(w_{ij}, 0) dw_{ij} = 1
\label{normPij0}$$ we find from (\[Pij\]) and (\[normTij\]), by induction, that $$\int P_{ij}(w_{ij},n) dw_{ij} = 1
\label{normPijn}$$ for all $n$. From (\[normTij\]) and (\[Pij\]) it follows that $$\begin{aligned}
\fl P_{ij}(w_{ij},n+1) - P_{ij}(w_{ij}, n) =
\int \ldots \int [T_{ij}(w_{ij}|\{w_{ik}'\}) P_{ij}(w_{ij}',n) \nonumber \\
- T_{ij}(w_{ij}'|w_{ij},\{w_{ik}'\}_{k\neq j}) P_{ij}
(w_{ij},n)]\displaystyle{\prod_{k\neq j}} [P_{ik}(w_{ik}',n)dw_{ik}']dw_{ij}'
\nonumber \\ \qquad \qquad (i=1,\ldots, N; j \in V_i)
\label{master}\end{aligned}$$ which is the so-called Discrete Master Equation for the weights $w_{ij}$. It masters the evolution of the weights $w_{ij}$ as a function of $n$, and determines the values of the weights in the long run.
In order to obtain an expression for the expectation value of the weights after infinitely many learning steps, we first consider the expectation value at time step $n$: $$\langle w_{ij} \rangle_n := \int P_{ij}(w_{ij},n) w_{ij}dw_{ij} \qquad
(j\in V_i)
\label{expvalue1}$$ The latter expression yields, using the Master Equation (\[master\]), $$\begin{aligned}
\fl \langle w_{ij} \rangle_{n+1} - \langle w_{ij} \rangle_n =
\int \ldots \int
w_{ij} [T_{ij}(w_{ij}|\{w_{ik}'\}) P_{ij}(w_{ij}',n) \nonumber \\
- T_{ij}(w_{ij}'|w_{ij},\{w_{ik}'\}_{k\neq j}) P_{ij}
(w_{ij},n)]\displaystyle{\prod_{k\neq j}} [P_{ik}(w_{ik}',n)dw_{ik}']dw_{ij}'
dw_{ij}
\label{masterw1}\end{aligned}$$ or, interchanging the primed and unprimed variables $w_{ij}$ and $w_{ij}'$ in the first term on the right hand side, $$\begin{aligned}
\fl \langle w_{ij} \rangle_{n+1} - \langle w_{ij} \rangle_n =
\int \ldots \int (w_{ij}'-w_{ij}) T_{ij}(w_{ij}'|w_{ij},\{w_{ik}'\}_{k\neq j})
P_{ij}(w_{ij},n) \nonumber \\
\times \displaystyle{\prod_{k\neq j}} [P_{ik}(w_{ik}',n)dw_{ik}']dw_{ij}'
dw_{ij}
\label{masterw2}\end{aligned}$$ or, with (\[transition\]) and integrating over $w_{ij}'$, $$\fl \langle w_{ij} \rangle_{n+1} - \langle w_{ij} \rangle_n =
\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p({\mbox{\boldmath $x $}}) \int \ldots \int
\Delta w_{ij}({\mbox{\boldmath $x $}}, {\mbox{\boldmath $w $}}_i)
\displaystyle{\prod_{k=1}^N} P_{ik}(w_{ik},n) dw_{ik} \qquad (j \in V_i)
\label{masterw3}$$ or, with (\[normPijn\]) and (\[expvalue1\]), $$\langle w_{ij} \rangle_{n+1} - \langle w_{ij} \rangle_n =
\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p({\mbox{\boldmath $x $}})\Delta w_{ij}({\mbox{\boldmath $x $}},
\langle {\mbox{\boldmath $w $}}_i \rangle_n), \qquad (j \in V_i)
\label{masterw4}$$ where we used that $\Delta w_{ij}$ is linear in the $w_{ik}\ (k=1,\ldots, N)$ to replace ${\mbox{\boldmath $w $}}_i$ by the expectation value $\langle {\mbox{\boldmath $w $}}_i \rangle_n$ in the expression for $\Delta w_{ij}$. If we assume that the expectation values of the synaptic connections $\langle w_{ij} \rangle_n$ converge to finite values, $\langle w_{ij} \rangle_{\infty}$, for $n$ tending to infinity, we can solve this equation for $n\rightarrow \infty$. This is the subject of the next section.
Final values for the weights {#solutionsec}
============================
If we suppose that the left-hand side of (\[masterw4\]) vanishes in the limit of $n$ tending to infinity, we have $$\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p({\mbox{\boldmath $x $}}) \Delta w_{ij}({\mbox{\boldmath $x $}}, \langle
{\mbox{\boldmath $w $}}_i \rangle_{\infty}) = 0
\label{masterw5}$$ At this point, we need an expression for the increment $\Delta w_{ij}(n)$, in the $n$-th learning step. We take the biologically motivated learning rule [@Hee99] $$\Delta w_{ij}(n) = \eta_i [\kappa - \gamma_i({\mbox{\boldmath $x $}},n)](2x_i-1)x_j
\qquad (i=1, \ldots, N; j \in V_i)
\label{leerregelHee}$$ where $\eta_i$ is the learning rate, $\kappa$ the margin parameter and $\gamma_i({\mbox{\boldmath $x $}}, n)$ the stability coefficient given by $$\gamma_i({\mbox{\boldmath $x $}},n) = (2x_i -1)[h_i({\mbox{\boldmath $x $}},n) -\theta_i]
\label{stabiliteit}$$ \[cf. eq. (\[stabcoef\])\]. Here, $h_i({\mbox{\boldmath $x $}},n)$ is the membrane potential of neuron $i$ at step $n$ of the learning process $$h_i({\mbox{\boldmath $x $}}, n) = \displaystyle{\sum_{k=1}^N} w_{ik}(n) x_k
\label{potential}$$ and $\theta_i$ the threshold potential of neuron $i$. It should be noted that in (\[potential\]) $x_k$ is the state of neuron $k$ at step $n$ of the learning procedure. Substituting (\[leerregelHee\]) with (\[stabiliteit\]) and (\[potential\]) into (\[masterw5\]) we find, using the fact that $(2x_i-1)^2=1$, $$\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}}p({\mbox{\boldmath $x $}})[\kappa(2x_i -1) -
\bigl(\displaystyle{\sum_{k=1}^N}
\langle w_{ik} \rangle_{\infty} x_k -\theta_i\bigr)]x_j = 0
\label{nlarge1}$$ where we divided by the learning rate $\eta_i$.
Up to now, the precise form of the probability distribution $p({\mbox{\boldmath $x $}})$ has been left unspecified. At this point, let us specify our probability distribution $p({\mbox{\boldmath $x $}})$ to be such that the chosen patterns ${\mbox{\boldmath $x $}}$ are centered around representative patterns ${\mbox{\boldmath $\xi $}}^{\mu}$. To that end, we choose our probability distribution $p({\mbox{\boldmath $x $}})$ such that it is a sum of $p$ equally probable, individually independent probability distributions, *i.e.*, $$p({\mbox{\boldmath $x $}}) = \frac{1}{p}\displaystyle{\sum_{\mu = 1}^p} p^{\mu}({\mbox{\boldmath $x $}})
\label{pxvoorp}$$ where $p^{\mu}({\mbox{\boldmath $x $}})$ is factorizable, $$p^{\mu}({\mbox{\boldmath $x $}}) = \displaystyle{\prod_{i=1}^N}p^{\mu}_i(x_i)
\label{pmufactor}$$ *i.e.*, the neurons behave independently from one another. The quantity $p_i^{\mu}(x_i)$ is the probability that, once the pattern index $\mu$ is chosen, neuron $i$ is in the state $x_i$. One therefore has $$p_i^{\mu}(0) + p_i^{\mu}(1) = 1
\label{imunorm}$$ In a learning process, at every step $n$, the index $\mu$ is drawn from a collection of $p$ equally probable pattern indices, thus fixing the probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$ according to which the pattern ${\mbox{\boldmath $x $}}$ is chosen for that learning step $n$.
Let us denote averages with respect to the probability $p^{\mu}({\mbox{\boldmath $x $}})$ by $\overline{\phantom{x_i}}^{\mu}$ $$\begin{aligned}
\displaystyle{\sum_{x_i=0,1}} p^{\mu}_i(x_i) x_i = \overline{x_i}^{\mu}, \qquad
&\displaystyle{\sum_{x_i=0,1}} p^{\mu}_i(x_i) x_i^2 = \overline{(x_i^2)}^{\mu}
\label{average1}\end{aligned}$$ implying, in view of (\[pmufactor\]) and (\[imunorm\]), $$\begin{aligned}
\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p^{\mu}({\mbox{\boldmath $x $}}) x_i =
\overline{x_i}^{\mu}, \qquad &
\displaystyle{\sum_{{\mbox{\boldmath $x $}} \in \Omega}} p^{\mu}({\mbox{\boldmath $x $}}) x_i^2 =
\overline{(x_i^2)}^{\mu}
\label{average2}\end{aligned}$$ Thus a bar with an index $\mu$ indicates an average with respect to the probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$. With the choice (\[pxvoorp\]), the result (\[nlarge1\]) can be rewritten in terms of these averages, where we must take be aware that a term $\overline{(x_j^2)}^{\mu}$ appears in the sum over $k$: $$\begin{aligned}
\fl \langle w_{ij} \rangle_{\infty} \displaystyle{\sum_{\mu=1}^p}
\left[\overline{(x_j^2)}^{\mu}
- \left(\overline{x_j}^{\mu}\right)^2 \right] =
\displaystyle{\sum_{\mu =1}^p}
\left[\kappa(2\overline{x_i}^{\mu} -1) - (\displaystyle{\sum_{k=1}^N}
\langle w_{ik} \rangle_{\infty} \overline{x_k}^{\mu} - \theta_i)\right]
\overline{x_j}^{\mu} \ \ \ j \in V_i
\label{nlarge2}\end{aligned}$$ The latter result can be rewritten as $$p\sigma^2_j \langle w_{ij} \rangle_{\infty}
= -\displaystyle{\sum_{k \in V_i}} (A_i)_{jk} \langle w_{ik} \rangle_{\infty}
+ B_{ij}, \ \
j \in V_i
\label{nlarge3}$$ where we abbreviated $$\begin{aligned}
\lo{\sigma^2_j} = \frac{1}{p}\displaystyle{\sum_{\mu=1}^p}
\overline{(x_j-\overline{x_j}^{\mu})^2}^{\mu} \nonumber \\
= \frac{1}{p} \displaystyle{\sum_{\mu=1}^p} \left[\overline{(x_j^2)}^{\mu} -
\left(\overline{x_j}^{\mu}\right)^2 \right]
\label{variance}\end{aligned}$$ and where we split up the sum over all $k$ in a sum over $V_i$ and a sum over its complement $V_i^C$: $$\begin{aligned}
\fl (A_i)_{jk} := \displaystyle{\sum_{\mu=1}^p} \overline{x_j}^{\mu}
\overline{x_k}^{\mu}, & i=1,\ldots, N; \ \ j,k \in V_i
\label{Aij}
\\
\fl B_{ij} := \displaystyle{\sum_{\mu=1}^p}
[ \kappa (2\overline{x_i}^{\mu}-1)-
(\displaystyle{\sum_{k \in V_i^C}} \langle w_{ik}
\rangle_0\overline{x_k}^{\mu} - \theta_i)]\,
\overline{x_j}^{\mu}, \qquad &i=1,\ldots, N; \ \ j \in V_i
\label{Bij} \end{aligned}$$ Note that the matrix $A_i$ is a symmetric matrix, the dimension of which equals the number of indices in $V_i$, *i.e.*, the number of adaptable afferent synaptic connections of neuron $i$. In the matrix $B$, we could write $\langle w_{ik} \rangle_0$ rather than $\langle w_{ik} \rangle_{\infty}$, since $\langle w_{ik} \rangle_0 = \langle w_{ik} \rangle_{\infty}$ for $k \in V_i^C$. It is easy to solve the equation (\[nlarge3\]). First, rewrite it as $$\displaystyle{\sum_{k \in V_i}} \left[(D_i)_{jk} + (A_i)_{jk}\right] \langle w_{ik}
\rangle_{\infty} = B_{ij}
\label{nlarge4}$$ where $D_i$ is the diagonal matrix $$(D_i)_{jk} := p\sigma^2_j \delta_{jk}, \qquad j,k \in V_i
\label{Dij}$$ The matrix $D_i + A_i$ is non-singular, and can be inverted. Inserting the explicit form of $B_{ij}$ (\[Bij\]), we then find $$\fl \langle w_{ij} \rangle_{\infty} = \displaystyle{\sum_{k \in V_i}} (D_i + A_i)^{-1}_{jk}\displaystyle{\sum_{\mu=1}^p} [ \kappa (2\overline{x_i}^{\mu}-1)-
(\displaystyle{\sum_{l \in V_i^C}} \langle w_{il}
\rangle_0\overline{x_l}^{\mu} - \theta_i)]\,
\overline{x_k}^{\mu} \ \ \ j \in V_i
\label{solutionwij}$$ where we used that $D_i$ and $A_i$ are symmetric matrices. In the usual treatments of noiseless recurrent neural networks ($\sigma_j=0$ for all $j$), one finds for the $w_{ij} (\infty)$ the so-called pseudo-inverse solution [@DieOp], [@Hee99], which reads, in our notation, $$\fl w_{ij}^{PI} = w_{ij}(0) + \displaystyle{\sum_{\nu, \mu =1}^p}
(C_i^{-1})^{\mu \nu}[\kappa(2\xi_i^{\mu} -1) - (\displaystyle{\sum_{k=1}^N}
w_{ik}(0)\xi_k^{\mu} - \theta_i)]\xi_j^{\nu} \ \ \ j \in V_i
\label{PI}$$ where $C_i^{-1}$ is the inverse of the correlation matrix $C_i^{\mu \nu} = \sum_{k \in V_i} \xi_k^{\mu} \xi_k^{\nu}$. Apparently, our result (\[solutionwij\]) is not a simple generalization of the standard result for noiseless recurrent neural networks. Note that the usual pseudo-inverse solution (\[PI\]) depends on the initial values $w_{ij}(0)$ of all the weights, whereas our solution (\[solutionwij\]) depends only on $w_{ij}(0)$ for $j \in V_i^C$ and not on the initial value $w_{ij}(0)$ of the changing weights ($j \in V_i$). Apparently, a little bit of noise completely wipes out the effect of the initial state of changing connections, since the result (\[solutionwij\]) is true for any value of the noise unequal zero.
In the limit that all $\sigma_j$ (\[variance\]) vanish, the set of equations (\[nlarge4\]) becomes under-determined for $p<N$, since the matrix $A_i$ is then singular. Hence, the solution (\[solutionwij\]) does not exist for a noiseless net. Explicitly, this can be seen as follows. Let us suppose that the $p$ average patterns $\{ \overline{x_1}^{\mu},
\overline{x_2}^{\mu}, \ldots, \overline{x_p}^{\mu} \}$, $(\mu=1, \ldots, p)$, span a $p$-dimensional vector space. Then, for $r>p$ there are coefficients $\alpha_{rl}$ such that $$\overline{x_r}^{\mu} = \displaystyle{\sum_{l=1}^p} \alpha_{rl} \overline{x_l}^{\mu}$$ for all $\mu=1, \ldots, p$. It follows that every column $(A_i)_{jr}$, ($i$ fixed, $j$ a running index of $V_i$ and $r$ a fixed number larger than $p$) is a linear combination of the first $p$ columns of $(A_i)_{js}$ ($i$ fixed, $j$ a running index of $V_i$ and $s$ smaller than or equal to $p$). Consequently, the matrix $A_i$ has a vanishing determinant, and is not invertible. Therefore, in case the average squared deviation (\[variance\]) would vanish, the unique solution (\[solutionwij\]) would not exist. The fact that for vanishing variances $\sigma_j$ our set of equations for the final weights is under-determined has been mentioned already in the introduction, in the text under equation (\[Ourequation\]).
In [@Hee00] the occurrence of the average squared deviation (\[variance\]) has been overlooked. This enabled the authors to solve the Master Equation (\[masterw4\]) in the usual way. By means of the so-called Gauss-Seidel procedure they obtained a modified version of the usual pseudo-inverse solution for the connections, rather than the expression (\[solutionwij\]).
Intermediate values for the weights {#intermediatesec}
===================================
Since our approach was simply based on the assumption of convergence of the $\langle w_{ij} \rangle_n$ for $n\rightarrow \infty$, we had no knowledge of the intermediate values of the weights $\langle w_{ij} \rangle_n$ for finite $n$. However, we can predict the evolution of the weights through an iterative procedure. If we repeat the derivation in section \[solutionsec\], starting from (\[masterw4\]) in stead of (\[masterw5\]), we find $$\begin{aligned}
\fl \langle w_{ij} \rangle_{n+1} = \langle w_{ij} \rangle_n +
\frac{\eta_i}{p}
\displaystyle{\sum_{\mu=1}^p}\kappa(2\overline{x_i}^{\mu}-1)
\overline{x_j}^{\mu} \nonumber \\
- \frac{\eta_i}{p} \displaystyle{\sum_{\mu=1}^p}
\Bigl(\displaystyle{\sum_{k=1}^N}
\langle w_{ik} \rangle_n \overline{x_k}^{\mu} - \theta_i\Bigr)
\overline{x_j}^{\mu}
- \eta_i \sigma_j^2 \langle w_{ij} \rangle_n \ \ \ j \in V_i
\label{recurs}\end{aligned}$$ In the limit $n \rightarrow \infty$, equation (\[recurs\]) implies (\[nlarge2\]), provided that the weights $\langle w_{ij} \rangle_n$ converge.
Using the relation (\[recurs\]), one can find, by numerical iteration, the quantities $\langle w_{ij} \rangle_n$ for any $n$, given the starting values $\langle w_{ij} \rangle_0$. Hence, we can verify numerically that the $\langle w_{ij} \rangle_n$ are independent of these starting values. Moreover, one can study the convergence of the learning procedure. In order to do so, one must make a particular choice for the probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$, which, up to now, was left unspecified. For our choice \[see (\[pjmub\])\], this distribution will depend on a so-called noise parameter $b$ $(0 \leq b \leq 1)$, such that $\overline{x_j}^{\mu} = \xi_j^{\mu}$ and $\sigma_j^2=0$ if the noise parameter $b$ vanishes $(b=0)$. Through the parameter $b$ we can tune the amount of noise during the learning process. Numerical calculations show that the $\langle w_{ij} \rangle_n$ do indeed converge in the limit $n\rightarrow \infty$, for arbitrary $b$, including $b=0$, if $\eta_i$ is small enough. Interestingly, convergence times to the final values (\[solutionwij\]) diverge for a decreasing noise parameter $b$ (*i.e.*, $b \rightarrow 0$), but the time of convergence drops to a small value if $b=0$ (see figure \[convergfigure\]), indicating that something peculiar happens in this limit.
In other words, if one demands existence of the solution (\[solutionwij\]), one may choose $b$ arbitrarily small, but not zero, and convergence to the solution is faster for larger values of $b$. If one puts $b$ to zero in the iterative application of (\[recurs\]), one observes rapid convergence of the weights, to the pseudo-inverse values (\[PI\]). Maybe surprisingly, these values have no continuous relation with the values for finite $b$, despite the fact that the expressions (\[recurs\]), the difference equations that determine the weights $w_{ij}$, do depend continuously on the $\sigma^2_j$, and, hence \[see equation (\[sigmab\]) below\], on $b$. In view of the difference in the solutions $\langle w_{ij} \rangle_{\infty}$ for the cases $b=0$ \[eq. (\[solutionwij\])\] versus $b\neq0$ \[eq. (\[PI\])\], this is obvious: there cannot be a continuous relationship between them, since the pseudo-inverse solution (\[PI\]) depends on all the initial values $w_{ij}(0)$, whereas our solution (\[solutionwij\]) is independent of the initial values of the weights $w_{ij}(0)$ for $j \in V_i$.
In the next section we investigate whether our final solution for the weights corresponds to the storage of patterns in a stable way.
Stability {#stabilitysection}
=========
It is well-known that a neural net with fixed weights $w_{ij}$ (in our case this will be after the learning phase) and deterministic neuron dynamics evolves, in the course of time, to limit cycles of finite length $n$ $(n=1,2, \ldots, 2^N)$. Cycles with $n=1$, or fixed points, are of particular interest in neural network theory. If a pattern ${\mbox{\boldmath $\xi $}}^{\mu}$ is a fixed point of the dynamics of a neural net for a given set of weights $w_{ij}$, the stability coefficients (\[stabcoef\]) or (\[stabiliteit\]) are positive for all $i$, in which case the system remains in the pattern ${\mbox{\boldmath $\xi $}}^{\mu}$, *i.e.*, the system is stable [@KinOpp91]. Besides the fact that $\gamma_i({\mbox{\boldmath $\xi $}}^{\mu})$ is a measure for the size of the basin of attraction of fixed point ${\mbox{\boldmath $\xi $}}^{\mu}$ [@Gar88], it is plausible that it is also a measure that determines to what extent the network state ${\mbox{\boldmath $x $}}$ remains in a neighborhood of ${\mbox{\boldmath $\xi $}}^{\mu}$ when the deterministic evolution of this neuron state ${\mbox{\boldmath $x $}}$ is replaced by a stochastic version of this evolution. In order to get an idea of the effectiveness of the learning process discussed in the preceding section, we will therefore consider the expectation value of the stability coefficients (\[stabcoef\]) in the limit $n\rightarrow \infty$: $$\langle \gamma_i({\mbox{\boldmath $\xi $}}^{\mu}) \rangle_{\infty} = (2\xi_i^{\mu}-1)
(\displaystyle{\sum_{j=1}^N}
\langle w_{ij} \rangle_{\infty} \xi_j^{\mu} - \theta_i)
\label{gammaksi2}$$ Once a set of patterns $\{ {\mbox{\boldmath $\xi $}}^{\mu} \}$ is known, these quantities can be explicitly calculated with the help of the expression (\[solutionwij\]). In this section we will attempt to derive, analytically, an approximation of (\[gammaksi2\]) by averaging over sets of patterns $\{ {\mbox{\boldmath $\xi $}}^{\mu} \}$ with given mean activity $a$. If, however, we would calculate the average of $\langle w_{ij} \rangle_{\infty}$ over these patterns directly, we would lose all dependence on neuron indices $j$ and pattern indices $\mu$, such that the correlations with the $\xi_j^{\mu}$ would disappear. We therefore have to take a different route.
An approximation for the expectation value (\[gammaksi2\]) is $$\langle \gamma_i({\mbox{\boldmath $\xi $}}^{\mu}) \rangle_{\infty} \approx
(2\xi_i^{\mu}-1)( \langle \overline{h_i}^{\mu} \rangle_{\infty} - \theta_i)
\label{gammaksi3}$$ where $\overline{h_i}^{\mu}$ is the membrane potential of neuron $i$ averaged with respect to $p^{\mu}({\mbox{\boldmath $x $}})$: $$\overline{h_i}^{\mu} = \displaystyle{\sum_{j=1}^N} w_{ij} \overline{x_j}^{\mu}
\label{potentiaalmu1}$$ In fact, the approximation (\[gammaksi3\]) would be exact if $\overline{x_j}^{\mu}$ would be equal to $\xi^{\mu}_j$ for all $j$, *i.e.*, in the limit that the probability function is such that $\overline{x_j}^{\mu}$ equals $\xi_j^{\mu}$, for all $j$. The average potential occurring in (\[gammaksi3\]) can be found from (\[nlarge2\]). Indeed, multiplying by $\overline{x_j}^{\mu}$ and summing with respect to $j \in V_i$, we find from this equation: $$\fl \displaystyle{\sum_{j \in V_i}} \overline{x_j}^{\mu} \langle w_{ij}
\rangle_{\infty}
p\sigma_j^2 = \displaystyle{\sum_{\nu=1}^p}\Bigl[ \kappa
(2\overline{x_i}^{\nu} -1)
- (\displaystyle{\sum_{k=1}^N} \langle w_{ik} \rangle_{\infty}
\overline{x_k}^{\nu}
- \theta_i)\Bigr] \displaystyle{\sum_{j \in V_i}}\overline{x_j}^{\nu}
\overline{x_j}^{\mu}
\label{nlarge5}$$ where we also used (\[variance\]).
The average square deviation $\sigma_j^2$ occurring in this equation depends on the neuron $j$. In this article we will consider the case in which all neurons have the same standard deviation $\sigma_j$ $$\sigma_j = \sigma \qquad j=1, \ldots, N \label{jindep1}$$ *i.e.*, the probability $p^{\mu}_j(x_j)$ is supposed to be such that the uncertainty to find neuron $j$ in a state $\xi_j^{\mu}$ is the same for all neurons of the neural net. Using the identity $$\displaystyle{\sum_{j \in V_i}} \overline{x_j}^{\mu} \langle w_{ij}
\rangle_{\infty} p\sigma^2 =
\Bigl(\displaystyle{\sum_{j=1}^N} \overline{x_j}^{\mu}
\langle w_{ij} \rangle_{\infty} - \displaystyle{\sum_{j \in V_i^C}}
\overline{x_j}^{\mu} \langle w_{ij} \rangle_{\infty} \Bigr)p\sigma^2
\label{replace}$$ we find from (\[nlarge5\]) $$\fl \Bigl( \langle \overline{h_i}^{\mu} \rangle_{\infty} -
\displaystyle{\sum_{j \in V_i^C}}
\langle w_{ij} \rangle_0 \overline{x_j}^{\mu}\Bigr) p\sigma^2 =
\displaystyle{\sum_{\nu=1}^p} [\kappa(2\overline{x_i}^{\nu} -1) + \theta_i -
\langle \overline{h_i}^{\nu} \rangle_{\infty}] \displaystyle{\sum_{j \in V_i}}
\overline{x_j}^{\nu} \overline{x_j}^{\mu}
\label{nlarge6}$$ where we used $\langle w_{ij} \rangle_0 = \langle w_{ij} \rangle_{\infty}$ for $j \in V_i^C$. An alternative form for (\[nlarge6\]) reads $$\displaystyle{\sum_{\nu=1}^p} [p\sigma^2{\mathbbm{1}}+ C_i]^{\mu \nu} \langle \overline{h_i}^{\nu}
\rangle_{\infty} = \displaystyle{\sum_{\nu=1}^p} f_i^{\nu}C_i^{\nu \mu} + g_i^{\mu}
\label{nlarge7}$$ where ${\mathbbm{1}}$ is the $p\times p$ unit matrix and where $C_i^{\mu \nu}$, the correlation matrix for averaged neuron states, is defined by $$C_i^{\nu \mu} := \displaystyle{\sum_{j \in V_i}} \overline{x_j}^{\nu} \overline{x_j}^{\mu}
\label{Cmunu}$$ Furthermore, we abbreviated $$\begin{aligned}
f_i^{\mu} = \kappa(2\overline{x_i}^{\mu} -1) + \theta_i \label{deff}
\\
g_i^{\mu} = p \sigma^2 \displaystyle{\sum_{j \in V_i^C}} \langle w_{ij} \rangle_0
\overline{x_j}^{\mu} \label{defg}\end{aligned}$$ Multiplying both sides of the matrix equation (\[nlarge7\]) by the inverse of the (symmetric) matrix occurring on its left-hand side we obtain the solution $$\langle \overline{h_i}^{\mu} \rangle_{\infty} =
\displaystyle{\sum_{\nu, \lambda=1}^p} f_i^{\nu} C_i^{\nu \lambda}
[(p\sigma^2 {\mathbbm{1}}+ C_i)^{-1}]^{\lambda \mu} + \displaystyle{\sum_{\nu=1}^p}
g_i^{\nu}[(p\sigma^2 {\mathbbm{1}}+ C_i)^{-1}]^{\nu \mu}
\label{himu}$$
Once a particular probability distribution $p^{\mu}({\mbox{\boldmath $x $}})$ of patterns centered around ${\mbox{\boldmath $\xi $}}^{\mu}$ is given, we can evaluate $f_i^{\nu}$ and $C^{\mu \nu}$, and, hence, via (\[himu\]), the average stability coefficient (\[gammaksi3\]).
In contrast to our expression (\[solutionwij\]) for the expectation value of the final value of the weights $\langle w_{ij} \rangle_{\infty}$, the result for the expected average potential $\langle \overline{h_i}^{\mu} \rangle_{\infty} = \sum_j \langle w_{ij} \rangle_{\infty}\overline{x_j}^{\mu} $ does exist for vanishing $\sigma$. This is clear, already, from (\[himu\]), in which the existence of the inverse $(p\sigma^2 {\mathbbm{1}}+ C_i)^{-1}$ does not depend on the presence of the extra term $p\sigma^2{\mathbbm{1}}$ as long as the average patterns $\overline{{\mbox{\boldmath $x $}}}^{\mu}$ are linearly independent, since then $C_i^{\mu \nu}$ is invertible. Using (\[solutionwij\]) and (\[deff\]), and assuming $\langle w_{ij} \rangle_0 =0$ for all $j \in V_i^C$ we may write the average potential $\langle \overline{h_i}^{\mu} \rangle_{\infty}$ as $$\langle \overline{h_i}^{\mu} \rangle_{\infty} =
\displaystyle{\sum_{\nu}\sum_{k,j \in V_i}} (D_i + A_i)_{jk}^{-1}
\overline{x_j}^{\mu} \overline{x_k}^{\nu} f_i^{\nu}
\label{hisubst}$$ Comparing this to (\[himu\]) with (\[gimu\]), we obtain the identity $$\displaystyle{\sum_{k,j \in V_i}} (D_i + A_i)_{jk}^{-1}
\overline{x_j}^{\mu} \overline{x_k}^{\nu} =
\displaystyle{\sum_{\lambda}} C_i^{\mu \lambda}[(p\sigma^2{\mathbbm{1}}+ C_i)^{-1}]^
{\lambda \nu}$$ Hence, though the matrix $(D_i + A_i)^{-1}$ occurring in (\[hisubst\]) itself does not exist for $b=0$, the above combination clearly does: it reduces to $\delta^{\mu \nu}$, as we see from the right hand side for $\sigma=0$, implying that in the limit of vanishing noise $$\langle \overline{h_i}^{\mu} \rangle_{\infty} = f_i^{\mu}$$ which is already clear from (\[himu\]) and is equivalent to —the average of— eq. (\[PIequations\]). Thus, although the values of the weights themselves do not have a continuous relation with the values corresponding to the pseudo-inverse solution, the average values for the membrane potentials, and, therefore, of the stability coefficients, do.
In the following we suppose that $w_{ij}=0$ for all $j \in V_i^C$. This corresponds to a neural net in which all existing connections are of adaptable strength and the only connections with constant strength are the non-existing connections. For $j \in V_i^C$, we then have $w_{ij}(0)=0$, and, hence, $\langle w_{ij} \rangle_0 = 0$, implying that $$g_i^{\mu} = 0
\label{gimu}$$ Let us choose the probability distribution $$p_j^{\mu}(x_j) = (1-b)\delta_{x_j, \xi_j^{\mu}} + b\delta_{x_j,1-\xi_j^{\mu}}
\qquad (j=1,\ldots,N)
\label{pjmub}$$ which fulfills (\[imunorm\]), and from which $p^{\mu}({\mbox{\boldmath $x $}})$ follows by the prescription (\[pmufactor\]). The noise parameter $b$ is a probability ($0 \leq b \leq 1$). More specifically, for given $\mu$ $(\mu=1,2,\ldots,p)$, $1-b$ is the probability that the activity $x_j$ of neuron $j$ equals that of the pattern $\xi_j^{\mu}$. We suppose that $b$ is small compared to unity. As follows from (\[pjmub\]), the noise parameter $b$ is related to the width of the distribution of input patterns around each pattern ${\mbox{\boldmath $\xi^{\mu} $}}$. We can immediately calculate the average neuron state (\[average2\]) associated with the distribution (\[pjmub\]) $$\overline{x_j}^{\mu}(\xi_j^{\mu}) = (1-b)\delta_{1,\xi_j^{\mu}} +
b\delta_{1,1-\xi_j^{\mu}}
\label{xjavmu}$$ the coefficient (\[deff\]) $$f_i^{\mu} = (2\xi_i^{\mu}-1)(1-2b)\kappa + \theta_i
\label{fimub}$$ as well as the average squared deviation (\[variance\]) $$\sigma^2 = b(1-b)
\label{sigmab}$$ The fact that $\sigma^2$ is $j$-independent is a consequence of the particular choice (\[pjmub\]) for $p_j^{\mu}(x_j)$, *i.e.*, of the fact that all neurons $j$ are supposed to have the same uncertainty to be in state $\xi_j^{\mu}$. Hence, the supposition (\[jindep1\]) is satisfied.
Let us suppose that the probability that $\xi_j^{\mu}=1$ is $a$, for each $j$ independent of any other neuron index $k$, and, hence, that the probability that $\xi_j^{\mu}=0$ is $(1-a)$, for all of the patterns ${\mbox{\boldmath $\xi $}}^1, \ldots, {\mbox{\boldmath $\xi $}}^p$. We can now use this to arrive at an estimate value for the average potential of neuron $i$, eq. (\[himu\]), which is exact in the limit of $p \rightarrow \infty$, for $\alpha=p/N$ fixed, and smaller than $1$. From (\[Cmunu\]) we find, for $\mu \neq \nu$, $$\begin{aligned}
\fl C_i^{\mu \nu} \approx \displaystyle{\sum_{j \in V_i}} \bigl\{ a^2
\overline{x_j}^{\mu}(1) \overline{x_j}^{\nu}(1) +
a(1-a)\overline{x_j}^{\mu}(1) \overline{x_j}^{\nu}(0) + \nonumber \\
(1-a)a\overline{x_j}^{\mu}(0) \overline{x_j}^{\nu}(1) +
(1-a)^2\overline{x_j}^{\mu}(0) \overline{x_j}^{\nu}(0) \bigr\}
\label{Cmunuu}\end{aligned}$$ while for $\mu=\nu$ we get $$C_i^{\mu \nu} \approx
\displaystyle{\sum_{j \in V_i}}\{a\overline{x_j}^{\mu}(1)^2
+ (1-a)\overline{x_j}^{\mu}(0)^2\}
\label{Cmunue}$$ Defining the dilution $d$ as the average fraction of neurons from which an arbitrary neuron does not have an incoming connection, each neuron has on the average $N(1-d)$ incoming connections. Hence, using (\[xjavmu\]), we find from (\[Cmunuu\]) and (\[Cmunue\]) $$\begin{aligned}
\fl C_i^{\mu \nu} \approx N(1-d)\{a(1-a)(1-2b)^2\delta^{\mu \nu} \nonumber \\
+ [a^2(1-b)^2 + 2ab(1-a)(1-b) + (1-a)^2b^2]\}
\label{Cmunu2}\end{aligned}$$ We thus have achieved that the correlation matrix $C_i^{\mu \nu}$ for an $N$ neuron net has been expressed in parameters typical for the network, namely the dilution $d$, the mean activity $a$ and the noise $b$. An alternative way to write (\[Cmunu2\]) is $$C_i^{\mu \nu} \approx l\delta^{\mu \nu} + m
\label{Cmunu3}$$ where $l$ and $m$ are shorthand for combinations of the typical network parameters $a$, $b$ and $d$ $$\begin{aligned}
l & := & N(1-d)a(1-a)(1-2b)^2 \nonumber \\
m & := & N(1-d)[a^2(1-b)^2 + 2ab(1-a)(1-b) + (1-a)^2b^2]
\label{deflm}\end{aligned}$$ With (\[Cmunu3\]), the matrix occurring in (\[himu\]) can be cast into the form $$(p\sigma^2{\mathbbm{1}}+ C_i)^{\mu \nu} \approx (p\sigma^2 + l)\delta^{\mu \nu} + m
\label{iplusC}$$ The inverse of a $p$-dimensional matrix $A$ with elements $A^{\mu \nu} = x\delta^{\mu \nu} + y$ is given by the matrix $A^{-1}$ with elements $$(A^{-1})^{\mu \nu} = [\delta^{\mu \nu}(x+py) - y]/x(x+py)
\label{Ainvers}$$ From (\[Cmunu3\]), and (\[Ainvers\]) applied to (\[iplusC\]), we find $$\begin{aligned}
\displaystyle{\sum_{\lambda}}C_i^{\nu \lambda}[(p\sigma^2 {\mathbbm{1}}+ C_i)^{-1}]
^{\lambda \mu} \approx
\frac{
l[l+p(\sigma^2 +m)]\delta^{\mu \nu}
+ mp\sigma^2 }
{(l+p\sigma^2)[l+ p(\sigma^2+m)]}
\label{CmunuCinvers}\end{aligned}$$ Substituting this result, together with (\[gimu\]), into the expression (\[himu\]), yields for the average potential of neuron $i$ in pattern $\mu$ the expression $$\begin{aligned}
\langle \overline{h_i}^{\mu} \rangle_{\infty} \approx
\frac{
\{l[l+p(\sigma^2 + m)]+mp\sigma^2 \}f_i^{\mu} + mp\sigma^2
\Sigma_{\nu \neq \mu}f_i^{\nu}}
{(l+p\sigma^2)[l+p(\sigma^2 + m)]}
\label{himufinal}\end{aligned}$$ The sum over the $f_i^{\nu}$ occurring in this expression can be calculated with (\[fimub\]), $$\displaystyle{\sum_{\nu \neq \mu}} f_i^{\nu} \approx (p-1)[(2a-1)(1-2b)\kappa +
\theta_i]
\label{sumfi}$$ where we used that the average value of the $p-1$ neuron activities $\xi_i^{\mu}$ can be approximated by $a$, the average activity of the net. We can now write down the final result for the stability parameters (\[gammaksi3\]), which is a function of the network parameters $d$, $a$ and $b$, the number of patterns $p$, the number of neurons of the net $N$, and the neuron properties $\kappa$ and $\theta_i$: $$\begin{aligned}
\fl \langle \gamma_i(\xi_i^{\mu}) \rangle_{\infty}
\approx
\frac{
\{l[l+p(\sigma^2 + m)]+mp\sigma^2 \}[(1-2b)\kappa + \theta_i(2\xi_i^{\mu}-1)]}
{(l+p\sigma^2)[l+p(\sigma^2 + m)]} \nonumber \\
+ \frac{
mp\sigma^2(p-1)[(2a-1)(1-2b)\kappa +
\theta_i](2\xi_i^{\mu}-1)}{(l+p\sigma^2)[l+p(\sigma^2 + m)]}
-\theta_i(2\xi_i^{\mu}-1)
\label{finalabdpN}\end{aligned}$$ Note that for $\sigma^2=0$ we immediately recover $\langle \overline{h_i}^{\mu} \rangle_{\infty} = f_i^{\mu}$ and $\langle \gamma_i(\xi_i^{\mu}) \rangle_{\infty} = \kappa$, as we should, from the equations (\[himufinal\]) and (\[finalabdpN\]) respectively.
The final average stability coefficient of neuron $i$ takes two different values respectively, depending on whether $\xi_i^{\mu} = 1$ or $\xi_i^{\mu}=0$.
In figure (\[figuur2\]) we plotted this quantity for a chosen average activity $a=0.5$, as a function of $b$. It is clear that the stability coefficient can be expected to remain positive. In the same figure we plotted the actual values of $\langle \gamma_i(\xi_i^{\mu}) \rangle_{\infty}$, as obtained by choosing randomly a set of patterns $\{ {\mbox{\boldmath $\xi $}}^{\mu} \}$ with given mean activity $a=0.5$, and using (\[gammaksi2\]) with (\[solutionwij\]) for the calculation.
The difference between the curves is evident, and indicates that we must be careful not to overestimate the accuracy of our result as an indication for $\langle \gamma_i(\xi_i^{\mu}) \rangle_{\infty}$. In fact, in a large region, the storage of undisturbed patterns is better than our estimate suggests by a factor $2$ to $3$, as can be concluded from the figure. With this in mind, we may assume that after the noisy learning process, the patterns ${\mbox{\boldmath $\xi $}}^{\mu}$ are indeed fixed points under the deterministic network dynamics for a small noise parameter $b$.
Retrieval and basins of attraction {#basinssection}
==================================
In this section we address the question what happens to the average size of the basins of attraction if noiseless learning (training parameter $b=0$) is compared to noisy learning ($b\neq 0$). After the network has been trained with patterns ${\mbox{\boldmath $x $}}$ with noise $b \ (b\neq0)$, we numerically check the retrieval capacity of the net by presenting patterns with noise $b^{\star}$, *i.e.*, patterns chosen according to a probability distribution of the form (\[pjmub\]), in which $b$ has been replaced by $b^{\star}$. The presented patterns evolve under deterministic parallel dynamics $x_i(n+1) = \Theta(h_i(n)-\theta_i)$. The attempt to retrieve a pattern is successful if the network state ${\mbox{\boldmath $x $}}$ runs into a fixed point equal to the clean, undistorted pattern ${\mbox{\boldmath $\xi $}}^{\mu}$ of which a noisy version was the initial state. The result can be read off from figure \[basins\]. Since the curves obtained via noisy learning lie above the curve with noiseless learning, the basins of attraction are, apparently, enlarged in the presence of noise during the learning stage.
The result is in agreement with earlier studies by Gardner et al [@GSW89], and Wong & Sherrington [@WS90a], [@WS90b], [@WS93].
In [@GSW89], like in our case, noise is added to patterns during a training stage. However, the algorithm is of a different kind, because it includes an error-mask, *i.e.*, the weights are updated if and only if, upon presentation of a noisy pattern during the learning stage, the membrane potential $h_i$ has the wrong sign. In this way, if the learning algorithm converges, retrieval of patterns for which the amount of noise is equal to that of the training patterns, is guaranteed.
In [@WS90a], [@WS90b] and [@WS93], various retrieval properties of a neural network are discussed. It is argued that optimizing (by finding the optimal weights) the overlap of a noisy pattern with its corresponding training representative after one retrieval step is, in fact, a way of noisy training [@WS90a]. The optimal network is sought for via a replica-calculation that minimizes a cost-function, thus optimizing the first step retrieval. No explicit learning rule is used in these articles. An explicit expression for the final values of the weights is not given.
Our approach is different from those discussed above, in the sense that we start from an explicit learning rule, which is biologically acceptable: it is derived from the principle that energy cost for synaptic adaptation is minimal [@Hee99]; it is a function of local variables; it does not contain error masks; neurons are assumed to be noisy. Though in our network the basins of attraction are not optimal (it was not our goal to optimize the basins of attraction), we do have an explicit learning algorithm as well as a final expression for the expectation value of weights.
Conclusion {#conclusionsec}
==========
We have shown that learning with noise leads to final values for the weights $w_{ij}$ which are different from those found in the corresponding situation without noise. Surprisingly, the solution for the values of the weights $w_{ij}$ of a noisy system in the limit of vanishing noise, does *not* converge to the values of the solution of the system without noise.
Moreover, in a system without noise the values of the final weights depend on the initial values of all weights, whereas in a noisy system the initial values of the changing weights, the $w_{ij}$ for $j \in V_i$, are wiped out in the course of time.
Our noisy trained networks have larger basins of attraction than noiselessly trained networks. This is in agreement with earlier findings in the literature. The exact dependence of the retrieval properties on various parameters, such as the mean activity $a$ and the memory load $\alpha = p/N$ is still to be elucidated.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors are indebted to Wouter Kager for carefully reading this manuscript and suggesting some improvements.\
\
References {#references .unnumbered}
==========
[99]{} Müller B, Reinhardt J and Strickland M T 1995 [*Neural Networks: An Introduction*]{} (Berlin: Springer) Personnaz L, Guyon I and Dreyfus G 1985 [*J. Physique Lett.*]{} [**46**]{} L359. Diederich S and Opper M 1987 [*Phys. Rev. Lett.*]{} [**58**]{} 949. Heerema M and van Leeuwen W A 1999 [*J. Phys. A: Math. Gen.*]{} [**32**]{} 263–86. Gardner E J, Stroud N and Wallace D J 1989 [*J. Phys. A:Math. Gen.*]{} [**22**]{} 2019–30. Wong K Y M and Sherrington D 1990 [*J. Phys. A: Math. Gen.*]{} [**23**]{} L175–82. Wong K Y M and Sherrington D 1990 [*J. Phys. A: Math. Gen.*]{} [**23**]{} 4659–72. Wong K Y M and Sherrington D 1993 [*Phys. Rev. E.*]{} [**47**]{} 4465–82 Heerema M and van Leeuwen W A 2000 [*J. Phys. A: Math. Gen.*]{} [**33**]{} 1781–95. Kinzel W and Opper M 1991 [*Models of Neural Networks*]{} ed Domany E (Berlin: Springer) p 152 Gardner E 1988[*J. Phys. A: Math. Gen*]{} [**21**]{} 257–70.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Color compatibility is important for evaluating the compatibility of a fashion outfit, yet it was neglected in previous studies. We bring this important problem to researchers’ attention and present a compatibility learning framework as solution to various fashion tasks. The framework consists of a novel way to model outfit compatibility and an innovative learning scheme. Specifically, we model the outfits as graphs and propose a novel graph construction to better utilize the power of graph neural networks. Then we utilize both ground-truth labels and pseudo labels to train the compatibility model in a weakly-supervised manner. Extensive experimental results verify the importance of color compatibility alone with the effectiveness of our framework. With color information alone, our model’s performance is already comparable to previous methods that use deep image features. Our full model combining the aforementioned contributions set the new state-of-the-art in fashion compatibility prediction.'
author:
- |
Heming Zhang^1^, Xuewen Yang^2^, Jianchao Tan^3^,\
Chi-Hao Wu^1^, Jue Wang^4^, C.-C. Jay Kuo^1^
bibliography:
- 'egbib.bib'
title: Learning Color Compatibility in Fashion Outfits
---
Introduction
============
While many factors contribute to the compatibility of a set of fashion items, the color compatibility plays a pivotal role. When verifying whether a set of fashion items form a compatible outfit, the first check that comes to one’s mind is whether their colors are compatible or not. While each human body may fit to a different silhouette, color compatibility is more universal and applicable to a wider audience. Learning color compatibility for fashion will help us better identify compatible outfits and provide recommendation with maximum color synergy. Previous fashion studies largely neglected the importance of color compatibility. They either focus on the impact of each individual color [@[2019arxiv]Toward] or mix colors together with other information [@[2019SIGIR]Interpretable]. Instead, we would like to bring the importance of color compatibility to audience’s attention and propose a novel solution to it. The colors of an outfit can be compatible for different reasons, e.g., similar colors versus contrasting colors. Identifying what colors are compatible in the fashion industry will help us choose compatible colors for outfits. Therefore, instead of making a simple yes or no prediction on the compatibility of a given set of items, we would like to go one step further to learn the different templates of color compatibility in fashion. Existing methods are not necessarily applicable to real world problems, especially to the fashion problems. The color compatibility templates developed by Matsuda *et al.* [@matsuda1995color; @tokumaru2002color] as shown in Figure \[fig:matsuda\] were later adopted for various purposes, e.g., natural image color harmonization [@cohen2006color; @paletteHarmonization]. Nevertheless, O’Donovan *et al.* [@o2011color] noticed that this theory can deviate significantly from the color data collected from human users. Moreover, those methods for natural images do not specifically apply to fashion domain. Besides, the fashion trend is notoriously volatile and changes quickly with time. Instead of using hand-weaved rules on color, it would be more preferable to learn color compatibility templates directly from fashion data.
For the reasons above, we aim to learn compatibility in fashion outfits and apply it to fashion colors for the purpose of analyzing color compatibility templates in fashion problems. Our proposed data-driven approach for learning compatibility in fashion is illustrated in Figure \[fig:our\_approach\]. Given fashion outfit data, three consecutive problems are studied using a compatibility model: 1) compatibility prediction 2) compatibility templates extraction 3) compatibility template-based recommendation. The compatibility model is the key to learn the compatibility from fashion outfits data. In our work, we propose a novel compatibility model, and an innovative learning scheme to train this model.
To better learn the compatibility of fashion outfits, we first model the outfits with graphs. Most previous works did not utilize graph models as they focus on learning the pairwise compatibility between two items rather than the compatibility of entire outfits [@[2018ECCV]Learning; @[2019AAAI]TransNFCM; @[2019CVPR]Context; @[2019SIGIR]Interpretable; @[2019TKDE]Explainable]. They obtained the outfit compatibility by simple aggregation (e.g., averaging) over all the pairwise compatibility scores. However, the pairwise compatibility does not necessarily guaranty compatibility across a entire set of items. A few attempts have been made to model the entire outfits. A Recurrent Neural Network (RNN) was used to model the entire outfits in [@[2017MM]Learning], which requires a predefined order to convert each set of items into a sequence.
A graph representation is more suitable in modelling of fashion outfits. Each graph is orderless ensemble of nodes and their interaction (i.e. edges), thus offering a natural abstraction of the compatibility among items of a fashion outfit. Although a graph model was used to model the entire dataset in [@[2019CVPR]Context], the outfit compatibility was still obtained from pairwise item compatibility. Graphs are used to model outfits in the work of Cui [@[2019WWW]Dressing]. Both of them [@[2019CVPR]Context; @[2019WWW]Dressing] took a straightforward approach of modelling each item as a node in the graphs and fed such graphs into Graph Neural Networks (GNN) or Graph Convolutional Networks (GCN). However, most GNNs/GCNs are designed for tasks such as node classification or recommendation [@[2017NIPS]Inductive; @[2018ICLR]Few; @[2018KDD]Graph], thus focus on node information aggregation. On the contrary, in the compatibility problem one should focus on the relation among items, which is represented by the edge information in the graphs in previous studies. Such mismatch leads to the poor results in [@[2019CVPR]Context; @[2019WWW]Dressing], comparing to the works without graphs. Instead of modelling fashion items as nodes, we model each pairwise relation between items as a node to utilize the full power of GNN methods to aggregate the outfit compatibility from pairwise compatibilities. This novel graph construction is more suitable for compatibility learning using GNNs/GCNs.
Popular fashion datasets only provide binary compatibility infomation, i.e. compatible or not compatible. It is not a trivial task to learn distinct compatibility templates from such binary data. To learn compatibility templates from such data, we apply a clustering algorithm to the compatible outfits to segregate outfits with distinct compatibility templates. However, a simple application of some clustering operation to the compatibility templates may not yield the segregation meaningful for human interpretation we expected. Therefore, we propose to integrate the clustering process into the compatibility learning through an iterative process, which ensures that the obtained outfit clusters are meaningful for compatibility interpretation. The obtained compatibility model is also improved, as demonstrated in the experimental results.
To summarize, our contributions are three-fold:
1. We raise the important problem of color compatibility in fashion outfits, which was largely neglected in previous work in fashion. We take a data-driven approach, which overcomes the drawbacks of previous color compatibility studies in natural images and is more applicable to the fashion domain.
2. We propose a new way to model the compatibility of outfits. We utilize graphs to model the entire fashion outfits with a novel graph construction method that is more suitable for compatibility learning. Our compatibility model is not limited to color compatibility, but applies to various compatibility problems.
3. We propose an innovative learning scheme to integrate compatibility prediction and outfit clustering, which helps the interpretation of the fashion compatibility as well as improves the prediction performance.
Related Work
============
Color Compatibility
-------------------
Most previous studies in color compatibility are not suitable for our problem. Color compatibility templates such as the hue templates proposed by Matsuda *et al.* in [@matsuda1995color] and other variants were frequently used by later work [@tokumaru2002color; @cohen2006color; @paletteHarmonization]. However, experiments on large-scale color theme datasets [@o2011color] have demonstrated an obvious mismatch between the templates and real-world data. Instead, O’Donovan *et al.* [@o2011color] proposed to learn a regression model for color compatibility using color theme data collected from human users. The limitation of this model is that it only evaluates fixed number of colors. Besides, those color themes are for general purposes and do not necessarily apply to fashion problems.
Fashion Outfit Compatibility {#subsec:fashion_compatibility}
----------------------------
Most approaches only modelled pairwise compatibility between fashion items, whereas the compatibility of an entire outfit was either not considered or predicted as simple average of the pairwise compatibilities. Some previous work used the distances of item embeddings as the measure of pairwise compatibility [@[2015ICCV]Learning; @[2018AAAI]Dress; @[2018ECCV]Learning; @[2019CVPR]Complete; @[2019KDD]POG; @[2019WWW]Enhancing] and [@[2019AAAI]TransNFCM] is a generalization of this approach. Furthermore, [@[2011IJCAI]Fashion; @[2017MM]NeuroStylist; @[2019arxiv]Toward; @[2019SIGIR]Interpretable; @[2019TKDE]Explainable; @[2019ICCV]Learning; @[2019ICIP]Learning; @[2018ICMR]Interpretable; @[2019arxiv]Coherent; @[2019CVPR]Context; @[2020CVPR]Fashion] directly modelled the pairwise compatibility. Jagadeesh *et al.* [@[2014KDD]Large] used color representations as clues to retrieve fashion items that are matched to the query.
It is challenging to model the compatibility of entire outfits with various number of items. In [@[2017TMM]Mining], item features were fused by global pooling. Similarly, in [@[2017ICCVworkshop2018WACV]], items are divided into six categories and the six features are concatenated. RNNs were utilized in [@[2017MM]Learning; @[2018KDDworkshop]Outfit], but a pre-defined order is required to model a set of items as a sequence. Although [@[2019CVPR]Context] directly modelled the pairwise compatibility, context information from other items were used. The entire dataset is modelled as a graph, where each node represents an item and nodes that appear in the same outfit are connected. During inference, the neighbors of each node contains items that appear in other outfits and provide rich context information. Such setting is different from previous work where each outfit is evaluated separately and no information from other outfits is provided. Experiments indicate that the method in [@[2019CVPR]Context] performs poorly in the common setting.
Closely related to our work, Cui[@[2019WWW]Dressing] proposed to model each outfit as a graph where each node represents an item with a specific category. However, the proposed GNN model [@[2019WWW]Dressing] learned on such graphs focuses more on item category information and neglected the relation among items. Instead, we propose a novel graph construction that is more suitable for the compatibility problem.
Deep Learning on Graphs
-----------------------
Graph neural networks (GNN) and graph convolutional networks (GCN) have been evolved over the years [@bruna2013spectral; @henaff2015deep; @defferrard2016convolutional; @kipf2016semi; @[2017NIPS]Inductive; @[2018ICLR]Graph]. A common drawback of the existing models is that they only utilized the node features. In their work, edge features were either ignored or only the 1-D edge weights were used.
A few attempts were made to overcome this drawback. In [@[2017CVPR]Dynamic], an edge-conditioned convolution was proposed where edge information was utilized to filter the nodes. Moreover, edge features were directly utilized for node feature aggregation functions in [@[2018arxiv]Exploiting]. In these works, edge features are only considered as an auxiliary for node feature aggregations and do not get updated throughout the network inference. Monti et al. [@[2018arxiv]Dual] proposed a Dual-Primal GCN framework, which alternated between the original graph (primal) and its line graph (dual). The dual convolutional layer produces features on the edges of the primal graph, and the primal edge features were used in the primal convolutional layer to produce primal node features. In [@[2019CVPR]Edge], 2-D edge features defined as (dis)similarities between pairs of nodes were updated following each update of nodes. Five options for updating edge features were discussed in [@[2019TOG]Dynamic]. The edge features in these works are only used for node feature updates. For fashion compatibility, however, direct aggregation and update of edge features are desired.
Learning Color Compatibility in Fashion Outfits {#sec:method}
===============================================
The fashion compatibility prediction problem takes a set of fashion items $\mathbf{x}$ as inputs and a binary compatibility label $y$ as the prediction target. For each fashion item, an image is provided, and additional information such as category annotations or text descriptions may also be available.
We formulate the compatibility template learning problem as a clustering problem. The compatible outfits are grouped into clusters, where each cluster obtained should identify an outfit compatibility template.
An overview of our color compatibility learning framework is shown in Figure \[fig:overview\]. We first extract color features from the items, which will be described in Section \[subsec:palette\]. As introduced in Section \[subsec:graph\], we then embed the outfit into an embedding feature, which will be used for compatibility prediction. We will explain our proposed learning scheme for the compatibility model in Section \[subsec:joint\].
Feature Extraction {#subsec:palette}
------------------
For the problem of color compatibility learning, we propose to use a color palette, collection of dominant colors, as the color feature for an item. Previous work [@[2013ICCV]Paper; @[2014MM]What; @[2015MM]Collaborative] on fashion study usually use color histogram as features. Here we use the color palette instead for several reasons: First, our goal is to study the compatibility among colors, rather than the statistics of color distribution; Second, noises such as illumination variation, are encoded in the color histogram but color palette is more invariant to such variations; Third, color palettes are more abstractive, from which we can better learn the compatibility templates.
To extract a color palette from a fashion item image, we convert the image color space from $\mathit{RGB}$ to $\mathit{Lab}$. In the $\mathit{Lab}$ color space, the distance between two colors indicates their visual difference in humans perception. For each image, we apply the K-means algorithm to cluster the pixels with their Lab colors as features. Each cluster centroid represent a dominant color in the image and the cluster centroids together form the palette of the corresponding fashion item. The color feature of the image is obtained by concatenating the cluster centroids. In our work, we cluster the colors into 3 clusters. The resulting color feature has 9 elements (the details are in the supplemental material).
Although we use color features for the purpose of color compatibility learning, our proposed compatibility learning framework can be applied to other representations as well. In section \[sec:experiment\], we demonstrate the effectiveness of our framework with other fashion item representations such as image features extracted from ResNet [@resnet].
Outfit Embedding {#subsec:graph}
----------------
Our outfit embedding consists of two steps. We first construct an outfit graph to model the given set of items, then embed the outfit graph into an outfit embedding.
As explained in previous sections, we would like to overcome the mismatch between outfit graphs and existing GNN models. Rather than designing a network structure specific to our problem, we choose to design a new graph construction method that fits to most existing models that aggregate the node information. Since the outfit compatibility is the global relation among outfit items, we propose to use nodes to represent pairwise item relations instead of individual items. Thus the global outfit relation can be aggregated from local pairwise relations.
For a given outfit of $N$ items, we construct a graph $G=(V,E)$ with $\binom{N}{2}$ nodes where each node represents the pairwise relation between two items in the outfit. For each pair of nodes with shared item, we connect them with an edge. An illustration is shown in Figure \[fig:graph\]. Comparing to the graph construction method in [@[2019WWW]Dressing], our proposed graph can be viewed as its line graph (also known as the edge-to-node dual graph).
For the node $\mathcal{V}_{(i,j)} \in V$ that represents the relation between the $i$th and the $j$th items, the node feature is obtained from features of these two items as $$f_{\mathcal{V}_{(i,j)}} = Z(f_i) \odot Z(f_j),$$ where $Z(\cdot)$ is an embedding module, $\odot$ denotes the Hadamard product and $f$ denotes the feature.
For an outfit graph constructed from an outfit, we then embed the outfit graph into an embedding of length $d$, where $d$ is a hyper-parameter. We can apply various models to aggregate the node information and embed the graph. In this work, we use GraphSage [@[2017NIPS]Inductive] for demonstration. At $l$-th layer, the hidden representation of node $\mathcal{V}_{(i,j)}$ in $G$ is updated as $$\begin{split}
m_{\mathcal{V}_{(i,j)}}^{l+1} &= \text{MEAN} \left(\{h_{\mathcal{V}'}^l, \forall \mathcal{V}' \in \mathcal{N}(\mathcal{V}_{(i,j)}) \} \right), \\
h_{\mathcal{V}_{(i,j)}}^{l+1} &= \sigma \left(W^l [h_{\mathcal{V}_{(i,j)}}^l, m_{\mathcal{V}_{(i,j)}}^{l+1}]+b\right),
\end{split}$$ where $\mathcal{N}_{\mathcal{V}_{(i,j)}}$ is the neighborhood of $\mathcal{V}_{(i,j)}$, $[\cdot,\cdot]$ denotes concatenation, $W$ is the weight matrix, $b$ is the bias term and $\sigma$ is a nonlinear activation function. We first obtain the new message $m_{\mathcal{V}_{(i,j)}}^{l+1}$ by aggregating information from the neighbors of node $\mathcal{V}_{(i,j)}$, then the updated hidden representation $h_{\mathcal{V}_{(i,j)}}^{l+1}$ is obtained from the new message and the previous hidden representation of node $\mathcal{V}_{(i,j)}$. The node features at the last layer are then globally pooled into a graph embedding.
Given the graph embedding, we can apply a fully connected layer and take a softmax to predict the compatibility from the outfit graph embedding. To train this layer together with the GNN model, we propose a novel learning scheme.
Joint Learning Scheme {#subsec:joint}
---------------------
There are different reasons for a set of colors to be considered compatible, e.g. similar colors or different colors with a synergy. To further understand such differences and explore compatibility templates, we cluster the compatible outfits in the embedding space such that the outfits in the same cluster should be close in the embedding space. Note that the incompatible samples are all simply assigned to a single “incompatible" cluster. We do not care why a set is not compatible so there is no need of finer clustering for them. (Detailed discussion of a single incompatible cluster is in the supplemental material).
We expect each outfit cluster to represent a distinct and tangible reason for the containing outfits to be considered compatible. Therefore, each cluster can be interpreted as a compatibility template. However, it is not trivial to cluster the outfits such that compatibility templates could be interpreted. If we first train the embedding with outfit compatibility labels and then cluster the outfits using their trained embeddings, the clusters obtained may not be meaningful for compatibility prediction and thus may not help us to interpret compatibility templates.
Therefore, we propose a joint learning scheme that brings the clustering process into the training loop of the compatibility model. It is an iterative learning scheme that alternatively generates pseudo labels from the outfit clusters and updates the network with pseudo labels.
### Pseudo label generation.
Originally a sample $i$ contains the outfit input $\mathbf{x}_i$ and a binary outfit compatibility label $y_i \in \{0,1\}$. For each sample $i$, we generate a multi-class pseudo label $\tilde{y}_i=c$, where $c \in \{0,1,\cdots,C\}$ and $C$ is the number of compatible clusters. The pseudo label generation procedure consists of four steps and is illustrated in Figure \[fig:joint\]:
**1. Collecting graph embeddings**. We collect the outfit embeddings of all outfit samples in the training set using the current model.
**2. Clustering**. The current embeddings of outfits are clustered to $C+1$ clusters and the clusters are indexed from 0 to C. For cluster $c \in \{0,1,\cdots, C\}$, we count the number of compatible outfits $m_{c,pos}$ and the number of incompatible outfits $m_{c,neg}$ it contains. Then we calculate the compatible ratio as $r_c = m_{c,pos} / m_{c,neg}$ for each cluster $c$.
**3. Assigning cluster labels**. In this step, we assign a cluster label $z_c \in \{0,1\}$ to cluster $c \in \{0,1,\cdots,C\}$. These labels are only assigned to the clusters and do not represent for the outfit sample they contain. The obtained clusters are ranked by the ratios $r_c$ obtained in the previous step. Then the top $C$ clusters are assigned with cluster labels of 1, whereas the rest cluster is assigned with a cluster label of 0.
**4. Generating pseudo sample labels**. For each data sample $(\mathbf{x}_i,y_i)$, we assign a pseudo multi-class label $\tilde{y}_i \in {0,1,\cdots,C}$. The Euclidean distance between the $i$th sample and the centroid of cluster $c$ in the embedding is denoted as $d(i,c)$. The pseudo label is assigned as $$\begin{split}
\tilde{y}_i = \underset{c}{\arg\min} \; d(i,c) \quad
\text{s.t.} \; y_i = z_c,
\end{split}$$ which is the index of the closest cluster that has the same label.
The unsupervised learning method DeepCluster [@[2018ECCV]Deep] has a similar concept which iteratively clusters the features obtained from the network, and update the network using the clusters. In our method, we make use of the available compatibility labels and obtain the pseudo labels by fusing the ground-truth compatibility labels with the cluster assignments.
### Network update with pseudo labels.
We combine two loss functions to update the network with both the pseudo and ground-truth labels. For each compatible outfit, we randomly sample an incompatible outfit to pair with it.
The first loss is a global loss that is applied to entire outfits. We use the categorical cross-entropy loss to update the network parameters with pseudo labels: $$\begin{split}
l_1 =\sum_i [l_{CCE}(\left[g(\mathbf{x}_{i,pos})\right]_{\tilde{y}_{i,pos}}, \, \tilde{y}_{i,pos})
+ l_{CCE}(\left[g(\mathbf{x}_{i,neg})\right]_{\tilde{y}_{i,neg}}, \, \tilde{y}_{i,neg})],
\end{split}
\label{eq:l1}$$ where $l_{CCE}(\cdot,\cdot)$ is the categorical cross-entropy loss, $g[\cdot]_c$ is the predicted softmax score for the class $c \in \{0,1,\cdots, C\}$, and the subscriptions $_{i,pos}$ and $_{i,neg}$ denote the positive and the negative outfits of the $i$th pair of outfits, respectively.
Additionally, if a positive/negative pair of outfits share common items, we apply a local loss to subsets of the outfits. It helps the network learn more subtle differences between item relations. For each positive/negative pair of outfits that share items, we remove their common items from both outfits and obtain a pair of outfit subsets. Then for this pair of outfit subsets, we apply a binary cross-entropy loss with the compatibility labels: $$\begin{split}
l_2 = \sum_i [l_{BCE}(g(\hat{\mathbf{x}}_{i,pos}),\, y_{i,pos})
+ l_{BCE}(g(\hat{\mathbf{x}}_{i,neg}),\, y_{i,neg})],
\end{split}
\label{eq:l2}$$ where $l_{BCE}(\cdot,\cdot)$ denotes the binary cross-entropy loss and $\hat{\mathbf{x}}$ denotes the subset of $\mathbf{x}$ obtained by removing the common items. The intuition is that a subset of a compatible outfit should also be compatible, and a subset of an incompatible outfit is most likely to be incompatible.
The final loss function takes the weighted sum of $l_1$ and $l_2$ with a hyper-parameter $\lambda$ $$l = l_1 + \lambda l_2.
\label{eq:loss}$$
More details of the loss functions are in the supplemental material.
Experimental Analysis {#sec:experiment}
=====================
Datasets
--------
**Polyvore Outfits** [@[2018ECCV]Learning]. It contains 53,306, 5000, and 10,000 outfits for training, validation, and testing, respectively. It provides one image per fashion item and some items also come with text descriptions. Two tasks are provided for benchmarking. One task is the compatibility prediction (CP), in which the compatibility predicted for given outfits are evaluated using area-under-curve (AUC). The other task is the fill-in-the-blank (FITB), where one fashion item out of four candidates that fits a given outfit subset best is chosen and the evalutation metric is the accuracy.
**Maryland Polyvore** [@[2017MM]Learning]. This dataset is similar to the Polyvore Outfits dataset, but much smaller in size. It has 17,316, 1,407, and 3076 outfits for training, validation and testing, respectively. It also has the two tasks (CP and FITB) for benchmarking. The test set provided by [@[2017MM]Learning] is not challenging enough as it contains negative samples that can be easily identified using category information. We follow the previous studies [@[2018ECCV]Learning; @[2019ICCV]Learning] and evaluate our model on a more challenging test set proposed in [@[2018ECCV]Learning].
Implementation Details
----------------------
At the preprocessing step, we first separate fashion items from image backgrounds using [@opencv]. Then we extract a 3-color palette from each fashion item.
We use one fully connected layer for node feature embedding and a 3-layer GraphSage [@[2017NIPS]Inductive] with ReLU activation and global max-pooling for graph embedding. The hidden representations are 60-D and the final embedding is 20-D. We utilize K-means with cosine similarity for clustering and re-assign the pseudo labels to outfits in every other 25 epochs of training. We tuned the hyper-parameters and chose $C=4$ and $\lambda=0.5$ for the best performance. Detailed analysis can be found in the supplemental material.
Since our compatibility model also applies to other representations, we follow [@[2018ECCV]Learning; @[2019ICCV]Learning] and extract image features from ResNet18 [@resnet], as well as text features from HGLMM Fisher vectors [@hglmm] of word2vec [@word2vec] in some experiments.
Comparison with Previous Work
-----------------------------
We compare with the results of the following methods:
**Bi-LSTM** [@[2017MM]Learning]. It models the outfits as sequences and the orders are pre-defined according to item categories. The image features are obtained from Inception V3 [@inceptionv3]. It utilizes the text descriptions as regularization during training.
**Siamese Network** [@[2015ICCV]Learning]. It models the pairwise compatibility. We compare with its results reported in [@[2018ECCV]Learning].
**Type-aware** [@[2018ECCV]Learning]. It considers the category differences when learning the pairwise compatibility, which additionally requires the item category annotation. Same as Bi-LSTM, the text features are used as regularization during training.
**Outfit generation** [@[2019arxiv]Coherent]. It learns the pairwise compatibility. BERT features [@bert] are extracted from text descriptions and used for both training and testing.
**Context-aware** [@[2019CVPR]Context]. It models the entire dataset as a graph. Its best results are obtained by utilizing context information from other outfits, which is a different setting from most previous works. We only compare to its results obtained without extra context information from other outfits. Its image features are extracted from ResNet50 [@resnet].
**NGNN** [@[2019WWW]Dressing]. It models each outfit as an graph where each item is modelled as a node. Item category annotation is required for graph construction. The authors only reported the results on the easy test set of the small Maryland Polyvore dataset. We conduct the experiments using their original source code on the challenging test sets provided by [@[2018ECCV]Learning].
**SCE-Net** [@[2019ICCV]Learning]. It models pairwise compatibility and requires text features for both training and testing.
**CSA-Net** [@[2020CVPR]Fashion]. It models pairwise compatibility and its category-based subspace attention network requires item category annotation.
**Ours**. We extract color features as described in Section \[subsec:palette\]. To further verify the effectiveness of our compatibility system, we also utilize ResNet18 features and HGLMM features following most previous works. Our framework does not require item category annotation and the text features are only used for training as regularization.
The comparisons are conducted in two groups. In the first group, the images are used as inputs for both training and testing. The image representations used are either color palettes extracted using our method, or deep image features extracted from convolutional neural networks. Among the first group, Type-Aware[@[2018ECCV]Learning], NGNN[@[2019WWW]Dressing] and CSA-Net[@[2020CVPR]Fashion] also requires item categories as inputs. The results of the first group are listed in Table \[tb:img\_results\]. With color palettes of 3 colors as image representation alone, our results already outperform most previous methods with deep features on the large-scale Polyvore Outfits dataset. This observation verifies our claim of the importance of color compatibility. Using the same deep features further boosts our results. Only the CSA-Net, which requires extra category annotation, outperforms ours. It demonstrates the effectiveness of the proposed compatibility learning method. The other graph method NGNN suffers from the bias on item categories. Its performance on the Maryland Polyvore drop 30%-40% on both tasks from the easy test set to the hard test set. Our graph method better utilizes individual item characteristics.
In the second group, the text descriptions are also used for the network training. Unlike Bi-LSTM, Type-aware and ours, Outfit gen. and SCE-Net also need text descriptions during testing. Similar conclusion can be drawn from the results of the second group listed in Table \[tb:img\_text\_results\]. When switching from the small-scale Maryland Polyvore dataset to the large-scale Polyvore Outfits dataset, the performances of Bi-LSTM and Type-aware drop dramatically, indicating potential overfitting. On the contrary, the proposed method gives consistent good results on both datasets. Although SCE-Net utilizes text description during inference, our method still shows comparable results on the CP task, while improving the FITB task by a significant margin of 4%-6%. Our results are only outperformed by Outfit gen., which uses a more powerful text features both for training and inference. Although the text descriptions provide more useful information during inference, it is less applicable to real scenarios where text descriptions are unavailable.
Ablation Studies {#subsec:ablation}
----------------
We compare the following modules as ablation studies to evaluate the contribution of each proposed module: NG denotes the node graph where each node represents an item in the outfit. LG denotes our proposed line graph where each node represents a pairwise relation between items. $l_{BCE}$ is the binary cross-entropy loss. $l_1$ and $l_2$ are the two losses proposed in (\[eq:l1\]) and (\[eq:l2\]), respectively. The ablation results are evaluated on the validation set of the large-scale Polyvore Outfits dataset and listed in Table \[tb:ablation\]. All methods compared use color palettes.
Each proposed module contributes to the final performance. Replacing the node graph with our proposed line graph, the overall performance is dramatically improved, supporting our argument that our graph better represents the fashion outfits for compatibility learning. By replacing the ground-truth binary labels to generated multi-class pseudo labels, the accuracy of the FITB task is increased by 2%. It shows that incorporating outfit clustering into the training loop not only provides us meaningful clusters but also helps the model better understand the compatibility. With the entire joint learning scheme, the accuracy of the FITB task is boosted by 6%. It verifies that our joint learning scheme helps the network learn the subtle difference of item relations.
Outfit Cluster Visualization {#subsec:cluster_viz}
----------------------------
To gain more insights of our color compatibility model, we visualize the compatible clusters obtained from learning the fashion outfits data. For each outfit sample in the validation set of the Polyvore Outfits dataset, we obtain its predicted cluster using our model. Due to the limited space, only two clusters with 35 outfit samples per cluster are displayed in Figure \[fig:sample\_color\]. More visualization results are found in the supplemental material. For each cluster, two outfit samples are enlarged and displayed on the right.
Each cluster has its own color compatibility template. The visualization in Figure \[fig:sample1\] suggests that the items in each outfit share similar colors. The cluster in Figure \[fig:sample2\] demonstrates the combination of two major colors that are distinct besides of the neutral colors with low chroma such as black or white. As illustrated via the visualization, our model does not simply group outfits together that share the similar colors, such as one cluster with green colors and the other with yellow colors. The outfits sharing similar color compatibility templates are grouped together through our model, which is the purpose of integrating the clustering process into the training loop.
It is not surprising that we can also find some color compatibility templates in the visualization of the model trained on deep image features. One example of the clusters is shown in Figure \[fig:sample\_it\], which displays the color compatibility template of similar colors. It again verifies our claim that color is an essential impact factor for outfit compatibility.
Fashion Recommendation based on Color compatibility Templates
-------------------------------------------------------------
While we can select one single overall best fit, as with the existing methods. Our innovative clustering method enables more versatile, customized recommendations from different compatibility templates. We made use of the outfit queries from the FITB task in the Polyvore Outfits dataset, and randomly sampled 2000 items as candidates. For each outfit query, we compare all the candidates and generate recommendation based on the cluster prediction. Figure \[fig:recommendation12\] shows the recommendation from clusters 1 and 2. More results are in the supplemental material. For each outfit query, three recommendations per cluster are displayed. When the query outfit has similar colors, recommendation from cluster 1 also suggests a similar color, whereas cluster 2 suggests a distinct color. When the query outfit has mixed colors, cluster 1 matches one of those colors, whereas cluster 2 either matches one of those colors or suggests a neutral color.
Conclusion
==========
In this paper, we addressed an important problem, namely learning color compatibility in fashion outfits. It plays an essential role in outfit compatibility, yet it has been neglected in previous work. We proposed a compatibility learning method including a novel compatibility model and an innovative learning scheme. Our approach provides more insights and possibility for customized recommendation. The proposed framework is not only suitable for color compatibility learning task, but also extendable to general outfit compatibility learning tasks. We conducted comprehensive experiments to demonstrate the importance of color compatibility, and the effectiveness of the proposed method.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
address:
- |
University of Chicago\
Chicago, IL 60637
- |
University of Massachusetts—Amherst\
Amherst, MA 01003
author:
- 'Pramod N. Achar'
- 'Eric N. Sommers'
bibliography:
- 'pramod.bib'
date: 14 December 2001
title: Local Systems on Nilpotent Orbits and Weighted Dynkin Diagrams
---
Introduction
============
Let $G$ be a reductive algebraic group over the complex numbers, $B$ a Borel subgroup of $G$, and $T$ a maximal torus of $B$. We denote by $\Weights = \Weights(G)$ the weight lattice of $G$ with respect to $T$, and by $\DomWeights = \DomWeights(G)$ the set of dominant weights with respect to the positive roots defined by $B$. Let ${\mathfrak{g}}$ be the Lie algebra of $G$, and let $\nilcone$ denote the nilpotent cone in $\Lie{g}$.
Now, let $e \in \nilcone$ be a nilpotent element, and let $\orb_e$ be the orbit of $e$ in ${\mathfrak{g}}$ under the adjoint action of $G$. We write $G^e$ for the centralizer of $e$ in $G$. Let $\nilcone_{o}$ denote the set of nilpotent orbits in ${\mathfrak{g}}$, and ${\nilcone_{o, r}}$ the set of $G$-conjugacy classes of pairs $$\{(e, \tau) \mid \text{$e \in \nilcone$ and $\tau$ an irreducible
rational representation of $G^e$} \}.$$
Lusztig [@lusztig:cells-aff-4] conjectured the existence of a bijection ${\nilcone_{o, r}}\leftrightarrow \DomWeights$ using his work on cells in affine Weyl groups. From the point of view of Harish-Chandra modules, Vogan also conjectured a bijection between ${\nilcone_{o, r}}$ and $\DomWeights$. Such a bijection has been established by Bezrukavnikov in two preprints (the bijections in each preprint are conjecturally the same) [@bezrukavnikov:tensor], [@bezrukavnikov:quasi-exc]. Bezrukavnikov’s second bijection is closely related to Ostrik’s conjectural description of the bijection [@ostrik:equiv-k-theory] (see also [@chm-ostrik:dist-inv]). In the case of $G=GL(n, \C)$, the first author [@achar:thesis] described an explicit combinatorial bijection between ${\nilcone_{o, r}}$ and $\DomWeights$ from the Harish-Chandra module perspective. At present, it is not known how all of these bijections are related (Bezrukavnikov’s two candidates; Ostrik’s conjectural candidate; and the first author’s candidate in type $A$). In this paper, we work in the context of [@achar:thesis], which we now review.
Let $K_G(\nilcone)$ denote the Grothendieck group of $G$-equivariant coherent sheaves on $\nilcone$. On the one hand, $K_G(\nilcone)$ has a natural basis indexed by elements of $\DomWeights$, denoted $AJ(\lambda)$ in [@ostrik:equiv-k-theory] (but unlike in [@ostrik:equiv-k-theory], we are not utilizing the $\C^*$ action on $\nilcone$ here). The algebraic global sections of $AJ(\lambda)$ are isomorphic as a $G$-module to $\Ind_T^G \lambda$. Thus, the space of global sections of any element ${\mathcal{F}}\in K_G(\nilcone)$ is given as a $G$-module by a unique expression of the form $$\label{eqn:vogan-thm}
\Sections(\nilcone, {\mathcal{F}}) = \sum_{\lambda \in \DomWeights}
m_\lambda \Ind_T^G \lambda,$$ where the $m_\lambda \in \Z$, and $m_\lambda \ne 0$ for only finitely many $\lambda$. (This fact was communicated to us by Vogan.)
Let us now fix a pair $(e,\tau) \in {\nilcone_{o, r}}$. We want to consider all elements ${\mathcal{F}}\in K_G(\nilcone)$ whose support is contained in $\overline{\orb}_e$, and whose restriction to $\orb_e$ is the vector bundle arising from $\tau$. For each such ${\mathcal{F}}$, there is at least one $\lambda$ of maximal length occurring in the expression (\[eqn:vogan-thm\]) (where we have fixed a $W$-invariant positive-definite symmetric bilinear form on the real span of $\Weights$, so that we can speak of the length of a weight of $G$). Define $\gamma: {\nilcone_{o, r}}\to
\DomWeights$ by $$\gamma(e,\tau) = \text{the \emph{smallest} such largest $\lambda$,
over all possible choices of ${\mathcal{F}}$}.$$ The following conjecture was made in [@achar:thesis]; moreover, it was proved in the case of $G = GL(n,\C)$ in [*op. cit*]{}.
The map $\gamma$ is well-defined and is a bijection. Moreover, there is a basis $\{M(e,\tau)\}$ for $K_G(\nilcone)$, indexed by ${\nilcone_{o, r}}$, such that
1. the support of $M(e,\tau)$ is $\overline{\orb}_e$,
2. the restriction of $M(e,\tau)$ to $\orb_e$ is the locally free sheaf arising from $\tau$, and
3. there is an upper-triangular relationship between this basis and the one indexed by $\DomWeights$: $$\Sections(\nilcone, M(e,\tau))
= \pm\Ind_T^G \gamma(e,\tau) + \sum_{{\|\mu\|^2} < {\|\tau\|^2}}
m_\mu \Ind_T^G \mu.$$
In this paper, we study the problem of computing $\gamma$ when $\tau$ gives rise to a local system on $\orb_e$ (for semisimple groups, this means that $\tau$ is trivial on the identity component of $G^e$). Let us denote by ${\Weights^{loc}_{+}}$ the image of $\gamma$ when $\tau$ corresponds to a local system. Here is an outline of the paper. In section 2, we compute ${\Weights^{loc}_{+}}$ explicitly for $G=GL(n)$. In section 3, we state a precise conjecture, for general $G$, that the Dynkin weights for the Langlands dual group of $G$ are a subset of ${\Weights^{loc}_{+}}$. In section 4, we show how to associate to $\gamma(e, \tau)$ a sub-bundle of the cotangent bundle of $G/B$. Then for $GL(n)$, we are able to prove that when $e$ is in a fixed nilpotent orbit, the cohomology of the sub-bundles (with coefficients in the structure sheaf) are independent of the local system $\tau$.
We thank David Vogan, Viktor Ostrik, and Roman Bezrukavnikov for helpful conversations. The second author was supported by NSF grant DMS-0070674.
Computing $\gamma$ for a local system in $GL(n)$
================================================
For a partition ${\mathbf{d}}$ of $n$, let $\orb_{\mathbf{d}}$ denote the nilpotent orbit in $\gln(n)$ indexed by ${\mathbf{d}}$. If ${\mathbf{d}}= [k_1^{a_1},
k_2^{a_2}, \ldots , k_l^{a_l}]$, then we know that $\pi_1(\orb_{\mathbf{d}}) \simeq \Z/c\Z$, where $c$ is the greatest common divisor of the $k_i$’s [@col-mcg:nilp]. Let $G^{\mathbf{d}}$ denote the isotropy group of an element in this orbit, and ${G_\mathrm{red}}^{\mathbf{d}}$ the reductive part thereof. Following the notation of [@col-mcg:nilp], we have $${G_\mathrm{red}}^{\mathbf{d}}\simeq GL(a_1)_\Delta^{k_1} \times \cdots \times
GL(a_l)_\Delta^{k_l}.$$ Here, $H_\Delta^m$ denotes the diagonal copy of $H$ in the product of $m$ copies of itself. Now, in the case of simply connected semisimple groups, we identify local systems on an orbit with representations of the component group of the centralizer of an element in the orbit. In the setting of $GL(n)$, we will first produce a list of representations of ${G_\mathrm{red}}^{\mathbf{d}}$, and then we will examine their restrictions to $SL(n)$ to determine the correspondence with local systems.
Let $A_1 \in GL(a_1)$, $A_2 \in GL(a_2)$, etc. The set of matrices $(A_1, \ldots, A_l)$ determines an element $\tilde A$ of $$GL(a_1)_\Delta^{k_1} \times \cdots \times GL(a_l)_\Delta^{k_l};$$ moreover, we have $$\det \tilde A = (\det A_1)^{k_1}\cdots(\det A_l)^{k_l}.$$ Since every $k_i$ is divisible by $c$, we can write this as $$\det \tilde A = \big((\det A_1)^{k_1/c}\cdots(\det A_l)^{k_l/c}\big)^c.$$ Let “$\det^{1/c}$” denote the character ${G_\mathrm{red}}^{\mathbf{d}}\to \C^\times$ given by $$(\det A_1)^{k_1/c}\cdots(\det A_l)^{k_l/c},$$ and for any $p \in \Z$, let $\det^{p/c}$ denote the $p$-th power of this map.
In $SL(n)$, the isotropy group $SL(n)^{\mathbf{d}}$ is the subgroup of $G^{\mathbf{d}}$ consisting of matrices of determinant $1$. The function $\det^{1/c}$ takes only finitely many values when restricted to this subgroup, namely, the $c$-th roots of unity. Indeed, the component group of $SL(n)^{\mathbf{d}}$ is just $\Z/c\Z$, and the irreducible representations of $\Z/c\Z$ are just powers of the function $\det^{1/c}$. We have, then, that the irreducible representations of $\pi_1(\orb_{\mathbf{d}})$ come from $$\det^{0/c}, \det^{1/c}, \ldots, \det^{(c-1)/c}.$$
Now we need to compute $\gamma(\orb_{\mathbf{d}}, \det^{p/c})$ for $0 \le p \le
c-1$. Suppose that the dual partition of ${\mathbf{d}}$ is ${\mathbf{d}}\trans =
[j_1^{b_1}, \ldots, j_s^{b_s}]$. Following the notation of [@achar:thesis], we let $L_{\mathbf{d}}$ denote the Levi subgroup $$L_{\mathbf{d}}= GL(j_1)^{b_1} \times \cdots \times GL(j_s)^{b_s},$$ and we let $P_{\mathbf{d}}$ be the standard (block-upper-triangular) parabolic subgroup whose Levi factor is $L_{\mathbf{d}}$. Recall that every nilpotent orbit in $GL(n)$ is Richardson; in particular, $\orb_{\mathbf{d}}$ is Richardson for $P_{\mathbf{d}}$. We fix a choice of $e \in \orb_{\mathbf{d}}$ such that ${G_\mathrm{red}}^{{\mathbf{d}}} \subset L_{\mathbf{d}}$ and $G^{\mathbf{d}}\subset P_{\mathbf{d}}$. Explicitly, the embedding of $G^{\mathbf{d}}$ in $P_{\mathbf{d}}$ is such that the factor $GL(a_i)^{k_i}_\Delta$ sits diagonally across the product of the first $k_i$ factors of $L_{\mathbf{d}}$. Let $\rho_{\mathbf{d}}$ denote the half-sum of the positive roots of $L_{\mathbf{d}}$.
We make the observation that any irreducible representation of $G^{\mathbf{d}}$ must be trivial on its unipotent radical, hence is completely determined by its restriction to ${G_\mathrm{red}}^{{\mathbf{d}}}$. Therefore, it makes sense to refer to irreducible representations of $G^{\mathbf{d}}$ by their highest weights. We do this at various points below.
In [@achar:thesis], the computational device of “weight diagrams” was introduced and employed to carry out the computation of $\gamma$. For our present purposes, however, we are only concerned with a small collection of representations for each orbit, and we do not need to use the cumbersome weight diagrams. Instead, we make use of the following auxiliary result.
\[prop:gamma\] Let $(\pi_\lambda, V_\lambda)$ denote the irreducible $L_{\mathbf{d}}$-representation of highest weight $\lambda$, regarded as a $P_{\mathbf{d}}$-representation by letting the unipotent part of $P_{\mathbf{d}}$ act trivially, and let $(\tau, W_\tau)$ be an irreducible representation of $G^{\mathbf{d}}$. Suppose that $\tau$ occurs as a summand in the restriction of $\pi_\lambda$ to $G^{\mathbf{d}}$, and, moreover, that $${\|\lambda + 2\rho_{\mathbf{d}}\|^2}$$ is minimal among all irreducible $P_{\mathbf{d}}$-representations whose restriction to $G^{\mathbf{d}}$ contains $\tau$ as a summand. Then $\gamma(\orb_{\mathbf{d}}, \tau) = \lambda + 2\rho_{\mathbf{d}}$, made dominant for $GL(n)$.
The next proposition describes $\gamma(\orb_{\mathbf{d}}, \det^{p/c})$ explicitly in terms of the standard basis. The remainder of the section will be devoted to establishing this proposition. In Section 4, we shall use this explicit description to prove the main result.
Let $\omega_p = \gamma(\orb_{\mathbf{d}}, \det^{p/c})$. Writing $\omega_0$ in the standard basis, let “block ${\mathrm{B}}_a$” refer to the collection of coordinate positions which contain the entry $a$ in $\omega_0$, for $a\in\Z$. Let $\mu_k$ be the multiplicity of $k$ in ${\mathbf{d}}\trans$. (Thus, if there is some $j_i$ such that $j_i = k$, then $\mu_k = b_i$; otherwise, $\mu_k = 0$.) Note that every $b_i$ and every $\mu_k$ must be a multiple of $c$.
\[prop:omega\] The length of block ${\mathrm{B}}_a$ is $\sum_{k \geq 0}
\mu_{a+2k+1}$ when $a\geq0$ and the length of ${\mathrm{B}}_{a}$ and ${\mathrm{B}}_{-a}$ are the same. Moreover, if we write the length of block ${\mathrm{B}}_a$ as $m_ac$, then $\omega_p$ is obtained by replacing the first $m_ap$ entries in block ${\mathrm{B}}_a$ with $a+1$.
Consider the orbit labeled by ${\mathbf{d}}= [6,3,3]$ in $GL(12)$. Then $c=3$, ${\mathbf{d}}\trans = [3,3,3,1,1,1]$, and $\mu_3 =3$ and $\mu_1=3$.
We illustrate the preceding proposition by listing here all the $\omega_p$. The first sentence of the proposition describes $\omega_0$ by giving the lengths of blocks, and the second sentence tells us how to obtain the other $\omega_p$’s by modifying $\omega_0$. $$\begin{aligned}
\omega_0 &= (\overbrace{2,2,2}^{{\mathrm{B}}_2},
\overbrace{0,0,0,0,0,0}^{{\mathrm{B}}_0},
\overbrace{-2,-2,-2}^{{\mathrm{B}}_{-2}}) \\
\omega_1 &= (3,2,2,1,1,0,0,0,0,-1,-2,-2) \\
\omega_2 &= (3,3,2,1,1,1,1,0,0,-1,-1,-2)\end{aligned}$$
Let $\lambda_p$ be the appropriate dominant weight of $L_{\mathbf{d}}$ as described in Proposition \[prop:gamma\], such that $\omega_p$ is just $\lambda_p + 2\rho_{\mathbf{d}}$, made dominant for $GL(n)$. In what follows, we shall be careless about saying “made dominant” every time; the reader should fill in those words wherever appropriate.
Let us begin with the trivial representation of $\pi_1(\orb_{\mathbf{d}})$, namely $\det^{0/c}$. Whatever $\lambda_0$ is, it must be dominant for $L_{\mathbf{d}}$, so that $\langle
\lambda_0, 2\rho_{\mathbf{d}}\rangle \ge 0$. Therefore $${\|\lambda_0 + 2\rho_{\mathbf{d}}\|^2} = {\|\lambda_0\|^2} + {\|2\rho_{\mathbf{d}}\|^2} +
2\langle \lambda_0, 2\rho_dd \rangle \ge {\|2\rho_{\mathbf{d}}\|^2}.$$ Now $0$ is a weight of $L_{\mathbf{d}}$ with the right restriction to ${G_\mathrm{red}}^{\mathbf{d}}$, and taking $\lambda_0 = 0$ obviously minimizes ${\|\lambda_0 +
2\rho_{\mathbf{d}}\|^2}$ (the above inequality becomes an equality). We therefore have $\omega_0 = 2\rho_{\mathbf{d}}$. What does $2\rho_{\mathbf{d}}$ look like? For each $GL(j_i)$ factor of $L_{\mathbf{d}}$, we get a part of $2\rho_{\mathbf{d}}$ that looks like $$(j_i-1, j_i-3, \ldots, 1-j_i).$$ Thus, in the total $2\rho_{\mathbf{d}}$, a particular coordinate $a$ with $a \geq 0$ occurs once for each factor $GL(j_i)$ with $j_i = a + 2k + 1$ for some $k$ with $k \geq 0$. It follows that the length of block ${\mathrm{B}}_a$ is precisely $\sum_{k \ge
0} \mu_{a+2k+1}$, as desired. It is also clear that the length of block ${\mathrm{B}}_a$ and block ${\mathrm{B}}_{-a}$ are equal.
Next, we consider the case $p \ne 0$. We will consider the first factor $GL(a_1)_\Delta^{k_1}$ of ${G_\mathrm{red}}^{\mathbf{d}}$ individually; the other factors would be treated identically. The factor $GL(a_1)_\Delta^{k_1}$ of ${G_\mathrm{red}}^{\mathbf{d}}$ sits diagonally across various $GL(j_i)^{b_i}$ factors of $L_{\mathbf{d}}$; indeed, it sits across $k_1$ of of them. Now, given a weight $\lambda$ of $L_{\mathbf{d}}$, we obtain the coordinates of the restriction $\lambda|_{{G_\mathrm{red}}^{\mathbf{d}}}$ by summing up coordinates of $\lambda$ according to the diagonal embedding of the factors of ${G_\mathrm{red}}^{\mathbf{d}}$ in $L_{\mathbf{d}}$. (See [@achar:thesis] for a detailed account of how to restrict weights from $L_{\mathbf{d}}$ to ${G_\mathrm{red}}^{\mathbf{d}}$.) In any case, we add up $k_1$ distinct coordinates of $\lambda$ to obtain each coordinate of the $GL(a_1)_\Delta^{k_1}$ part of the restriction of $\lambda$. Now, the highest (and only) weight of ${G_\mathrm{red}}^{\mathbf{d}}$ on the representation $\det^{p/c}$ is $$\Bigg(\underbrace{\Big(\frac{pk_1}{c},\ldots,\frac{pk_1}{c}\Big)}
_{GL(a_1)_\Delta^{k_1}},
\ldots,\underbrace{\Big(\frac{pk_l}{c},\ldots,\frac{pk_l}{c}\Big)}
_{GL(a_l)_\Delta^{k_l}}\Bigg);$$ in particular, every coordinate in the $GL(a_1)_\Delta^{k_1}$ part of the weight is equal to $pk_1/c$. If we want ${\|\lambda\|^2}$ to be minimal, it is clear that the $k_1$ coordinates we add up to obtain this coordinate should consist of $pk_1/c$ $1$’s and $(c-p)k_1/c$ $0$’s.
Repeating this argument for every factor of ${G_\mathrm{red}}^{\mathbf{d}}$, we see that ${\|\lambda\|^2}$ is minimized if, among all its coordinates, there are $pn/c$ $1$’s and $(c-p)n/c$ $0$’s. But to compute $\gamma$, we need to minimize ${\|\lambda + 2\rho_{\mathbf{d}}\|^2}$, not ${\|\lambda\|^2}$. We have $${\|\lambda + 2\rho_{\mathbf{d}}\|^2} = {\|\lambda\|^2} + {\|2\rho_{\mathbf{d}}\|^2} +
2\langle \lambda , 2\rho_{\mathbf{d}}\rangle.$$ So among possible $\lambda$’s of minimal size, we could try to choose one so as to minimize $\langle \lambda, 2\rho_{\mathbf{d}}\rangle$. In fact, we can arrange for the latter inner product to be $0$.
Among the factors $GL(j_i)^{b_i}$ of $L_{\mathbf{d}}$, take the weight $(1,\ldots,1)$ ([*i.e.*]{}, the determinant character) on $pb_i/c$ of the factors, and $(0,\ldots,0)$ on the remaining $(c-p)b_i/c$ of them. Concatenating these weights together, we obtain a weight $\lambda_p$ of $L_{\mathbf{d}}$. It is easy to see that $\lambda_p$ has the right number of $0$’s and $1$’s to have the desired restriction to ${G_\mathrm{red}}^{\mathbf{d}}$ as well as to be of minimal size. Moreover, as promised, we have that $\langle \lambda_p, 2\rho_{\mathbf{d}}\rangle = 0$.
Therefore, $\omega_p = \lambda_p + 2\rho_{\mathbf{d}}$. This looks very similar to $\omega_0$, except that in each block ${\mathrm{B}}_a$, a proportion $p/c$ of the coordinates have been increased by $1$. Thus $\omega_p$ has precisely the desired form.
A conjecture about Dynkin diagrams
==================================
Let $\Langdual{G}$, $\Langdual{B}$, $\Langdual{T}$ be the data of the Langlands dual group corresponding to $G$, $B$, $T$, respectively. Let $\Langdual{{\mathfrak{g}}}$, $\Langdual{{\mathfrak{b}}}$, $\Langdual{{\mathfrak{h}}}$ denote the Lie algebras of $\Langdual{G}$, $\Langdual{B}$, $\Langdual{T}$, respectively. The weights of $T$ can be identified with the elements $h \in \Langdual{{\mathfrak{h}}}$ such that $\alpha\postcheck(h)$ is integral for all coroots $\alpha\postcheck$ of $G$ (which are the roots of $\Langdual{G}$). This identification allows us to associate to a nilpotent orbit $\Langdual{\orb}$ in $\Langdual{{\mathfrak{g}}}$ a dominant weight for $G$. Namely, we can choose $e \in \Langdual{\orb}$ and let $e, h,
f$ span an $\Lie{sl}_2$-subalgebra of $\Langdual{\mathfrak{g}}$ with $h \in
\Langdual{{\mathfrak{h}}}$. Then by Dynkin-Kostant theory, $h$ is well-defined up to $W$-conjugacy, and by $\Lie{sl}_2$-theory, $h$ takes integral values at the coroots of $G$. Hence, $h$ determines an element of $\DomWeights$. We refer to the dominant weight of $G$ thus obtained as the Dynkin weight of $\Langdual{\orb}$, and we denote by $\mathcal
D \subset \DomWeights$ the set of Dynkin weights of all nilpotent orbits in $\Langdual{{\mathfrak{g}}}$. In the previous section, we saw that $\omega_0 = 2\rho_{\mathbf{d}}$ (made dominant for $GL(n)$). This is nothing more than the Dynkin weight of the nilpotent orbit $\orb_{{\mathbf{d}}\trans}$ (we are identifying $G$ with $\Langdual G$), since $\orb_{{\mathbf{d}}\trans}$ intersects the Lie algebra of $L_{\mathbf{d}}$ in the regular orbit. This result and a similar one for Richardson orbits in other groups (along with calculations in groups of low rank) have led a number of people to conjecture that $\mathcal D$ is a subset of ${\Weights^{loc}_{+}}$ (see [@chm-ostrik:dist-inv]). We wish to state a precise conjecture about how $\mathcal D$ sits inside of ${\Weights^{loc}_{+}}$. To this end, we assign to each $\Langdual{\orb}$ in $\Langdual{{\mathfrak{g}}}$ a finite cover of $\orb= d(\Langdual{\orb})$, where $\orb$ is the nilpotent orbit of ${\mathfrak{g}}$ dual to $\Langdual{\orb}$ under Lusztig-Spaltenstein duality. Our conjecture essentially says that if we write our putative finite cover of $\orb$ as $G/ K$, then for some $\tau$ which is trivial on $K$, $\gamma(\orb, \tau)$ equals the Dynkin weight of $\Langdual\orb$.
Let $A(\orb)$ denote the fundamental group of $\orb$ and let $\bar{A}(\orb)$ denote Lusztig’s canonical quotient of $A(\orb)$ (see [@lusztig:reductive], [@sommers:duality]). Let $\nilcone_{o,c}$ be the set of pairs $(\orb, C)$ consisting of a nilpotent orbit $\orb \in {\mathfrak{g}}$ and a conjugacy class $C \subset
A(\orb)$. We denote by ${\Langdual\nilcone}_{o}$ the set of nilpotent orbits in $\Langdual{\mathfrak{g}}$. In [@sommers:duality] a duality map $d:
\nilcone_{o,c} \to {\Langdual\nilcone}_{o}$ is defined which extends Lusztig-Spaltenstein duality. This map is surjective and the image of an element $(\orb, C) \in \nilcone_{o,c}$, denoted $d_{(\orb, C)}$, depends only on the image of the conjugacy class $C$ in $\bar{A}(\orb)$.
Now given $\Langdual{\orb} \in \Langdual{{\mathfrak{g}}}$, Proposition 13 of [@sommers:duality] exhibits an explicit element $(\orb, C) \in \nilcone_{o,c}$ such that $d_{(\orb, C)}=\Langdual{\orb}$. The orbit $\orb$ also satisfies $\orb= d(\Langdual{\orb})$ (in particular, $\orb$ is special). Consider the image of $C$ in $\bar{A}(\orb)$, which we will also denote by $C$. We suspect that this conjugacy class coincides with one that Lusztig associates to $\Langdual{\orb}$ using the special piece for $\Langdual{\orb}$ (see Remark 14 in [@sommers:duality]).
The canonical quotient $\bar{A}(\orb)$ of $A(\orb)$ is always of the form $S_3, S_4, S_5$ or a product of copies of $S_2$. Hence, it is possible to describe $\bar{A}(\orb)$ as a Coxeter group of type $A$ and then to associate to each conjugacy class $C$ in $\bar{A}(\orb)$ a subgroup $H_C$ of $\bar{A}(\orb)$ which is well-defined up to conjugacy in $\bar{A}(\orb)$. Lusztig did this for the exceptional groups in [@lusztig:notes] and we now do it for the classical groups.
First we need to describe $\bar{A}(\orb)$ as a Coxeter group in the classical groups (where $\bar{A}(\orb)$ is a product of copies of $S_2$). We use the description of $\bar{A}(\orb)$ in [@sommers:duality]. Let $\lambda = [\lambda_1^{a_1},
\lambda_2^{a_2}, \ldots, \lambda_k^{a_k}]$ be the partition corresponding to $\orb$ in the appropriate classical group of type $B, C,$ or $D$. Let $\mathcal{M}$ be the set of integers $m$ equal to some $\lambda_i$ such that $$\begin{aligned}
\lambda_i \text{ is odd and } \nu_i \text{ is odd} \text{ in type $B_n$} \\
\lambda_i \text{ is even and } \nu_i \text{ is even} \text{ in type $C_n$} \\
\lambda_i \text{ is odd and } \nu_i \text{ is even} \text{ in type $D_n$}
\end{aligned}$$ where $\nu_i = \sum^{i}_{j=1} a_j$. Then from section 5 of [@sommers:duality], we know that the elements of $\bar{A}(\orb)$ are indexed by subsets of $\mathcal{M}$ in type $C$ and by subsets of $\mathcal{M}$ of even cardinality in types $B$ and $D$. In type $C$ we choose our set of simple reflections in $\bar{A}(\orb)$ to correspond to subsets of $\mathcal{M}$ with a single element. In type $B$ and $D$ we choose our set of simple reflections in $\bar{A}(\orb)$ to correspond to subsets $\{ a, b \}$ of $\mathcal{M}$ with $a > b$ and where no element of $\mathcal{M}$ is both less than $a$ and greater than $b$. Thus given a conjugacy class $C$ of $\bar{A}(\orb)$ (which consists of a single element, $w$, since the group is abelian), we can write $w$ minimally as a product of simple reflections. The simple reflections used are unique, and we define $H_C$ to be the subgroup of $\bar{A}(\orb)$ generated by those simple reflections. Consider the surjection $\pi: G^{e} \to
\bar{A}(\orb)$ where $e \in \orb$ and define $K = \pi^{-1}(H_C)$ in $G^{e}$. We can now make our conjecture.
Given $\Langdual\orb$ in $\Langdual{\mathfrak{g}}$, we have assigned a conjugacy class $C$ in $\bar{A}(\orb)$ where $\orb = d(\Langdual\orb)$ and then a subgroup $K$ in $G^e$ where $e \in \orb$. Consider the finite cover $\tilde{\orb}= G/ K$ of $\orb$. Let $\C[\tilde{\orb}]$ denote the global algebraic functions on $\tilde{\orb}$. It is immediate that $\C[\tilde{\orb}] = \sum \Gamma( \orb, \mathcal L)$ where the sum is over the irreducible local systems $\mathcal L$ (counted with multiplicity) which arise from the irreducible representations of $A(\orb)$ appearing in $\Ind_{(G^e)^{0}}^{K} \C$, where $(G^e)^0$ is the identity component of $G^e$. Hence (\[eqn:vogan-thm\]) implies that as a $G$-module $\C[\tilde{\orb}]$ can be written as $\sum_{\lambda \in \DomWeights} m_\lambda \Ind_T^G
\lambda$. Let $\mu$ be a weight of largest length with $m_{\mu} \neq
0$.
\[conjecture:dynkin\] The weight $\mu$ is unique and is the Dynkin weight of $\Langdual{\orb}$.
We have verified the conjecture in a number of cases, although a general proof is elusive at the moment.
It seems likely that when $\tau$ gives rise to a local system, denoted ${\mathcal L}_{\tau}$, that $M(e, \tau)$ is just the direct image of ${\mathcal L}_{\tau}$ from $\orb_e$ to the whole nilpotent cone and so $\Sections(\nilcone, M(e, \tau)) = \Sections(\orb_e, {\mathcal
L}_{\tau})$ (this would be consistent with Bezrukavnikov’s and Ostrik’s work). Since $\C[\tilde{\orb}] = \sum \Gamma( \orb, \mathcal L)$, Conjecture \[conjecture:dynkin\] would then state that the Dynkin weight of $\Langdual{\orb}$ occurs as $\gamma(\orb, \tau)$ for some irreducible representation $\tau$ of $G^e$ which is trivial on $K$.
Specifying exactly what $\tau$ is seems to be more difficult. For example, let $\Langdual\orb$ be the subregular orbit in type $B_n$. Then $\orb = d(\Langdual {\orb})$ is the smallest non-zero special orbit in type $C_n$. This orbit is Richardson, coming from the parabolic subgroup with Levi factor of type $C_{n-1}$, and the parabolic subgroup gives rise to a $2$-fold cover of $\orb$. Thus it is clear that the Dynkin weight of $\Langdual{\orb}$ comes from this $2$-fold cover of $\orb$ (which is, in fact, the one specified by our conjecture) since $\Langdual{\orb}$ is regular in a Levi factor of type $B_{n-1}$. However, when $n$ is odd, the Dynkin weight will correspond to the trivial local system on $\orb$, but when $n$ is even, the Dynkin weight will correspond to the non-trivial local system on $\orb$ (see the calculations in [@chm-ostrik:dist-inv]).
Cohomology of the associated sub-bundles
========================================
For an element $h \in {\mathfrak{h}}$, we can define a subspace ${\mathfrak{n}}_{h}$ of the nilradical ${\mathfrak{n}}$ of ${\mathfrak{b}}$ as follows. We set $${\mathfrak{n}}_{h} = \bigoplus_
{\substack{\text{$\alpha$ a positive root}\\
\alpha(h) \ge 2}} \Lie{g}_\alpha$$ where $\Lie{g}_\alpha$ is the $\alpha$-eigenspace of the root $\alpha$. As in the previous section, we may identify $\DomWeights(\Langdual G)$ with a subset of ${\mathfrak{h}}$. Then for $\lambda \in \DomWeights(\Langdual G)$ we get a subspace ${\mathfrak{n}}_{\lambda}$ of ${\mathfrak{n}}$. Our definition is motivated by the fact that if $\lambda \in {\mathfrak{h}}$ happens to be a Dynkin weight for a nilpotent orbit $\orb \in {\mathfrak{g}}$, then by work of McGovern, $\C[G \times_B {\mathfrak{n}}_{\lambda}] \simeq \C[\orb]$ [@mcgovern:regfn] and moreover, by work of Hinich and Panyushev, $H^i(G/B, S^j({\mathfrak{n}}_{\lambda}^{*})) = 0$ for all $j \geq 0$ and $i>0$ [@hinich:van], [@panyushev:van]. Hence, it seems reasonable, especially given Conjecture \[conjecture:dynkin\], to pick a general element $\lambda \in {\Weights^{loc}_{+}}(\Langdual G)$ and study $H^i(G/B, S^j({\mathfrak{n}}_{\lambda}^{*}))$. Note that $\sum_{j \geq 0}
H^0(G/B, S^j({\mathfrak{n}}_{\lambda}^{*})) \simeq \C[G \times_B {\mathfrak{n}}_{\lambda}]$.
We begin with the definition
Two $B$-representations $V, {\tilde V}$ are called $G/B$-equivalent if $$H^i(G/B, S^j(V^*)) \simeq H^i(G/B, S^j({\tilde V}^*)).$$ for all $i,j \geq 0$.
Our main result is for $GL(n)$, and we identify $G$ with $\Langdual G$. We hope in future work to say something interesting for other groups.
\[main\] Fix the partition ${\mathbf{d}}$ and let ${\mathfrak{n}}_{{\mathbf{d}},p} = {\mathfrak{n}}_{\omega_p}$. For any $p$ the $B$-representations ${\mathfrak{n}}_{{\mathbf{d}},p}$ are all mutually $G/B$-equivalent to each other. Thus, we have for any $p$ that $$H^i(G/B, S^j({\mathfrak{n}}_{{\mathbf{d}},p}^*)) \simeq
H^i(G/B, S^j({\mathfrak{n}}_{{\mathbf{d}},0}^*))$$ and the latter equals $0$ if $i >0$ and equals $\C^j[\orb_{{\mathbf{d}}\trans}]$ if $i=0$ (the algebraic functions on the dual orbit of degree $2j$).
The last sentence is just a re-formulation of the Hinich, McGovern, and Panyushev results. Before proving the theorem, we prove two lemmas which rely on the following proposition.
\[prop:demazure\] Let ${\tilde V} \subset V$ be representations of $B$ such that $V/{\tilde V} \simeq \C_{\mu}$ is a one-dimensional representation of $B$ corresponding to the character $\mu$. Let $\alpha$ be a simple root, and let $P_\alpha$ be the parabolic subgroup containing $B$ and having $\alpha$ as the only positive root in its reductive part. If $V$ extends to a $P_\alpha$-representation and $\langle\alpha\postcheck, \mu\rangle = -1$, then $V, {\tilde V}$ are $G/B$-equivalent.
Consider the exact sequence $0 \to {\tilde V} \to V \to \C_{\mu} \to 0$ and its dual $0 \to \C_{-\mu} \to V^* \to {\tilde V}^* \to 0$. By Koszul, we have $0 \to S^{j-1}(V^*) \otimes \C_{-\mu}
\to S^j(V^*) \to S^j({\tilde V}^*) \to 0$ is exact for all $j \geq 0$. By the lemma of Demazure [@demazure:bott-simple], $H^*(G/B, S^{j-1}(V^*) \otimes \C_{-\mu}) = 0$ for all $j \geq 1$ since $V^*$ is $P_{\alpha}$-stable and $\langle\alpha\postcheck, -\mu\rangle = 1$ (here our $B$ corresponds to the positive roots, hence the difference with Demazure’s convention). The result follows from the long exact sequence in cohomology.
We need to introduce notation to describe the $B$-stable subspaces of ${\mathfrak{n}}$. It is clear that if ${\mathfrak{g}}_{\alpha}$ belongs to a $B$-stable subspace $U$ of ${\mathfrak{n}}$, then so does ${\mathfrak{g}}_{\beta}$ for all positive roots $\beta$ with $\alpha \preceq \beta$ (where $\preceq$ denotes the usual partial order on positive roots). Hence it is enough to describe $U$ by the positive roots $\alpha$ such that ${\mathfrak{g}}_{\alpha} \subset U$ and $\alpha$ is minimal among all positive roots with this property. In this case, we say that $\alpha$ is minimal for $U$.
List the simple roots of $G$ as $\alpha_1, \dots, \alpha_{n-1}$. Then any positive root of $G=GL(n)$ is of the form $\alpha_i +
\alpha_{i+1} + \cdots + \alpha_j$, which we denote by $[ i , j ]$. We can express the usual partial order on the positive roots as $[ i ,
j ] \preceq [i', j']$ if and only if $i' \le i$ and $j \le j'$. We can then specify $U$ by its minimal positive roots, namely a collection of intervals $[i,j ]$ such that for any two intervals $[
i,j ]$ and $[i' , j' ]$ with $i \leq i'$, we have $j \geq j'$. We will say that $U$ is partially specified by the interval $[i,j]$ if the root $[i,j]$ is minimal for $U$ (although there may be other minimal roots). Let us also say that $U$ is $i$-stable if $U$ is stable under the action of the parabolic subgroup $P_{\alpha_i}$.
Let $U$ be a $B$-stable subspace of ${\mathfrak{n}}$ which is partially specified by the interval $[a,b]$ and which is either $(a\!-\!1)$-stable or $(b\!+\!1)$-stable. Let $U'$ be the subspace of ${\mathfrak{n}}$ which is specified by the same intervals as $U$ except that $[a,b]$ is replaced by the two intervals $[a-1, b]$ and $[a, b+1]$. Then $U$ and $U'$ are $G/B$-equivalent. This is simply an application of Proposition \[prop:demazure\] where $\mu$ is the root $[a,b]$ and $\alpha$ is either $\alpha_{a-1}$ or $\alpha_{b+1}$. We refer to the $G/B$-equivalence of $U$ and $U'$ as the *basic move* for $a-1$ (respectively, for $b+1$). We now state two lemmas which rely solely on the basic move.
\[lemma1\] Let $U$ be a subspace of ${\mathfrak{n}}$ which is partially specified by an interval $[a,b]$ and such that $U$ is $i$-stable for $b < i < d$ for some $d$. Let $U'$ be the subspace of ${\mathfrak{n}}$ specified by the same intervals as those defining $U$, but replacing $[a,b]$ with the collection of intervals $$\{ [a-j+1, d-j] \ | \ 1 \leq j \leq d-b \}.$$ Then $U$ is $G/B$-equivalent to $U'$.
Applying the basic move for $b+1$ replaces $[a,b]$ with the intervals $[a-1, b] \cup [a, b+1]$. Now apply the basic move repeatedly on the right (for $i$ in the range $b+2 \leq i <d$) to the interval $[a, b+1]$ and we are left with the two intervals $[a-1, b] \cup [a, d-1]$. The general result follows by induction on $d-b$: we apply the proposition to the interval $[a-1, b]$ with $d$ replaced by $d-1$. The base case $d-b=1$ is trivially true.
\[lemma2\] Let $U$ be a subspace of ${\mathfrak{n}}$ partially specified by the intervals $$[b_0, b_1] \cup [b_1+1,b_2] \cup \dots \cup
[b_{l-1}+1, b_l] \cup [b_l+1, b_{l+1}]$$ and $$[b_1, b_2-1] \cup \dots \cup [b_{l-1}, b_l-1]$$ where $b_j \leq b_{j+1}-2$ for $1 \leq j \leq l-1$. Assume that no interval of $U$ is of the form $[b_l, a_1]$ but that there are intervals of $U$ of the form $[b_0-1, a_2]$ and $[a_3, b_{l+1}+1]$. Let $U'$ be the subspace of ${\mathfrak{n}}$ specified by the same intervals as $U$, except that $[b_0, b_1]$ is replaced by $[b_0, b_1-1]$ and $[b_l+1, b_{l+1}]$ is replaced by $[b_l, b_{l+1}]$. Then $U$ is $G/B$-equivalent to $U'$. A similar statement holds even if the we omit the interval $[b_l+1, b_{l+1}]$ or $[b_0, b_1]$ from the specification of $U$.
Let $U_1$ be specified by the same intervals as $U$, except with the above intervals replaced by the intervals $$[b_0, b_1] \cup [b_1+1,b_2-1] \cup [b_2+1, b_3-1]\cup \dots \cup [b_{l-1}+1, b_l-1] \cup [b_l+1, b_{l+1}]$$ Then $U_1$ is seen to be $G/B$-equivalent to $U$ by applying the the basic move to the roots $[b_1+1,b_2-1], [b_2+1, b_3-1], \dots, [b_{l-1}+1, b_l-1]$ since $U_1$ is stable for $b_2$, $b_3$, $\dots$ , $b_l$. The stability for $b_l$ follows from the assumption that no interval of $U$ is of the form $[b_l, a_1]$.
Let $U_2$ be specified by the same intervals as $U_1$, except we replace the interval $[b_0, b_1]$ with $[b_0, b_1-1]$ and the interval $[b_l+1, b_{l+1}]$ with $[b_l, b_{l+1}]$. Then $U_2$ is $G/B$-equivalent to $U_1$. This can be seen by applying the basic move to $[b_0, b_1-1]$ (as $U_2$ is stable for $b_1$) and applying the basic move (in reverse) to $[b_l, b_{l+1}]$ (as $U_1$ is $b_l$-stable). Here we are using the fact that $U$ contains intervals of the form $[b_0-1, a_2]$ and $[a_3, b_{l+1}+1]$.
The proof is completed by observing that $U_2$ is stable for $b_1$, $b_2$, $\dots$, $b_{l-1}$, so we can apply the basic move to the roots $[b_1+1,b_2-1], [b_2+1, b_3-1]$, $\dots$, $[b_{l-1}+1, b_l-1]$, arriving at $U'$.
For an orbit corresponding to ${\mathbf{d}}$ and a local system corresponding to $p$, we have computed $\omega_p$ in Section 2. We recall that $c$ is the g.c.d. of the parts of ${\mathbf{d}}$ and $m_a$ was defined by the equation $m_a c= \sum_{i \geq 0} \mu_{a+2i+1}$, where $\mu_i$ is the multiplicity of $i$ as a part in ${\mathbf{d}}\trans$. Write $k$ for one less than the largest part of ${\mathbf{d}}\trans$, and set $s_i = \sum_{j \geq i}
m_j c$.
Then by Proposition \[prop:omega\], ${\mathfrak{n}}_{{\mathbf{d}},p} =
{\mathfrak{n}}_{\omega_p}$ is specified by the set of intervals $$I_i =
[s_{i+1} + p m_{i}, s_{i} + p m_{i-1}]$$ where $k \geq i \geq -k+1$. The difference between the left endpoint of $I_i$ and the right endpoint of $I_{i+2}$ will be denoted $\Delta_i$. So $$\Delta_i =
(c-p)m_{i+1} + p m_{i}.$$ For $i \geq 1$ we have $\Delta_i \leq \Delta_{-i} \leq \Delta_{i-2}$ since $m_{-i}=m_i$ and $m_i \leq m_{i-2}$ when $i \geq 1$.
*Step 1.* Application of Lemma \[lemma1\] to each of the intervals $I_i$ shows that ${\mathfrak{n}}_{{\mathbf{d}},p}$ is $G/B$-equivalent to the subspace $W$ of ${\mathfrak{n}}$ defined by the set of intervals $$I_{i,j} = [s_{i+1}+pm_i -
j+1, s_{i-1}+pm_{i-2} -j]$$ where $1 \leq j \leq \Delta_i$ if $k \geq
i \geq 1$ and where $1 \leq j \leq \Delta_{i-2}$ if $0 \geq i \geq
-k+1$. In the former case, if $\Delta_{i} < j \leq \Delta_{i-2}$, then the intervals $I_{i,j}$ are not minimal, so we may omit them from our specification of $W$.
*Step 2.* Now for each $r$ in the range $k \geq r \geq 1$, starting with $r=k$ and working down to $r=1$, we will modify the intervals $I_{r+1,j}$ and $I_{-r+1, j}$ for $j > pm_{r+1}$ (in the cases where those intervals are defined) to obtain a new subspace of ${\mathfrak{n}}$ which is $G/B$-equivalent to $W$.
First, we will modify $I_{-r+1, j}$ for $\Delta_{r+1} < j \leq
\Delta_{-r-1}$. Consider the intervals $$I_{r-1,j} \cup I_{r-3,j}
\cup \dots \cup I_{-r+5,j} \cup I_{-r+3, j}$$ for $\Delta_{-r-1} < j
\leq \Delta_{r-1}$ and the intervals $$I_{r-1,j} \cup I_{r-3,j} \cup \dots \cup
I_{-r+3,j} \cup I_{-r+1, j}$$ for $\Delta_{r+1} < j \leq
\Delta_{-r-1}$. These yield a situation where we can apply Lemma \[lemma2\] a total number of $\Delta_{r-1} -\Delta_{-r-1} = p
(m_{r-1} - m_{-r-1})$ times to each of the intervals $I_{-r+1,j}$ for $\Delta_{r+1} < j \leq \Delta_{-r-1}$. This will replace $I_{-r+1,j}=[s_{-r+2} + p m_{-r+1} - j+1, s_{-r} + p m_{-r-1}-j]$ with $[s_{-r+2} +pm_{-r-1} -j +1, s_{-r} +pm_{-r-1}-j]$. Second, we will modify both $I_{r+1,j}$ and $I_{-r+1, j}$ for $pm_{r+1}<j \leq \Delta_{r+1}$ Consider the intervals $$I_{r-1,j} \cup
I_{r-3} \cup \dots \cup I_{-r+5,j} \cup I_{-r+3, j}$$ for $\Delta_{r+1} < j \leq (\Delta_{r-1} -\Delta_{-r-1})+\Delta_{r+1}$ and the intervals $$I_{r+1,j} \cup I_{r-1,j} \cup \dots \cup I_{-r+3,j}
\cup I_{-r+1,j}$$ for $pm_{r+1} < j \leq \Delta_{r+1}$. We again apply Lemma \[lemma2\] a total number of $\Delta_{r-1}
-\Delta_{-r-1}$ times by modifying the pair of intervals $I_{r+1,j}
\cup I_{-r+1,j}$. Then for $pm_{r+1} < j \leq \Delta_{r+1}$ we replace $I_{r+1,j} \cup I_{-r+1,j}$ with $[s_{r+2} + pm_{r+1} -j
+1,s_{r} +pm_{r+1} -j] \cup [s_{-r+2} + pm_{-r-1} -j +1, s_{-r}
+pm_{-r-1}-j]$. We do this for all $r$ starting with $r=k$ and working backward to $r=1$ (note we haven’t done anything to the intervals $I_{1,j}$).
*Step 3.* At this point our subspace of ${\mathfrak{n}}$ is specified by the following intervals: $[s_{i} - l +1, s_{i-2} - l]$ for $i \geq 2$ with $1 \leq l
\leq (c-p)m_{i}$ and for $1 \geq i$ with $1 \leq l \leq (c-p)m_{i-2}$, and those $I_{i,j}$ where $j \leq pm_i$ for $i \geq 1$ and $j \leq
pm_{i+2}$ for $i \leq 0$. Fix $r \geq 1$. We may assume by induction on $r$ that our subspace is $G/B$-equivalent to one partially specified by the following intervals (the case $r=1$ is already true): $[s_{i} - l +1, s_{i-2} - l]$ for $1 \leq i \leq r$ and $1 \leq l \leq cm_{i}$; and $[s_{i} - l +1, s_{i-2} - l]$ for $0 \geq i \geq -r+2$ and $1 \leq l \leq cm_{i-2}$. We want to show that the our subspace is $G/B$-equivalent to one partially specified by the previous intervals together with the cases where $i=r+1$ and $i=-r+1$ in the above formulas. Let $J_{i,l} = [s_{i+1} - l +1, s_{i-1} - l]$. Consider the intervals $$I_{r+1,j} \cup J_{r-2, j + (c-p)m_{r-1}} \cup \dots
\cup J_{-r+2 ,j + (c-p)m_{r-1}} \cup I_{-r+1,j}$$ for $1 \leq j \leq pm_{r+1}$ and the intervals $$J_{r-2, l} \cup \dots \cup J_{-r+2,l}$$ for $1 \leq l \leq (c-p)m_{r-1}$. We apply Lemma \[lemma2\] a total number of $(c-p)(m_{r-1} - m_{r+1})$ times to the pair of intervals $I_{r+1,j} \cup I_{-r+1,j}$ for $1 \leq j \leq pm_{r+1}$. We thus replace $I_{r+1,j} \cup I_{-r+1,j}$ with $[s_{r+2} + pm_{r+1} -j +1, s_{r-1}+(p-c)m_{r+1}-j] \cup
[s_{-r+1}+(p-c)m_{-r-1} -j+1, s_{-r}+pm_{-r-1}-j]$. These intervals are just $[s_{r+1} -l +1, s_{r-1}-l] \cup
[s_{-r+1} -l +1, s_{-r-1} -l]$ for $(c-p)m_{r+1} <l \leq cm_{r+1}$. Hence, by induction on $r$ we see that our original subspace is $G/B$-equivalent to the subspace specified by the intervals $J_{i,l}$ where $1 \leq l \leq cm_{i+1}$ for $i \geq 1$ and where $1 \leq l \leq cm_{i-1}$ for $i \leq 0$. This subspace is independent of $p$ which completes the proof.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
Plutinos are Kuiper-belt objects that share the 3:2 Neptune resonance with Pluto. The long-term stability of Plutino orbits depends on their eccentricity. Plutinos with eccentricities close to Pluto (fractional eccentricity difference $\Delta e/e_p=|e-e_p|/e_p\simless 0.1$) can be stable because the longitude difference librates, in a manner similar to the tadpole and horseshoe libration in coorbital satellites. Plutinos with $\Delta
e/e_p\simgreat 0.3$ can also be stable; the longitude difference circulates and close encounters are possible, but the effects of Pluto are weak because the encounter velocity is high. Orbits with intermediate eccentricity differences are likely to be unstable over the age of the solar system, in the sense that encounters with Pluto drive them out of the 3:2 Neptune resonance and thus into close encounters with Neptune. This mechanism may be a source of Jupiter-family comets.
author:
- Qingjuan Yu and Scott Tremaine
nocite:
- '[@M93; @M95]'
- '[@M96]'
title: The dynamics of Plutinos
---
=cmcsc10
\#1\#2[.5pt]{} \#1\#2 \#1
Introduction
============
The orbit of Pluto has a number of unusual features. It has the highest eccentricity ($e_p=0.253$) and inclination ($i_p=17.1^\circ$) of any planet in the solar system. It crosses Neptune’s orbit and hence is susceptible to strong perturbations during close encounters with that planet. However, close encounters do not occur because Pluto is locked into a 3:2 orbital resonance with Neptune, which ensures that conjunctions occur near Pluto’s aphelion ([@CH65]). More precisely, the critical argument $3\lambda_p-2\lambda_n-\varpi_p$ librates around $180^\circ$ with a period of $1.99\times 10^4$ yr and an amplitude of $82^\circ$; here $\lambda_p$ and $\lambda_n$ are the mean longitudes of Pluto and Neptune and $\varpi_p$ is Pluto’s longitude of perihelion. Other resonances are present with longer periods: for example, Pluto’s argument of perihelion librates with an amplitude of $23^\circ$ and a period of 3.8 Myr. In part because of its rich set of resonances, Pluto’s orbit is chaotic, although it exhibits no large-scale irregular behavior over Gyr timescales (see [@MW97] for a comprehensive review of Pluto’s orbit). For reference, Pluto’s semimajor axis and orbital period are $a_p=39.774\au$ and $P_p=250.85$ yr.
The most compelling explanation for Pluto’s remarkable orbit was given by Malhotra (1993, 1995). Malhotra argues that Pluto formed in a low-eccentricity, low-inclination orbit in the protoplanetary disk beyond Neptune. Subsequent gravitational scattering and ejection of planetesimals in the disk by all four giant planets caused Neptune’s orbit to migrate outwards ([@FI84]). As its orbit expands, Neptune’s orbital resonances sweep through the disk, first capturing Pluto into the 3:2 resonance and then pumping up its eccentricity. If Pluto’s orbit was circular before capture, its present eccentricity implies that it was captured when Neptune’s semimajor axis was 0.814 times its current value or $24.6\au$ (eq. \[eq:renu\]). This process may also excite Pluto’s inclination although the details are less certain ([@M98]).
Malhotra’s argument predicts that most Kuiper belt objects with $30\au\simless
a\simless 50\au$ should also be captured—and presently located—in Neptune resonances ([@M95]). This prediction has proved to be correct: of the $\sim 90$ Kuiper belt objects with reliable orbits as of 1999 January 1, over 30% have semimajor axes within 1% of the 3:2 resonance (although this number is exaggerated by observational selection effects). These objects have come to be called Plutinos, since they share the 3:2 resonance with Pluto (see [@MDL99] for a recent review of the Kuiper belt).
Almost all studies of the dynamics of the Kuiper belt so far have neglected the gravitational influence of Pluto, because of its small mass ($M_p/M_\odot=7.40\times 10^{-9}$ for the Pluto-Charon system, [@S92; @TW97]). However, like the Trojan asteroids and Jupiter, or the Saturn coorbital satellites Janus and Epimetheus, Pluto and the Plutinos share a common semimajor axis and hence even the weak gravitational force from Pluto can have a substantial influence on the longitude of a Plutino relative to Pluto. A crude illustration of the importance of Pluto’s gravity is to note that the half-width of the 3:2 resonance, $(\Delta a/a)_{\rm res}\simeq 0.01$ for $0.2\simless e\simless 0.3$ (the maximum fractional amplitude of stable libration in semi-major axis, [@M96]), is only a few times larger than the Hill radius of Pluto, $(\Delta a/a)_{\rm H}=(M_p/3M_\odot)^{1/3}=0.0014$.
The goal of this paper is to explore the dynamical interactions between Pluto and Plutinos and their consequences for the present structure of the Kuiper belt. Section 2 provides an approximate analytical description of the interactions, §3 describes the results of numerical orbit integrations, and §4 contains a discussion.
Analysis
========
We examine a simplified model solar system containing only the Sun and Neptune, with masses $M_\odot$ and $M_n$; we assume that Neptune’s orbit is circular and neglect all orbital inclinations. We describe the motion of the Plutino using the canonical variables $$\begin{aligned}
x_1=(GM_\odot a)^{1/2},\qquad\qquad & & y_1=\lambda, \nonumber \\
x_2=(GM_\odot a)^{1/2}[1-(1-e^2)^{1/2}],\qquad & &
y_2=-\varpi,
\label{eq:canvar}\end{aligned}$$ where $a$ and $e$ are the semi-major axis and eccentricity, $\lambda$ and $\varpi$ are the mean longitude and longitude of perihelion. The same variables for Neptune or Pluto are denoted by adding a subscript “n” or “p”.
We consider only the resonant perturbations exerted by Neptune, which can only depend on angles as an even function of the combination $3y_1-2y_{n1}+y_2$. Thus the Hamiltonian of a Plutino may be written H\_0(,,t)=H\_K(x\_1)+A(,3y\_1-2y\_[n1]{}+y\_2)+B(,,t), where $H_K(x_1)=-\half(GM_\odot)^2/x_1^2$ is the Kepler Hamiltonian, $A$ is the resonant potential from Neptune, and $B(\bfx,\bfy,t)$ is the potential from Pluto. The same Hamiltonian describes the motion of Pluto if we set $B=0$.
Now impose a canonical transformation to new variables $(\bfJ,\bfw)$ defined by the generating function S(,,t)=J\_1(3y\_1-2y\_[n1]{}+y\_2)+J\_2(y\_2-y\_[p2]{}). Thus $w_1=\partial S/\partial J_1=3y_1-2y_{n1}+y_2$, $w_2=\partial S/\partial J_2=\frac{1}{3}(y_2-y_{p2})$, while $x_1=\partial S/\partial y_1=3J_1$, $x_2=\partial S/\partial y_2=J_1+\frac{1}{3}J_2$, and the new Hamiltonian is H(,,t)=H\_0+[St]{}= H\_K(3J\_1)-2y\_[n1]{}J\_1-y\_[p2]{}J\_2+A(3J\_1,J\_1+J\_2,w\_1)+B, \[eq:ham\] where $\dot y_{n1}$ is the mean motion of Neptune and $-\dot y_{p2}$ is the apsidal precession rate of Pluto.
To proceed further we make use of the fact that $B/A=\hbox{O}(M_p/M_n)\ll
1$. Thus we can divide the motion of a Plutino into “fast” and “slow” parts. The fast motion is determined by the Kepler Hamiltonian $H_K$ and the resonant potential $A$ (this is the opposite of normal usage, where the resonant perturbations from Neptune are regarded as “slow” compared to non-resonant perturbations). The slow variations are caused by the Pluto potential $B$.
First we examine the fast motion. We drop the potential $B$, so that $w_2$ is ignorable and $J_2=3x_2-x_1=(GM_\odot a)^{1/2}[2-3(1-e^2)^{1/2}]$ is a constant of the motion. Thus if Pluto was captured into resonance from a circular orbit with semimajor axis $a_i$, its present semimajor axis and eccentricity are related by e\^2=-(a\_i/a)\^[1/2]{}-(a\_i/a). \[eq:renu\]
We write $J_1=J_{1r}+\Delta
J_1$ where $J_{1r}$ is chosen to satisfy the resonance condition for the Kepler Hamiltonian, 2y\_[n1]{}=3y\_1=3 [dH\_Kdx\_1]{} J\_[1r]{}=x\_[1r]{}=[(GM\_)\^[2/3]{}(18y\_[n1]{})\^[1/3]{}]{}. \[eq:chosen\] Since $A/H_K=\hbox{O}(M_n/M_\odot)\ll1$ we expect $|\Delta J_1|\ll J_{1r}$, so we can expand $H_K$ to second order in $\Delta J_1$; dropping unimportant constant terms the fast motion is determined by the Hamiltonian $$\begin{aligned}
H_f(\Delta J_1,w_1) & = & \frac{9}{2}\left({d^2H_K\over
dx_1^2}\right)_{x_{1r}} (\Delta J_1)^2+A(3J_{1r}+3\Delta J_1,J_{1r}+\Delta
J_1+\ffrac{1}{3}J_2,w_1) \nonumber \\ & = & -{27\over 2a^2}(\Delta
J_1)^2+A(3J_{1r}+3\Delta J_1,J_{1r}+\Delta J_1+\ffrac{1}{3}J_2,w_1).
\label{eq:pend}\end{aligned}$$ The Hamiltonian is autonomous and hence has a conserved energy $E_f=H_f$ and action $I=(2\pi)^{-1}\oint \Delta J_1 dw_1$. The motion is along the level surfaces of $H_f$ in the $(\Delta J_1,w_1)$ plane and typically consists of either libration ($w_1$ oscillates between fixed limits) or circulation ($w_1$ increases or decreases without reversing), just as in the case of the pendulum Hamiltonian. The stable equilibrium solutions (i.e. zero-amplitude libration) are given by J\_1=[a\^29]{}([Ax\_1]{}+), =0, <0. \[eq:eq\] The slow motion is determined by averaging the Hamiltonian (\[eq:ham\]) over the fast motion: H\_s(J\_2,w\_2,t)=E\_f(J\_2)-y\_[p2]{}J\_2+B. \[eq:hamslow\] Here $\langle\cdot\rangle$ indicates an average over one period of the fast motion. The fast energy $E_f$ depends on $J_2$ through the constraint that the fast action $I$ is adiabatically invariant.
Solutions with zero-amplitude libration
---------------------------------------
The solutions to the fast and slow equations of motion are particularly simple in the case where the fast libration amplitude is zero for both Pluto and the Plutino. This approximation is not particularly realistic—the libration amplitude of Pluto is $82^\circ$—but illustrates the principal features of the Plutino motions.
In this case the fast energy is E\_f=A(3J\_[1r]{},J\_[1r]{}+J\_2,w\_[1r]{}) where $w_{1r}$ is the equilibrium angle given by equation (\[eq:eq\]) and we have dropped much smaller terms that are O$(A^2)$. For simplicity we shall assume that there is only one stable equilibrium point, that is, one solution to equations (\[eq:eq\]) for given $J_2$. At the equilibrium point the fast action is $I=0$, and the slow Hamiltonian (\[eq:hamslow\]) is H\_s(J\_2,w\_2,t)=A(3J\_[1r]{},J\_[1r]{}+J\_2,w\_[1r]{}) -y\_[p2]{}J\_2 + B(3J\_[1r]{},J\_[1r]{}+J\_2,y\_1,y\_2,t)+(A\^2), where $y_1=\frac{1}{3}(w_{1r}+2y_{n1}-y_2)$. Since Pluto also is assumed to have zero libration amplitude, $y_{p1}=\frac{1}{3}(w_{1r}+2y_{n1}-y_{p2})$. Thus 3(y\_[p1]{}-y\_1)=y\_2-y\_[p2]{}=3w\_2 (2); \[eq:resrel\] that is, the difference in longitude of perihelion between Pluto and the Plutino is three times the difference in mean longitude. The same result will hold true on average even if the libration amplitudes are non-zero.
Now let us assume in addition that the eccentricities of both Pluto and the Plutino are small. Since their semi-major axes are the same, the gravitational potential from Pluto at the Plutino may be written B=GM\_p([\_p|\_p|\^3]{}-[1|-\_p|]{}) = [GM\_pa]{}. Using equation (\[eq:resrel\]) this simplifies to B=[GM\_pa]{}(w\_2-[12|w\_2|]{} ). Moreover =[Ax\_2]{}(3J\_[1r]{},J\_[1r]{}+J\_[p2]{},w\_[1r]{}) +(A\^2); thus the slow Hamiltonian can be rewritten as $$\begin{aligned}
H_s(J_2,w_2)=A(3J_{1r},J_{1r}+\ffrac{1}{3}J_2,w_{1r})
-\ffrac{1}{3}J_2{\p A\over\p x_2}(3J_{1r},J_{1r}+\ffrac{1}{3}J_{p2},w_{1r})
\nonumber \\
+{GM_p\over a}\left(\cos w_2-{1\over 2|\sin\half w_2|} \right)
+\hbox{O}(A^2).\end{aligned}$$ The interesting behaviour occurs when the actions $J_2$ and $J_{p2}$ are similar (i.e. the eccentricities of Pluto and the Plutino are similar), so we write $J_2=J_{p2}+\Delta J_2$ and expand $A$ to second order in $\Delta
J_2$. Dropping unimportant constants and terms of O$(A^2)$ we have H\_s(J\_2,w\_2)=J\_2\^2A\_[22]{} (3J\_[1r]{},J\_[1r]{}+J\_[p2]{},w\_[1r]{}) +[GM\_pa]{}(w\_2-[12|w\_2|]{} ). \[eq:tro\] where $A_{22}=\p^2A/\p x_2^2$.
This Hamiltonian is strongly reminiscent of the Hamiltonian for a test particle coorbiting with a satellite, H\_c(x,w)=-[32a\^2]{}x\^2 +[GMa]{}(w-[12|w|]{} ); \[eq:hamc\] here $\Delta x=x_1-x_{s1}$ and $w=\lambda_1-\lambda_{s1}$ are conjugate variables, and the subscript $s$ denotes orbital elements of the satellite. In this case the torques from the satellite lead to changes in semi-major axis; for a Plutino the semi-major axis is locked to Neptune’s by the resonance, so torques from Pluto lead to changes in the eccentricity instead.
When $A_{22}<0$ many of the features of orbits in the slow Hamiltonian (\[eq:tro\]) follow immediately from the analogy with the Hamiltonian (\[eq:hamc\]), which has been studied by many authors (e.g. Yoder et al. 1983, Namouni et al. 1999). The trajectories are determined by the level surfaces of the Hamiltonian. The equilibrium solutions correspond to the triangular Lagrange points in the coorbital case: $\Delta J_2=0$, $w_2=\pm
60^\circ$, $H_s=-\half GM_p/a$; the eccentricity of the Plutino equals the eccentricity of Pluto, the mean longitude leads or lags by $60^\circ$, and the perihelia are $180^\circ$ apart. These solutions are maxima of the potential from Pluto. For smaller values of $H_s$, the orbits librate around the triangular points (“tadpole orbits”). Small-amplitude tadpole librations have frequency $\omega$ given by \^2=-A\_[22]{}[GM\_pa]{}. \[eq:freq\] The tadpole orbits merge at the separatrix orbit, $H_s=-\frac{3}{2}GM_p/a$; for this orbit the minimum separation is $w_{2,\rm min}=23.91^\circ$. Even smaller values of $H_s$ yield “horseshoe” orbits, with turning points at $\pm w_{2,\rm min}$ where $H_s=(GM_p/a)(\cos
w_{2,\rm min}-\half|\sin \half w_{2,\rm min}|^{-1})$. For all tadpole and horseshoe orbits, the maximum and minimum values of $\Delta J_2$ occur at $w_2=\pm 60^\circ$, and are given by J\_2=\^[1/2]{}. \[eq:sep\] Eventually the theory breaks down, when $w_{2,\rm min}$ is small enough that adiabatic invariance is no longer a valid approximation.
The resonant potential from Neptune
-----------------------------------
For quantitative applications we must evaluate the resonant Neptune potential $A(\bfx,w_1)$. For small eccentricities, the potential can be derived analytically, $$\begin{aligned}
A(\bfx,w_1) &= & -{GM_n\over a}\bigg[\half b_{1/2}^{(0)}(\alpha)
+\ffrac{1}{8}e^2(2\alpha D+\alpha^2D^2)b_{1/2}^{(0)}(\alpha) +
\half e(5+\alpha D)b_{1/2}^{(2)}(\alpha)\cos w_1 \nonumber \\
& & +\ffrac{1}{8}e^2(104 +22\alpha D+\alpha^2D^2)b_{1/2}^{(4)}(\alpha)
\cos 2w_1
+\hbox{O}(e^3)\bigg],
\label{eq:anal} \end{aligned}$$ where $\alpha=a_n/a<1$, $D=d/d\alpha$, b\_[1/2]{}\^[(j)]{}()=[1]{}\_0\^[2]{} [dj(1-2+\^2)\^[1/2]{}]{}is a Laplace coefficient, Neptune is assumed to be on a circular orbit, and inclinations are neglected.
Unfortunately, in the present case the high eccentricity of the Plutino orbits makes this expansion invalid. However, we may determine $A(\bfx,w_1)$ numerically for given actions $\bfx$ by averaging the gravitational potential from Neptune over $y_2$ at fixed $w_1$, that is, A(,w\_1)=-[GM\_na]{}F(,w\_1), \[eq:resdef\] where F(,w\_1)= [a6]{}\_0\^[6]{} dy\_2([1|\_n-|]{}-[\_n|\_n|\^3]{})\_[,w\_1,y\_2]{}+ ; \[eq:AF\] the unimportant constant is chosen so that $A=0$ for circular orbits, i.e. $F(x_1,0,w_1)=0$. It can be shown analytically that the contribution from the second (indirect) term in the integrand vanishes.
We shall also write A\_[22]{}=[\^2 Ax\_2\^2]{}-[M\_nM\_a\^2]{} F\_[22]{}(,w\_1) F\_[22]{}(,w\_1)=x\_1\^2[\^2F(,w\_1)x\_2\^2]{}. \[eq:twotwo\]
Figure \[fig:contone\] plots the contours of $F(\bfx,w_1)$ at the resonant semimajor axis $x_{1r}$, as obtained from equation (\[eq:AF\]). The potential is singular for collision orbits, which for small eccentricity satisfy w\_1=[a-a\_nea]{}. \[eq:coll\] The conditions (\[eq:eq\]) for stable zero-amplitude libration are satisfied if and only if $w_1=w_{1r}=\pi$ or $w_1=0$ and $e>1-a_n/a=0.237$. Figure \[fig:corot180\] and \[fig:corot0\] show examples of these two solutions, plotted in a reference frame corotating with Neptune. Orbits of the first kind are similar to Pluto’s, although with smaller libration amplitude (compare Fig. 4 of [@MW97]). Orbits of the second kind (Fig. \[fig:corot0\]) were discussed by Malhotra (1996), who calls them “perihelion librators”. We shall not discuss these further, since they do not appear to form naturally during resonance capture of initially circular orbits; moreover for $e\simgreat 0.35$ they are likely to be unstable, since they cross Uranus’s orbit and thus are subject to close encounters and collisions with that planet.
Figure \[fig:approx\] plots $F(x_{1r},x_2,\pi)$ and $F_{22}(x_{1r},x_2,\pi)$; at the eccentricity of Pluto, corresponding to $x_2/(GM_\odot a)^{1/2}=0.0325$, we have $F(x_{1r},x_2,\pi)=-0.313$ and $F_{22}(x_{1r},x_2,\pi)=79.3$. Thus, for example, the eccentricity oscillation in the separatrix orbit that marks the boundary between tadpole and horseshoe orbits has amplitude $\Delta e =0.007$ (eq. \[eq:sep\]) and the period of libration of small tadpole orbits is $2\pi/\omega=9.1\times
10^{7}$ yr (eq. \[eq:freq\]).
For our purposes it is sufficient to work with the following numerical approximation to the resonant potential: F(,w\_1)=-|1-4.222ew\_1|, \[eq:F\] where $x_2=x_{1r}[1-(1-e^2)^{1/2}]$. This approximation is chosen to match the resonant potential at the resonant semimajor axis $x_1=x_{1r}$; the dependence on the relative semimajor axes of Neptune and the Plutino is suppressed since the effects of this potential are only important near resonance. The logarithmic factor is chosen to reproduce the singularity in the resonant potential near the collision orbits defined approximately by equation (\[eq:coll\]). The approximation formula also matches the analytic formula (\[eq:anal\]) to O$(e^2)$ at $w_1=\pi$.
Figure \[fig:conttwo\] shows the contour plot analogous to Figure \[fig:conttwo\] for the approximate resonant potential $\widetilde F$, and the triangles in Figure \[fig:approx\] show $\widetilde F$ and $\widetilde
F_{22}$. The agreement is very good, especially considering that errors are amplified by taking the two derivatives required to generate $\widetilde
F_{22}$.
Numerical experiments
=====================
We follow the orbital evolution of Pluto and a Plutino in a simplified version of the Sun-Neptune-Pluto-Plutino four-body system that isolates the resonant potential from Neptune. Neptune is assumed to have a circular orbit that migrates outward according to the rule ([@M93]) a\_n(t)=a\_f-a(-t/), where $a_f=30.17\au$ is Neptune’s present semimajor axis, $\Delta a=6\au$, and $\tau=1.5$ Myr. Thus Neptune’s initial semimajor axis is $24.17\au$ and the initial location of the 3:2 orbital resonance is $31.67\au$.
The initial eccentricity of Pluto is taken to be zero and its initial semimajor axis is $33\au$, as required so that its present eccentricity matches the observed value (eq. \[eq:renu\]). We followed 160 test particles, with initial semimajor axes distributed uniformly in the range $[31\au,39\au]$ and eccentricities distributed uniformly in the range $[0,0.03]$. The inclinations of Pluto and the test particles are chosen randomly in the range $[0,3^\circ]$ and their angular elements are chosen randomly from $[0,2\pi]$. Pluto and the test particles feel the resonant potential from Neptune, as defined by equations (\[eq:resdef\]) and (\[eq:F\]), but no other Neptune forces. The effects of the resonant Neptune potential on the orbital elements of Pluto and the test particles are followed using Lagrange’s equations.
The test particles do not influence Pluto or one another. However, they are subject to the gravitational potential from Pluto, B(,,t)=-GM\_p([1|-\_p|]{}-[\_p|\_p|\^3]{}); \[eq:B\] the effects of this potential on the orbital elements of the test particles are followed using Gauss’s equations.
The evolution of Pluto and the test particles is followed for 0.45 Gyr or 10% of the age of the solar system.
Results
=======
Of the 160 test particles, all but 12 are captured into the 3:2 resonance with Neptune, in the sense that their final semimajor axes are close to $(3/2)^{2/3}a_n$ and their eccentricities are near the prediction of equation (\[eq:renu\]), as shown in Figure \[fig:renu\]. The 12 particles that are not captured lie inside the location of Neptune’s 3:2 resonance at the start of the calculation, $(3/2)^{2/3}\times
24.17\au=31.67\au$, and would presumably be captured into other resonances if we used the full Neptune potential to work out their motion. We have verified this presumption by conducting shorter integrations ($1\times10^7 \yr$) using the same initial conditions but the complete Neptune potential. In this case all but 15 of the 160 particles were captured into the 3:2 resonance; the remainder were captured into the 4:3, 5:3 or 7:5 resonances.
The behavior of the test particles in the 3:2 resonance (henceforth Plutinos) falls into the following broad classes:
- Tadpole orbits (5 particles): these have longitude difference $y_1-y_{p1}$ and differences in longitude of perihelion $y_{p2}-y_2$ that librate around the leading or trailing Lagrange point of Pluto (Figure \[fig:tadpole\]). (Note that the libration center for the orbit in this Figure is $y_1-y_{p1}\simeq 100^\circ$, not $60^\circ$ as implied by the analysis in §2.1. This discrepancy arises because Pluto has a high eccentricity, while our analysis is only valid for near-circular orbits. Similarly, the perihelion difference librates around $\varpi-\varpi_p\simeq 300^\circ$, three times the difference in mean longitude as required by eq. \[eq:resrel\].) The tadpoles show no evidence of chaotic behavior or secular evolution over the length of our integration. The analysis in §2.2 suggests that the maximum eccentricity difference for these orbits is $\Delta e\simeq 0.007$; this requires in turn that their initial semimajor axes must have been close to Pluto’s, as is seen to be the case in Figure \[fig:renu\].
- Horseshoe orbits (19 particles): the longitude difference oscillates around $180^\circ$, with jumps in the Plutino eccentricity at the extrema of the longitude oscillation, as predicted by the analysis of §2.1 (Figures \[fig:horseshoe1\], \[fig:horseshoe2\]). The motion appears stable over the length of our integration although there are significant variations in semimajor axis oscillations during the course of the integration, and some horseshoes may evolve into transitional orbits over longer time intervals.
- Transitional orbits (43 particles): these show irregular behavior or transitions between libration and circulation of the longitude difference (Figure \[fig:trans\]). When the longitude difference circulates, the particles are no longer protected from close encounters with Pluto. However, the particles remain in the 3:2 Neptune resonance in the sense that the resonant angle $w_1$ continues to librate.
- Doubly transitional orbits (2 particles): Like transitional orbits, these show libration-circulation transitions in the longitude difference, but in addition they show irregular behavior in the resonant angle $w_1$, leading eventually to a transition of $w_1$ from libration to circulation (Figure \[fig:dtrans\]). Although only 2 particles in our simulation exhibit this behavior, a number of others show growing amplitude in the $w_1$ libration and will probably move into this class in less than the age of the solar system. Such orbits are normally short-lived since once $w_1$ circulates, they are no longer protected from close encounters with Neptune.
- Irregular circulating orbits (17 particles): the longitude difference circulates throughout the integration. Pluto induces irregular behavior (Figure \[fig:unaffect2\]), but the Neptune resonance is preserved in the sense that $w_1$ continues to librate, at least over the span of our integration.
- Regular circulating orbits (62 particles): the longitude difference circulates throughout the integration, but the orbits appear fairly regular (Figure \[fig:unaffect1\]). Generally, the orbits with larger eccentricity differences are more regular, because the encounter velocity with Pluto is higher so the perturbations from close encounters are smaller.
These classes represent a sequence in eccentricity difference: the typical eccentricity difference $|e-e_p|$ is smallest for tadpoles and largest for orbits unaffected by Pluto. Because the Plutino eccentricity is determined by the semimajor axis at the time of resonant capture (eq. \[eq:renu\]), the classes also reflect the initial semimajor axes of the Plutinos: the tadpoles and horseshoes all have initial semimajor axes in the range 32.2–34.2 (i.e. close to Pluto’s initial semimajor axis of 33). The transitional and irregular circulating Plutinos mostly have initial semimajor axes in the range 31.7–36, and the regular circulating Plutinos have initial semimajor axes concentrated in the range 35–39.
Discussion
==========
Test particles captured into the 3:2 Neptune resonance (Plutinos) have a complex range of dynamical interactions with Pluto. The strength of the interaction depends on the difference in eccentricity between the test particle and Pluto, and thus on the difference in initial semi-major axis if the initial orbits were circular and capture occurred through outward migration of Neptune. Plutinos are stable only if the eccentricity difference $\Delta e$ is small ($\Delta e\simless 0.02$ from Figure \[fig:renu\]), in which case the Plutinos librate on tadpole or horseshoe orbits; or if the eccentricity difference is large ($\Delta e\simgreat 0.06$), in which case the longitude difference circulates but relative velocity at encounter is high enough that Pluto has little effect. Unstable orbits at intermediate $\Delta
e$ can be driven out of the 3:2 Neptune resonance by interactions with Pluto, and thereafter are short-lived because of close encounters with Neptune. Thus we expect that the population of Plutinos has decayed over time, although determining the survival fraction will require integrations over the lifetime of the solar system using the full Neptune potential.
The long-term behavior of orbits in the 3:2 Neptune resonance is central to the origin of Jupiter-family comets. The usual explanation is that slow chaotic diffusion and collisional kicks drive Plutinos out of the 3:2 resonance, after which they are subjected to close encounters with the giant planets and eventually evolve into Jupiter-family comets ([@mo97]). Our results suggest that Pluto-induced evolution of Plutinos onto Neptune-crossing orbits may contribute to or even dominate the flux of Jupiter-family comets.
Our results also enhance the motivation to obtain accurate orbits for Kuiper-belt objects, and fuel speculation that the formation of the Pluto-Charon binary may be linked to interactions between Pluto and Plutinos.
This research was supported in part by NASA Grant NAG5-7310. We thank Matt Holman, Renu Malhotra, and Fathi Namouni for discussions and advice.
[99]{}
Cohen, C. J., & Hubbard, E. C. 1965, AJ, 70, 10
Fernández, J. A., & Ip, W.-H. 1984, Icarus, 58, 109
Malhotra, R. 1993, Nature, 365, 819
Malhotra, R. 1995, AJ, 110, 420
Malhotra, R. 1996, AJ, 111, 504
Malhotra, R. 1998, 29th Annual Lunar and Planetary Science Conference, Houston, TX, abstract no. 1476
Malhotra, R., & Williams, J. G. 1997, in Pluto and Charon, eds. S. A. Stern and D. J. Tholen (Tucson: University of Arizona Press), 127
Malhotra, R., Duncan, M., & Levison, H. 1999, in Protostars and Planets IV.
Morbidelli, A. 1997, Icarus, 127, 1
Namouni, F., Christou, A. A., & Murray, C. D. 1999, astro-ph/9904016
Stern, S. A. 1992, ARAA, 30, 185
Tholen, D. J., & Buie, M. W. 1997, in Pluto and Charon, eds. S. A. Stern and D. J. Tholen (Tucson: University of Arizona Press), 193
Yoder, C. F., Colombo, G., Synnott, S. P., & Yoder, K. A. 1983, Icarus, 53, 431
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- |
Alexander Astashkevich$^{1}$, Nikita Nekrasov$^{2}$ and Albert Schwarz$^{1}$\
\
[^1]\
\
$^{1}$ Department of Mathematics, UC Davis\
$^{2}$ Lyman Laboratory of Physics, Harvard University\
$^{2}$ Institute of Theoretical and Experimental Physics, Moscow
title: |
\[0pt\]
[l]{} hep-th/9810147
\
On noncommutative Nahm transform.
---
\[section\] \[section\] \[section\] \[section\] \[section\]
\[section\] \[section\] \[section\]
\[section\] \[section\] \[section\] ¶[[P]{}]{}
Motivated by the recently observed relation between the physics of $D$-branes in the background of $B$-field and the noncommutative geometry we study the analogue of Nahm transform for the instantons on the noncommutative torus.
Introduction {#sec1}
============
It is shown in [@cds] that noncommutative geometry can be successfully applied to the analysis of M(atrix) theory. In particular, it is proven that one can compactify M(atrix) theory on a noncommutative torus; later, compactifications of this kind were studied in numerous papers. In present paper, we analyze instantons in noncommutative toroidal compactifications. This question is important because instantons can be considered as BPS states of compactified M(atrix) model. Instantons on a noncommutative ${\rm I\!R}^4$ were considered earlier in [@ns]. It is shown in [@ns] that these instantons give some insight in the structure of (2,0) super-conformal six dimensional theory; the instantons on a noncommutative torus also should be useful in this relation. The main mathematical tool used in [@ns] is the noncommutative analogue of ADHM construction of instantons. The present paper is devoted to the noncommutative analogue of Nahm transform (recall that the Nahm transform can be regarded as some kind of generalization of ADHM construction). We prove that some of important properties of Nahm transform remain correct in noncommutative case.
Preliminaries. {#sec2}
==============
In this section we recall several notions related to the theory of noncommutative tori. We roughly discuss the ideas behind the noncommutative Nahm transform and formulate our main results. A more formal approach to the noncommutative Nahm transform is taken in the next section.
\[tor\] An $n$-dimensional noncommutative torus ${\cal A}_{\theta}$ is a $C^*$-algebra having unitary generators $U_{\bf i}$, ${\bf i}\in{\rm Z\!\!Z}^n$ obeying $$U_{\bf i}U_{\bf j}=e^{\frac{i}{2}\theta({\bf i,
j})}U_{\bf i+j},
{}~~~~~~~
U_{\bf 0}=1;$$ where $\theta({\bf \cdot , \cdot})$ is a skew-symmetric bilinear form on ${\rm Z\!\!Z}^n$.
We can naturally consider $\theta$ as a skew-symmetric bilinear form on ${{\rm I\!R}}^n$. Any element of $\At$ can be uniquely represented as a sum $\sum_{{\bf i}\in{\rm Z\!\!Z}^n} c_{\bf i}U_{\bf i}$, where $c_{\bf i}$ are complex numbers. Let $e_k$, $1\leq k\leq n$ be the natural base in ${\rm Z\!\!Z}^n$. The transformations $\delta _l
U_{e_k}=\delta _{l,k}U_{e_k}$, $1\leq k,l\leq n$ generate an abelian Lie algebra $L_{\theta}$ of infinitesimal automorphismes of ${\cal A}_{\theta}$. We use $L_{\theta}$ to define the notion of connection in a ${\cal
A}_{\theta}$-module following [@Con1] (we do not need the general notion of connection [@Con2]).
Any element from $\At$ can be considered as a function on the $n$-dimensional torus whose Fourier coefficients are $c_{\bf
i}$ (see above). The space of smooth functions on $T^n$ forms a subalgebra of $\At$. We denote it by ${\At}^{smooth}$ and call it the smooth part of $\At$. If $E$ is a projective finitely generated $\At$ module one can define its smooth part $E^{smooth}$ in a similar manner (see [@Rf1]). Now we can define the notion of $\At$ connection.
\[def0\] [*$\At$ connection on a right $\At$ module $E$ is a linear map $\nabla$ from $L_{\theta}$ to the space $\mbox{\rm End}_{\At}E$ of endomorphisms of $E^{smooth}$, satisfying the condition $$\n_{\delta}(ea)=(\n_{\delta}e)a+e(\delta(a)),$$ where $e\in E^{smooth}$, $a\in\At^{smooth}$, and $\delta\in
L_{\theta}$. The curvature $F_{\mu,\nu}=[\n_{\mu},\n_{\nu}]$ of connection $\n$ is considered as a two-form on $\Lt$ with values in endomorphisms of $E$.*]{}
We always consider Hermitian modules and Hermitian connections. This means that if $E$ is a right $\At$ module it is equipped with $\At$ valued Hermitian inner product $\langle\cdot,\cdot\rangle$ (for the detailed list of properties see [@Bl]) and all connections that we will consider should be compatible with this inner product.
If $E$ is endowed with a $\At$-connection, then one can define a Chern character $$\label{eqchern}
\mbox{\rm ch}(E)=\sum_{k=0}\frac{\hat{\tau}(F^k)}{(2\pi
i)^k k!},$$ where $F$ is a curvature of a connection on $E$, and $\hat{\tau}$ is the canonical trace on $\hat{A}=\mbox{\rm End}_{\At}(E)$ (we use that $\At$ is equipped with a canonical trace $\tau=c_0$). One can consider $\mbox{\rm ch}(E)$ as an element in the Grassmann algebra $\Lambda(\Lt^*)$. However, it is convenient to use the fact that this Grassmann algebra can be identified with cohomology $H( T,{\rm I\!\!\!C})$ where $ T$ stands for the Lie group of automorphisms of the algebra $\At$ corresponding to the Lie algebra $L_\theta$. (In other words $T=L_\theta / D$ where $D$ is a lattice.) In the commutative case ${\rm ch}(E)$ is an integral cohomology class. In noncommutative case this is wrong, but there exists an integral cohomology class $\mu (E) $ $\in H(T,{\rm Z\!\!Z})$ related to $\mbox{\rm ch}(E)$ by the following formula (see [@Ell], [@Rf1]) $$\label{eqrelchmu}
\mbox{\rm ch}(E)=e^{\iota(\theta)}\mu(E),$$ Here ${\iota(\theta)}$ stands for the operation of contraction with ${\theta}$ considered as an element of two-dimensional homology group of $T$. In particular, formula (\[eqrelchmu\]) means that $e^{-\iota(\theta)}\mbox{\rm ch}(E)\in\ H(T,{\rm Z\!\!Z})$.
One can regard $\mu(E)$ as a collection of integer quantum numbers characterizing topological class of a gauge field on noncommutative torus (or from mathematical viewpoint as a K-theory class of projective module $E $.)
The formula (\[eqrelchmu\]) is familiar to physicists [@D-WZ] in the following ($T$-dual) form: $$\label{phys}
{\cal L}_{WZ}= \int_{X} \nu\wedge \tilde C, \,\,\ \nu =
\mbox{\rm
ch}(E)
e^{- \frac{B}{2\pi i}}
\sqrt{{\hat A}(X)}$$ The element $\nu$ of the cohomology group $H^{\rm
even}(X, {{\rm I\!R}})$ is called the generalized Mukai vector. Here ${\cal L}_{WZ}$ describes the so-called Wess-Zumino couplings on the worldvolume $X$ of a $D$-brane (more precisely of a stack of $D$-branes), $E$ is the Chan-Paton bundle (or sheaf), $\tilde C$ is the collection of all Ramond-Ramond potentials, ${\hat A}(X)$ is the A-roof genus and $B$ is the Neveu-Schwarz $B$-field. The formula (\[phys\]) captures the effect of the non-trivial topology of the Chan-Paton bundle on the $D$-brane charges induced on the brane. Since in our case $X$ is a torus with flat metric then ${\hat A}(X) =
1$ and we arrive at (\[eqrelchmu\]) provided that we performed a $T$-duality transformation which maps $B$-field two form into a bivector $\theta$ and also exchanges ${\rm ch}_{k}$ with ${\rm ch}_{\frac{n}{2}-k}$.
\[inst\] An instanton is a connection such that the self-dual part of its curvature is a scalar operator, i.e., $F^+_{\alpha\beta}$ is a multiplication operator by the scalar that we denote $-\omega^+_{\alpha\beta}$.
We are interested in instantons on four-dimensional noncommutative torus. In the framework of supersymmetric gauge theory they can be interpreted as BPS-fields. Notice that in the definition of Hodge dual $\star F$ we need an inner product on the Lie algebra $\Lt$; we fix such a product.
As in commutative case we can prove that the minimum of euclidean action $A(E,\nabla)=\frac{1}{8\pi^2} \hat{\tau}(F\wedge \star F)[T]$ for connections in module $E$ (i.e. for gauge fields with given topological numbers) is achieved on instantons or anti-instantons (connections where antiselfdual part of curvature is a scalar). Here $[T]$ stands for the fundamental homology class of $T$.
\[action\] The expression of the instanton action in terms of $\mu$ and $\theta$ has the following form : $$\label{eqforact}
A(E,\nabla)=\left(\mu_2(E)-
\frac{i}{2\pi }
\frac{\left(\mu_1(E)^{+}+\left(
\iota(\theta)\mu_2(E)\right)^{+}\right)^{2}}{\mu_{0}(E) +
\iota(\theta)\mu_1(E) +
\frac{1}{2} \iota(\theta)^2 \mu_2(E)} \right)[T],$$ where $\mu_k(E)$ is $2k$-dimensional component of $\mu(E)$.
To prove this formula we notice that if $(E,\nabla)$ is an instanton ([*i.e.*]{} $F^++\omega^+=0$) the we can express $*F$ as follows $$\label{eqinst}
\star F=-F-2\omega^+.$$ Using (\[eqinst\]) we obtain $$\hat{\tau}(F\wedge \star F)=\hat{\tau}(F\wedge
[-F-2\omega^+])=
-\hat{\tau}(F\wedge F)-2\hat{\tau}(F)\wedge\omega^+,$$ since $\omega$ is complex valued two-form hence $$\hat{\tau}(F\wedge\omega^+)=\hat{\tau}(F)\wedge\omega^+ = -
\hat{\tau}(1) \omega^+ \wedge \omega^+.$$ One can easily obtain from the formula (\[eqrelchmu\]) that $\hat{\tau}(F\wedge F)=-8\pi^2\mu_2(E)$, $\hat{\tau}(F)=2\pi i (\mu_1(E)+\iota(\theta)\mu_2(E))$. $\Box$.
We will construct a generalization of Nahm’s transform [@Na], [@Dk] relating connections on ${\cal
A}_{\theta}$-modules with connections on ${\cal
A}_{\hat{\theta}}$-modules. (Here ${\cal A}_ {\theta}$ and ${\cal A}_{\hat
{\theta}}$ are two four-dimensional noncommutative tori.) To define a noncommutative generalization of Nahm’s transform we need a $({\cal A}_{\theta}, {\cal A}_{\hat {\theta}})$-module ${\cal P}$ with ${\cal A}_{\theta}$-connection $\nabla$ and ${\cal A}_{\hat
{\theta}}$-connection $\hat{\nabla}$. Speaking about $({\cal
A}_{\theta},{\cal A}_{\hat {\theta}})$-module ${\cal P}$ we have in mind that ${\cal P}$ is a left ${\cal A}_{\theta}$-module and a right ${\cal
A}_{\hat {\theta}}$-module; and $(ax)b=a(xb)$ for $a\in {\cal
A}_{\theta},\ b\in {\cal A}_{\hat {\theta}},\ x\in
{\cal P}$. We assume that the commutators $[\nabla _{\alpha}, \nabla _{\beta}],\ [\hat
{\nabla}_{\mu},
\hat {\nabla}_{\nu}],\ [\nabla _{\alpha},\hat
{\nabla}_{\mu}]$ are scalars: $$[\nabla _{\alpha}, \nabla _{\beta}]=\omega
_{\alpha\beta},\ [\hat
{\nabla}_{\mu}, \hat {\nabla}_{\nu}]=\hat {\omega}_{\mu\nu},\
[\nabla
_{\alpha},\hat {\nabla}_{\mu}]=\sigma _{\alpha\mu}.$$ One more assumption is that $\nabla _{\alpha}$ commutes with the multiplication by the elements of ${\cal A}_{\hat {\theta
}}$ and $\hat
{\nabla} _{\alpha}$ commutes with the multiplication by the elements of ${\cal
A}_{\theta }$. One can reformulate the above conditions saying that $\P$ is an ${\cal A}_{\theta\oplus(-\hat{\theta})}$-module, and $\nabla,\hat{\nabla}$ give us a constant curvature connection on it. For every right ${\cal A}_{\theta}$-module $R$ with connection $\nabla^R$ we consider Dirac operator $D=\Gamma ^{\alpha}(\nabla
_{\alpha}^R+\nabla _{\alpha})$ acting on the tensor product $$(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S$$ (or more precisely on its smooth part). To define $\Gamma$-matrices we introduce an inner product in $L_{\theta}$. $S$ is a ${\rm Z\!\!Z}/2{\rm Z\!\!Z}$ graded vector space $S=S^{+}\oplus S^-$ and Dirac operator is an odd operator. Thus, we can consider $$\label{eq10}
D^+:(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes
S^+\rightarrow
(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^-,\\
D^-:(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes
S^-\rightarrow
(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^+.$$
The Dirac operator commutes with the multiplication by the elements of ${\cal
A}_{\hat {\theta }}$, therefore the space of zero modes of $D$ can be regarded as ${\rm Z\!\!Z}/2{\rm Z\!\!Z}$ graded ${\cal A}_{\hat {\theta}}$-module; we denote it by ${\hat{R}}$. We will see later that sometimes it is reasonable to modify this definition of $\hat{R}$.
Next we would like to get a connection on $\hat{R}$. Using $\At$ Hermitian inner product on $R$, $\At\times\Aht$ Hermitian inner product on $\P$, and a canonical trace $\tau$ on $\At$ we can define an $\Aht$ Hermitian inner product on $(R\otimes_{{\cal A}_{\theta}}{\cal
P})\otimes S$. We assume that we have an orthogonal projection (with respect to the $\Aht$ Hermitian inner product) $P$ of $(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S$ onto $\hat{R}$. We will prove its existence in the next section under some additional conditions. The connection $\hat {\nabla}$ induces a connection $\hat
{\nabla}^{\prime}$ on $$(R\otimes _{{\cal A}_{\theta}}{\cal P})\otimes S;$$ we obtain a connection $\nabla^{\hat{R}}$ on $\hat {R}$ as $P\hat
{\nabla}^{\prime}$.
The above construction can be regarded as a generalized Nahm’s transform. To prove that its properties are similar to the properties of standard Nahm’s transform we should impose additional conditions on module ${\cal P}$ and connections $\nabla,\, \hat {\nabla}$. We assume that $\cal P$ is a projective Hermitian module over ${\cal A}_{\theta\oplus{(-\hat{\theta})}}$ and that $\sigma _{\alpha \mu}$ determines a non-degenerate pairing between $L_{\theta}$ and $L_{\hat {\theta}}$. Then we can use this pairing to define an inner product in $L_{\hat{\theta}}$; this allows us to talk about instantons on $\Aht$. Under certain conditions we prove that the Nahm transform of an instanton is again an instanton. More precisely, if $F^+_{\alpha\beta}+\omega^+_{\alpha\beta}=0$ where $\omega^+_{\alpha\beta}$ stands for the selfdual part of $\omega_{\alpha\beta}$, then $\hat{F}^+_{\alpha\beta}-
\hat{\omega}^+_{\alpha\beta}=0$. Here $F_{\alpha\beta}$ (correspondingly $\hat{F}_{\alpha\beta}$) is the curvature of the connection $\n^R$ (correspondingly $\nabla^{\hat{R}}$) on $R$ (correspondingly $\hat{R}$) and $^+$ stands for the self-dual part of it.
Notice that by taking the trace of the curvature of the connection $\nabla\oplus\hat{\nabla}$ on $\P$ we can express $\omega$ in terms of topological quantum numbers: $$\omega={2\pi i}\frac{\left(
\frac{1}{6}\iota(\tilde{\theta})^3\mu_4(\P)+
\frac{1}{2}\iota(\tilde{\theta})^2\mu_3(\P)+
\iota(\tilde{\theta})\mu_2(\P)+
\mu_1(\P)\right)|_{\Lt}}{\frac{1}{24}\iota(\tilde{\theta})^4\mu_4(\P)+
\frac{1}{6}\iota(\tilde{\theta})^3\mu_3(\P)+
\frac{1}{2}\iota(\tilde{\theta})^2\mu_2(\P)+
\iota(\tilde{\theta}) \mu_1(\P) + \mu_0(\P)} .$$ Here we use the notation $\tilde{\theta}$ for $\theta\oplus(-\hat{\theta})$, $\mu_k(\P)$ for $2k$-dimensional component of $\mu(\P)$.
In general, the Nahm transform defined above is not bijective (even in commutative case, i.e. when $\theta=\hat{\theta}=0$). However the commutative Nahm transform is bijective if $\P$ is “Poincare module” (the module corresponding to the Poincare line bundle). Strictly speaking, the term “Nahm transform” is used only in this situation. It is natural to define the Nahm transform in noncommutative case using an $\Adt$ module having the same topological numbers as Poincare module. (We will prove in Section \[sec5\] that the deformed Poincare module can be equipped with constant curvature connection and hence it can be used to define the Nahm transform.) One should expect that in noncommutative case the Nahm transform is a bijection (and its square is $(-1)^*$). We can give only heuristic proof of this conjecture.
[**Remark.**]{} The relation between topological quantum numbers of $\hat{R}$ and $R$ in the case when $\P$ is deformed Poincare module is the same as in commutative case (the relation between discrete quantities cannot change under continuous deformation): $$\begin{array}{l}
\mu(R)=p+\frac{1}{2}a_{ij}\, \alpha^{i}\wedge\alpha^{j}+q\,\,
\alpha^1\wedge
\alpha^2\wedge\alpha^3\wedge\alpha^4,\\[12pt]
\mu(\hat{R})=q-\frac{1}{4}
\epsilon^{ijkl}a_{ij}\, \beta_k\wedge\beta_l+
p\,\, \beta_1\wedge\beta_2\wedge\beta_3\wedge\beta_4
\end{array}$$ where $\epsilon^{ijkl}$ is completely antisymmetric tensor and the bases $\alpha^i$ and $\beta_i$ are the standard dual bases of $H^2(T,{\rm Z\!\!Z})$ and $H^2(\hat{T},{\rm Z\!\!Z})$.
We now present the formulae which relate $\omega$, $\hat{\omega}$, $\theta$, $\hat{\theta}$, etc. Recall that in the commutative case $L_{\theta} \approx L_{\hat\theta}^{\prime}$. The integral class $\mu(\P) = \exp {\sum_k \alpha^{k} \wedge
\beta_k}$ is the same as in the commutative case and it allows to identify $L_{\theta}^{\prime} \approx L_{\hat\theta}$ in the noncommutative case too. It is convenient to think of the forms $\omega, \hat\omega, \sigma$ as of the operators: $$\omega: L_{\theta} \to L_{\theta}^{\prime} \approx
L_{\hat\theta}, \,
\, \hat\omega: L_{\hat\theta} \to L_{\theta}, \,\, \sigma:
L_{\hat\theta} \to L_{\hat\theta}.$$ In the same way the bivectors $\theta,
\hat\theta$ are viewed as skew-symmetric operators: $\theta:
L_{\hat\theta} \to
L_{\theta}$, $\hat\theta:
L_{\theta} \to L_{\hat\theta}$. From (\[eqrelchmu\]) using the Wick theorem we get: $$\label{eqrelthom}
\begin{array}{l}
\mbox{\rm dim}\,\P= ~~
\sqrt{{\rm Det}(1 - \hat\theta\theta)},\\[12pt]
\omega = ~~\hat\theta ( 1 - \theta\hat\theta)^{-1} \\[12pt]
\hat{\omega} = ~~- \theta ( 1 - \hat\theta\theta)^{-1}\\[12pt]
\sigma = ~~( 1 - \hat\theta\theta)^{-1}\\[12pt]
\end{array}$$
Definitions. {#sec3}
============
We assume that all modules are Hermitian modules, all connections are Hermitian connections, and noncommutative tori $\At$ and $\Aht$ are four dimensional. Let $(\P,\n,\hat{\n})$ be a finitely generated projective $\Adt$ module equipped with constant curvature connection $\n\oplus\hat{\n}$. The curvature of $\n\oplus\hat{\n}$ is an element of $\wedge^2(\Lt)'\oplus(\Lt\otimes\Lht)'\oplus\wedge^2(\Lht)'$. We denote by $\omega_{\alpha\beta}$ the $\wedge^2(\Lt)'$ part of the curvature of $\n\oplus\hat{\n}$, by $\sigma_{\alpha\mu}$ the $(\Lt\otimes\Lht)'$ part, and by $\hat{\omega}_{\mu\nu}$ the $\wedge^2(\Lht)'$ part. We fix an inner product on $\Lt$. The inner product on $\Lht$ is obtained from the inner product on $\Lt$ via the pairing $\sigma_{\alpha\mu}$. Let $\{\alpha_i\}$ be an orthonormal basis of $\Lt$.
\[def1\] A connection $\CR$ on $\At$-module $R$ is called $\P$-irreducible iff there exists an inverse operator $G$ to the Laplacian $\Delta=\sum_{i}(\CR_{\alpha_i}+\n_{\alpha_i})(\CR_{\alpha_i}+\n_{\alpha_i})$ and $G$ is bounded operator.
\[lem1\] If $\CR$ is $\P$-irreducible connection on $R$ and $F^+
+\omega^+\cdot 1=0$ then $\mbox{\rm ker}(D^+)=0$, and $$\label{eq11}
D^- D^+=\Delta.$$
[**Proof:**]{} The proof is the same as in the commutative case. $\Box$
Now we can define a noncommutative Nahm transform. Let $R$ be a projective Hermitian module over $\At$ with $\P$-irreducible connection $\CR$ such that its curvature $F$ satisfies the condition $F^+
+\omega^+\cdot 1=0$. Denote by $\hat{R}$ the closure of the kernel of $D^-: (R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes
S^-\rightarrow
(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^+.$ Notice that $(R\otimes_{{\cal A}_{\theta}}{\cal
P})\otimes S^-$ is Hermitian $\Aht$ module and that $D^-$ commutes with $\Aht$ action. Therefore, $\hat{R}$ is a Hermitian $\Aht$ module (submodule of $(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^-$). Let us denote by $P$ the projection operator (with respect to the $\Aht$ Hermitian inner product) from $(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^-$ onto $\hat{R}$. In other words, $P$ is Hermitian, $P^2=P$, and $\mbox{\rm
Im}\,P=\hat{R}$. Its existence is proven in the theorem \[thm1\] below. We denote by $\ChR$ the composition $P\circ \hat{\n}$.
\[thm1\] $\hat{R}$ is a finitely generated projective Hermitian $\Aht$ module and $\ChR$ is a Hermitian $\Aht$ connection on $\hat{R}$.
[**Proof:**]{} The projection operator on the kernel of $D^-$ can be defined by the following explicit formula $P=1-D^+ G D^-$. We can check that $P$ is hermitian, $P^2=P$, and $\mbox{\rm Im}\,P=\mbox{\rm Ker}\,D^-$ by means of formal algebraic manipulations using $D^-D^+=\Delta$. We claim that $P$ is “compact” operator over $\Aht$ (that is a limit of the operators of the type $\sum_i x_i(y_i,\cdot)$). This follows immediately as usual from the fact that $D^+$ admits a parametrix (over $\Aht$, see Appendix A) and $G D^-$ is left inverse to $D^+$. Since $P$ is “compact” over $\Aht$ it follows from the general theory (see Appendix A or [@Bl]) that $P=\sum_j u_j(v_j,\cdot)$ is a projection on a finitely generated projective $\Aht$ module which proves the first statement. $\hat{R}$ inherits $\Aht$ valued inner product from $(R\otimes_{{\cal A}_{\theta}}{\cal P})\otimes S^-$.
The second statement follows immediately from the fact that $P$ commutes with the action of $\Aht$ (since $D^+$, $D^-$, and $G$ commute with $\Aht$ action). $\Box$
\[def3\] The noncommutative Nahm transform $\N$ of $(R,\CR)$ is a pair $(\hat{R},\ChR)$ of projective Hermitian $\Aht$ module $\hat{R}$ and connection $\ChR$.
Properties of noncommutative Nahm transform. {#sec4}
============================================
Now we can find a formula for the curvature of $\ChR$.
\[thm2\] We have the following formula for the curvature $\hat{F}$ of connection $\ChR$ $$\label{eq15}
\hat{F}_{\mu\nu}=\hat{\omega}_{\mu\nu}+PG\left(\sum_{\alpha,\beta}
(\Gamma^\alpha\Gamma^\beta-\Gamma^\beta\Gamma^\alpha)\sigma_{\alpha\mu}
\sigma_{\beta\nu}\right).$$
\[rem1\] It follows from (\[eq11\]) that $D^-D^+$ commutes with $\Gamma^\alpha$. Therefore $\Gamma^\alpha$ commutes with G. We will use this in the proof.
[**Proof:**]{} Since the Lie algebra $\Lht$ is commutative the curvature $\hat{F}_{\mu\nu}=[\ChR_\mu,\ChR_\nu]$. The definition of $\ChR$ gives us that $\hat{F}_{\mu\nu}=P\hat{\n}_\mu P \hat{\n}_\nu-
P \hat{\n}_\nu P\hat{\n}_\mu$. Let us simplify this expression using the definition of $P=1-D^+ G D^-$. If we replace the middle $P$ by this expression we obtain $$\label{eq16}
\begin{array}{l}
\hat{F}_{\mu\nu}=P [\hat{\n}_\mu,\hat{\n}_\nu]+P
(\hat{\n}_\nu D^+ G D^-\hat{\n}_\mu-\hat{\n}_\mu D^+ G
D^-\hat{\n}_\nu)
=\\[12pt]
{}~~~~~~~~~~=\hat{\omega}_{\mu\nu}+P
(\hat{\n}_\nu D^+ G D^-\hat{\n}_\mu-\hat{\n}_\mu D^+ G
D^-\hat{\n}_\nu).
\end{array}$$ Next, let us notice that $PD^+=0$ and $D^-$ equals zero on $\hat{R}$. Therefore, we can rewrite (\[eq16\]) as $$\label{eq17}
\hat{F}_{\mu\nu}=\hat{\omega}_{\mu\nu}+P
([\hat{\n}_\nu,D^+]G[D^-,\hat{\n}_\mu]-[\hat{\n}_\mu,D^+]G[D^-,\hat{\n}_\nu]).$$ To proceed further we need the following commutation relation $$\label{eq18}
[D,\hat{\n}_\lambda]=\sum_{\alpha}\Gamma^{\alpha}\sigma_{\alpha\lambda}.$$ (The proof is straightforward calculation using the definition of $D= \hfill \\
\sum_{\alpha}\Gamma^{\alpha}(\CR_\alpha+\n_{\alpha})$ and the fact that $\CR$ commutes with $\hat{\n}$.) The immediate consequence of (\[eq18\]) is that $$\hat{F}_{\mu\nu}=\hat{\omega}_{\mu\nu}+P
\left(-\{\sum_{\beta}\Gamma^{\beta}\sigma_{\beta,\nu}\}G
\{\sum_{\alpha}\Gamma^{\alpha}\sigma_{\alpha\mu}\}+
\{\sum_{\alpha}\Gamma^{\alpha}\sigma_{\alpha\mu}\}G
\{\sum_{\beta}\Gamma^{\beta}\sigma_{\beta\nu}\}\right).$$ Now the theorem follows from the remark that $\Gamma^\lambda$ commutes with $G$ and very simple formal manipulations. $\Box$\
Next, we would like to compute $\hat{F}^+$ (the self-dual part of the curvature). Let us remind that the inner product on $\Lht$ came from a non-degenerate pairing between $\Lt$ and $\Lht$ given by $\sigma_{\alpha\mu}$.
\[lem3\] The selfdual part of $\left(\sum_{\alpha,\beta}
(\Gamma^\alpha\Gamma^\beta-\Gamma^\beta\Gamma^\alpha)\sigma_{\alpha\mu}
\sigma_{\beta\nu}\right)$ on $S^-$ is equal to zero.
[**Proof:**]{} Let us choose an orthonormal basis $\{\alpha_i\}$ in $\Lt$. Let $\{\mu_j\}$ be the dual basis in $\Lht$ with respect to the pairing given by $\sigma_{\alpha\mu}$. Then, the element $$\left(\sum_{i,j}
(\Gamma^{\alpha_i}\Gamma^{\alpha_j}-\Gamma^{\alpha_j}\Gamma^{\alpha_i})
\sigma_{\alpha_i,\mu_k}\sigma_{\alpha_j,\mu_l}\right)=
\Gamma^{\alpha_k}\Gamma^{\alpha_l}-\Gamma^{\alpha_l}\Gamma^{\alpha_k}$$ operator It is well-known that this element is antiselfdual on $S^-$. Thus, the selfdual part of it is equal to zero. $\Box$\
As an immediate corollary we obtain that noncommutative Nahm transform is similar to a commutative Nahm transform in the following relation
\[thm2.5\] Let $(R,\CR)$ be a projective Hermitian module over $\At$ with $\P$-irreducible connection $\CR$ which satisfy the condition $F^++\omega^+ \cdot
1=0$, where $F$ is the curvature of $\CR$. Let $(\hat{R},\ChR)$ be a noncommutative Nahm transform of $(R,\CR)$. Then the curvature $\hat{F}$ of $\ChR$ satisfies the equation $\hat{F}^+-\hat{\omega}^+\cdot 1=0$.
[**Proof:**]{} The statement immediately follows from Proposition \[thm2\] and the previous lemma. $\Box$\
Let ${\hat{R}}^*$ be the left $\Aht$ module dual to $\hat{R}$, [*i.e.*]{}, ${\hat{R}}^*=\mbox{\rm Hom}_{\Aht}(\hat{R},\Aht)$. Notice that as a vector space ${\hat{R}}^*$ is isomorphic to $\hat{R}$ since $\hat{R}$ is a projective Hermitian $\Aht$ module. Consider the tensor product $\hat{R}\otimes_{\Aht}{\hat{R}^*}$. Since $\hat{R}$ is a finitely generated projective $\Aht$ module the algebra $\hat{R}\otimes_{\Aht}{\hat{R}^*}$ is naturally isomorphic to the algebra $\mbox{\rm End}_{\Aht}(\hat{R})$. Let $e$ be an identity element in $\mbox{\rm End}_{\Aht}(\hat{R})$. By abuse of notation we denote its image in $\hat{R}\otimes_{\Aht}{\hat{R}^*}$ by the same letter $e$.
\[rem2\] Notice that $e$ is a finite sum $\sum_i x_i\otimes y_i$, where $x_i\in{\hat{R}}$ and $y_i\in{\hat{R}^*}$, because $\hat{R}$ is a finitely generated projective Hermitian module over $\Aht$.
The module $\hat{R}$ was defined as a submodule of $(\RtP)\otimes S^-$. Therefore, we have a canonical embedding $\hat{R}\otimes_{\Aht}{\hat{R}^*}$ into $\left((\RtP)\otimes S^-\right)\otimes_{\Aht}{\hat{R}^*}$. Denote by $\Psi$ the image of $e$ in $\left((\RtP)\otimes S^-\right)\otimes_{\Aht}{\hat{R}^*}$. Let us notice that the module $\left((\RtP)\otimes S^-\right)\otimes_{\Aht}{\hat{R}^*}$ is canonically isomorphic to $(\RtP\otimes_{\Aht}{\hat{R}^*})\otimes
S^-$. We use the latter one everywhere instead of the former.
Denote by $R^*$ the left $\Aht$ module dual to ${R}$, $R^*=\mbox{\rm Hom}_{\At}({R},\At)$. Any element $f\in
R^*$ gives us a map from $(\RtP\otimes_{\Aht}{\hat{R}^*})\otimes S^-$ to $(\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-$. $$(\RtP\otimes_{\Aht}{\hat{R}^*})\otimes S^-\ni
(x\otimes p \otimes y)\otimes s\mapsto (f(x)p\otimes
y)\otimes s
\in (\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-,$$ where $x\in R$, $p\in\P$, $y\in{\hat{R}^*}$, and $s\in S^-$. We denote this map by $\bf f$.
Notice, that $G$ (inverse to $D^-D^+$) commutes with the action of $\Aht$. Therefore, it acts on $(R\otimes_{\At}\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-$. We would like to consider a canonical element $G\Psi\in
(R\otimes_{\At}\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-$. Strictly speaking the spinor spaces in the definitions of $D_{R}$ and $D_{\hat{R}^*}$ are different (one of them is constructed using $\Lt$, another one using $\Lht$). However, we may use $\sigma_{\alpha\mu}$ to identify Euclidean spaces $\Lt$ and $\Lht$ and hence the corresponding spinor spaces. Thus we can consider the Dirac operator $D^-_{\hat{R}^*}$ as a map from $(\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-$ to $(\P\otimes_{\Aht}{\hat{R}^*})\otimes S^+$. Notice that this identification does not respect the duality used in (\[eqrelthom\]).
\[prop2\] For any element $f\in R^*$ the element ${\bf f}(G\Psi)\in
(\P\otimes_{\Aht}{\hat{R}^*})\otimes S^-$ lies in the kernel of $D^-_{\hat{R}^*}$.
[**Proof:**]{} First, we need two lemmas.
\[lem4\] Let $\{\alpha_i\}$ be an orthonormal basis in $\Lt$ and $\{\mu_i\}$ be the dual basis (also orthonormal) in $\Lht$ (the pairing between $\Lt$ and $\Lht$ is given by $\sigma_{\alpha\mu}$). Then we have $$\label{eq21}
[\hat{\n}_{\mu_i},G]=2G(\CR_{\alpha_i}+\n_{\alpha_i})G.$$
[**Proof:**]{} Recall that $G$ was defined as an inverse operator to $D^-D^+$. From Lemma \[lem1\] we know that $D^-D^+=\sum_i(\CR_{\alpha_i}+\n_{\alpha_i})^*(\CR_{\alpha_i}+\n_{\alpha_i})=
\sum_i(\CR_{\alpha_i}+\n_{\alpha_i})(\CR_{\alpha_i}+\n_{\alpha_i})$, since all connections are by definition Hermitian, [*i.e.*]{}, selfadjoint. From (\[eq18\]) we obtain $$\label{eq22}
[D^-D^+,\hat{\n}_{\mu_j}]=[\sum_i(\CR_{\alpha_i}+\n_{\alpha_i})
(\CR_{\alpha_i}+\n_{\alpha_i}),\hat{\n}_{\mu_j}]=
2(\CR_{\alpha_j}+\n_{\alpha_j}),$$ since $\sigma_{\alpha_k,\mu_l}=\delta_{k,l}$. Multiplying the formula (\[eq22\]) by $G$ from the left and by $G$ from the right and using the fact that $G$ is inverse to $D^-D^+$, we obtain (\[eq21\]). $\Box$\
Let $e=\sum_i x_i\otimes y_i$ be an element of $\hat{R}\otimes_{\Aht}{\hat{R}}^*$ as in Remark \[rem2\]. Then, any element $z\in\hat{R}^*$ can be written as $$z=\sum_i (z,x_i)y_i$$ where $(z,x_i)=z(x_i)$ is an element of $\Aht$ and we think about $z$ as a homomorphism from $\hat{R}$ to $\Aht$. The goal of the next lemma is to describe connection $(\ChR)^*$ on $\hat{R}^*$.
\[lem5\] For any smooth element $z\in{\hat{R}^*}$ we have $$\label{eq23}
(\ChR_{\mu_i})^*z=-\sum_{k,l}(z,\hat{\n}_{\mu_i}x_k-
\Gamma^{\alpha_l}\Gamma^{\alpha_i}(\CR_{\alpha_l}+\n_{\alpha_l})Gx_k)y_k,$$ where the basises $\{\alpha_i\}$ and $\{\mu_i\}$ are chosen as in Lemma \[lem4\].
[**Proof:**]{} The proof is the following tedious trivial calculation. $$\begin{array}{l}
(\ChR_{\mu_i})^*z=\sum_k ((\ChR_{\mu_i})^*z,x_k)y_k=
-\sum_k (z,\ChR_{\mu_i}x_k)y_k
=-\sum_k (z,P\hat{\n}_{\mu_i}x_k)y_k=\\[12pt]
-\sum_k (z,(1-D^+GD^-)\hat{\n}_{\mu_i}x_k)y_k=
-\sum_k
(z,\hat{\n}_{\mu_i}x_k-D^+G[D^-,\hat{\n}_{\mu_i}]x_k)y_k,
\end{array}$$ since $D^-x_k=0$. From (\[eq18\]) and the choice of the basises $\{\alpha_i\}$ and $\{\mu_i\}$ it follows that $[D^-,\hat{\n}_{\mu_i}]=\Gamma^{\alpha_i}$. Therefore, we obtain $$\begin{array}{c}
(\ChR_{\mu_i})^*z=
-\sum_k
(z,\hat{\n}_{\mu_i}x_k-D^+G\Gamma^{\alpha_i}x_k)y_k=\\[12pt]
{}~~~~=-\sum_{k,l}
(z,\hat{\n}_{\mu_i}x_k-\Gamma^{\alpha_l}
\Gamma^{\alpha_i}(\CR_{\alpha_l}+\n_{\alpha_i})Gx_k)y_k,
\end{array}$$ since $G$ commutes with $\Gamma^\alpha$ (and we replaced $D^+$ by its definition $D^+=\sum_l
\Gamma^{\alpha_l}(\CR_{\alpha_l}+\n_{\alpha_l})$). $\Box$\
Now we prove the proposition. Let us choose the bases $\{\alpha_i\}$ and $\{\mu_i\}$ as in lemma \[lem4\]. First, recall that $D^-_{\hat{R}^*}=\sum_j \Gamma^{\mu_j}\left(\hat{\n}_{\mu_j}+
(\ChR_{\mu_j})^*\right)$. Since we identified the spinors for $\Aht$ with the spinors for $\At$ $\Gamma^{\mu_j}=\Gamma^{\alpha_j}$ and $D^-_{\hat{R}^*}=\sum_j
\Gamma^{\alpha_j}\left(\hat{\n}_{\mu_j}+
(\ChR_{\mu_j})^*\right)$. Second, recall that $\Psi=\sum_k x_k\otimes
y_k$, therefore $${\bf f}(G\Psi)={\bf f}(G\sum_k x_k\otimes y_k)=
\sum_k {\bf f}(Gx_k\otimes y_k),$$ since $G$ acts only on the first argument. Third, notice that $D^-_{\hat{R}^*}$ commutes with ${\bf f}$ therefore we obtain $$D^-_{\hat{R}^*}({\bf f}(G\Psi))={\bf f}\left[\sum_{j,k}
\Gamma^{\alpha_j}\left(\hat{\n}_{\mu_j}+
(\ChR_{\mu_j})^*\right)(Gx_k\otimes y_k)\right].$$ We continue our manipulations $$\label{eq24}
D^-_{\hat{R}^*}({\bf f}(G\Psi))={\bf f}\left[\sum_{j,k}
\Gamma^{\alpha_j}\left(\hat{\n}_{\mu_j}Gx_k\otimes y_k+
Gx_k\otimes(\ChR_{\mu_j})^* y_k\right)\right].$$ Using lemma \[lem5\] we can rewrite $$\begin{array}{l}
\sum_k Gx_k\otimes(\ChR_{\mu_j})^* y_k=
\\[12pt]
=\sum_k Gx_k\otimes(
-\sum_{l,m}(y_k,\hat{\n}_{\mu_j}x_m-
\Gamma^{\alpha_l}\Gamma^{\alpha_j}(\CR_{\alpha_l}+\n_{\alpha_l})Gx_m)y_m)=
\\[12pt]
=-\sum_{k,l,m} Gx_k ((y_k,\hat{\n}_{\mu_j}x_m-
\Gamma^{\alpha_l}\Gamma^{\alpha_j}(\CR_{\alpha_l}+\n_{\alpha_l})Gx_m)
\otimes y_m=\\[12pt]
=-\sum_{l,m}G(\hat{\n}_{\mu_j}x_m-
\Gamma^{\alpha_l}\Gamma^{\alpha_j}(\CR_{\alpha_l}+\n_{\alpha_l})Gx_m)\otimes
y_m=\\[12pt]
=-\sum_{k,l} \left(G\hat{\n}_{\mu_j}x_k-
\Gamma^{\alpha_l}\Gamma^{\alpha_j}G(\CR_{\alpha_l}+\n_{\alpha_l})Gx_k\right)
\otimes y_k,
\end{array}$$ in the last line we replaced everywhere $m$ by $k$. Substituting this into the formula (\[eq24\]) we get $$\label{eq25}
\begin{array}{l}
D^-_{\hat{R}^*}({\bf f}(G\Psi))={\bf f}\\[12pt]
\left[\sum_{j,k}
\Gamma^{\alpha_j}\left(\hat{\n}_{\mu_j}Gx_k\otimes y_k
-\sum_{k,l} \left(G\hat{\n}_{\mu_j}x_k-
\Gamma^{\alpha_l}\Gamma^{\alpha_j}G(\CR_{\alpha_l}+\n_{\alpha_l})Gx_k\right)
\otimes y_k\right)\right]\\[12pt]
={\bf
f}\left[\sum_k\left(\sum_j\Gamma^{\alpha_j}[\hat{\n}_{\mu_j},G]+
\sum_{j,l}\Gamma^{\alpha_j}\Gamma^{\alpha_l}\Gamma^{\alpha_j}
G(\CR_{\alpha_l}+\n_{\alpha_l})G\right)x_k\otimes
y_k\right].
\end{array}$$ Using lemma \[lem4\] we substitute $2G(\CR_{\alpha_j}+\n_{\alpha_j})G$ instead of $[\hat{\n}_{\mu_j},G]$ and obtain $$\label{eq26}
D^-_{\hat{R}^*}({\bf f}(G\Psi))=
{\bf f}\left[\sum_{k,l}\left(2\Gamma^{\alpha_l}+\sum_j
\Gamma^{\alpha_j}\Gamma^{\alpha_l}\Gamma^{\alpha_j}\right)
G(\CR_{\alpha_l}+\n_{\alpha_l})Gx_k\otimes y_k\right].$$ The proposition follows from the fact that $2\Gamma^{\alpha_l}+\sum_j
\Gamma^{\alpha_j}\Gamma^{\alpha_l}\Gamma^{\alpha_j}$ equals zero for all $l$, since the dimension is four. $\Box$
We interpret this proposition as existence of $\At$ homomorphism from $R^*$ to the kernel of $D^-_{\hat{R}^*}$. If we knew that ${\hat{R}^*}$ is $\P$-irreducible we could say that we have a $\At$ homomorphism from $R^*$ to $\widehat{{\hat{R}^*}}$.
Examples. {#sec5}
=========
In our previous discussion we assumed that we have a module $\cal P$ over ${\A}_{\theta\oplus(-\hat{\theta})}$ with some properties. In this section we give examples of such modules. Although the modules are quite simple the constructions of connections with desired properties are quite technical, and tedious.
At first we consider so called elementary finitely generated projective modules over ${\A}_{\theta\oplus{(-\hat{\theta})}}$ (see [@Rf1]). The second example shows how to deform the commutative Poincare module. Essentially we will show that the commutative Poincare module can be viewed as an elementary finitely generated projective module (in the language of [@Rf1]). This allows us to deform Poincare module and constant curvature connection on it using the constructions of the first part of the section.
In the first part we follow mainly the exposition of paper [@Rf1] with some minor modifications. We can think about $\theta$ and $\hat{\theta}$ as skew-symmetric bilinear forms on ${\rm Z\!\!Z}^4$. We embed ${\rm Z\!\!Z}^4$ in a usual way into ${\rm I\!R}^4$ and extend the forms $\theta$ and $\hat{\theta}$ to be skew-symmetric bilinear forms on ${\rm I\!R}^4$. For simplicity we assume that $\theta$ and $\hat{\theta}$ are non-degenerate symplectic forms. We can take their direct sum $\theta\oplus(-\hat{\theta})$ and consider it as a two-form on ${\rm I\!R}^8$. Now take another ${\rm I\!R}^8={\rm I\!R}^4\oplus({\rm I\!R}^4)'$ with a canonical skew-symmetric bilinear form $\Omega$ given by $\Omega
\left(\left(x_1,y_1\right),\left(x_2,y_2\right)\right)=
\langle y_2,x_1\rangle-\langle y_1,x_2\rangle$, where $x_1,x_2\in{\rm I\!R}^4$ and $y_1,y_2\in({\rm I\!R}^4)'$. Let $\eta$ be an arbitrary integral two form (by definition, it takes integer values on the lattice ${\rm Z\!\!Z}^8$) on ${\rm I\!R}^8$. Let $T$ be a linear map from ${\rm I\!R}^8$ to ${\rm I\!R}^4\oplus({\rm I\!R}^4)^{\prime}$ such that $\theta\oplus(-\hat{\theta})+\eta=T^*(\Omega)$. Such map always exists but it is never unique (any two maps of this kind are conjugate by an element from a symplectic group). Now we can describe some examples of modules $\P$.
We can realize $\P$ as a space of functions on ${\rm I\!R}^4$; the smooth part $\P^{smooth}=S({\rm I\!R}^4)$ is the space of Schwartz functions on ${\rm I\!R}^4$. Let us describe the left action of $\At$ and the right action of $\Aht$ (which commute with each other). First, notice that ${\rm I\!R}^4\oplus({\rm I\!R}^4)'$ acts on functions as follows: $\left(\left(x,y\right)f\right)\left(z\right)=e^{2\pi i
\langle y,z\rangle}f(z+x)$. The left action of an element $U_\nu\in \At$, $\nu\in{\rm Z\!\!Z}^4$ is given by $$(U_\nu f)(z)=(T(i_1(\nu))f)(z),$$ where $i_1$ is the canonical inclusion of ${\rm Z\!\!Z}^4\hookrightarrow {\rm I\!R}^4
\hookrightarrow {\rm I\!R}^4\oplus{\rm I\!R}^4={\rm I\!R}^8$ in the first ${\rm I\!R}^4$, [*i.e.*]{}, $i_1(\nu)=(\nu,0)$. Similarly, we define $i_2$ as the canonical inclusion of ${\rm Z\!\!Z}^4$ in the second ${\rm I\!R}^4$ in ${\rm I\!R}^8$, [*i.e.*]{}, $i_2(\nu)=(0,\nu)$. We define the right action of $\hat{U}_\nu\in \Aht$, $\nu\in{\rm Z\!\!Z}^4$ as $$(\hat{U}_\nu f)(z)=(T(i_2(\nu))f)(z).$$ A straightforward calculation shows that in this way we actually get an $(\At,\Aht)$ module. For the proof that $\P$ is projective $\Adt$ module see [@Rf1].
Next we define connections $\n$ and $\hat{\n}$ on $\P$. Notice that we can identify the Lie algebra $\Lt$ with $({\rm I\!R}^4)'$. If $\alpha\in ({\rm I\!R}^4)'$ then the corresponding derivation $\delta_\alpha$ acts on $U_\nu$, $\nu\in{\rm Z\!\!Z}^4$ as a multiplication by $2\pi i\langle \alpha,\nu\rangle$. Similarly, we can identify the Lie algebra $\Lht$ with $({\rm I\!R}^4)'$. For $\alpha\in ({\rm I\!R}^4)'$, we have $\delta_\alpha(\hat{U}_\nu)=2\pi i
\langle \alpha,\nu\rangle \hat{U}_\nu$.
Now, let us define for any $(x,y)\in {\rm I\!R}^4\oplus({\rm I\!R}^4)'$ an operator $Q_{(x,y)}$ on the smooth functions on ${\rm I\!R}^4$ as follows $$(Q_{(x,y)}f)(z)=2\pi i \langle y,z\rangle f(z)+
\frac{d(f(z+tx))}{dt}|_{t=0}.$$ The straightforward calculation shows that $$\begin{array}{l}
[Q_{(x,y)},U_\nu]=2\pi i
\Omega((x,y),T((\nu,0)))U_\nu,\\[12pt]
[Q_{(x,y)},\hat{U}_\nu]=2\pi i
\Omega((x,y),T((0,\nu)))\hat{U}_\nu.
\end{array}$$
We define connection $\n_\alpha$ as $Q_{V_1(\alpha)}$, where $V_1$ is the unique map from $\Lt=({\rm I\!R}^4)'$ such that $$\Omega(V_1(\alpha),T((\nu,\mu)))=\langle
\alpha,\nu\rangle,$$ where $\alpha\in \Lt=({\rm I\!R}^4)'$, and $\nu,\mu\in{\rm I\!R}^4$. Since $\Omega$ is a non-degenerate bilinear form, $T$ is an isomorphism and the existence and uniqueness of $V_1$ is obvious. From this definition it is clear that $\n$ commutes with $\Aht$. Similarly, we define $\hat{\n}_\alpha$ as $Q_{V_2(\alpha)}$, where $V_2$ is the unique map from $\Lht=({\rm I\!R}^4)'$ such that $$\Omega(V_2(\alpha),T((\nu,\mu)))=\langle
\alpha,\mu\rangle,$$ where $\alpha\in \Lht=({\rm I\!R}^4)'$, and $\nu,\mu\in{\rm I\!R}^4$. From this definition we immediately see that $\hat{\n}$ commutes with $\At$.
To compute the curvature of the connections $\n$ and $\hat{\n}$ and to compute the commutators $[\n_\alpha,\hat{\n}_\beta]$ we need to know the commutator $[Q_{(x_1,y_1)},Q_{(x_2,y_2)}]$. It is easy to see that $$[Q_{(x_1,y_1)},Q_{(x_2,y_2)}]=2\pi i
\Omega\left((x_1,y_1),(x_2,y_2)\right)\Id.$$ As an immediate corollary we obtain that connections $\n$ and $\hat{\n}$ have constant curvature and that $[\n_\alpha,\hat{\n}_\beta]=
2\pi i \Omega(V_1(\alpha),V_2(\beta))$. It is not hard to check that generically this pairing will be non-degenerate. Therefore, for generic $\theta$ and $\hat{\theta}$ we get many non-trivial examples of module $\P$.
The most interesting example of module $\P$ can be obtained by deforming the “Poincare module” - the space of sections of Poincare line bundle (see [@bbr], [@Dk]).
Let us remind one of the possible definitions of the Poincare module Let ${\rm Z\!\!Z}^4$ be a lattice in ${\rm I\!R}^4$. Let $({\rm Z\!\!Z}^4)'$ be the dual lattice in $({\rm I\!R}^4)'$. The Poincare module $\P$ consists of functions $f(x,\hat{x})$ on ${\rm I\!R}^4\oplus({\rm I\!R}^4)'$ which satisfy the following condition: $$\label{Poicbun}
f(x+\lambda,\hat{x}+\hat{\lambda})=e^{-2\pi
i\langle\hat{\lambda},x\rangle}
f(x,\hat{x}),$$ for all $\lambda\in{\rm Z\!\!Z}^4$ and $\hat{\lambda}\in({\rm Z\!\!Z}^4)'$. The algebra of functions on the torus considered as the algebra of periodic functions on ${\rm I\!R}^4\oplus({\rm I\!R}^4)'$; it acts on the module by multiplication.
We will use another construction of Poincare module (this construction is similar to the construction of modules over two-dimensional torus given in [@Ho]).
Notice that for fixed $\hat{x}$ the function $f(x,\hat{x})$ is a periodic function in $x$ which can be written as a Fourier series with $\hat{x}$-dependent coefficients: $$\label{Four}
f(x,\hat{x})=\sum_{\mu\in ({\rm Z\!\!Z}^4)'} e^{2\pi
i\langle\mu,x\rangle}
c_{\mu}(\hat{x}).$$ The coefficients $c_{\mu}(\hat{x})$ are smooth functions of $\hat{x}$ if the original function was smooth since they are given by the integrals over torus (which is compact). For convenience we denote $c_0(\hat{x})$ by $\phi(\hat{x})$. The property (\[Poicbun\]) of function $f(x,\hat{x})$ gives us that $$c_{\mu}(\hat{x})=c_0(\hat{x}+\mu).$$ Therefore, we can rewrite (\[Four\]) as $$\label{decomp}
f(x,\hat{x})=\sum_{\mu\in ({\rm Z\!\!Z}^4)'} e^{2\pi
i\langle\mu,x\rangle}
\phi(\hat{x}+\mu).$$ Moreover, it is not hard to see that function $f(x,\hat{x})$ given by the formula (\[decomp\]) is smooth if and only if the function $\phi(\hat{x})$ belongs to the Schwartz space.
The algebra ${\cal A}_0$ is generated by the elements $U_{\mu}=e^{2\pi i\langle\mu,x\rangle}$ and $\hat{U}_{\nu}=e^{2\pi i\langle\nu,\hat{x}\rangle}$, where $\nu\in{\rm Z\!\!Z}^4$ and $\mu\in ({\rm Z\!\!Z}^4)'$. We can rewrite the action of the operators $U_{\mu}$ and $\hat{U}_{\nu}$ in terms of their action on $\phi(\hat{x})$. One easily obtains $$\label{actionn}
\hat{U}_{\nu}(\phi(\hat{x}))=e^{2\pi
i\langle\nu,\hat{x}\rangle}\phi(\hat{x})
{}~~~~~~\mbox{\rm and}~~~~~~~
U_{\mu}(\phi(\hat{x}))=\phi(\hat{x}-\mu).$$
Thus, we realized the Poincare module as the space of functions on $({\rm I\!R}^4)'$ with the action given by the formula (\[actionn\]).
The constant curvature connection on the Poincare line bundle is given by $$\nabla=d+2\pi i\langle \hat{x},dx\rangle$$ in the first realization of Poincare module. We can easily rewrite it in terms of its action on the function $\phi(\hat{x})$. If $y\in{\rm I\!R}^4$ and $\hat{y}\in({\rm I\!R}^4)'$ then $$\label{connectn}
\nabla_y\phi(\hat{x})=2\pi i\langle
y,\hat{x}\rangle\phi(\hat{x})~~~~~~
\mbox{\rm and}~~~~~~~
\nabla_{\hat{y}}\phi(\hat{x})=\frac{d\,\phi(\hat{x}+t\hat{y})}{dt}|_{t=0}.$$
One can also easily rewrite the Hermitian inner product $\langle f(x,\hat{x}),\tilde{f}(x,\hat{x})\rangle$ in terms of $\phi(\hat{x})$ and $\tilde{\phi}(\hat{x})$. In particular, the trace of the inner product of $f(x,\hat{x})$ and $\tilde{f}(x,\hat{x})$ equals $$\label{innerpr}
\tau(\langle f(x,\hat{x}),\tilde{f}(x,\hat{x})\rangle)=
\int_{-\infty}^\infty
\phi(\hat{x})\overline{\tilde{\phi}(\hat{x})}d\hat{x}.$$
From the above discussion we see that Poincare module fits very well in the picture in the first example. More precisely, the map $T$ from the first example is $$T(x,y)=(x,y).$$ Here, we identified the dual space $({\rm I\!R}^4)'$ with ${\rm I\!R}^4$ using some fixed standard integral bases. We can easily calculate $\theta$, $\hat{\theta}$, and $\eta$ and we obtain that $\theta=\hat{\theta}=0$ and $\eta$ is the two form on ${\rm I\!R}^8$ given by $$\eta((x_1,y_1),(x_2,y_2))=\langle y_2,x_1\rangle-
\langle y_1,x_2\rangle,$$ where $x_i\in{\rm I\!R}^4\stackrel{(Id,0)}{\hookrightarrow}{\rm I\!R}^4\oplus{\rm I\!R}^4={\rm I\!R}^8$ and $y_i\in{\rm I\!R}^4\stackrel{(0,Id)}{\hookrightarrow}{\rm I\!R}^4\oplus{\rm I\!R}^4={\rm I\!R}^8$, $i=1,2$. The inner product $\langle\cdot,\cdot\rangle$ comes from the above identification of ${\rm I\!R}^4$ with $({\rm I\!R}^4)'$. It is obvious that $\eta$ is integral two form in this case.
We can easily deform the map $T$ so that the form $\eta$ will be preserved but the forms $\theta$ and $\hat{\theta}$ will be deformed. In this way we obtain a deformation of the commutative Poincare module. One can easily check that this deformation is always possible for small $\theta$ and $\hat{\theta}$. Using the above construction we obtain a constant curvature connection in the deformed Poincare module.
Explicitly, the equation (\[actionn\]) is deformed to: $$\label{actionnn}
\hat{U}_{\nu}(\phi(\hat{x}))=e^{2\pi
i\langle C\nu,\hat{x}\rangle}\phi(\hat{x} - A \nu)
{}~~~~~~\mbox{\rm and}~~~~~~~
U_{\mu}(\phi(\hat{x}))=e^{2\pi
i\langle B\mu,\hat{x}\rangle} \phi(\hat{x}-D \mu).$$ where $A,B, C, D$ are the operators: ${\rm I\!R}^4 \to (
{\rm I\!R}^4)^{\prime}$ such that: $$B^{t} D - D^{t} B = \frac{\theta}{2\pi i}, \,\, C^{t} A -
A^{t} C =
\frac{\hat\theta}{2\pi i}, \,\, B^{t} A - D^{t} C\in {\rm
GL}(4, {\rm Z\!\!Z})$$
Appendix A. {#sec7}
===========
In this section we will discuss under which conditions on the module $\P$ the kernel of $D^-$ (we assume that the kernel of $D^+$ is trivial) is a finitely generated projective module over $\Aht$. First, we will explain how to construct a parametrix to $D^-$ ($D$). More precisely, what we want to construct is a “compact” operator $Q$ commuting with the action of $\Aht$ such that $DQ=1+K_l$ and $QD=1+K_r$ and $K_l,K_r,DK_r,DK_l,K_rD,K_lD$ are “compact” operators. This is enough for all our purposes.
We would like to construct a parametrix to $D:\,
E\otimes_{\At}\P\otimes S
\rightarrow E\otimes_{\At}\P\otimes S$, where $E$ is a finitely generated projective module over $\At$ and $\P$ is a finitely generated projective module over $\At\times \Aht$. First, we can reduce the problem to the case when $E$ is a free $\At$ module. Indeed, since $E$ is a finitely generated projective $\At$ module it is a direct summand in $\At^k$ for some $k$. We have: $\At^k=E\oplus \tilde{E}$. Let $P_E$ be an orthogonal projection on $E$. Choose any $\At$ connection on $\tilde{E}$. Consider the Dirac operator $\hat{D}:
(E\oplus\tilde{E})\otimes_{\At}\P\otimes S\rightarrow
(E\oplus\tilde{E})\otimes_{\At}\P\otimes S$. Notice that $\hat{D}=D\oplus \tilde{D}$, where $\tilde{D}$ is the Dirac operator on $\tilde{E}\otimes_{\At}\P\otimes S$. Moreover, $P_E
\tilde{D}=\tilde{D}P_E=0$. Therefore, if $\hat{Q}$ is a parametrix to $\hat{D}$ then it is easy to see that $Q=P_E \hat{Q} P_E$ is a parametrix to $D$. More precisely, if $\hat{Q}\hat{D}=1+\hat{K}_1$ and $\hat{D}\hat{Q}=1+\hat{K}_2$, then $QD=1+K_1$ and $DQ=1+K_2$, where $K_1=P_E\hat{K}_1 P_E$ and $K_2=P_E\hat{K}_2 P_E$. Notice that $K_1$ and $K_2$ preserve the properties of $\hat{K}_1$ and $\hat{K}_2$.
Similarly, we can assume that $\P$ is a free module over $\At\times\Aht$. Therefore, we need to construct parametrix only in the case of Dirac operator $D$ on $(\At\times\Aht)^k\otimes S$. One can consider $(\At\times\Aht)$ as the space of functions on the product of tori $T^n\times T^n$. The Dirac operator $D$ in this case becomes a sum of a usual commutative Dirac operator $D_c$ along the first torus plus a bounded operator $A$ (preserving the space of smooth functions), that is $D=D_c+A$. Let $Q_c$ be a parametrix to $D_c$. Then, $Q_c D=Q_c D_c+Q_c A$. It follows immediately from the properties of $Q_c$ that $Q_c A$ is bounded operator from $H_{(0)}$ to $H_{(1)}$, where $H_{(l)}$ are Sobolev spaces of functions along the first torus. Thus, for any given natural number $l$ we can easily construct an operator $Q$ (maps $H_{(0)}$ to $H_{(1)}$) such that $QD=1+K_1$ and $K_1$ maps $H_{(0)}$ to $H_{(l)}$. Similarly, for any given natural number $l$ we can construct an operator $Q$ (maps $H_{(0)}$ to $H_{(1)}$) such that $DQ=1+K_2$ and $K_2$ maps $H_{(0)}$ to $H_{(l)}$. Moreover, one can see that $Q$ can be chosen to be the same. Such an operator $Q$ is enough for all our purposes. We call it parametrix here.
Second, let us prove the following lemma.
\[lemA1\] Let $P$ be an operator on $L^2({\rm I\!R})\times \Aht$ commuting with the action of $\Aht$. Assume that $P^*=P$, $P^2=P$ and $P$ is “compact” ([*i.e.*]{}, $P\in {\cal K}
\hat{\otimes}\Aht$, where $\cal K$ is the algebra of compact operators on $L^2({\rm I\!R})$). Then $P$ is a projection on a finitely generated projective $\Aht$ module.
[**Proof:**]{} First, we can approximate $P$ by a self-adjoint operator $A=\sum_j \langle f_j,\cdot\rangle g_j\times
a_j$, where $f_j,g_j\in L^2({\rm I\!R})$ and $a_j\in\Aht$ (since $P$ is self-adjoint). That is, $P=A+\epsilon$, where $\epsilon$ is a self-adjoint operator of norm less then $\frac{1}{100}$. It is obvious that $P$ is a projection on the kernel of $1-P$. We can write $1-P=1-\epsilon-A=(1-A(1-\epsilon)^{-1})(1-\epsilon)$. If $(1-P)x=0$ it means that $(1-A(1-\epsilon)^{-1})(1-\epsilon)x=0$. Denote by $y=(1-\epsilon)x$. Then $(1-A(1-\epsilon)^{-1})y=0$. From this we see that $y=\sum_j \langle
f_j,(1-\epsilon)^{-1}y\rangle
g_j\times a_j$. In particular this means that $y\in \sum_j
g_j\times\Aht$. The module $\sum_j g_j\times\Aht$ is a free finitely generated $\Aht$ submodule of $L^2({\rm I\!R})\times \Aht$. We denote it by $U$. Denote by $M$ the set $(1-\epsilon)^{-1}U$. Since $(1-\epsilon)^{-1}$ is invertible operator commuting with $\Aht$ we see that $M$ is $\Aht$ module. It is not hard to check (using the fact that $(1-\epsilon)^{-1}$ is self-adjoint and close to identity) that $M$ is a free hermitian finitely generated $\Aht$ module.
From the above it is obvious that if $(1-P)x=0$ then $x\in M$. Therefore, we can restrict projection $P$ on $M$ and we will get the same image. What is left to check is that $P|_M$ is a projection operator. We see that if $m\in M$ then $(1-P)Pm=
P m-P^2 m=Pm-Pm=0$. This implies that $P|_M$ is projection operator. Now, since $M$ is finitely generated free $\Aht$ module and $P:M\rightarrow M$ is a projection lemma is proved. $\Box$
We cannot directly apply the above lemma to the module $E\otimes_{\At}\P\otimes S$. But since $E\otimes_{\At}\P\otimes S$ can be considered a direct summand in $(\At\times\Aht)^k\otimes S$ apply Lemma \[lemA1\] to $(\At\times\Aht)^k\otimes S$.
Acknowledgments. {#last}
================
We are grateful to G. Elliott, D. Fuchs, A. Gorokhovsky, G. Kuperberg, A. Lawrence, A. Losev, M. Rieffel, S. Shatashvili and I. Zakharevich for discussions. The research of N. N. was supported by Harvard Society of Fellows, partially by NSF under grant PHY-98-02-709, partially by RFFI under grant 98-01-00327 and partially by grant 96-15-96455 for scientific schools. The research of A. S. was partially supported by NSF under grant DMS-9500704.
[10]{}
[^1]: ast@math.ucdavis.edu, nikita@string.harvard.edu, schwarz@math.ucdavis.edu
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The conversion of neutron matter to strange matter in a neutron star have been studied as a two step process. In the first step, the nuclear matter gets converted to two flavour quark matter. The conversion of two flavour to three flavour strange matter takes place in the second step. The first process is analysed with the help of equations of state and hydrodynamical equations, whereas, in the second process, non-leptonic weak interaction plays the main role. Velocities and the time of travel through the star of these two conversion fronts have been analysed and compared.'
title: 'The conversion of Neutron star to Strange star : A two step process'
---
[Abhijit Bhattacharyya$^{\textrm{1}}$, Sanjay K. Ghosh$^{\textrm{2,3}}$, Partha S. Joardar$^{\textrm{3}}$, Ritam Mallick$^{\textrm{2}}$and Sibaji Raha$^{\textrm{2,3}}$]{}
$^{\textrm{1}}$Department of Physics, University of Calcutta, 92, A.P.C Road, Kolkata - 700009, INDIA
$^{\textrm{2}}$Department of Physics, Bose Institute, 93/1, A.P.C Road, Kolkata - 700009, INDIA
$^{\textrm{3}}$Centre for Astroparticle Physics and Space Science, Bose Institute, 93/1, A.P.C Road, Kolkata - 700009, INDIA
I. Introduction {#i.-introduction .unnumbered}
================
It has been conjectured that strange quark matter, consisting of almost equal numbers of u, d and s quarks, may be the true ground state of strongly interacting matter [@key-1; @key-2] at high density and/or temperature. This conjecture is supported by bag model calculations [@key-3] for certain range of values for the strange quark mass and the strong coupling constant. By considering realistic values for the strange quark mass (150 - 200 MeV [@key-4]), it may be shown that the strangeness fraction in a chemically equilibrated quark matter is close to unity for large baryon densities. Such bulk quark matter would be referred to as ‘"strange quark matter (SQM)‘" in what follows.
The above hypothesis may lead to important consequences both for laboratory experiments as well as for astrophysical observations. Normal nuclear matter at high enough density and/or temperature, would be unstable against conversion to two flavour quark matter. The two flavour quark matter would be metastable and would eventually decay to SQM, releasing a finite amount of energy in the process. Such conversion may take place in the interior of a neutron star where the densities can be as high as (8-10)$\rho_{0}$ with $\rho_{0}$ being the nuclear matter density at saturation [@key-5; @key-6]. If Witten’s conjecture [@key-1] is correct, the whole neutron star may convert to a strange star with a significant fraction of strange quarks in it. (Neutron star may also become a hybrid star with a core of SQM in case the entire star is not converted to a strange star. Such a hybrid star would have a mixed phase region consisting of both the quark matter and the hadronic matter [@key-7].) Hadron to quark phase transition inside a compact star may also yield observable signatures in the form of Quasi-Periodic Oscillations (QPO) and the Gamma ray bursts [@key-7a; @key-7b].
There are several ways in which the conversion may be triggered at the centre of the star. A few possible mechanisms for the production of SQM in a neutron star have been discussed by Alcock [*et al*]{} [@key-8]. The conversion from hadron matter to quark matter is expected to start as the star comes in contact with a seed of external strange quark nugget. Such a seed would then grow by ’eating up’ baryons in the hadronic matter during its travel to the centre of the star, thus converting the neutron star either to a strange star or a hybrid star. Another mechanism for the initiation of the conversion process was given by Glendenning [@key-7] . It was suggested there that a sudden spin down of the star may increase the density at its core, thereby triggering the conversion process spontaneously.
Conversion of neutron matter to strange matter has been studied by several authors. Olinto [@key-9] viewed the conversion process to proceed via weak interactions as a propagating slow-combustion (i.e. a deflagration) front and derived the velocity of such a front. Olsen and Madsen [@key-10] and Heiselberg [*et al*]{} [@key-11] estimated the speed of such conversion front to range between 10 m/s to 100 km/s. The combustive conversion front was assumed to have a microscopic width of a few tens to a few hundreds of fm in these calculations.
Collins and Perry [@key-12], on the other hand, assumed that the hadronic matter gets converted first to a two flavour quark matter that eventually decays to a three flavour strange matter through weak interactions. Lugones [*et al*]{} [@key-13] argued that the hadron to SQM conversion process may rather proceed as a detonation than as a deflagration even in the case of strangeness production occuring through seeding mechanisms [@key-8].
Horvarth and Benvenuto [@key-14] examined the hydrodynamic stability of the combustive conversion in a non-relativistic framework. These authors inferred that a convective instability may increase the velocity of the deflagration front, so that a transition from slow combustion to detonation may occur. They argued that such a detonation may as well be responsible for the type II supernova explosions[@key-15]. In a relativistic framework, Cho [*et al*]{} [@key-16] examined the conservation of the energy-momentum and the baryonic density flux across the conversion front. Using Bethe-Johnson [@key-16a] and Fermi-Dirac neutron gas [@key-16b] equations of state (EOS) for the nuclear matter (NM) and the Bag model for SQM, they found that the conversion process was never a detonation but a slow combustion only for some special cases. Recently, Tokareva [*et al*]{} [@key-17] modelled the hadron to SQM conversion process as a single step process. They argued that the mode of conversion would vary with temperature of SQM and with the value of bag constant in the Bag model EOS. Berezhiani [*et al*]{} [@key-18], Bombaci [*et al*]{} [@key-19] and Drago [*et al*]{} [@key-20], on the other hand, suggest that the formation of SQM may be delayed if the deconfinement process takes place through a first order transition [@key-20a] so that the purely hadronic star can spend some time as a metastable object.
In this paper, we model the conversion of nuclear matter to SQM in a neutron star as occuring through a two step process. Deconfinement of nuclear matter to a two (up and down) flavour quark matter takes place in the first step in strong interaction time scale. The second step concerns with the generation of strange quarks from the excess of down quarks via a weak process. We may add here that this is the first instance where a realistic nuclear matter EOS is used to study the nuclear matter to SQM conversion as two step process. Drago [*et al*]{} [@key-20], on the other hand, studied the burning of nuclear matter directly to SQM in detail by using the conservation conditions and the compact star models.
To study the conversion of nuclear matter to a two-flavour quark matter, we here consider relativistic EOSs describing the forms of the matter in respective phases. Along with such EOSs, we would also consider hydrodynamical equations depicting various conservation conditions to examine such conversion process in a compact neutron star. Development of the conversion front, as it propagates radially through the model star, would be examined. We would next study the conversion of two-flavour quark matter to a three-flavour SQM through a non-leptonic weak interaction process by assuming $\beta$ equilibrium for the SQM. The paper is organised as follows. In section II, we discuss the EOSs used for the present work. In section III, we discuss the conversion to two-flavour quark matter. Conversion to three flavour SQM is discussed in section IV. In section V, we summarise the results. Conclusions that may be drawn from these results regarding the actual conversion process that may take place in a neutron star are also presented in this final section.
II. The equation of state {#ii.-the-equation-of-state .unnumbered}
=========================
The nuclear matter EOS has been evaluated using the nonlinear Walecka model [@key-21]. The Lagrangian density in this model is given by:
$$\mathcal{L}(x)=\sum_{i}\bar{\psi_{i}}(i\gamma^{\mu}\partial_{\mu}-m_{i}
+g_{\sigma i}\sigma+g_{\omega i}\omega_{\mu}\gamma^{\mu}
-g_{\rho i}\rho_{\mu}^{a}\gamma^{\mu}T_{a})\psi_{i}-
\frac{1}{4}\omega^{\mu\nu}\omega_{\mu\nu}$$
$$+\frac{1}{2}m_{\omega}^{2}\omega_{\mu}\omega^{\mu}
+\frac{1}{2}(\partial_{\mu}\sigma\partial^{\mu}\sigma
-m_{\sigma}^{2}\sigma^{2})-\frac{1}{4}\rho_{\mu\nu}^{a}\rho_{a}^{\mu\nu}
+\frac{1}{2}m_{\rho}^{2}\rho_{\mu}^{a}\rho_{a}^{\mu}$$
$$-\frac{1}{3}bm_{n}(g_{\sigma N} {\sigma})^{3}-\frac{1}{4}C(g_{\sigma N} {\sigma})^{4}
+\bar{\psi_{e}}(i\gamma^{\mu}\partial_{\mu}-m_{e})\psi_{e}\qquad\qquad
\label{1}$$
The Lagrangian in equation (1) includes nucleons (neutrons and protons), electrons, isoscalar scalar, isoscalar vector and isovector vector mesons denoted by $\psi_{i}$, $\psi_{e}$, $\sigma$, $\omega^{\mu}$ and $\rho^{a,\mu}$, respectively. The Lagrangian also includes cubic and quartic self interaction terms of the $\sigma$ field. The parameters of the nonlinear Walecka model are meson-baryon coupling constants, meson masses and the coefficient of the cubic and quartic self interaction of the $\sigma$ mesons (b and c, respectively). The meson field interacts with the baryons through linear coupling. The $\omega$ and $\rho$ meson masses have been chosen to be their physical masses. The rest of the parameters, namely, nucleon-meson coupling constant ($\frac{g_{\sigma}}{m_{\sigma}},\frac{g_{\omega}}{m_{\omega}}$ and $\frac{g_{\rho}}{m_{\rho}}$) and the coefficient of cubic and quartic terms of the $\sigma$ meson self interaction (b and c, respectively) are determined by fitting the nuclear matter saturation properties, namely, the binding energy/nucleon (-16 MeV), baryon density ($\rho_{0}$=0.17 $fm^{-3}$), symmetry energy coefficient (32.5 MeV), Landau mass (0.83 $m_{n}$) and nuclear matter incompressibility (300 MeV).
In the present paper, we first consider the conversion of nuclear matter, consisting of only nucleons (*i.e.* without hyperons) to a two flavour quark matter. The final composition of the quark matter is determined from the nuclear matter EOS by enforcing the baryon number conservation during the conversion process. That is, for every neutron two down and one up quarks and for every proton two up and one down quarks are produced, electron number being same in the two phases. While describing the state of matter for the quark phase we consider a range of values for the bag constant. Nuclear matter EOS is calculated at zero temperature, whereas, the two flavour quark matter EOS is obtained both at zero temperature as well as at finite temperatures.
III. Conversion to two flavour matter {#iii.-conversion-to-two-flavour-matter .unnumbered}
=====================================
In this section we discuss the conversion of neutron proton (n-p) matter to two flavour quark matter, consisting of u and d quarks along with electrons for the sake of ensuring charge neutrality. We heuristically assume the existence of a combustion front. Using the macroscopic conservation conditions, we examine the range of densities for which such a combustion front exists. We next study the outward propagation of this front through the model star by using the hydrodynamic (*i.e.* Euler) equation of motion and the equation of continuity for the energy density flux [@key-23]. In this study, we consider a non-rotating, spherically symmetric neutron star. The geometry of the problem effectively reduces to a one dimensional geometry for which radial distance from the centre of the model star is the only independent spatial variable of interest.
Let us now consider the physical situation where a combustion front has been generated in the core of the Neutron star. This front propagates outwards through the neutron star with a certain hydrodynamic velocity, leaving behind a u-d-e matter. In the following, we denote all the physical quantities in the hadronic sector by subscript 1 and those in the quark sector by subscript 2.
Condition for the existence of a combustion front is given by [@key-23a] $$\epsilon_{2}(p,X)<\epsilon_{1}(p,X),\label{1a}$$ where $p$ is the pressure and $X=(\epsilon+p)/{n_{B}}^{2}$, $n_{B}$ being the baryon density. Quantities on opposite sides of the front are related through the energy density, the momentum density and the baryon number density flux conservation. In the rest frame of the combustion front, these conservation conditions can be written as [@key-17; @key-23; @key-23a1]:
$$\omega_{1}v_{1}^{2}\gamma_{1}^{2}+p_{1}=\omega_{2}v_{2}^{2}\gamma_{2}^{2}+p_{2},
\label{2}$$
$$\omega_{1}v_{1}\gamma_{1}^{2}=\omega_{2}v_{2}\gamma_{2}^{2},
\label{3}$$
and $$n_{1}v_{1}\gamma_{1}=n_{2}v_{2}\gamma_{2}.
\label{4}$$
In the above three conditions $v_{i}$ (i=1,2) is the velocity, $p_{i}$ is the pressure, $\gamma_{i}=\frac{1}{\sqrt{1-v_{i}^{2}}}$ is the Lorentz factor, $\omega_{i}=\epsilon_{i}+p_{i}$ is the specific enthalpy and $\epsilon_{i}$ is the energy density of the respective phases.
Besides the conservation conditions given in (\[1a\]-\[4\]), the condition of entropy increase across the front puts an additional constraint on the possibility of the existence of the combustion front. This entropy condition is given by [@key-23b], $$s_{1}v_{1}\gamma_{1}\leq s_{2}v_{2}\gamma_{2}\label{4a}$$ with $s_{i}$ being the entropy density.
The velocities of the matter in the two phases, given by equations (\[2\]-\[4\]), are written as [@key-23]:
$$v_{1}^{2}=\frac{(p_{2}-p_{1})(\epsilon_{2}+p_{1})}{(\epsilon_{2}
-\epsilon_{1})(\epsilon_{1}+p_{2})},
\label{5}$$
and $$v_{2}^{2}=\frac{(p_{2}-p_{1})(\epsilon_{1}+p_{2})}{(\epsilon_{2}
-\epsilon_{1})(\epsilon_{2}+p_{1})}.
\label{6}$$
It is possible to classify the various conversion mechanisms by comparing the velocities of the respective phases with the corresponding velocities of sound, denoted by $c_{si}$, in these phases. Thes conditions are [@key-23c],
strong detonation : $v_{1}>c_{s1},\qquad v_{2}<c_{s2}$,
Jouget detonation : $v_{1}>c_{s1},\qquad v_{2}=c_{s2}$,
supersonic or weak detonation: $v_{1}>c_{s1},\qquad v_{2}>c_{s2}$,
strong deflagration : $v_{1}<c_{s1},\qquad v_{2}>c_{s2}$,
Jouget deflagration : $v_{1}<c_{s1},\qquad v_{2}=c_{s2}$,
weak deflagration : $v_{1}<c_{s1},\qquad v_{2}<c_{s2}$.
For the conversion to be physically possible, velocities should satisfy an additional condition, namely, $0\leq v_{i}^{2}\leq 1$. We here find that the velocity condition, along with the eq.(\[1a\]), puts severe constraint on the allowed equations of state.
To examine the nature of the hydrodynamical front, arising from the neutron to two flavour quark matter conversion, we plot, in fig.1, the quantities $v_{1},v_{2},c_{s1}$ and $c_{s2}$ as functions of the baryon number density ($n_{B}$). As mentioned earlier, the u and d quark content in the quark phase is kept same as the one corresponding to the quark content of the nucleons in the hadronic phase. With these fixed densities of u and d quarks and electrons, the EOS of the two flavour matter has been evaluated using the bag model prescription. We find that the energy condition (eqn.(\[1a\])) and velocity condition ( ${v_{i}}^{2}>0$ ) both are satisfied only for a small window of $\approx\pm5.0MeV$ around the bag pressure $B^{1/4}=160MeV$. The constraint imposed by the above conditions results in the possibility of deflagration, detonation or supersonic front as shown in the figs.(1-3).
In fig.1, we considered both the phases to be at zero temperature. A possibility, however, exist that a part of the internal energy is converted to heat energy, thereby increasing the temperature of the two flavour quark matter during the exothermic combustive conversion process. Instead of following the prescription for the estimation of temperature as given in references [@key-15; @key-20], we study the changes in the properties of combustion with the temperature of the newly formed two flavour quark phase in the present paper. In figures 2 and 3, we plot the variation of velocities with density at two different temperatures, namely, $T=50MeV$ and $T=100MeV$, respectively. These figures show that the range of values of baryon density, for which the flow velocities are physical, increases with temperature. Figure 4 shows the variation of velocities with temperature for values of baryon number densities given by $n_{B} \approx 3\rho_{0}$ and $7\rho_{0}$, respectively. In this figure, the difference between velocities $v_{1}$ and $v_{2}$ increases with temperature of the two flavour quark matter. In the present paper we have considered only the zero temperature nuclear matter EOS. On the other hand, equation of state of quark matter has a finite temperature dependence and hence the difference between ${v_{1}}$ and ${v_{2}}$, varies with temperature.
The preceding discussion is mainly a feasibility study for the possible generation of the combustive phase transition front and its mode of propagation. Having explored such possibilities, we now study the evolution of the hydrodynamical combustion front with position as well as time. This might give us some insight regarding the actual conversion of a hadronic star to a quark star and the time scale involved in such a process. To examine such an evolution, we move to a reference frame in which the nuclear matter is at rest. The speed of the combustion front in such a frame is given by ${v}_{f}={-v}_{1}$ with $v_{1}$ being the velocity of the nuclear matter in the rest frame of the front.
In the present work, we use the special relativistic formalism to study the evolution of combustion front as it moves outward in the radial direction inside the model neutron star. The relevant equations are the equation of continuity and the Euler’s equation, that are given by [@key-17]:
$$\frac{1}{\omega}(\frac{\partial\epsilon}{\partial\tau}
+v\frac{\partial\epsilon}{\partial r})+
\frac{1}{W^{2}}(\frac{\partial v}{\partial r}
+v\frac{\partial v}{\partial\tau})+2\frac{v}{r}=0
\label{7}$$
and $$\frac{1}{\omega}(\frac{\partial p}{\partial r}+
v\frac{\partial p}{\partial\tau})+
\frac{1}{W^{2}}(\frac{\partial v}{\partial\tau}+
v\frac{\partial v}{\partial r})=0,
\label{8}$$ where, $v=\frac{\partial r}{\partial\tau}$ is the front velocity in the nuclear matter rest frame and $k=\frac{\partial p}{\partial\epsilon}$ is taken as the square of the effective sound speed in the medium.
Substituting these expressions for $v$ and $k$ in equations (\[7\]) and (\[8\]) we get $$\frac{2v}{\omega}\frac{\partial\epsilon}{\partial r}
+\frac{1}{W^{2}}\frac{\partial v}{\partial r}(1+v^{2})+
\frac{2v}{r}=0
\label{9}$$
and $$\frac{n}{\omega}\frac{\partial\epsilon}{\partial r}(1+v^{2})
+\frac{2v}{W^{2}}\frac{\partial v}{\partial r}=0
\label{10}$$
Equations (\[9\]) and (\[10\]) ultimately yield single differential equation that is written as:
$$\frac{dv}{dr}=\frac{2vkW^{2}(1+v^{2})}{r[4v^{2}-k(1+v^{2})^{2}]}.
\label{11}$$
The equation (\[11\]) is integrated, with respect to r(t), starting from the centre towards the surface of the star. The nuclear and quark matter EOS have been used to construct the static configuration of compact star, for different central densities, by using the standard Tolman-Oppenheimer-Volkoff (TOV) equations [@key-17a]. The velocity at the centre of the star should be zero from symmetry considerations. On the other hand, the 1/r dependence of the dv/dr, in eq.(\[11\]) suggests a steep rise in velocity near the centre of the star.
Our calculation proceeds as follows. We first construct the density profile of the star for a fixed central density. Equations (\[5\]) and (\[6\]) then specify the respective flow velocities $v_{1}$ and $v_{2}$ of the nuclear and quark matter in the rest frame of the front, at a radius infinitesimally close to the centre of the star. This would give us the initial velocity of the front ($-v_{1}$), at that radius, in the nuclear matter rest frame. We next start with equation (\[11\]) from a point infinitesimally close to the centre of the star and integrate it outwards along the radius of the star. The solution gives us the variation of the velocity with the position as a function of time of arrival of the front, along the radius of the star. Using this velocity profile, we can calculate the time required to convert the whole star using the relation $v=dr/d\tau$.
In figure 5, we show the variation of the velocity for values of the central baryon densities 3$\rho_{0}$, 4.5$\rho_{0}$ and 7$\rho_{0}$, respectively. The respective initial velocities corresponding to such central densities are taken to be 0.43, 0.67 and 0.68. The figure shows that the velocity of the front, for all the central densities, shoots up near the centre and then saturates at a certain velocity for higher radius. Such a behaviour of velocity near the central point is apparent from the equation (\[11\]) above. The numerically obtained saturation velocity varies from 0.92 for central baryon density 3$\rho_{0}$ to 0.98 for 7$\rho_{0}$. The existence of a saturation velocity, at large r, is apparent from the asymptotic behaviour of equation (\[11\]). A comparison with fig.1 shows that for the densities 3$\rho_{0}$ and 4.5$\rho_{0}$, the conversion starts as weak detonation and stays in the same mode throughout the star. On the other hand, for 7$\rho_{0}$, initial detonation front changes over to weak detonation and the velocity of front becomes almost 1 as it reaches the outer crust. The corresponding time taken by the combustion front to propagate inside the star is plotted against the radius in figure 6. The time taken by the front to travel the full length of the star is of the order of few milliseconds. According to the present model, the initial neutron star thus becomes a two flavour quark star in about $10^{-3}$ sec. The results discussed above correspond to the case in which both nuclear as well as quark matter are at zero temperature. For finite temperature quark matter results vary only by a few percent of the front velocities for the quark matter at zero temperature.
We would like to mention here that in the above discussions, the equations governing the conversion of nuclear to quark matter are purely hydrodynamic. There is no dissipative process, so that the combustion front continues to move with a finite velocity depending on the density profile. Furthermore, there is no reaction rate involved here as the deconfinement process occurs in the strong interaction time scale and hence can be taken to be instantaneous. This is certainly very different from the second step process, to be discussed in the next section, where the two flavour matter converts to a three flavour matter. Here, the governing rate equations are weak interaction rates which play a decisive role in the conversion. Comparing the total time ( $\equiv10^{-3}$ sec) taken by the combustion front to travel through the star with the weak interaction time scale ($10^{-7}-10^{-8}$ sec), it is evident that the second step may start before the end of the first step process. In that case, perhaps, one should ideally considers two fronts, separated by a finite distance, moving inside the star [@key-23]. In the present paper, we have taken a much simplified picture and considered the conversion of a chemically equilibrated two flavour to three flavour quark matter as the second step process. Our results may provide us with more information regarding the necessity of considering two fronts.
IV. The conversion to three flavour SQM {#iv.-the-conversion-to-three-flavour-sqm .unnumbered}
=======================================
In this section we discuss the conversion of two flavour quark matter to three flavour SQM in a compact star. Similar to the discussion above, we assume the existence of a conversion front at the core of the star that propagates radially outward leaving behind the SQM as the combustion product. This conversion is governed by weak interactions that take place inside the star.
For a three flavour quark matter, the charge neutrality and the baryon number conservation conditions yield
$$3n_{B}=n_{u}+n_{d}+n_{s}\label{12}$$
$$2n_{u}=n_{d}+n_{s}+3n_{e}\label{13}$$
where $n_{i}$ is the number density of particle i (i= u, d, s and e).
The weak reactions which govern the conversion of excess down quark to strange quark can be written as,
$$d\rightarrow u+e^{-}+\overline{\nu}_{e-};\;~~~s\rightarrow u+e^{-}+
\overline{\nu}_{e-};\;~~~d+u\rightarrow s+u$$
We assume that the neutrinos escape freely from the site of reaction and the temperature of the star remains constant. The non-leptonic weak interaction in such a case becomes the governing rate equation. The semi-leptonic weak decays, then, are solely responsible for the chemical equilibration which can be incorporated through the relations given below.
$$\mu_{e^{-}}=\mu_{d}-\mu_{u};\;~~~\mu_{d}=\mu_{s},\label{14}$$
where $\mu_{i}$ is the chemical potential of the i’th particle. The number densities ($n_{i}$) of the quarks and electrons are related to their respective chemical potentials by,
$$n_{i}=g_{i}\int_{0}^{\infty}d^{3}p/(2\pi)^{3}[f^{+}-f^{-}],\label{15}$$
where $f^{+}$and $f^{-}$ are given by
$$f^{+}=\frac{1}{exp[(E_{p}-\mu)/T]+1}\qquad \qquad f^{-}=\frac{1}{exp[(E_{p}+\mu)/T]+1}.
\label{16}$$
In equations (\[15\]) and (\[16\]), $g_{i}$ is the degeneracy factor and $T$ the temperature. Eqns.(\[12\]-\[16\]) can be solved self consistently to calculate the number densities of quarks and electrons.
The conversion to SQM starts at the centre ($r=0$) of the two flavour star. Assuming that the reaction region is much smaller than the size of the star, we have considered the front to be one dimensional. Moreover, as we are considering spherical static stars only, there is no angular dependence. The combustion front, therefore, moves radially towards the surface of the star. As the front moves outwards, excess d quarks get converted to s quark through the non leptonic weak process. The procedure employed in the present work is somehow similar to that of ref.[@key-9], although the physical boundary conditions are different.
We now define a quantity,
$$a(r)=[n_{d}(r)-n_{s}(r)]/2n_{B}\label{17}$$
such that, $a(r=0)=a_{0}$ at the core of the star. The quantity $a_{0}$ is the number density of the strange quarks, at the centre, for which the SQM is stable and its value lies between 0 and 1. For equal numbers of d and s quarks, $a(r)=0$. Ideally at the centre of the star $a_{0}$ should be zero for strange quark mass $m_{s}=0$. Since s quark has a mass $m_{s}\sim150$MeV, at the centre of the star $a_{0}$ would be a small finite number, depending on the EOS. The s quark density fraction, however, decreases along with the decrease of the baryon density towards the surface of the star, so that, $a(r\rightarrow R)\rightarrow1$ with R being the radius of the star. At any point along the radius, say $r=r_{1}$, initial $a(r_{1})$, before the arrival of the front, is decided by the initial two flavour quark matter EOS. The final $a(r_{1})$, after the conversion is obtained from the equilibrium SQM EOS at the density corresponding to $r_{1}$.
The conversion to SQM occurs via decay of down quark to strange quark ($u+d\rightarrow s+u$) and the diffusion of the strange quark from across the front [@key-9]. The corresponding rate of change of $a(r)$ with time is governed by following two equations: $$\frac{da}{dt}=-R(a),\label{18}$$ and $$\frac{da}{dt}=D\frac{d^{2}a}{dr^{2}},\label{19}$$ In the equation (\[18\]) $R(a)$ is the rate of conversion of d to s quarks. Equation (\[19\]) yields the rate of change of $a(r)$ due to diffusion of s quarks, with $D$ being the diffusion constant. Following Olinto [@key-9], assuming the one dimensional steady state solution and using equations (\[18\]) and \[19\] we get: $$Da^{''}-va^{'}-R(a)=0,\label{20}$$ where $v$ is the velocity of the fluid. In equation (\[20\]) $a^{'}=\frac{da}{dr}$.
Conservation of baryon number flux at any position yields $n_{q}v_{q}=n_{s}v_{s}$. The subscripts q and s denote the two flavour quark matter phase and SQM phase, respectively. The baryon flux conservation condition yields the initial boundary condition at any point $r$ along the radius of the star: $$a'(r)=-\frac{v}{D}({a_{i}}(r)-{a_{f}}(r)),\label{20a}$$ where ${a_{i}}(r)$ and ${a_{f}}(r)$ are the values for the $a(r)$ before and after the combustion, respectively.
The reaction rate for the non-leptonic weak interaction $u+d\rightarrow u+s$ is in general a five dimensional integral for non zero temperature and $m_{s}$ [@key-25; @madsen]. Here, instead, we have taken the zero temperature, small $a$ limit [@key-9]. $$R(a)\approx\frac{16}{15\pi}{G_{F}}^{2}{cos^{2}}\theta_{c}{sin^{2}}\theta_{c}{\mu_{u}}^{5}
{\frac{a}{3}}^{3},\label{20b}$$ where, $G_{F}$ is the weak coupling constant and $\theta_{c}$ is the Cabibbo angle. The above equation can be written in the following form: $$R(a)\approx\frac{a^{3}}{\tau},\label{20c}$$ where $\tau=\frac{16}{(3^{3})15\pi}{G_{F}}^{2}{cos^{2}}\theta_{c}{sin^{2}}\theta_{c}{\mu_{u}}^{5}$, here, depends on the position of the front.
Following the line of arguments given in ref.[@key-9], we write down the analytic expressions for $D$ and $v$ as:
$$D=\frac{\lambda\overline{v}}{3}\simeq10^{-3}(\frac{\mu}{T})^{2}cm^{2}/s,\label{22}$$
and $$v=\sqrt{{\frac{D}{\tau}\frac{{{a_{f}}(r)}^{4}}{2({a_{i}}(r)-{a_{f}}(r))}}}.\label{23}$$
Our calculation proceeds as follows. First we get the star characteristics for a fixed central baryon density $\rho_{c}$. For a given $\rho_{c}$, number densities of u, d and s quarks, in both the two and three flavour sectors, are known at any point inside the star. That means ${a_{i}}(r)$ and ${a_{f}}(r)$ is fixed. Eqns.(\[20b\]-\[23\]) are then used to get the diffusion constant and hence the radial velocity of the front.
The central baryon densities considered here are same as those of section III. Assuming that the neutrinos leave the star, the temperature is kept constant at some small temperature so that we can use the equation (\[20b\]), evaluated in the zero temperature limit. The variation of $a(r)$ with the radius of the star is given in figure 7. The plot shows that $a(r)$ increases radially outward, which corresponds to the fact that as density decreases radially, the number of excess down quark which is being converted to strange quark by weak interaction also decreases. Hence, it takes less time to reach a stable configuration and hence the front moves faster, as shown in the figs. 8 and 9.
In Fig.8, we have plotted the variation of velocity along the radius of the star. The velocity shows an increase as it reaches sufficiently low density and then drops to zero near the surface as $d\rightarrow s$ conversion rate becomes zero. Fig.9 shows the variation of time taken to reach a stable configuration at different radial position of the star. The total time needed for the conversion of the star, for different central densities, is of the order of 100 seconds, as can be seen from the figure.
V. Summary and discussion {#v.-summary-and-discussion .unnumbered}
=========================
We have studied the conversion of a neutron star to strange star. This conversion takes place in two stages. In the first stage a detonation wave is developed in the hadronic matter (containing neutrons, protons and electrons). We have described this hadronic matter with a relativistic model. For such an equation of state the density profile of the star is obtained by solving Tolman-Openheimer-Volkoff equations. The corresponding quark matter equation of state is obtained by using the bag model. However, this quark matter equation of state is not equilibrated and contains two flavours. Matter velocities in the two media, as measured in the rest frame of the front, have been obtained using conservation conditions. These velocities have been compared with the sound velocity in both phases.
For a particular density inside the star, flow velocities of the matter on the two sides of the front is now fixed. Starting from a point, infinitesimally close to the centre, hydrodynamic equations are solved radially outward. The solution of the hydrodynamic equations gives the velocity profiles for different central densities. The velocity of the front shoots up very near to the core and then saturates at a value close to 1. The mode of combustion is found to be weak detonation for lower central densities. For higher central densities, the initial detonation becomes weak detonation as the the front moves radially outward inside the star. This result is different from that of ref.[@key-20], where the conversion process always correspond to a strong deflagration. The time required for the conversion of the neutron star to a two flavour quark star is found to be of the order of few milliseconds. After this front passes through, leaving behind a two flavour matter a second front is generated. This second front converts the two flavour matter via weak interaction processes. The velocity of the front varies along the radius of the star. As the front moves out from the core to the crust, its velocity increases, implying faster conversion. The time for the second conversion to take place comes out to be $\sim100$ seconds. This is comparable to the time scale obtained in ref.[@key-9].
The comparison of the time of conversion from neutron star to two flavour quark star and the weak interaction time scale suggests that at some time during the passage of the first combustion front, the burning of two flavour matter to strange matter should start. This means that at some point of time, there should be two fronts moving inside the star. On the other hand, our results show that, inside the model star, the burning of the nuclear matter to two flavour quark matter takes much smaller time compared to the conversion from two flavour quark matter to SQM. However, the consideration of two fronts might provide us with some more information regarding the conversion of neutron star to a final stable strange star. In the present case we have considered a two step process, there being only one type of front, inside the star, at any instant of time.
Here we would also like to mention that ideally the second step should start with a non-equilibrated two flavour matter [@key-24; @key-25]. Since this is a numerically involved calculation, in the present case we have taken the simplified picture of equilibrated quark matter.
Finally, the burning of the nuclear matter to two flavour quark matter is studies using special relativistic hydrodynamic equations. The actual calculation should involve general relativity, taking into account the curvature of the front for the spherical star. We propose to explore all these detailed features in our subsequent papers.
Acknowledgements {#acknowledgements .unnumbered}
================
R.M. would like to thank CSIR for financial support. S.K.G., S.R. and P.S.J. ,in particular, thanks DST for financial support under IRHPA scheme.
[10]{} E. Witten, Phys. Rev. [**D30**]{} (1984 ) 272. P. Haesel, J. L. Zdunik and R. Schaeffer. Astron. Astrophys. [**160**]{} (1986) 121. E. Farhi and R.L. Jaffe, Phys. Rev. [**D30**]{} ( 1989 ) 2379. A. Chodos, R.L. Jaffe, K. Jhonson, C.B. Thorn and V.F. Weisskopf, Phys. Rev. [**D9**]{} (1974 ) 3471. A. R. Bodmer, Phys. Rev. [**D4**]{} (1971) 1601. N. Itoh, Prog. Theor. Phys [**44**]{} (1970) 291. N. K. Glendenning, Nucl. Phys. (Proc. Suppl.) [**B24**]{} (1991) 110; Phys. Rev. [**D46**]{} (1992) 1274 A. Bhattacharyya and S. K. Ghosh, astro-ph/0506202. A. Bhattacharyya, S. K. Ghosh and S. Raha, Phys.Lett. [**B635**]{} (2006) 195. C. Alcock, E. Farhi and A. Olinto, Astrophys. J. [**310**]{} (1986) 261. A. V. Olinto, Phys. Lett. [**B192**]{} (1987) 71; Nucl.Phys.Proc.Suppl. [**B24**]{} (1991) 103. M. L. Olsen and J. Madsen, Nucl. Phys. (Proc. Suppl.) [**B24**]{} (1991) 170. H. Heiselberg, G. Baym and C. J. Pethick, Nucl. Phys. (Proc. Suppl.) [**B24**]{} (1991) 144. J. Collins and M. Perry, Phys. Rev. Lett. [**34**]{} (1975) 1353. G. Lugones, O. G. Benvenuto and H. Vucetich, Phys. Rev. [**D50**]{} (1994) 6100. J. E. Horvarth and O. G. Benvenuto, Phys. Lett. [**B213**]{} (1988) 516. O. G. Benvenuto, J. E. Horvarth and H. Vucetich, Int. J. Mod. Phys. [**A4**]{} (1989) 257; O. G. Benvenuto and J. E. Horvarth, Phys. Rev. Lett. [**63**]{} (1989) 716. H. T. Cho, K. W. Ng and A. D. Speliotopoulos, Phys. Lett. [**B326**]{} (1994) 111. S. L. Shapiro and S. A. Teukolsky, *Black Holes, White Dwarfs, and Neutron Stars*, Chapter 8 (John Wiley & Sons, New York, 1983). L. D. Landau and E. M. Lifshitz, *Statistical Physics, Part 1, 3rd ed.*, Chapter 5 (Pergamon Press, New York 1980). I. Tokareva, A. Nusser, V. Gurovich and V. Folomeev, Int. J. Mod. Phys. [**D14**]{} (2005) 33. Z. Berezhiani, I. Bombaci, A. Drago, F. Frontera and A. Lavagno, Astrophys. J. [**586**]{} (2003) 1250. I. Bombaci, I. Parenti and I. Vidana, Astrophys. J. [**614**]{} (2004) 314. A. Drago, A. Lavagno and G. Pagliara, Phys. Rev. [**D69**]{} (2004) 057505. J. Alam, S. Raha, B. Sinha, Phys. Rep, [**273**]{} (1996) 243. J. Ellis,J. I Kapusta and K. A. Olive, Nucl. Phys. [**B348**]{} (1991) 345. L. D. Landau and E. M. Lifshitz, Fluid Mechanics (1987), Pergamon Press. A. M. Anile, Relativistic fluids and Magneto-fluids : with application in Astrophysics and Plasma Physics (1989), Cambridge University Press, U.K. A. M. Gleeson and S. Raha, Phys. Rev. [**C26**]{} (1982) 1521. M. Gyulassy, K. Kajantie, H. Kurki-Suonio and L. McLerran, Nucl. Phys. [**B237**]{} (1984) 477. M. Laine, Phys. Rev. [**D49**]{} (1994) 3847. J. R. Oppenheimer and G. M. Volkoff, Phys. Rev. [**55**]{} (1939) 374. Sanjay K. Ghosh, S. C. Phatak and P. K. Sahu, Nucl. Phys. [**A596**]{} (1996) 67. J. Madsen, Phys. Rev. [**D47**]{} (1993) 325. Z. Dai, T. Lu and Q. Peng, Phys. Lett. [**B319**]{} (1993) 199.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Theory and simulations suggest that magnetic fields from radio jets and lobes powered by their central super massive black holes can be an important source of magnetic fields in the galaxy clusters. This is paper II in a series of studies where we present self-consistent high-resolution adaptive mesh refinement cosmological magnetohydrodynamic (MHD) simulations that simultaneously follow the formation of a galaxy cluster and evolution of magnetic fields ejected by an active galactic nucleus (AGN). We studied 12 different galaxy clusters with virial masses ranging from 1 $\times$ 10$^{14}$ to 2 $\times$ 10$^{15}$ M$_{\odot}$. In this work we examine the effects of the mass and merger history on the final magnetic properties. We find that the evolution of magnetic fields is qualitatively similar to those of previous studies. In most clusters, the injected magnetic fields can be transported throughout the cluster and be further amplified by the intra-cluster medium (ICM) turbulence during the cluster formation process with hierarchical mergers, while the amplification history and the magnetic field distribution depend on the cluster formation and magnetism history. This can be very different for different clusters. The total magnetic energies in these clusters are between 4 $\times$ 10$^{57}$ and $10^{61}$ erg, which is mainly decided by the cluster mass, scaling approximately with the square of the total mass. Dynamically older relaxed clusters usually have more magnetic fields in their ICM. The dynamically very young clusters may be magnetized weakly since there is not enough time for magnetic fields to be amplified.'
author:
- 'Hao Xu, Hui Li, David C. Collins, Shengtai Li, and Michael L. Norman'
title: 'Evolution and Distribution of Magnetic Fields from AGNs in Galaxy Clusters II. The Effects of Cluster Size and Dynamical State'
---
Introduction
============
The increasing detections of radio emission from galaxy clusters, called radio halos and relics [see @Carilli02; @Ferrari08; @Giovanini09] suggest that the ICM is permeated with magnetic fields. Radio halos are generally diffuse and extended over $\ge 1$ Mpc, covering the whole clusters, while radio relics, which are often observed at the edges at clusters, can extend as long as 2 Mpc [e.g. @Weeren10]. By assuming that the total energy in relativistic electrons is comparable to the magnetic energy, the magnetic fields in the cluster halos can be estimated at $0.1-1.0$ $\mu$G and the total magnetic energy can be as high as $10^{61}$ erg [@Feretti99].
Magnetic fields in galaxy clusters are also extensively studied through Faraday rotation measurements (FRM). Distribution of FRM, combined with the ICM density measurement, often yields cluster magnetic fields of a few to ten $\mu$G level, mostly in the cluster centers [@Carilli02]. More interestingly, FRM was investigated to suggest that cluster magnetic fields may have a Kolmogorov-like turbulent spectrum in core regions [@Vogt05], with an energy spectrum peak at several kpc. Other studies [for example @Taylor93; @Eilek02] have suggested that the coherence scales of magnetic fields can range from a few kpc to a few hundred kpc, implying large amounts of magnetic energy and fluxes in the ICM [@Colgate00]. Recently, study of magnetic fields by FRM [@Govoni06; @Guidetti08; @Bonafede10] are extended to the outer part of clusters by observations of more radio galaxies behind or embedded in these clusters. It is expected that the Extened Very Large Array (EVLA) will provide unprecedented new observations on the magnetic fields in the ICM.
It is unlikely that magnetic fields have been dynamically important during the cluster formation. But, it is suggested that the strength and geometry of magnetic fields in clusters may play a crucial role in cluster formation through other processes, such as heat transport, which consequently affect the applicability of clusters as sensitive probes for cosmological parameters [@Voit05]. In addition, since magnetic fields are closely related to synchrotron emission and FRM, the distribution of magnetic fields is important to the understanding the radio observations of the ICM.
Magnetic field evolution is highly nonlinear during cluster formation and difficult to be studied analytically. Cosmological MHD simulations are being used to study the properties of magnetic fields in the ICM. Such simulations can be very useful in interpreting the magnetic field information from the observation results. Since the origin of cluster magnetic fields is still being debated [@Widrow02], various initial magnetic fields are used in MHD cluster formation simulations. Simulations were done with initial magnetic fields either from some random or uniform fields at high redshifts [@Dolag02; @Dubois08; @Dubois09] or from the outflows of normal galaxies [@Donnert08]. All these simulations have found that fields can be further amplified by cluster merger [@Roettiger99] and turbulence in addition to collapse, and their findings roughly match results from observations. On the other hand, very small seed fields from some first principle mechanism, like Biermann battery effect, are also studied [@Kulsrud97; @Xu09b] in galaxy cluster simulations. @Kulsrud97 and @Ryu08 suggested that very week seed fields can be amplified by dynamo processes in clusters, though current simulations have not been able to model such processes self-consistently.
Magnetic fields in large scale radio jets and lobes from AGNs serve as one very intriguing source of cluster magnetic fields because observations show that they can carry large amounts of magnetic energy and flux [@Burbidge59; @Kronberg01; @Croston05; @McNamara07]. The magnetization of the ICM and the wider inter-galactic medium (IGM) by AGNs has been suggested on the energetic grounds [@Colgate00; @Furlanetto01; @Kronberg01], without details of the physical processes of magnetic field transportation and amplification. Through cosmological MHD simulations, @Xu09 showed that magnetic fields injected from a single AGN into a local region can be sufficient to magnetize the whole cluster to the micro Gauss level by the operation of small-scale dynamo [@Brandenburg05; @Subramanian06]. Recently, @Sur10 used MHD simulations to show that same process can also operate during the formation of first stars. Small-scale dynamo may be an important process to generate magnetic fields in various cosmic objects [@Schleicher10].
In the first paper [@Xu10] of this study, we studied the magnetic field evolution in a single massive, relaxed galaxy cluster with different AGN injected energies and injection redshifts. That paper found that, as long as the magnetic fields are injected before the active merger period during the cluster formation history, the AGN magnetic fields can be spread throughout the whole cluster and get substantial amplification. The final magnetic fields are weakly dependent on the amount of initial magnetic fields. But the behavior of magnetic fields in clusters of different masses and at various dynamical states, young (unrelaxed) or old (relaxed) clusters, is still not well studied.
In this paper, we perform a series of high resolution adaptive mesh refinement (AMR) MHD simulations of 12 galaxy clusters of different masses with initial magnetic fields injected by an AGN. This allows us to investigate the robustness of magnetizing the ICM using the AGN magnetic fields and address additional questions that are not examined with a single cluster. We explore the properties of magnetic field distribution and their evolution in the ICM of those clusters in different dynamical states during their formation histories. The organization of this paper is as follows. In Section \[sec:model\], we provide the details of the simulation setup, including the cluster formation and the magnetic fields injection. We then summarize the main results in Section \[sec:result\]. We present the detailed spatial distribution of magnetic fields, evolution of the magnetic energy and the radial profiles of magnetic fields strength. We also present and discuss the properties of synthetic Faraday rotation measurement of our simulated clusters. In Section 4, we present a summary of the main findings and conclusions.
Basic Model and Simulations {#sec:model}
===========================
We have performed self-consistent adaptive mesh refinement (AMR) cosmological MHD galaxy cluster formation simulations of a set of clusters with initial magnetic fields injected by an AGN, using cosmological MHD code ENZO+MHD [@Collins09]. The simulation setup is the same as that in @Xu10. The initial conditions are generated from an @Eisenstein99 power spectrum. The simulations use a $\Lambda$CDM model with parameters $h=0.73$, $\Omega_{m}=0.27$, $\Omega_{b}=0.044$, $\Omega_{\Lambda}=0.73$, and $\sigma_{8}=0.77$. These parameters are close to the values from recent WMAP 3 results [@Spergel07]. Initial conditions from different random seeds are used to generate different clusters. The simulation volume of each run is $256$ $h^{-1}$Mpc on a side, and it uses a $128^3$ root grid and $2$ level nested static grids in the Lagrangian region where the cluster forms. This gives an effective root grid resolution of $512^3$ cells ($\sim$ 0.69 Mpc) and dark matter particles of mass resolution of $1.07
\times 10^{10}M_{\odot}$. During simulations, $8$ levels of refinement are allowed beyond the root grid, for a maximum spatial resolution of $7.8125$ $h^{-1}$kpc.
The simulations are evolved from redshift $z=30$ to $z=0$ using an adiabatic equation of state, with gamma=5/3. The Simulations do not include other physics, such as radiative cooling or star formation feedback, as they are not important to the majority of the dynamics of cluster formation. Here, we have simulations of 12 galaxy clusters with mass ranging from 9.9 $\times$ 10$^{13}$ to 1.9 $\times$ 10$^{15}$ $M_\odot$ at $z=0$. In previous studies [@Xu10], we have shown that the magnetic field evolution is neither sensitive to the injection redshift between z=3 and z=1 nor to the injected AGN magnetic energy, so in this study we have only one injected AGN magnetic energy and a single injection time. We “turn on” the AGN with magnetic field injection at redshift $z=3$ in the most massive halo (or the second most massive one in some runs). The initial magnetic fields are injected into the ICM locally assuming that they are from an AGN (see description in @Xu08a [@Xu09]), basing on a magnetic tower model [@Li06]. The injected magnetic energy is $\sim$ 6 $\times$ 10$^{59}$ erg for all runs. Based on the dynamical states of these simulated clusters, these 12 clusters fall into 2 groups at z=0, each has 6 clusters. We consider the clusters, which have more than half of their final masses by z=0.5, as relaxed clusters, and the clusters, which gain more than half of their final masses from z=0.5 to z=0, as unrelaxed clusters. Group one is the relaxed clusters. They are labeled as R1 to R6, in the order of their final masses. Group two is the unrelaxed clusters. They are labeled as U1 to U6, also in the order of their final masses. For two of the unrelaxed clusters (U1 and U2) we have performed two simulations, with magnetic energy injected into different progenitor halos. They are labeled as U1a, U1b, U2a and U2b, respectively. The reason that we have injections in different locations is because the choice of initial magnetized halo turns out to play an important role in the final magnetism of these two clusters. So we have a total of 14 simulations. The properties of the injection halos and the final clusters are summarized in Table \[table:halos\], while the injection magnetic energies are listed in Table \[table:energy\].
\[table:halos\]
------ ---------------- -------------------- -------------------- ---------------- -------------------- --------------------
Name $r_{200}(Mpc)$ $M_{vir}(M_\odot)$ $M_{gas}(M_\odot)$ $r_{200}(Mpc)$ $M_{vir}(M_\odot)$ $M_{gas}(M_\odot)$
R1 2.161 1.252e15 1.863e14 0.194 1.577e13 2.296e12
R2 1.909 8.633e14 1.215e14 0.171 1.111e13 1.495e12
R3 1.590 4.985e14 7.116e13 0.160 9.132e12 1.223e12
R4 1.364 3.149e14 5.000e13 0.191 1.548e13 2.109e12
R5 1.148 1.877e14 2.864e13 0.190 1.536e13 2.160e12
R6 0.927 9.897e13 1.491e13 0.171 1.113e13 1.541e12
U1a 2.498 1.934e15 2.661e14 0.191 1.552e13 2.278e12
U1b 2.468 1.866e15 2.733e14 0.143 6.540e12 9.612e11
U2a 1.743 6.572e14 8.697e13 0.089 1.578e12 1.940e11
U2b 1.718 6.292e14 7.963e13 0.142 6.349e12 9.033e11
U3 1.709 6.198e14 8.105e13 0.116 3.460e12 4.269e11
U4 1.675 5.829e14 7.665e13 0.126 4.449e12 5.954e11
U5 1.626 5.338e14 7.370e13 0.133 5.318e12 7.007e11
U6 1.447 3.763e14 5.312e13 0.146 6.908e12 9.070e11
------ ---------------- -------------------- -------------------- ---------------- -------------------- --------------------
: Properties of simulated galaxy clusters.
The AMR settings are the same as in @Xu10. The AMR is applied only in a region of ($\sim$ 50 Mpc)$^3$ where the galaxy cluster forms. During the course of cluster formation before the magnetic fields are injected, the refinement is controlled by baryon and dark matter overdensity. After magnetic field injection, all the regions where magnetic field strength is higher than $5 \times 10^{-8}$ G are refined to the highest level, with 7.81 h$^{-1}$kpc resolution. The data analysis in this paper was performed using yt[^1] for Enzo [@Turk11].
There are limitations with our current simulations, which should be keep in mind when interpreting the results from these simulations. The highest resolution of $\sim$10 kpc, which already made these simulations very big and very computationally expensive, is close to the characteristic scale of the ICM magnetic fields obtained from observations. So our simulations miss some small scale features of the magnetic fields. Higher resolution simulations are underway. In addition, the MHD treatment without any kinetic effects in our simulations are not enough to completely understand the magnetic properties in the ICM. In some situations, the difference introduced by kinetic effects may be dramatic. For example, when effects of anisotropic pressure is included in a weakly-collisional plasma, growth rate of magnetic field strength may be much higher [@Schekochihin06; @Schekochihin08]. Unfortunately, the microphysical processes are still impossible to be simulated self-consistently in cosmological simulations with current computational capability.
Results {#sec:result}
=======
Formation of the Galaxy Clusters and Evolution of Magnetic Field Distribution {#sec:clusterformation}
-----------------------------------------------------------------------------
We first briefly show the formation histories of our simulated clusters. The hierarchy formation histories of our simulated clusters are presented by plotting the projected gas densities at various redshifts from z=3 to z=0. Relaxed clusters are plotted in Figure \[fig:density\_r\], and the unrelaxed clusters are shown in Figure \[fig:density\_u\]. The clusters usually undergo numerous mergers at redshift between 3 and 0.5. For redshift less than $\sim$ 0.5, the clusters start to relax unless some rare big mergers happen. The dynamically old and relaxed clusters in group 1 have experienced major mergers several Gyr before the end of the simulation, and have been dynamically relaxed for some time before z=0. In contrast, the younger, unrelaxed clusters of Group 2 have experienced major mergers closer to the end of the simulations. As we have mentioned at previous section, the clusters in group 1 have more than half of their final masses at z=0.5, while the clusters in group 2 more than double their masses from z=0.5 to z=0 by big mergers. From the shapes of the projected density, some clusters of group 2 look relaxed , e.g. cluster U1, though they have just finished a very big merger and their ICM motions are still quite active. The cluster U2 is undergoing active mergers after z=0.25 and is the dynamically youngest cluster in our simulations. This makes the evolution of magnetic fields in U2a and U2b very different from other runs.
Figure \[fig:med\_r\] and \[fig:med\_u\] show the projected magnetic energy densities. Magnetic fields mostly follow the motion of the halos in which magnetic fields are injected and may move a long distance during the cluster formation. No matter where the magnetic fields are injected, the final magnetic fields tend to distribute around the cluster centers at z=0. These plots clearly show that mergers, especially major mergers, spread the magnetic fields throughout the clusters. So the distribution of magnetic fields are effected by when these big mergers happen during the cluster formation history. There is a clear trend that the sizes of the magnetized areas and their associated magnetic energy densities are proportional to the sizes of their host halos. This is no surprise since larger clusters, which are formed by more mergers and have higher turbulence level to diffuse, amplify and maintain their magnetic fields.
Magnetic fields have already been spread to a large area of the whole clusters when simulations stop at $z=0$ in all runs, except for runs U2a and U2b. But the distribution of magnetic fields can be quite different from that of the ICM gas if major cluster mergers have occurred recently. In such case, the magnetic field strength may peak away from the cluster centers. Magnetic fields in larger clusters, which are less affected by late time small mergers, are distributed more regularly.
In most runs, the halo with the initial magnetic injection grows fast and becomes the major one progenitor of the clusters during their formation histories, so their magnetic field evolution and distribution are similar within their groups. But that is not true in three clusters, R2, U1 and U2. For cluster R2, though it is well formed before $z=0.5$ and has relaxed at final output, its magnetized halo does’t merge with the major one until z $\approx$ 1. So its magnetic fields can’t spread to a large volume and get amplified in the early time. This makes its magnetic fields weak and only locally distributed as a large relaxed cluster. For the two unrelaxed clusters U1 and U2, they don’t have a single major halo during their cluster formation until very low redshift. Initial magnetic fields are then injected into two different halos to see the different magnetic filed evolution. The cluster U1 is finally formed by the merging of two similar size sub-clusters at z $\sim$ 0.2. In runs U1a and U1b, we initially magnetize one of these two sub-clusters, respectively. Their late major merger doesn’t have enough time to spread their magnetic fields to the other parts of the cluster. The cluster U2 is formed by active mergers of several sub-clusters at very low redshifts. Runs U2a and U2b have two of them magnetized initially. In both runs, the lack of big mergers in the early time causes the magnetic fields to have much less amplification, and their final magnetic fields are very weak and locally distributed at z=0. The magnetic fields in run U2b are so weak that they are barely seen in Figure \[fig:med\_u\]
Magnetic Energy Evolution
-------------------------
We present the evolution of the total magnetic energy of all the simulations in Figure \[fig:energy\], separated into two groups of different dynamical states. The magnetic energy generally decreases rapidly initially, for a few hundred million years, because of the rapid expansion of the initial magnetic structure. It typically takes about 1-2 Gyr for the magnetic fields to expand in the ICM to a large volume to catch the turbulent motions and to be amplified. Since we inject large amount magnetic energy into the system, which may be much more than the magnetic energy that some small halos, like in cluster R6, can maintain. In such cases, their magnetic energies are keep decreasing until they drop to the levels that their magnetic fields can be sustained in their host halos by the turbulence. After that the magnetic energy gradually increases due to the increasing size and the ICM turbulent level by continuous mergers until saturation occurs (due to lack of mergers) or the simulation ends. Although the precise evolution of magnetic fields is quite different for various runs, the fastest amplifications of magnetic fields generally happen a while (hundred million years) after the active mergers (big or continuous mergers) which assemble the major part of the clusters.
The final magnetic energies are between 4 $\times$ 10$^{57}$ and 1 $\times$ 10$^{61}$ erg. The magnetic energy is saturated (stops growing) [for a detailed discussion, see @Xu10] for several Gry for relaxed clusters in group 1. The only exception is cluster R2, whose weak magnetic fields, due to lack of early major mergers, are still being amplified long after the relaxation of the cluster. Almost all of the clusters in group 2 have major mergers at low z, so their magnetic energies keep increasing until simulations end, except for run U2b. For this run, its magnetic fields, which are in a very small halo before the final assembly, occupy too small a volume to get amplified effectively when the simulation ends.
Not surprisingly, the total magnetic energy is higher for the larger clusters, which have deeper potential well, and higher thermal and kinetic energies. We plot the total magnetic energy versus the virial mass of the clusters in Figure \[fig:me\_mass\] at $z=0$. The total magnetic energy is scaled as M$_{viral}^{2}$ for most of the runs, except runs U2a and U2b. The total magnetic energy and the virial mass can be a single proportionality parameter $\alpha$ as : $$\frac{E_M}{6.09 \times 10^{60} erg} = \alpha (\frac{M_{vir}}{10^{15} M_{\odot}})^2
\label{equ:alpha}$$ where $\alpha$ is within a range between 0.131 to 1.445 for all runs except runs U2a and U2b. The energy 6.09 $\times$ 10$^{60}$ is chosen to make $\alpha = 1$ for run R1. We summarize $\alpha$, magnetic energies and their ratio to kinetic energies inside R$_{200}$ of all simulations in Table \[table:energy\]. The coefficient $\alpha$ is mostly decided by the cluster formation history. They are generally higher for the relaxed clusters in group 1 than for the dynamically younger clusters in group 2. The magnetic energies of runs U2a and U2b of the youngest cluster are far away from the E$_{M}$ $\sim$ M$^{2}$ scaling. This is because their magnetic fields do not get amplified until the late active mergers, and their current magnetic energies are limited by their small halos before the major active mergers.
Though the processes of cluster formation and magnetic field evolution are highly nonlinear and complex, this scaling relation between magnetic energy and virial mass can be qualitatively understood. Using an overly-simplified isothermal model, the thermal and kinetic energies of galaxy clusters are proportional to M$_{vir}^{5/3}$ [@Bryan98] with a radial density profile as $\rho \propto r^{-2}$ and a flat velocity dispersion. These relations were numerically confirmed by @Vazza06 [@Vazza11] using both SPH and AMR codes. Since the clusters are not exactly isothermal, especially in the outer part of the clusters, these scalings should be modified somewhat. The profiles of the density and the velocity dispersion deviate from the isothermal distribution as $\rho \propto r^{-3}$ and $\sigma \propto r$ [@Navarro95; @Sunyaev03] at large radii. In such a case, we find that the total kinetic energy is approximately proportional to the square of the virial masses in our simulations. As the magnetic energies in clusters in similar dynamical state are proportional to the kinetic energies with small scattering related to formation and magnetization histories (since magnetic power spectra of all simulations have a same shape and their levels are simply determined by the levels of the kinetic ones (see Section. \[sec:spectrum\])), the magnetic energy is then proportional to the square of the cluster mass.
Since the magnetic energy in a halo is proportional to the square of it mass, the magnetic fields in a massive cluster are much larger than the magnetic fields residing in its progenitor halos. So the major contributing factor to the final cluster magnetic fields is from the dynamo process after majority of the cluster has formed. It also suggests that additional magnetic fields from more AGNs or regular galaxies in smaller halos may have only small impact on the final magnetic fields in a cluster.
As listed in Table \[table:energy\], the magnetic energy in a cluster is only a small fraction of its kinetic energy. For relaxed clusters, their magnetic energies are about 1% of their kinetic energies inside their virial radii. These ratios drop to about 0.1% for the dynamical younger clusters, and even smaller for the recently formed cluster U2. The local magnetic energy is much smaller than the kinetic energy in most of the cluster regions as well. To show this, we plot the volume histogram of kinetic $\beta$ (e$_{K}$/e$_{M}$) of all clusters at z=0 in Figure \[fig:beta\]. For relaxed clusters, kinetic $\beta$ peaks at between 50 and 100, while peaks of kinetic $\beta$ are bigger than 100 for unrelaxed clusters. Even for a relaxed cluster, only a small fraction of its volume has magnetic energy bigger than kinetic energy. So, the magnetic fields are dynamically unimportant in all these simulated clusters. Several factors may contribute to why the magnetic energy is only a small fraction of the corresponding kinetic energy. First, it is possible that simulation resolution is not high enough, as the corresponding numerical diffusion is big, so the turbulence decays before it has an opportunity to amplify the magnetic fields. Some of our higher resolution test runs show that there are somewhat larger amount of more magnetic fields, but the total magnetic energy is still much smaller than the kinetic energy. Second, it could be that there is not enough time during cluster formation for the magnetic fields to grow in large scales, of which the eddy turnover time is $\sim$ Gyr. Third, the driving from mergers is not constant and not homogeneous, and becomes weaker when the cluster gets bigger. Fourth, it is not known what the expected level of magnetic field saturation is for compressible turbulence. Simulations of supersonic and superalfvenic compressible MHD turbulence in @Kritsuk09 found that the saturated magnetic energy is much smaller than the kinetic energy if the initial fields are weak. A further study to understand why the magnetic fields saturate at a level much lower than the kinetic energy in the ICM is needed. Fifth, since we only have one seed source for magnetic fields, significant magnetic fields only fill a portion of the ICM (from 20% to 80% see later section on magnetic fields distribution). So in a case of more sources and a higher filling of magnetic fields in a galaxy cluster, its magnetic energy ratio could potentially be somewhat higher (but still much smaller that kinetic energy, at most a few percent).
\[table:energy\]
Name E$_M^{i}$ (erg)$^a$ E$_M^{200}$ (erg)$^b$ $\alpha$$^c$ $E_M^{200}/E_K^{200}$$^d$ $\gamma$$^e$
------ --------------------- ----------------------- -------------- --------------------------- --------------
R1 6.17e59 9.55e60 1.00 0.00814 0.44
R2 4.85e59 1.34e60 0.295 0.00306 0.65
R3 5.57e59 9.10e59 0.601 0.00240 0.57
R4 5.32e59 4.19e59 0.693 0.0122 0.3
R5 6.37e59 3.07e59 1.436 0.0113 0.46
R6 5.94e59 8.62e58 1.445 0.0113 0.45
U1a 6.32e59 2.99e60 0.131 0.00119 0.56
U1b 5.83e59 5.88e60 0.277 0.00248 0.55
U2a 3.32e59 8.90e58 3.38e-2 1.77e-4 0.67
U2b 6.01e59 3.68e57 1.53e-3 7.19e-6 0.81
U3 4.74e59 4.09e59 0.175 8.21e-4 0.43
U4 5.35e59 1.25e60 0.603 0.00513 0.74
U5 5.35e59 6.07e59 0.350 0.00224 0.59
U6 5.21e59 3.29e59 0.382 0.00259 0.53
: Properties of magnetic energy in simulated galaxy clusters
a: magnetic energy at the end of injection;\
b: magnetic energy inside R$_{200}$ at z=0;\
c: $\alpha$ in Equation \[equ:alpha\], proportionality parameter between magnetic energy and cluster mass square;\
d: ratio of magnetic to kinetic energy inside R$_{200}$ at z=0;\
e: best-fit of $\gamma$ in relation between the magnetic field strength and the gas density (B $\propto$ n$^{\gamma}$);
Magnetic Field Distribution over the ICM at $z=0$
-------------------------------------------------
### Radial Profiles of Magnetic Fields {#sec:profile}
In Figure \[fig:Bprofile\], we present the spherically rms averaged radial profiles of magnetic fields at z=0. The relaxed and unrelaxed clusters are plotted in two different panels, and the radii are normalized by their virial radii. It is clear that the profiles of magnetic field strength are determined both by the cluster sizes and their dynamical states.
The magnetic field strength is generally higher for the bigger clusters and/or dynamically older clusters. For relaxed cluster, micro Gauss fields are present at the centers of clusters bigger that 10$^{15}$ M$_\odot$, but fields drop to about 0.5 $\mu$G for the smallest cluster of 10$^{14}$ M$_\odot$. The magnetic field profiles of relaxed clusters in group 1 are more regular as a function of radius, while the magnetic field profiles of dynamically younger clusters in group 2 are obviously disturbed by their recent big mergers. For runs U2a and U2b, since the cluster is formed so late and no significant amplification of magnetic fields has been operated, their magnetic fields are much weaker and reside only in the inner part of the cluster.
Several other MHD simulations [e.g. @Dolag02; @Dubois09], as well as the study on the RM and the X-ray surface brightness correlation [@Dolag01], suggest that the decline of magnetic fields is correlated with the gas density as n$^{\gamma}$. Analysis of RM and X-ray brightness data in @Dolag01 yields the $\gamma = 0.9$ for A119, and $\gamma=0.5$ for 3C129 with large uncertainty. We fit our magnetic field radial profile with the gas density profile and list the best-fitted results in Table \[table:energy\] . The $\gamma$ in our simulations scatters between 0.3 and 0.81, which is consistent with the results from @Dolag01. Relaxed clusters usually have smaller $\gamma$, so have flatter magnetic field profiles as they have more time to amplify their magnetic fields in the outer part by the ICM turbulence after major mergers. However, the reliability of our results may be limited by the way we magnetized clusters, we need simulations with many AGNs in one cluster to get more reliable n$^{\gamma}$ profiles. In addition, since this result is obtained by fitting the averaged radial profiles of the magnetic field strength and the gas electron density, it doesn’t mean that there is a simple correlation between the magnetic field strength and the gas density throughout the cluster.
### Magnetic Field Spatial Distributions
We plot the volume histograms and complementary cumulative volume ratio histograms of magnetic field strength inside the virial radii for all simulations at z=0 in Fig. \[fig:filling\]. The magnetic field fillings are quite different for the relax and unrelaxed clusters. For dynamically older clusters, most volumes inside the virial radii are filled with magnetic fields. Magnetic fields stronger than 10$^{-8}$ G typically fill more than 75% of the cluster volume for group one except run R2, and less than 50% for group two except run U5. For run R2 of group one, since the magnetized halo merges with the bigger halo after z=1, its magnetic field volume filling is consequently smaller. Once the magnetic fields are well spread throughout the clusters and get amplified, magnetic field strength distribution peaks between 0.1 and 0.3 $\mu$G. Larger and older clusters have much higher peaks as well as longer tails of stronger magnetic fields. Runs U2a and U2b have very weak magnetic fields that only fill very small volumes.
We also plot the two-dimensional distribution of the magnetic field strength verses the gas density inside their virial radii in Figure \[fig:contour\], showing how magnetic fields distributed over the ICM plasma. There is no obvious correlation between the field strength and the gas density. The distribution is similar for clusters of similar sizes. Most of the magnetic field strength is between 0.1 to a few micro Gauss, and are mixed with gas over a wide range of densities. This casts doubts on cluster magnetic field modeling when simple correlation between $|B|$ and ICM density is assumed.
Kinetic and Magnetic Energy Density Power Spectra {#sec:spectrum}
-------------------------------------------------
The kinetic and magnetic energy density power spectra of all clusters at z=0 are shown in Figure \[fig:spectra\], which are computed from boxes of 512$^3$ cells ($\sim$(5.5 Mpc)$^3$) in the highest level enclosing the clusters. The ICM turbulence is represented by Kolmogorov-like spectra in kinetic energy. These kinetic spectra are also seen from pure hydrodynamics simulations of galaxy clusters [@Vazza09]. All magnetic spectra are in a similar shape and follow the k$^{3/2}$ Kazantsev law in the large scales. The Kazantsev spectrum, which is the prediction of small-scale dynamo theory in the kinematic regime [see @Brandenburg05], is also found in simulations of a collapsing Bonnor-Ebert sphere in @Federrath11. These magnetic spectra in our simulations show that the small-scale dynamo [see @Xu10 for detailed discussions] does operate in all our simulated clusters. For dynamically young clusters, some (like cluster U4 ) of their power spectra are disturbed by the recently mergers.
For the relaxed or dynamically older clusters, the magnetic fields have enough time to be amplified, their magnetic energy densities are close to the kinetic energy densities in small scales. One possible reason that the magnetic energies in smaller scales are smaller than the kinetic energies is because the magnetic fields do not fill all the space of the clusters. For the recently merged clusters, the magnetic fields do not have enough time to spread through out the newly formed clusters and be completely amplified by the ICM turbulence. So their magnetic energies are much smaller than the kinetic ones even in the small scales.
Faraday Rotation Measurement from Simulations
---------------------------------------------
The Faraday Rotation Measurement is a key method for measuring cosmic magnetic fields and has provided important information on cluster magnetic fields. The RM maps have been used not only to estimate the strength of magnetic fields (see review by @Carilli02) but also the turbulent structure of the fields [@Vogt05; @Ensslin06] in the ICM. We compute the RM maps using the magnetic field and free electron density distribution from our simulations by integrating 8 Mpc over the clusters centered at each cluster’s center along the y axis at z=0. We show our synthetic RM maps in Figure \[fig:rm\]. Though the basic features of the RM distribution are similar to those from previous study [@Xu09; @Xu10], clusters of different size and/or at different dynamical states have quite different RM distribution in their absolute values, cover areas, and structure scales. The morphologies of the RM maps reflect the dynamical states of the clusters and their merger histories. There is a clear trend that there are more small scale patterns in the dynamically old relaxed clusters than in young clusters. This is because they have more time for the magnetic field amplification in small scales after major mergers. On the other hand, there are more long filaments on the younger, unrelaxed clusters reflecting their recent mergers. There are very long filaments ($>$ 1 Mpc) in runs U1a and U1b associated with their recent large head-on mergers. It is interesting to see that the RM maps are quite different between these two runs, of which the pre-merger magnetic fields reside in different sub-clusters, though the gas dynamics of these two runs are almost identical. This shows that the magnetism history of the cluster also plays an important role in determining their RM distribution.
The 2-D azimuthally averaged radial profiles of the absolute values of these FRM ($|RM|$) are plotted in Figure \[fig:rm\_profile\], while the radial profiles of the standard deviation of the RM are shown in Figure \[fig:sigma\_rm\_profile\]. The $|RM|$ profiles resemble the magnetic field strength profiles with steep slopes, as the ICM gas density decreases with the radius. These $|RM|$ profiles are roughly similar to the pattern from observational data in @Clarke01. The $\sigma_{RM}$ profiles are also consistent with the recently observational results in @Govoni10
We also plot the area histograms of RM and the complementary cumulative histograms of $|RM|$ inside 500 kpc central circles of the clusters in Figure \[fig:rm\_histogram\]. Relaxed clusters generally have more areas covered by significant $|RM|$. For all cases, the distribution of positive and negative RM is roughly symmetric. This is because there is no net magnetic fluxes in the clusters, for the model of field injection we have used. The $RM$ area histograms are similar for clusters of similar sizes and dynamical states, though their RM maps may look quite different from their different formation or magnetism history. This suggests that the histograms of RM distribution are not sensitive to the cluster merger and magnetism history, which can be reflected by their RM filament and patchy structures.
Conclusions
===========
In this paper, we report an ensemble of simulations of magnetic field evolution in galaxy clusters with a wide range of masses between 1 $\times$ 10$^{14}$ and 2 $\times$ 10$^{15}$ M$_\odot$ in various dynamical states. With similar amounts of initial magnetic fields injected from a single AGN at a high redshift, all clusters are eventually filled with micro Gauss magnetic fields, except for dynamically very young clusters. The power spectra of kinetic energy density show that the ICM is in a turbulent state, while the spectra of magnetic energy density show that the small-scale dynamo process is being driven by by the ICM turbulence. This result, along with the previous study of a single cluster with the magnetic field injections of different amounts of magnetic energy and at different redshifts [@Xu10], suggests that magnetization of galaxy clusters by the operation of small-scale dynamo with the seed magnetic fields from AGN is very robust, and it produces magnetic fields consistent with observed magnetic fields.
The magnetic field evolution and distribution are decided both by the masses of the clusters and their dynamical formation histories. The total magnetic energy is approximately scaled as the square of the virial mass of the cluster, while the dynamically older (relaxed) clusters usually have more magnetic fields. This implies that the cluster magnetic fields are mostly determined by the dynamo process of the ICM turbulence generated by the hierarchical mergers. Additional magnetic fields from more AGNs or smaller cluster forming halos will not have a major impact on the final magnetic fields in a massive cluster. The $\gamma$ in the scaling relation between magnetic field and gas density radial profiles ($|B|$ $\propto$ n$^{\gamma}$) range between 0.3 and 0.81 for our simulated clusters, while relaxed clusters have flatter magnetic fields profiles. In addition, the relaxed clusters usually have self-similar magnetic field radial distribution, while the field distribution in younger unrelaxed clusters is disturbed by their recently mergers. Though our simulated clusters only have initial magnetic fields from a single local source, most volumes in the simulated clusters are well magnetized.
We also studied Faraday rotation measurements obtained from the magnetic field and gas density distribution in our simulated clusters. They are also determined by both the cluster sizes and their dynamical states. The radial profiles of $|RM|$ resemble the profiles of the magnetic field strength, and are consistent with the pattern from the observational data. The RM maps reflect the recent cluster mergers, as well as the cluster magnetism history. There are very long filaments in the RM maps in the recently merged clusters, while their small scale patchy bands reflect the ICM turbulence. Very different distributions of RM of the same cluster but with magnetic fields injected in different locations are observed. This suggests that RM distribution may be a good probe not only for the cluster formation but also its magnetism history. A detailed study on the relation of RM from simulated clusters with their ICM turbulence will be presented in a forthcoming paper.
This work was supported by the LDRD and IGPP programs at LANL and by DOE office of science via CMSO. Computations were performed using the institutional computing resources at LANL. ENZO$\_$MHD is developed at the Laboratory for Computational Astrophysics, UCSD with partial support from NSF AST-0808184 to M.L.N.
[53]{} natexlab\#1[\#1]{}
, A., [Feretti]{}, L., [Murgia]{}, M., [Govoni]{}, F., [Giovannini]{}, G., [Dallacasa]{}, D., [Dolag]{}, K., & [Taylor]{}, G. B. 2010, , 513, A30+
, A., & [Subramanian]{}, K. 2005, , 417, 1
, G. L., & [Norman]{}, M. L. 1998, , 495, 80
, G. R. 1959, , 129, 849
, C. L., & [Taylor]{}, G. B. 2002, , 40, 319
, T. E., [Kronberg]{}, P. P., & [B[ö]{}hringer]{}, H. 2001, , 547, L111
, S. A., & [Li]{}, H. 2000, in IAU Symposium, Vol. 195, Highly Energetic Physical Processes and Mechanisms for Emission from Astrophysical Plasmas, ed. P. C. H. [Martens]{}, S. [Tsuruta]{}, & M. A. [Weber]{}, 255–264
, D. C., [Xu]{}, H., [Norman]{}, M. L., [Li]{}, H., & [Li]{}, S. 2010, , 186, 308
, J. H., [Hardcastle]{}, M. J., [Harris]{}, D. E., [Belsole]{}, E., [Birkinshaw]{}, M., & [Worrall]{}, D. M. 2005, , 626, 733
, K., [Bartelmann]{}, M., & [Lesch]{}, H. 2002, , 387, 383
, K., [Schindler]{}, S., [Govoni]{}, F., & [Feretti]{}, L. 2001, , 378, 777
, J., [Dolag]{}, K., [Lesch]{}, H., & [M[ü]{}ller]{}, E. 2009, , 392, 1008
, Y., [Devriendt]{}, J., [Slyz]{}, A., & [Silk]{}, J. 2009, , 399, L49
, Y., & [Teyssier]{}, R. 2008, , 482, L13
, J. A., & [Owen]{}, F. N. 2002, , 567, 202
, D. J., & [Hu]{}, W. 1999, , 511, 5
, T. A., & [Vogt]{}, C. 2006, , 453, 447
, C., [Sur]{}, S., [Schleicher]{}, D. R. G., [Banerjee]{}, R., & [Klessen]{}, R. S. 2011, , 731, 62
, L. 1999, in Diffuse Thermal and Relativistic Plasma in Galaxy Clusters, ed. H. [Boehringer]{}, L. [Feretti]{}, & P. [Schuecker]{}, 3–8
, C., [Govoni]{}, F., [Schindler]{}, S., [Bykov]{}, A. M., & [Rephaeli]{}, Y. 2008, Space Sci. Rev., 134, 93
, S. R., & [Loeb]{}, A. 2001, , 556, 619
, G., [Bonafede]{}, A., [Feretti]{}, L., [Govoni]{}, F., [Murgia]{}, M., [Ferrari]{}, F., & [Monti]{}, G. 2009, , 507, 1257
, F., [Dolag]{}, K., [Murgia]{}, M., [Feretti]{}, L., [Schindler]{}, S., [Giovannini]{}, G., [Boschin]{}, W., [Vacca]{}, V., & [Bonafede]{}, A. 2010, , 522, A105+
, F., [Murgia]{}, M., [Feretti]{}, L., [Giovannini]{}, G., [Dolag]{}, K., & [Taylor]{}, G. B. 2006, , 460, 425
, D., [Murgia]{}, M., [Govoni]{}, F., [Parma]{}, P., [Gregorini]{}, L., [de Ruiter]{}, H. R., [Cameron]{}, R. A., & [Fanti]{}, R. 2008, , 483, 699
, A. G., [Ustyugov]{}, S. D., [Norman]{}, M. L., & [Padoan]{}, P. 2009, in Astronomical Society of the Pacific Conference Series, Vol. 406, Numerical Modeling of Space Plasma Flows: ASTRONUM-2008, ed. [N. V. Pogorelov, E. Audit, P. Colella, & G. P. Zank]{}, 15–+
, P. P., [Dufton]{}, Q. W., [Li]{}, H., & [Colgate]{}, S. A. 2001, , 560, 178
, R. M., [Cen]{}, R., [Ostriker]{}, J. P., & [Ryu]{}, D. 1997, , 480, 481
, H., [Lapenta]{}, G., [Finn]{}, J. M., [Li]{}, S., & [Colgate]{}, S. A. 2006, , 643, 92
, B. R., & [Nulsen]{}, P. E. J. 2007, , 45, 117
, J. F., [Frenk]{}, C. S., & [White]{}, S. D. M. 1995, , 275, 720
, K., [Stone]{}, J. M., & [Burns]{}, J. O. 1999, , 518, 594
, D., [Kang]{}, H., [Cho]{}, J., & [Das]{}, S. 2008, Science, 320, 909
, A. A., & [Cowley]{}, S. C. 2006, Physics of Plasmas, 13, 056501
, A. A., [Cowley]{}, S. C., [Kulsrud]{}, R. M., [Rosin]{}, M. S., & [Heinemann]{}, T. 2008, Physical Review Letters, 100, 081301
, D. R. G., [Banerjee]{}, R., [Sur]{}, S., [Arshakian]{}, T. G., [Klessen]{}, R. S., [Beck]{}, R., & [Spaans]{}, M. 2010, , 522, A115+
, D. N., [Bean]{}, R., [Dor[é]{}]{}, O., [Nolta]{}, M. R., [Bennett]{}, C. L., [Dunkley]{}, J., [Hinshaw]{}, G., [Jarosik]{}, N., [Komatsu]{}, E., [Page]{}, L., [Peiris]{}, H. V., [Verde]{}, L., [Halpern]{}, M., [Hill]{}, R. S., [Kogut]{}, A., [Limon]{}, M., [Meyer]{}, S. S., [Odegard]{}, N., [Tucker]{}, G. S., [Weiland]{}, J. L., [Wollack]{}, E., & [Wright]{}, E. L. 2007, , 170, 377
, K., [Shukurov]{}, A., & [Haugen]{}, N. E. L. 2006, , 366, 1437
, R. A., [Norman]{}, M. L., & [Bryan]{}, G. L. 2003, Astronomy Letters, 29, 783
, S., [Schleicher]{}, D. R. G., [Banerjee]{}, R., [Federrath]{}, C., & [Klessen]{}, R. S. 2010, , 721, L134
, G. B., & [Perley]{}, R. A. 1993, , 416, 554
, M. J., [Smith]{}, B. D., [Oishi]{}, J. S., [Skory]{}, S., [Skillman]{}, S. W., [Abel]{}, T., & [Norman]{}, M. L. 2011, , 192, 9
, R. J., [R[ö]{}ttgering]{}, H. J. A., [Br[ü]{}ggen]{}, M., & [Hoeft]{}, M. 2010, Science, 330, 347
, F., [Brunetti]{}, G., [Gheller]{}, C., [Brunino]{}, R., & [Br[ü]{}ggen]{}, M. 2011, , 529, A17+
, F., [Brunetti]{}, G., [Kritsuk]{}, A., [Wagner]{}, R., [Gheller]{}, C., & [Norman]{}, M. 2009, , 504, 33
, F., [Tormen]{}, G., [Cassano]{}, R., [Brunetti]{}, G., & [Dolag]{}, K. 2006, , 369, L14
, C., & [En[ß]{}lin]{}, T. A. 2005, , 434, 67
, G. M. 2005, Reviews of Modern Physics, 77, 207
, L. M. 2002, Reviews of Modern Physics, 74, 775
, H. 2009, PhD thesis, University of California, San Diego
, H., [Li]{}, H., [Collins]{}, D., [Li]{}, S., & [Norman]{}, M. L. 2008, , 681, L61
, H., [Li]{}, H., [Collins]{}, D. C., [Li]{}, S., & [Norman]{}, M. L. 2009, , 698, L14
—. 2010, , 725, 2152
[^1]: http://yt.enzotools.org
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'For each closed, positive $(1,1)$-current $\omega$ on a complex manifold $X$ and each $\omega$-upper semicontinuous function ${\varphi}$ on $X$ we associate a disc functional and prove that its envelope is equal to the supremum of all $\omega$-plurisubharmonic functions dominated by ${\varphi}$. This is done by reducing to the case where $\omega$ has a global potential. Then the result follows from Poletsky’s theorem, which is the special case $\omega=0$. Applications of this result include a formula for the relative extremal function of an open set in $X$ and, in some cases, a description of the $\omega$-polynomial hull of a set.'
author:
- Benedikt Steinar Magnússon
bibliography:
- 'bibref.bib'
title: 'EXTREMAL $\omega$-PLURISUBHARMONIC FUNCTIONS AS ENVELOPES OF DISC FUNCTIONALS '
---
Introduction {#intro}
============
If $\omega$ is a closed, positive $(1,1)$-current on a connected complex manifold $X$, then for every point $x_0 \in X$ we can find a neighbourhood $U$ of $x_0$ and a plurisubharmonic local potential $\psi$ for $\omega$, i.e., $dd^c \psi = \omega$ on $U$. Let $u\colon X \to \overline {{\mathbb R}}$ be a function on $X$ with values in the extended real line. If we can locally write $u = v - \psi$, where $v$ is plurisubharmonic, then we say the function $u$ is *$\omega$-plurisubharmonic*. We denote by ${{\operatorname{{\cal PSH}}}}(X,\omega)$ the set of all $\omega$-plurisubharmonic functions on $X$ which are not identically equal to $-\infty$ in any connected component of $X$.
If $\psi_1$ and $\psi_2$ are two local potentials for $\omega$ then their difference is pluriharmonic on their common set of definition. This implies that the singular set, *${{\operatorname{sing}}}(\omega)$*, of $\omega$ is well defined and locally given as $\psi^{-1}(\{-\infty\})$ for a local potential $\psi$ of $\omega$.
We say that a function ${\varphi}\colon X \to \overline {{\mathbb R}}$ is *$\omega$-upper semicontinuous* if ${\varphi}+ \psi$ is upper semicontinuous on $U \setminus {{\operatorname{sing}}}(\omega)$, extends to an upper semicontinuous function on $U$ for every local potential $\psi \colon U \to {{\mathbb R}}\cup \{-\infty\}$ of $\omega$, and for $a\in {{\operatorname{sing}}}(\omega)$ we have $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z \to a} u(z)=u(a)$.
An *analytic disc* is a holomorphic map $f\colon {{\mathbb D}}\to X$ from the unit disc ${{\mathbb D}}$ into $X$. It is said to be *closed* if it can be extended to a holomorphic map in some neighbourhood of the closed unit disc. We let $\mathcal O({{\mathbb D}},X)$ denote the set of all analytic discs and ${{\cal A}}_X$ denote the set of all closed analytic discs in $X$.
For every analytic disc we have a pullback $f^*\omega$ of $\omega$ which is a Borel-measure on ${{\mathbb D}}$. It is defined locally as the Laplacian of the pullback $f^*\psi$, of a local potential $\psi$, to an open subset of ${{\mathbb D}}$. We define $R_{f^*\omega}$ as the Riesz potential of this measure on ${{\mathbb D}}$.
The main result of this paper is the following
\[th\] Let $X$ be a connected complex manifold, $\omega$ be a closed, positive $(1,1)$-current on $X$, and ${\varphi}$ be an $\omega$-upper semicontinuous function on $X$ such that $\{u \in {{\operatorname{{\cal PSH}}}}(X,\omega) ; u \leq {\varphi}\}$ is nonempty. Then for $x \in X\setminus {{\operatorname{sing}}}(\omega)$ $$\sup\{u(x) ; u \in {{\operatorname{{\cal PSH}}}}(X,\omega), u \leq {\varphi}\} =
\inf\{ -R_{f^*\omega}(0) + \int_{{\mathbb T}}{\varphi}\circ f\, d\sigma ; f \in {{\cal A}}_X, f(0) = x \},$$ where $\sigma$ is the arc length measure on the unit circle ${{\mathbb T}}$ normalized to $1$. Furthermore, if $\{ u \in {{\operatorname{{\cal PSH}}}}(X,\omega) ; u \leq {\varphi}\}$ is empty then the right hand side is $-\infty$.
This theorem is a generalization of Poletsky’s theorem, which is the special case $\omega=0$, see Poletsky [@Pol:1991], Lárusson and Sigurdsson [@LarSig:1998] and [@LarSig:2003], and Rosay [@Ros:2003].
However, if $\omega$ has a global potential $\psi$, i.e. $\psi \in {{\operatorname{{\cal PSH}}}}(X)$ with $dd^c \psi = \omega$, then the formula above becomes $$\sup\{u(x) ; u \in {{\operatorname{{\cal PSH}}}}(X,\omega), u \leq {\varphi}\} + \psi(x) =
\inf\{ \int_{{\mathbb T}}(\psi + {\varphi}) \circ f\, d\sigma ; f \in {{\cal A}}_X, f(0) = x \},$$ which is a direct consequence of Poletsky’s theorem since $\psi + {\varphi}$ is an upper semicontinuous function. This case is handled in Theorem \[th\_stein\].
The general case follows from this case and an $\omega$-version of a reduction theorem (Theorem 1.2 in [@LarSig:2003]) proved by Lárusson and Sigurdsson, see Theorem \[th\_red\]. The reduction theorem states that the right hand side in Theorem 1.1 is $\omega$-plurisubharmonic on $X$ if all its pullbacks to a manifold with a global potential are, and if we can assume some continuity properties of it with respect to the discs in ${{\cal A}}_X$.
By applying Theorem \[th\] to the characteristic function of the complement of an open set $E$ we get a disc formula for the relative extremal function, which Guedj and Zeriahi introduce in [@GueZer:2005], Chapter 4. Our result is the following $$\begin{gathered}
\sup\{ u(x) ; u \in {{\operatorname{{\cal PSH}}}}(X,\omega), u|_E \leq 0 \text{ and } u\leq 1 \}\\
= \inf\{ -R_{f^*\omega}(0) + \sigma({{\mathbb T}}\setminus f^{-1}(E) ) ; f \in {{\cal A}}_X, f(0) = x \}.\end{gathered}$$ In certain cases this formula can give us a description of the $\omega$-polynomial hull of a set, which is a generalization of the polynomial hull in ${{\mathbb C}}^n$.
For more information about the recent development of $\omega$-psh functions we refer the reader to Guedj and Zeriahi [@GueZer:2005] and [@GueZer:2007], Harvey and Lawson [@HarLaw:2006], Ko[ł]{}odziej [@Kol:2003], Dinew [@Din:2007], and Branker and Stawiska [@BraSta:2009]. In these papers $X$ is usually assumed to be a compact [Kähler]{} manifold and $\omega$ a smooth current on $X$.
[**Acknowledgement:**]{} I would like to thank my teacher Professor Ragnar Sigurdsson for pointing out this subject to me and for helping me writing the paper. I would also like to thank Professor Jón Ingólfur Magnússon for his help and Institut Mittag-Leffler where a part of this work was done during the Several Complex Variables program in 2008.
Basic results on $\omega$-plurisubharmonicity {#sec:2}
=============================================
Here we will define $\omega$-plurisubharmonic functions and study their properties analogous to those of plurisubharmonic functions.
Assume $X$ is a complex manifold of dimension $n$ and is $\omega$ a closed positive $(1,1)$-current on $X$, i.e. $\omega$ acts on $(n-1,n-1)$-forms.
It follows from Proposition 1.19, Ch. III in [@Demailly], that locally there is a plurisubharmonic function $\psi$ such that $dd^c \psi = \omega$. Here $d$ and $d^c$ are the real differential operators $d = \partial + \overline \partial$ and $d^c = i(\overline \partial - \partial)$. Hence, in ${{\mathbb C}}$ then $dd^c u = \Delta u\, dV$, where $\Delta u$ is the Laplacian of $u$ and $dV$ is the standard volume form.
Note that the difference of two potentials for $\omega$ is a pluriharmonic function, thus $\mathcal C^\infty$. This implies that the singular set of $\omega$, ${{\operatorname{sing}}}(\omega)$, is well defined as the union of all $\psi^{-1}(\{-\infty\})$, for all local potentials $\psi$ of $\omega$.
In the case when $\omega$ has continuous local potentials we have no trouble with continuity. If $\psi$ is a continuous local potential for $\omega$ then $u + \psi$ is upper semicontinuous if and only if $u$ is. In general this is not always the case, and we do not want to exclude the case when $\psi$ takes the value $-\infty$, e.g., when $\omega$ is a current of integration, or if $\psi$ is discontinuous. This however forces us to define the value of $u+\psi$ at points $x \in {{\operatorname{sing}}}(\omega)$ where $\psi(x)=-\infty$ and possibly $u(x)=+\infty$. If $u+\psi$ is bounded above on $X\setminus {{\operatorname{sing}}}(\omega)$ in a neighbourhood of $x$, then this can be done by taking upper limits of $u+\psi$ as we approach points in ${{\operatorname{sing}}}(\omega)$, we therefore make the following definition.
A function $u\colon X \to [-\infty,+\infty]$ is called *$\omega$-upper semicontinuous* ($\omega$-usc) if for every $a \in {{\operatorname{sing}}}(\omega)$, $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} u(z) = u(a)$ and for each local potential $\psi$ of $\omega$, defined on an open subset $U$ of $X$, $u+\psi$ is upper semicontinuous on $U\setminus {{\operatorname{sing}}}(\omega)$, and locally bounded above around each point of ${{\operatorname{sing}}}(\omega)$.
Equivalently we could say that $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} u(z) = u(a)$ for every $a \in {{\operatorname{sing}}}(\omega)$ and $u+\psi$ extends as $$\limsup_{U\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} (u+\psi)(z), \qquad \text{for } a \in {{\operatorname{sing}}}(\omega)$$ to an upper semicontinuous function on $U$ with values in ${{\mathbb R}}\cup \{-\infty\}$. We will denote this extension by $(u+\psi)^\star$.
Note that the question whether $(u+\psi)^\star$ is usc does not depend at all on the values of $u$ at ${{\operatorname{sing}}}(\omega)$. The reason for the conditions on $u$ at ${{\operatorname{sing}}}(\omega)$ is to ensure that $u$ is Borel measurable and to uniquely determine the function from its values outside of ${{\operatorname{sing}}}(\omega)$.
It is easy to see that $u$ is Borel measurable from the fact that $u = (u+\psi)-\psi$ is the difference of two Borel measurable functions on $X\setminus {{\operatorname{sing}}}(\omega)$ and that $u$ restricted to the Borel set ${{\operatorname{sing}}}(\omega)$ is the increasing limit of usc functions. Hence it is Borel measurable.
A function $u\colon X\to [-\infty,+\infty]$ is called *$\omega$-plurisubharmonic* ($\omega$-psh) if it is $\omega$-usc and $u+\psi$ is psh on $U\setminus {{\operatorname{sing}}}(\omega)$ for every local potential $\psi$ of $\omega$ defined on an open subset $U$ of $X$. We let ${{\operatorname{{\cal PSH}}}}(X,\omega)$ denote the set of all $\omega$-psh functions on $X$ which are not identically equal to $-\infty$ on any connected component of $X$. When the manifold is one dimensional we say that these functions are $\omega$-subharmonic ($\omega$-sh) and denote the set of ${{\operatorname{{\cal PSH}}}}(X,\omega)$ by $\mathcal{SH}(X,\omega)$.
Equivalently we could say that $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} u(z) = u(a)$ for every $a \in {{\operatorname{sing}}}(\omega)$ and $(u+\psi)^\star$ is psh for every local potential $\psi$. We see that the $\omega$-psh functions are locally integrable because outside of the zero set ${{\operatorname{sing}}}(\omega)$ they can locally be written as the difference of two functions which are locally integrable on $X$.
Our approach depends on the fact that we can define the pullback of currents by holomorphic maps. This we can do in two very different cases, first if the map is a submersion and secondly if it is an analytic disc not lying in ${{\operatorname{sing}}}(\omega)$.
If $\Phi \colon Y \to X$ is a submersion and $\omega$ is a current on $X$ then we can define the inverse image $\Phi^*\omega$ of $\omega$ by its action on forms, $\langle \Phi^*\omega,\tau \rangle = \langle \omega, \Phi_*\tau \rangle$, where $\Phi_*\tau$ is the direct image of the form $\tau$. For more details see Demailly [@Demailly], 2.C.2 Ch. I.
If $f$ is an analytic disc in $X$, $f({{\mathbb D}}) \nsubseteq {{\operatorname{sing}}}(\omega)$ then we can define a closed, positive $(1,1)$ current $f^*\omega$ on ${{\mathbb D}}$ in the following way.
Let $a \in {{\mathbb D}}$ and $\psi \colon U \to {{\mathbb R}}\cup \{-\infty\}$ be a local potential on an open neighbourhood $U$ of $f(a)$, and let $V$ be the connected component of $f^{-1}(U)$ containing $a$. If $\psi \circ f \neq -\infty$ on $V$ then we define $f^*\omega = dd^c (\psi \circ f)$. If $\psi \circ f = -\infty$ on $V$ then we define $f^*\omega$ as the measure which sends $\emptyset$ to $0$ and $E$ to $+\infty$ for all $E \neq \emptyset$. We denote this measure by $+\infty$. Observe that if $\psi \circ f = -\infty$ on $V$ then the same applies for every other local potential. In fact, if $\psi$ and $\psi'$ are two potentials for $\omega$ on open sets $U$ and $U'$ respectively then on the open set $V \cap V'$ we have two subharmonic functions $f^*\psi$ and $f^*\psi'$ which differ by a harmonic function. Therefore, if one of them is equal to $-\infty$ the other one is also equal to $-\infty$.
Remember that every positive $(n,n)$-current (of order 0) on an $n$ dimensional manifold can be given by a positive Radon measure, and conversely every positive Radon measure defines a positive $(n,n)$-current. So when we pull $\omega$ back to ${{\mathbb D}}$ by an analytic disc it is possible look at it both as a $(1,1)$-current and as a Radon measure.
We let $R_{f^*\omega}$ be the Riesz potential of the positive measure $f^*\omega$. It is defined by $$\label{R_f}
R_{f^*\omega}(z) = \int_{{\mathbb D}}G_{{\mathbb D}}(z,\cdot)\, d(f^*\omega),$$ where $G_{{\mathbb D}}$ is the Green function for the unit disc, $G_{{\mathbb D}}(z,w) = \log (|z-w|/|1-z\overline w|)$. If $f^*\omega = +\infty$ then $R_{f^*\omega} = -\infty$. The Riesz potential of $f^*\omega$ is not identically $-\infty$ if and only if $f^*\omega$ satisfies the Blaschke condition (see [@Hor:convexity], Theorem 3.3.6) $$\int_{{\mathbb D}}(1-|\zeta|)\, d(f^*\omega)(\zeta) < +\infty.$$
If $f$ is a closed analytic disc not lying in ${{\operatorname{sing}}}(\omega)$, then this condition is satisfied since $f^*\omega$ is a Radon measure in a neighbourhood of the unit disc, thus with finite mass on ${{\mathbb D}}$.
Also if we have a local potential $\psi$ defined in a neighbourhood of $\overline{f({{\mathbb D}})}$ then the Riesz representation formula, (ibid.), at the point 0 gives $$\label{riesz}
\psi(f(0)) = R_{f^*\omega}(0) + \int_{{\mathbb T}}\psi \circ f\, d\sigma.$$
\[equi\] The following are equivalent for a function $u$ on $X$.
1. $u$ is in ${{\operatorname{{\cal PSH}}}}(X,\omega)$.
2. $u$ is $\omega$-usc and $h^*u \in {{\operatorname{{\cal SH}}}}({{\mathbb D}},h^*\omega)$ for all $h\in {{\cal A}}_X$ such that $h({{\mathbb D}}) \nsubseteq {{\operatorname{sing}}}(\omega)$.
Assume $u \in {{\operatorname{{\cal PSH}}}}(X,\omega)$, take $h \in {{\cal A}}_X$, $h({{\mathbb D}}) \nsubseteq {{\operatorname{sing}}}(\omega)$, and $a \in {{\mathbb D}}$. Let $\psi$ be a local potential for $\omega$ defined in a neighbourhood $U$ of $h(a)$. Note that $(u+\psi)^\star \circ h = (u\circ h + \psi \circ h)^\star$, i.e. the extension of $(u + \psi) \circ h$ over ${{\operatorname{sing}}}(h^*\omega)$ is the same as the extension of $u+\psi$ over ${{\operatorname{sing}}}(\omega)$ pulled back by $h$, for both functions are subharmonic and equal almost everywhere, thus the same. Since $(u + \psi)^\star \in {{\operatorname{{\cal PSH}}}}(U)$ and $(u + \psi)^\star \circ h = (u \circ h + \psi \circ h)^\star$ is subharmonic in a neighbourhood of $a$ we see that $u \circ h$ is $h^*\omega$-subharmonic.
Assume now that *(ii)* holds and let $\psi \in {{\operatorname{{\cal PSH}}}}(U)$ be a local potential for $\omega$. Then $(u+\psi)^\star$ is upper semicontinuous, and *(ii)* implies that $(u+\psi)^\star \circ h \in {{\operatorname{{\cal SH}}}}({{\mathbb D}})$ for
From the proposition we also see that $\omega$-plurisubharmonicity like plurisubharmonicity is a local property, so in condition *(ii)* it is sufficient to look at $h \in {{\cal A}}_U$ in a neighbourhood $U$ of a given point.
If $u_0 \neq -\infty$ is psh and ${\varphi}$ is an usc function then the family $\{ u \in {{\operatorname{{\cal PSH}}}}(X) ; u_0 \leq u \leq {\varphi}\}$ is compact in the $L^1_\text{loc}$ topology, which implies that $\sup\{ u(x) ; u \in {{\operatorname{{\cal PSH}}}}(X), u \leq {\varphi}\}$ is plurisubharmonic. We have a similar result for $\omega$-plurisubharmonic functions.
If ${\varphi}\colon X \to [-\infty,+\infty]$ is $\omega$-usc, $\mathcal F_{\omega,{\varphi}} = \{ u \in {{\operatorname{{\cal PSH}}}}(X,\omega) ; u \leq {\varphi}\}$ and $\mathcal F_{\omega,{\varphi}} \neq \emptyset$, then $\sup \mathcal F_{\omega,{\varphi}} \in {{\operatorname{{\cal PSH}}}}(X,\omega)$, and consequently $\sup \mathcal F_{\omega,{\varphi}} \in \mathcal F_{\omega,{\varphi}}$.
Define $s_{\varphi}= \sup\mathcal F_{\omega,{\varphi}}$, by definition $s_{\varphi}+ \psi \leq {\varphi}+ \psi$ outside of ${{\operatorname{sing}}}(\omega)$ for every local potential $\psi$ on $U$. Since ${\varphi}$ is $\omega$-usc, the upper semicontinuous regularization $(s_{\varphi}+\psi)^*$of $s_{\varphi}+\psi$ also satisfies $$(s_{\varphi}+ \psi)^* \leq {\varphi}+ \psi \text{ on } U\setminus {{\operatorname{sing}}}(\omega).$$ Note that the left hand side is plurisubharmonic. We define the function $S$ on $X \setminus {{\operatorname{sing}}}(\omega)$ by $$S(x) = (s_{\varphi}+ \psi)^*(x) - \psi(x),$$ where $\psi$ is a local potential for $\omega$ in some neighbourhood of $x$. Observe that since the difference between two local potentials is continuous it is clear that the function $S$ is well defined on $X \setminus {{\operatorname{sing}}}(\omega)$. We extend $S$ to a $\omega$-usc function on $X$ by taking the $\limsup$ at points in ${{\operatorname{sing}}}(\omega)$. Furthermore, it is then obvious that $S \in {{\operatorname{{\cal PSH}}}}(X,\omega)$ and $s_{\varphi}\leq S \leq {\varphi}$ follows from the inequality above, so $s_{\varphi}= S \in \mathcal F_{\omega,{\varphi}}$.
Disc functionals and their envelopes {#sec:3}
====================================
A *disc functional* $H$ is a function defined on some subset ${{\cal A}}$ of ${{\cal O}}({{\mathbb D}},X)$, the set of all analytic discs in a manifold $X$, with values in $[-\infty,+\infty]$. The *envelope* $EH$ of a disc functional $H$ is then a function defined on the set $X_{{\cal A}}= \{ x\in X ; x=f(0) \text{ for some } f \in {{\cal A}}\}$ by the formula $$EH(x) = \inf\{ H(f) ; f\in {{\cal A}}, f(0)=x \}, \quad x \in X_{{\cal A}}.$$
The Poisson disc functional is defined as $H_{\varphi}(f) = \int_{{\mathbb T}}{\varphi}\circ f\, d\sigma$ where ${\varphi}$ is an usc function and $\sigma$ is the normalized arc length measure on the unit circle ${{\mathbb T}}$. If $f \in {{\cal A}}_X$ is a closed analytic disc, $u \in {{\operatorname{{\cal PSH}}}}(X)$ and $u \leq {\varphi}$ then $$u(f(0)) \leq \int_{{\mathbb T}}u \circ f\, d\sigma \leq H_{\varphi}(f).$$ The envelope $EH_{\varphi}$ of $H_{\varphi}$ is plurisubharmonic by the Poletsky theorem [@LarSig:2003] and equal to the supremum of all plurisubharmonic functions less than or equal ${\varphi}$, that is $$\sup \{ u(x) ; u \in {{\operatorname{{\cal PSH}}}}(X), u \leq {\varphi}\} = \inf \Big\{ \int_{{\mathbb T}}{\varphi}\circ f\, d\sigma ; f \in {{\cal A}}_X, f(0) = x \Big\}.$$
We will now generalize the definition of the Poisson functional to $\omega$-usc functions and look at the largest $\omega$-psh minorant of ${\varphi}$. This functional will be denoted by $H_{\omega,{\varphi}}$.
Fix a $\omega$-usc function ${\varphi}$ on a complex manifold $X$ and a point $x\in X \setminus {{\operatorname{sing}}}(\omega)$, let $f \in {{\cal A}}_X$, $f(0)=x$ and assume there is a function $u \in {{\operatorname{{\cal PSH}}}}(X,\omega)$, $u \leq {\varphi}$. Then since $f(0)=x\notin {{\operatorname{sing}}}(\omega)$ the pullback $f^*\omega$ is a well defined Radon measure on ${{\mathbb D}}$. Remember that $R_{f^*\omega}$ is a global potential for $f^*\omega$ on ${{\mathbb D}}$ and equal to $0$ on the boundary, so by Proposition \[equi\], $$u(x) + R_{f^*\omega}(0) \leq \int_{{\mathbb T}}u\circ f\, d\sigma + \int_{{\mathbb T}}R_{f^*\omega}\, d\sigma.$$ Since $u \leq {\varphi}$ and $R_{f^*\omega} = 0$ on ${{\mathbb T}}$ we see that $$\label{fund_inequal}
u(x) \leq - R_{f^*\omega}(0) + \int_{{\mathbb T}}{\varphi}\circ f\, d\sigma.$$ The right hand side is independent of $u$, so we define the functional $H_{\omega,{\varphi}}$ for every $f\in {{\cal A}}_X$, $f(0)\notin {{\operatorname{sing}}}(\omega)$ by $$H_{\omega,{\varphi}}(f) = - R_{f^*\omega}(0) + \int_{{\mathbb T}}{\varphi}\circ f\, d\sigma.$$
Now take supremum on the left hand side of (\[fund\_inequal\]) over all $\omega$-psh function $u$ satisfying $u\leq {\varphi}$ and infimum on the right over all $f \in {{\cal A}}_X$ such that $f(0)=x$. Then we get the fundamental inequality $$\sup {{\cal F}}_{\omega,{\varphi}} \leq EH_{\omega,{\varphi}} \qquad \text{on } X\setminus {{\operatorname{sing}}}(\omega).$$ Theorem \[th\] states that this is actually an equality.
Since the disc functional is not defined for discs centered at $a \in{{\operatorname{sing}}}(\omega)$ we extend the envelope to a function on the whole space $X$ by $$EH_{\omega,{\varphi}}(a) = \limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z \to a} EH_{\omega,{\varphi}}(z).$$
In the following we let $D_r = \{ t \in {{\mathbb C}}; |t|<r\}$ and if $x$ is a point in $X$ then $H(x)$ will denote the value of $H$ at the constant disc $t \mapsto x$, the meaning should always be clear from the context.
Notice now that if we look at the constant discs in $X$, then $H_{\omega,{\varphi}}(x) = {\varphi}(x)$ and consequently $EH_{\omega,{\varphi}} \leq {\varphi}$. Therefore, if we show that $EH_{\omega,{\varphi}}$ is $\omega$-psh then it is in ${{\cal F}}_{\omega,{\varphi}}$ and we have an equality $\sup {{\cal F}}_{\omega,{\varphi}} = EH_{\omega,{\varphi}}$.
An immediate corollary of the main theorem is a formula for the *relative extremal function* of a set $E$ in $\Omega$, where $\Omega$ is an open subset of $X$. It is defined as $$h_{E,\Omega,\omega}(x) = \sup\{u(x) ; u \in {{\operatorname{{\cal PSH}}}}(\Omega,\omega), u|_E \leq 0 \text{ and } u\leq 1 \},$$ Now assume $E$ is open and apply Theorem \[th\] to $\Omega$ with ${\varphi}$ as the characteristic function for the complement of $E$. For $x\in \Omega \setminus {{\operatorname{sing}}}(\omega)$ it gives that $$h_{E,\Omega,\omega}(x) = \inf\{ -R_{f^*\omega}(0) + \sigma({{\mathbb T}}\setminus f^{-1}(E)) ; f \in {{\cal A}}_\Omega, f(0) = x \}.$$ When $\Omega = X$ we denote this function by $h_{E,\omega}$.
In the local theory, $\omega=0$ and $X \subset {{\mathbb C}}^n$, the relative extremal function can be used to describe the polynomial hull of a compact set. The result is due to Poletsky [@Pol:1993] and can also be found in the following form in [@LarSig:2007], Theorem 2. It states that for a compact set $K$ in ${{\mathbb C}}^n$, a point $a$ in ${{\mathbb C}}^n$, and $\Omega$ a pseudoconvex neighbourhood of $K$ and $a$, bounded and Runge, the following are equivalent.
1. $a$ is in the polynomial hull of $K$.
2. For every neighbourhood $U$ of $K$ and every $\epsilon > 0$ there is a $f \in {{\cal A}}_\Omega$ with $f(0)=a$ and $\sigma({{\mathbb T}}\setminus f^{-1}(U)) < \epsilon$.
If we now wish to use our formula to get a similar result on a general complex manifold we have to start by finding an alternative to the polynomial hull. It follows from Theorem 5.1.7 in [@Kli] that the polynomial hull in ${{\mathbb C}}^n$ is exactly the hull with respect to the psh functions in ${{\mathbb C}}^n$ of logarithmic growth. These functions correspond to the $\omega$-psh functions on ${{\mathbb P}}^n$ if $\omega$ is the integration current for the hyperplane at infinity. This motivates the following definition, which is similar to the definition given by Guedj [@Gue:1999] of the $\omega$-polynomial hull.
If $K \subset X$ is compact subset of a complex manifold $X$ and $\omega$ is a closed, positive $(1,1)$-current on $X$, we define the $\omega$-polynomial hull of $K$ as $$\hat K^\omega = \{ x \in X ; u(x) \leq \sup_K u \text{ for all } u \in {{\operatorname{{\cal PSH}}}}(X,\omega)\}.$$
Our goal is to use the disc formula above to describe this hull. To make that possible we have to be able to use the relative extremal function to describe the hull, that is there is an $\Omega \subset X$ such that $h^{-1}_{K,\Omega,\omega}(\{0\}) = \hat K^\omega$. In the local theory it is sufficient to have $\Omega \subset {{\mathbb C}}^n$ hyperconvex.
The disc formula only applies to open sets, here we are considering compact sets so we start by showing that it is enough to look at shrinking neighbourhoods of compact sets.
\[compact\_limit\] Assume $\omega$ is a closed, positive $(1,1)$-current on a complex manifold $X$ such that $\omega$ has continuous local potentials on $X\setminus {{\operatorname{sing}}}(\omega)$. Let $K_1 \supset K_2 \supset ...$ be sequence of compact subsets of an open set $\Omega \subset X \setminus {{\operatorname{sing}}}(\omega)$ and $K = \cap_{j=1}^\infty K_j$, then $$\lim_{j\to \infty} h_{K_j,\Omega,\omega} = h_{K,\Omega,\omega}.$$
The proof is the same as in the case $\omega=0$, see Klimek [@Kli] Proposition 4.5.10. We only have to note that the assumptions on $\omega$ imply that all $\omega$-psh functions are usc.
Next we derive the result with some assumptions on $\Omega$, below we see that in some cases we can take $\Omega=X$.
Let $K$ be a compact subset of $\Omega \subset X\setminus {{\operatorname{sing}}}(\omega)$, and assume $\omega$ has continuous local potentials and $\Omega$ satisfies $h^{-1}_{K,\Omega,\omega}(\{0\}) = \hat K^\omega$. Then a point $x \in \Omega$ is in $\hat K^\omega$ if and only if for every neighbourhood $U$ of $K$ in $\Omega$ and every $\epsilon >0$ there is an analytic disc $f \in {{\cal A}}_\Omega$ such that $f(0)=x$ and
$$-R_{f^*\omega}(0) + \sigma({{\mathbb T}}\setminus f^{-1}(U) ) < \epsilon.$$
Let $x \in \hat K^\omega$. Then $0 \leq h_{U,\Omega,\omega}(x) \leq h_{K,\Omega,\omega}(x)=0$, and by the disc formula for $h_{U,\Omega,\omega}$ there is a disc $f\in {{\cal A}}_\Omega$ such that $f(0)=x$ and $$-R_{f^*\omega}(0) + \sigma({{\mathbb T}}\setminus f^{-1}(U) ) < \epsilon.$$
Conversely, if such $f$’s exist then $h_{U,\Omega,\omega}(x) = 0$ for every neighbourhood $U$ of $K$. Let $\{K_j\}$ be a sequence of compact subsets of $\Omega$ such that $\cap_j K_j = K$ and $K_{j+1} \subset \mathring K_j$. Then $h_{K_j,\Omega,\omega}(x)=0$ and by Proposition \[compact\_limit\] we see that $h_{K,\Omega,\omega}(x) = 0$.
If ${{\operatorname{sing}}}(\omega) =\emptyset$ and $X$ is compact then $h^{-1}_{K,\omega}(\{0\}) = \hat K^\omega$.
Assume $x \in h^{-1}_{K,\omega}(\{0\})$ and let $u \in {{\operatorname{{\cal PSH}}}}(X,\omega)$. Note that if we let $\psi_j\colon U_j \to {{\mathbb R}}^+$ be positive local potentials for $\omega$, such that $\cup_j U_j = X$, then for every $j$ the function $u+\psi_j$ is usc and locally bounded on $U_j$, and by a compactness argument we then see that $\sup_X u < \infty$.
Now let $\tilde u= (u-\sup_K u)/(\sup_X u - \sup_K u)$ if $\sup_X u -\sup_K u > 1$, else if $\sup_X u - \sup_K u \leq 1$ then we let $\tilde u = u - \sup_K u$. Either way, $\tilde u$ is $\omega$-psh, $\tilde u \leq 1$ and $\tilde u|_K \leq 0$. Therefore $\tilde u \leq h_{K,\omega}$ and $\tilde u(x) \leq h_{K,\omega}(x) = 0$, that is $u(x) \leq \sup_K u$ and $x \in \hat K^\omega$.
Obviously $\hat K^\omega \subset h^{-1}_{K,\omega}(\{0\})$, so $h^{-1}_{K,\omega}(\{0\}) = \hat K^\omega$.
Proof of the main result {#sec:4}
========================
We start by restricting to the case when $\omega$ has a global potential. The general case then follows from the reduction theorem later on.
\[global\_potential\] Let $\omega$ be a closed positive $(1,1)$-current on Stein manifold $X$. If there is a current $\eta$ such that $d\eta = \omega$, then $\omega$ has a global plurisubharmonic potential $\psi \colon X \to {{\mathbb R}}\cup \{-\infty\}$, so in particular $dd^c \psi = \omega$.
Since $\omega$ is a positive current it is real, so $\eta$ can be assumed to be real, $\eta \in \Lambda'_1(X,{{\mathbb R}})$. Now write $\eta = \eta^{1,0} + \eta^{0,1}$, where $\eta^{1,0} \in \Lambda'_{1,0}(X,{{\mathbb C}})$ and $\eta^{0,1} \in \Lambda'_{0,1}(X,{{\mathbb C}})$. Note that $\eta^{0,1} = \overline{ \eta^{1,0}}$ since $\eta$ is real. We see, by counting degrees, that $\overline \partial \eta^{0,1} = \omega^{0,2} = 0$, then since $X$ is Stein there is a distribution $\mu$ on $X$ such that $\overline \partial \mu = \eta^{0,1}$. Then $$\eta = \overline{ \overline\partial \mu} + \overline\partial \mu
= \partial \overline \mu + \overline \partial \mu.$$ If we set $\psi = (\mu - \overline \mu)/2i$, then $$\omega = d\eta = d(\partial \overline \mu + \overline \partial \mu)
= (\partial + \overline \partial)(\partial \overline \mu
+ \overline \partial \mu)
= \partial \overline \partial (\mu - \overline \mu)
= dd^c \psi.$$ Finally, $\psi$ is a plurisubharmonic function since $\omega$ is positive.
\[th\_stein\] Let $\omega$ be a closed, positive $(1,1)$-current on a manifold $X$ and ${\varphi}\colon X \to [-\infty,+\infty]$ an $\omega$-usc function such that $\mathcal F_{\omega,{\varphi}} \neq \emptyset$. If $\omega$ has a global potential $\psi$ then $EH_{\omega,{\varphi}} \in {{\operatorname{{\cal PSH}}}}(X,\omega)$ and consequently $EH_{\omega,{\varphi}} = \sup \mathcal F_{\omega,{\varphi}}$ on $X\setminus {{\operatorname{sing}}}(\omega)$.
For $f \in {{\cal A}}_X$, $f({{\mathbb D}}) \nsubseteq {{\operatorname{sing}}}(\omega)$, the Riesz representation (\[riesz\]) of $f^*\psi$ gives the following $$H_{\omega,{\varphi}}(f) + \psi(f(0)) =
H_{\omega,{\varphi}}(f) + R_{f^*\omega}(0) + \int_{{\mathbb T}}\psi \circ f \, d\sigma =
\int_{{\mathbb T}}(\psi + {\varphi})^\star \circ f\, d\sigma = H_{\psi+{\varphi}}(f).$$ The equality in the middle follows from the fact that $\int_{{\mathbb T}}{\varphi}\circ f\, d\sigma + \int_{{\mathbb T}}\psi \circ f\, d\sigma = \int_{{\mathbb T}}({\varphi}+\psi)^\star\circ f\, d\sigma$ since $\sigma(f^{-1}({{\operatorname{sing}}}({\omega}))\cap {{\mathbb T}}) = 0$. Therefore $$EH_{\omega,{\varphi}}(x) + \psi(x) = \inf\{ H_{\omega,{\varphi}}(f) + \psi(x) ;
f\in {{\cal A}}_X, f(0) = x \} = EH_{(\psi+{\varphi})^\star}(x).$$ By Poletsky’s theorem $EH_{(\psi+{\varphi})^\star}$ is psh, hence $EH_{\omega,{\varphi}}$ is $\omega$-psh.
If $\Phi\colon Y \to X$ is a holomorphic map between complex manifolds and $H$ is a disc functional on $X$, then the pullback $\Phi^*H$ is a disc functional on $Y$ defined by $\Phi^*H(f) = H(\Phi\circ f)$ for $f\in {{\cal A}}_Y$. Since $\{ \Phi \circ f ; f\in {{\cal A}}_Y \} \subset {{\cal A}}_X$ we get the following.
\[pullback\] $\Phi^*EH \leq E\Phi^*H$ and equality holds if every disc in $Y$ is a lifting of a disc in $X$ by $\Phi$.
Moreover, for the Poisson functional $H_{\omega,{\varphi}}$ we have the following result.
\[submersion\] Assume $\Phi\colon Y \to X$ is a holomorphic submersion, then $\Phi^*H_{\omega,{\varphi}} = H_{\Phi^*\omega,\Phi^*{\varphi}}$.
By associativity of compositions we have $(\Phi_*f)^*\omega = f^*(\Phi^*\omega)$ for $f\in {{\cal A}}_Y$, $f({{\mathbb D}}) \nsubseteq \Phi^{-1}({{\operatorname{sing}}}(\omega))$, so $$\begin{aligned}
\Phi^*H_{\omega,{\varphi}}(f) = H_{\omega,{\varphi}}(\Phi_* f) &=&
-R_{(\Phi_* f)^*\omega}(0) + \int_{{\mathbb T}}{\varphi}\circ \Phi \circ f\, d\sigma \\
&=& -R_{f^*(\Phi^*\omega)}(0) + \int_{{\mathbb T}}(\Phi^* {\varphi}) \circ f\, d\sigma = H_{\Phi^*\omega,\Phi^*{\varphi}}(f)\end{aligned}$$
We will now state the reduction theorem which will enable us to prove Theorem \[th\] using Theorem \[th\_stein\].
\[th\_red\][(Reduction theorem):]{} Let $X$ be a complex manifold, $H$ a disc functional on ${{\cal A}}= \{ f\in {{\cal A}}_X ; f(0) \notin {{\operatorname{sing}}}(\omega)\}$ and $\omega$ a positive, closed $(1,1)$-current on $X$. The envelope $EH$ is $\omega$-plurisubharmonic if it satisfies the following.
1. $E\Phi^*H$ is $\Phi^*\omega$-plurisubharmonic for every holomorphic submersion $\Phi$ from a complex manifold where $\Phi^*\omega$ has a global potential and for every $a \in {{\operatorname{sing}}}(\omega)$ we have $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} EH(z) = EH(a)$.
2. There is an open cover of $X$ by subsets $U$, with $\omega$-pluripolar subsets $Z \subset U$ and local potentials $\psi$ on $U$, $\psi^{-1}(\{-\infty\}) \subset Z$, such that for every $h \in {{\cal A}}_U$, $h({{\mathbb D}}) \nsubseteq Z$ the function $t \mapsto (H(h(t)) + \psi(h(t)))^\star$ is dominated by an integrable function on ${{\mathbb T}}$.
3. If $h \in {{\cal A}}_X$, $h(0) \notin {{\operatorname{sing}}}(\omega)$, $t_0 \in {{\mathbb T}}\setminus h^{-1}({{\operatorname{sing}}}(\omega))$ and $\epsilon > 0$, then $t_0$ has a neighbourhood $U$ in ${{\mathbb C}}$ and there is a local potential $\psi$ in a neighbourhood of $h(U)$ such that for all sufficiently small arcs $J$ in ${{\mathbb T}}$ containing $t_0$ there is a holomorphic map $F\colon D_r \times U \to X$ such that $F(0,\cdot) = h|_U$ and $$\frac{1}{\sigma(J)} \int_J \big( H(F(\cdot,t))+\psi(F(0,t)) \big)^\star \ d\sigma(t) \leq (EH+\psi)(h(t_0)) + \epsilon.$$
Before proving the theorem we show that $H_{\omega,{\varphi}}$ satisfies the conditions *(i)-(iii)* and consequently that Theorem 1.1 follows from it.
[Proof of Theorem 1.1]{} Condition *(i)* follows from Lemma \[submersion\] since it implies $E\Phi^*H_{\omega,{\varphi}} = EH_{\Phi^*\omega,\Phi^*{\varphi}}$ and by Theorem \[th\_stein\], $EH_{\Phi^*\omega,\Phi^*{\varphi}}$ is $\Phi^*\omega$-psh.
Condition *(ii)* follows from the fact that $H_{\omega,{\varphi}}(h(t)) + \psi(h(t)) = {\varphi}(h(t)) + \psi(h(t))$, which extends to an upper semicontinuous function on ${{\mathbb T}}$ and is thus dominated by a continuous function.
Assuming $h$ and $t_0$ as in condition *(iii)* and $\epsilon >0$, set $x=h(t_0)$ and let $f \in {{\cal A}}_X$ such that $f(0) = x$ and $H_{\omega,{\varphi}}(f) \leq EH_{\omega,{\varphi}}(x) + \epsilon/2$. By Lemma 2.3 in [@LarSig:1998] there is an open neighbourhood $V$ of $x$ in $X$, $r>1$ and a holomorphic function $\tilde F\colon D_r \times V \to X$ such that $\tilde F(\cdot,x) = f$ on $D_r$ and $\tilde F(0,z) = z$ on $V$. Shrinking $V$ if necessary, we assume $\psi$ is a local potential for $\omega$ on $V$. Let $U = h^{-1}(V)$ and define $F\colon D_r \times U \to X$ by $F(s,t) = \tilde F(s,h(t))$. By the Riesz representation (\[riesz\]), $$\label{riesz2}
\big(H_{\omega,{\varphi}}(F(\cdot,t)) + \psi(F(0,t))\big)^\star = \int_{{\mathbb T}}({\varphi}+ \psi)^\star \circ F(s,t)\, d\sigma(s).$$ Since the integrand is usc on $D_r \times U$, then it is easily verified that (\[riesz2\]) is an usc function of $t$ on $U$. That allows us by shrinking $U$ to assume that $$\big(H_{\omega,\psi}(F(\cdot,t))+\psi(F(0,t))\big)^\star
\leq H_{\omega,{\varphi}}(F(\cdot,t_0)) + \psi(F(0,t_0)) + \frac \epsilon 2$$ for $t \in U$. Then by the definition of $f = F(\cdot,t_0)$ $$\big(H_{\omega,{\varphi}}(F(\cdot,t)) + \psi(F(0,t))\big)^\star < EH_{\omega,{\varphi}}(x) + \psi(x) + \epsilon, \quad \text{on } U.$$ Condition *(iii)* is then satisfied for all arcs $J$ in ${{\mathbb T}}\cap U$.
Finally, if $\mathcal F_{{\varphi},\omega} = \emptyset$ then the only function which is both dominated by ${\varphi}$ and satisfies the subaverage property is the constant function $-\infty$. We know that $EH_{{\varphi},\omega} \leq {\varphi}$ and the proof of the reduction theorem gives the subaverage property, this implies that $EH_{{\varphi},\omega} = -\infty$.
We now prove that $EH$ is $\omega$-psh if $H$ satisfies the three conditions in Theorem \[th\_red\]. The main work is to show that $h^*EH$ satisfies the subaverage property of $h^*\omega$-sh functions for a given analytic disc $h$. This implies by Proposition \[equi\] that $EH$ is $\omega$-psh and that concludes the proof of the reduction theorem.
Let $H$ be a disc functional on an $n$-dimensional complex manifold $X$ and $\omega$ a positive, closed $(1,1)$-current on $X$. If the envelope $E\Phi^*H$ is $\Phi^*\omega$-usc for every holomorphic submersion $\Phi$ from an $(n+1)$-dimensional polydisc into $X$ and for every $a \in {{\operatorname{sing}}}(\omega)$ we have $\limsup_{X\setminus {{\operatorname{sing}}}(\omega) \ni z\to a} EH(z) = EH(a)$, then $EH$ is $\omega$-usc.
First, let $U$ be a coordinate polydisc in $X$ such that there is a potential $\psi$ for $\omega$ defined on $U$. Define $\Phi\colon U \times {{\mathbb D}}\to U$ as the projection. By assumption and Lemma \[pullback\] $$EH(z) + \psi(z) = EH(\Phi(z,t)) + \psi(\Phi(z,t)) \leq E\Phi^*H(z,t) + \psi(\Phi(z,t))
< +\infty,$$ for every $z \in U\setminus {{\operatorname{sing}}}(\omega)$ and $t\in {{\mathbb D}}$. Then $EH < +\infty$ on $X\setminus {{\operatorname{sing}}}(\omega)$ and $EH + \psi$ is bounded above in some neighbourhood of every point in ${{\operatorname{sing}}}(\omega) \cap U$, where $\psi$ is any local potential for $\omega$.
Let $x \in X\setminus {{\operatorname{sing}}}(\omega)$ and $\beta > EH(x)+\psi(x)$. Assume $f \in {{\cal A}}_X$ is a holomorphic disc defined on $D_r$ such that $f(0)=x$ and $H(f) + \psi(x) < \beta$. By using a theorem of Siu [@Siu:1976] it is shown in the proof of Lemma 2.3 in [@LarSig:1998] that for $\tilde r \in ]0,r[$ there exists a neighbourhood $U$ of the graph $\{t,f(t)\}$ in $D_r \times X$ and a biholomorphism $$\Psi \colon U \to D_{\tilde r} \times {{\mathbb D}}^n$$ such that $\Psi(t,f(t)) = (t,0)$. Let $\pi\colon {{\mathbb C}}\times X \to X$ be the projection and define $\Phi = \pi \circ \Psi^{-1}$. Clearly $f = \Phi \circ \tilde f$, where $\tilde f \in A_{D_{\tilde r} \times {{\mathbb D}}^n}$ is the lifting $\tilde f(t) = (t,0)$. By assumption and the fact that $$E\Phi^*H(x)+\psi(x) \leq \Phi^*H(\tilde f)+\psi(x) = H(f)+\psi(x) < \beta$$ there is a neighbourhood $W$ of $0 \in D_{\tilde r} \times {{\mathbb D}}^n$ such that $$(E\Phi^*H +\Phi^*\psi)^\star < \beta, \quad \text{ on } W.$$ Then for every $z$ in the open set $\Phi(W)$ $$(EH(z)+\psi(z))^\star = (\Phi^*EH(\tilde z)+\psi(z))^\star \leq
(E\Phi^*H(\tilde z)+\Phi^*\psi(\tilde z))^\star < \beta,$$ where $\tilde z \in \Phi^{-1}(\{z\})$. This along with the definition of $EH$ at ${{\operatorname{sing}}}(\omega)$ shows that the envelope is $\omega$-usc.
Now we turn to the subaverage property of the envelope.
[Proof of reduction theorem]{} We have already shown that the envelope $EH$ is $\omega$-usc, so by Proposition \[equi\] we only need to show that for a local potential $\psi$ on an open set $U \subset X$ and every disc $h \in {{\cal A}}_U$ such that $h(0) \notin {{\operatorname{sing}}}(\omega)$ we have $$\label{in0}
EH(h(0)) + \psi(h(0)) \leq \int_{{\mathbb T}}(EH \circ h + \psi\circ h)^\star \, d\sigma.$$ Observe that this is automatically satisfied if $EH(h(0)) = -\infty$, so we may assume $EH(h(0))$ is finite. It is sufficient to show that for every $\epsilon > 0$ and every continuous function $v\colon U \to {{\mathbb R}}$ with $v \geq (EH+\psi)^\star$, there exists $g \in {{\cal A}}_X$ such that $g(0) = h(0)$ and $$H(g) + \psi(h(0)) \leq \int_{{\mathbb T}}v \circ h\, d\sigma + \epsilon.$$ Then by definition of the envelope $EH(h(0)) + \psi(h(0)) \leq \int_{{\mathbb T}}v\circ h \, d\sigma + \epsilon$ for all $v$ and $\epsilon$, and (\[in0\]) follows.
We assume that $h$ is holomorphic on $D_r$, $r>1$ and $h(0) \notin {{\operatorname{sing}}}(\omega)$. It is easily verified (see proof of Theorem 1.2 in [@LarSig:2003]) that a function satisfying the subaverage property for all holomorphic discs in $X$ not lying in a pluripolar set $Z$ is plurisubharmonic not only on $X \setminus Z$ but on $X$. We may therefore assume that $h(\overline {{\mathbb D}}) \nsubseteq Z$.
Note that $h({{\mathbb T}}) \setminus {{\operatorname{sing}}}(\omega)$ is dense in $h({{\mathbb T}})$ by the subaverage property of $\psi\circ h$ and the fact that $h(0) \notin {{\operatorname{sing}}}(\omega)$. Therefore by a compactness argument along with property *(iii)* we can find a finite number of closed arcs $J_1,\ldots,J_m$ in ${{\mathbb T}}$, each contained in an open disc $U_j$ centered on ${{\mathbb T}}$ and holomorphic maps $F_j\colon D_s \times U_j \to X$, $s \in ]1,r[$ such that $F_j(0,\cdot) = h|_{U_j}$ and, using the continuity of $v$, such that $$\label{in1}
\underline{\int_{J_j}} \Big( H(F_j(\cdot,t))+\psi (F(0,t)) \Big)^\star\, d\sigma(t)
\leq \int_{J_j} v\circ h\, d\sigma + \frac{\epsilon}{4}\sigma(J_j).$$ We may assume that the discs $U_j$ are relatively compact in $D_r$ and have mutually disjoint closure. By the continuity of $v$ and condition *(ii)* we may also assume that $$\label{in2}
\int_{{{\mathbb T}}\setminus \cup_j J_j} |v\circ h|\, d\sigma < \frac{\epsilon}{4}$$ and $$\label{in3}
\overline {\int_{{{\mathbb T}}\setminus \cup_j J_j}} \Big(H(h(w))+\psi(h(w))\Big)^\star\, d\sigma(w) < \frac{\epsilon}{4}.$$
We now embed the graph of $h$ in ${{\mathbb C}}^4 \times X$ as follows $$K_0 = \{ (w,0,0,0,h(w)) ; w \in \overline {{\mathbb D}}\}$$ and the graphs of the $F_j$’s as $$K_j = \{ (w,z,0,0,F_j(z,w)) ; w \in J_j, z\in \overline {{\mathbb D}}\}.$$ Let $\Phi\colon {{\mathbb C}}^4 \times X \to X$ denote the projection. This function restricted to a smaller subset will be our submersion.
What is needed to find the disc $g$ we are looking for is a Stein neighbourhood $V$ of the compact set $K = \cup_{j=0}^m K_j$ in ${{\mathbb C}}^4 \times X$ where we can solve $d\eta =\Phi^*\omega$. Then we have by Lemma \[global\_potential\] a global potential for the pullback $\Phi^*\omega$ and the $\Phi^*\omega$-plurisubharmonicity of $E\Phi^*H$, given by property *(i)*, then gives the existence of $g$.
For convenience we let $U_0 = D_r$ and $F_0(z,w) = h(z)$. In [@LarSig:2003], using Siu’s theorem [@Siu:1976] and slightly shrinking the $U_j$’s and the $s$, Lárusson and Sigurdsson define for $j = 0,\ldots,m$ a biholomorphism $\Phi_j$ from $U_j \times D_s^{n+3}$ onto its image in ${{\mathbb C}}^4 \times X$ satisfying $$\Phi_j(w,z,0) = (w,z,0,0,F_j(z,w)), \quad w\in U_j, z \in D_s$$ for $j=1,\ldots,m$, and $$\Phi_0(w,0) = (w,0,0,0,h(w)), \quad w \in D_r.$$ The image of each $\Phi_j$ is therefore biholomorphic to a $4+n$ dimensional polydisc. These extensions of the graphs above are defined such that the first coordinate is the identity map. This tells us that these images are mutually disjoint for $j\geq 1$ and that the intersection of the image of $\Phi_0$ and $\Phi_j$ is a subset of $U_j \times {{\mathbb C}}^{3}\times X$.
As in [@LarSig:2003] we let $U_j''$ and $U_j'$ be discs cocentric with $U_j$ such that $$J_j \subset U_j'' \subset\subset U_j' \subset\subset U_j,$$ and we assume our $\Phi_0$ is the $\Phi_0$ after the modification made in [@LarSig:2003] which are necessary to have all but the first coordinate of $\Phi_j^{-1} \circ \Phi_0$ close to the identity. This modification which is done by precomposing $\Phi_0$ with a holomorphic map is necessary for constructing the Stein neighbourhood. Importantly for our purpose it does not change the first coordinate.
Now, for each $w \in \overline{U_j'}$ there is an $\epsilon_w >0$ such that $\Phi_0(w,D_{\epsilon_w}^{n+3}) \subset \Phi_j(w,D_s^{n+3})$. This holds by continuity for $\epsilon_w/2$ on a neighbourhood of $w$ in $U_j$. By compactness of $\cup_{j=1}^m \overline{U'_j}$ there is an $\epsilon$ independent of $w$ such that $\Phi_0(w,D_{\epsilon}^{n+3}) \subset \Phi_j(w,D_s^{n+3})$ for $w \in \overline{U'_j}$. We now restrict $\Phi_0$ to $D_r \times D_\epsilon^{n+3}$, then the intersection of the images $\Phi_0(D_r \times D_\epsilon^{n+3})$ and $\Phi_j(U'_j \times D_s^{n+3})$ is $\Phi_0(U'_j \times D_\epsilon^{n+3} )$
We define $V_j = \Phi_j(U_j \times D_s^{n+3})$, $V_0 = \Phi_0(U_0 \times D_\epsilon^{n+3})$ and $U = \cup_{j=0}^m V_j$. To solve $d\eta = \Phi^*\omega$ on $U$ it then enough to show that the cohomology $H^2(U)$ is zero. This can be done using the exact Mayer-Vietoris sequence ([@Spivak:I] Ch. 11, Theorem 3), $$\ldots \to H^q(M\cup N) \to H^q(M) \oplus H^q(N) \to H^q (M \cap N) \to H^{q+1}(M\cup N) \ldots.$$ We start by letting $M=V_0$ and $N=V_1$, these sets and their intersection are biholomorphic to a polydisc, so they are smoothly contractable and then by Poincaré’s lemma $H^2(V_j)=H^1(V_0 \cap V_1)= 0$. Consequently, we see from the Mayer-Vietoris sequence $$\ldots \to H^1(V_0 \cap V_1) \to H^2(V_0 \cup V_1) \to H^2(V_0) \oplus H^2(V_1) \to \ldots$$ that $H^2(V_0 \cup V_1) = 0$. Next we let $M=V_0 \cup V_1$ and $N = V_2$, then $H^1( (V_0\cup V_1) \cap V_2) = H^1( V_0 \cap V_2)=0$ since $V_1$ and $V_2$ are disjoint. The sequence above tells us then that $H^2(V_0 \cup V_1 \cup V_2) = 0$. Iterating this process for all the $V_j$’s we finally see that $H^2(U) = H^2(\cup_{j=0}^m V_j) = 0$.
Next step is to find a Stein neighbourhood $V$ of $K$ which is a subset of $U$. Note that $K$ only relies on the holomorphic functions $h$ and $F_j$. Therefore $V$ can be constructed in exactly the same way as in [@LarSig:2003]. It is done by defining a continuous strictly plurisubharmonic exhaustion function $\rho$ on $U$. This function is positive and satisfies $K \subset \rho^{-1}[0,\frac 12]$. Finally $V$ is defined as $\rho^{-1}([0,1))$.
Then by Lemma \[global\_potential\] we have a global potential on $V$ for $\Phi^*\omega$. By property *(i)* the envelope $E\Phi^*H$ is then $\Phi^*\omega$-psh on $V$ and if $\tilde h \colon D_r \to V$ is the lifting $w\mapsto (w,0,0,0,h(w))$ of $h$, then $$E\Phi^*H(\tilde h(0)) + \Phi^*\psi(\tilde h(0))
\leq \int_{{\mathbb T}}(E\Phi^*H \circ \tilde h + \Phi^*\psi \circ \tilde h)^\star \, d\sigma.$$ Since $EH(h(0))\neq -\infty$ then $-\infty < \Phi^*EH(\tilde h(0)) \leq E\Phi^*H(\tilde h(0))$ and we may assume there is a disc $\tilde g \in {{\cal A}}_V$ such that $\tilde g(0) = \tilde h(0)$ and $\Phi^*H(\tilde g) \leq E\Phi^*H(\tilde g(0)) + \epsilon/4$. Define the disc $g = \Phi \circ \tilde g \in {{\cal A}}_X$, then $g(0) = h(0)$ and since $H(g)=\Phi^*H(\tilde g)$ and $\Phi^*\psi(\tilde h) = \psi(h)$, $$\label{in4}
H(g) + \psi(h(0)) \leq
\int_{{\mathbb T}}(E\Phi^*H \circ \tilde h + \psi\circ h)^\star\, d\sigma + \frac{\epsilon}{4}.$$
For $w \in J_j$, $1\leq j\leq m$ we have a lifting of $F_j(\cdot,w)$ by $\Phi$ given by $z \mapsto (w,z,0,0,F_j(z,w))$. Clearly $0 \mapsto \tilde h(w)$, so $E\Phi^*H(\tilde h(w)) \leq \Phi^*H(\tilde F_j(\cdot,w)) = H(F_j(\cdot,w))$. However, if $w \in {{\mathbb T}}\setminus \cup_j J_j$ then $E\Phi^*H(\tilde h(w)) \leq \Phi^*H(\tilde h(w)) = H(h(w))$. Therefore, $$\int_{{\mathbb T}}E\Phi^*H \circ \tilde h\, d\sigma
\leq \sum_{j=1}^m \underline{\int_{J_j}} H(F_j(\cdot,w))\, d\sigma(w)
+ \overline{\int_{{{\mathbb T}}\setminus \cup_j J_j}} H(h(w))\, d\sigma(w).$$ Adding the integral of $\psi(h)$ to both sides of this inequality and using the inequalities (\[in1\]) and (\[in3\]) we see that $$\int_{{\mathbb T}}(E\Phi^*H \circ \tilde h + \psi\circ h)^\star\, d\sigma \leq
\int_{\cup J_j} v\circ h\, d\sigma + \frac{\epsilon}{4}\sigma(\cup_j J_j) + \frac{\epsilon}{4}.$$ Then by using first (\[in4\]) and then (\[in2\]) we have finally $$H(g) + \psi(h(0)) \leq \int_{\cup J_j} v\circ h\, d\sigma + \frac{3}{4}\epsilon
< \int_{{\mathbb T}}v\circ h\, d\sigma + \epsilon.$$
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Hans-Christian Pauli'
date: 15 December 2003
title: The hadronic potential at short distance
---
Introduction: Highlights of a discussion
========================================
In the aftermath of recent conference contributions [@Pau03a], I had to survive serious discussions whose highlights should be interesting for a larger audience. In a way, the arguments of a fictitious anonymous represent the state of the art public opinion of the high energy physics community. I let play him the role of the devils advocate. Here are some of his points:
1. QCD interactions at short distances are scale invariant – up to the logarithmic asymptotic freedom corrections predicted from the running coupling. This near conformal scaling of QCD is well tested in quark-quark scattering and other jet physics measurements at colliders.
2. \[Brodsky:scale\] Certainly $\Lambda_{QCD}$ sets the scale of the hadronic parameters. It is relevant to QCD interactions at large distances.
3. Evidence is accumulating that the running coupling – as defined from the Landau gauge gluon propagator or observables – is well regulated at small $Q^2$. Some relevant papers are [@vonSmekal:1997is; @Howe:2002rb; @Mattingly:1994; @Brodsky:2002nb].
4. \[Brodsky:Fourier\] It might be worthwhile to look at the Fourier transform of ${\alpha_s(Q^2)}/{Q^2}$ where $\alpha_s(Q^2)$ is the Shirkov form, as given for example in the paper [@Baldicchi:2002].
5. \[Brodsky:Yukawa\] Pauli uses a cut-off $\lambda$ to motivate a modification of the short distance $r \to 0$ QCD interaction. — Let us compare the Yukawa ${1}/{(Q^2+\lambda^2)}$ and Coulomb ${1}/{Q^2}$ interactions. They are identical at large $Q^2$. — Their respective Fourier transforms potentials are the ${\mathrm{e}^{-\lambda r}}/{r}$ and $1/r$. They are identical at short distances $r \to 0$!
Here is a summary of my points:
1. Relating large momentum transfers to short distances is often only a figure of speech. The relation between a scattering amplitude $T(Q^2)$ and an interaction $V(r)$ is a highly non-trivial matter.
2. Colliders measure the scattering amplitude $\alpha_s(Q^2)/Q^2$ at large $Q^2$. The interaction is the Fourier transform of ${\alpha_s(Q^2)}/{Q^2}$, all $Q^2$ are needed, and not only the large ones.
3. We should be more careful about our terms of speaking, and specify by operational prescriptions what we mean by ‘interaction’ and/or ‘potential’.
4. The $q\bar q$ potential at short distance simply cannot go like $\alpha_s/r$, whatever the numerical value of the strong coupling ‘constant’ $\alpha_s$ is.
5. The literature has too many $\Lambda$’s! The same mathematical symbol refers to completely different physics.
In this short note, I will work out to some detail why and to which extent I will arrive at conclusions different from the public opinion.
I use this opportunity as the port of entry for presenting novel insights on the nature of the gauge field interaction at short distance. In [@Pau03b] I will present a possible solution to the problem of the non perturbative renormalization in a gauge theory. In [@Pau03c], the technical details of the fine and hyperfine interaction on the light cone will be discussed. In [@Pau03d] and [@Pau03e] I will calculate the ground state masses and mass spectra of all flavor off diagonal pseudo scalar and pseudo vector mesons, analytically, for a linear and for a quadratic potential.
Taking Fourier transforms {#sec:2}
=========================
Collider physics is consistent with asymptotic freedom, $$\begin{aligned}
\alpha_s(Q^2)= \frac{4\pi}{\beta_0\mathrm{ln}( Q^2/\Lambda^2_{QCD} )}
\,,\label{eq:1}\end{aligned}$$ at sufficiently large $Q^2$, with the famous $\beta_0=11n_c-2n_f$. What is its Fourier transform?
Let us be specific. Consider $$\begin{aligned}
V(r) &=& -\frac{1}{2\pi^2}\int\!d^3 q\;
\frac{\alpha_c(q^2)}{q^2}\;
\mathrm{e}^{-i\vec q \vec r}
,\label{eq:2}\end{aligned}$$ where $\alpha_c(q^2)=\frac43 \alpha_s(q^2)$ given by Eq.(\[eq:1\]) with the Feynman four momentum transfer $Q^2$ substituted by the three momentum transfer $q^2=(\vec k - \vec k')^2$. Integrating over the angles gives $$\begin{aligned}
V(r) &=& -\frac{1}{\pi}\int_{0}^{\infty}\!d q\;q^2\;
\frac{\alpha_c(q^2)}{q^2}\;
\int_{-1}^{+1}\!\!\!\!d (\cos\theta)\;
\mathrm{e}^{-iq r \cos\theta}
\nonumber\\ &=& -
\frac{2}{\pi}\int_{0}^{\infty}\!d q\;{\alpha_c(q^2)}\;
\frac{\sin qr}{qr}
.\label{eq:3}\end{aligned}$$ But this integral does not exist!
The problem is as absent for short distances $r\to0$, as long as $\alpha_c(q^2)$ is integrable. The available ‘effective’ $\alpha_c(q^2)$ in [@vonSmekal:1997is; @Howe:2002rb; @Mattingly:1994; @Brodsky:2002nb] may or may not allow for $\int_{0}\!d q\;{\alpha_c(q^2)}$.
The problem resides at the upper limit. A Fourier transform of a function is defined only if its *limes superior* exists, that is, it exists only if $$\begin{aligned}
\int^{\infty}\!\frac{d q}{q}\;{\alpha_c(q^2)}\;\end{aligned}$$ is a well defined expression. But Eq.(\[eq:3\]) it not well defined. It diverges at the upper limit, since: $$\begin{aligned}
\int\!\frac{d q}{q \ln q} = \ln \ln q
\,.\end{aligned}$$ Knowing that asymptotic freedom is too weak a regulator, I have not even attempted in the past, to carry out point \[Brodsky:Fourier\] of the advocate. Baldicchi and Prosperi [@Baldicchi:2002] also know about the problem, but they veil it by inserting a linear potential as a regulator, by hand and out of desperation.
One concludes that the integral, Eq.(\[eq:2\]), *must be regulated*. What are possible alternatives?
Cutting-off at the upper limit à la $\int^{q_0}\!d q$ makes not much sense sense. A sharp cut-off with the step function $R(q)=\Theta(q^2-\lambda^2)$ generates uncontrollable oscillations. A soft regulator function, as for example $$\begin{aligned}
R(q) = \frac{\lambda^2}{\lambda^2+q^2}
\,,\label{eq:5}\end{aligned}$$ satisfies however all important requirements, among them $R(q)\Longrightarrow 1$ for $\lambda\Longrightarrow\infty$ and no extra physical dimensions. It is dimensionless.
In consequence one replaces Eq.(\[eq:2\]) by $$\begin{aligned}
V(r) &=& -\frac{1}{2\pi^2}\int\!d^3 q\;
\frac{\alpha_c(q^2)}{q^2}\;R(q)\;
\mathrm{e}^{-i\vec q \vec r}
,\end{aligned}$$ and Eq.(\[eq:3\]) by $$\begin{aligned}
V(r) &=& -
\frac{2}{\pi}\int_{0}^{\infty}\!d q\;{\alpha_c(q^2)}\;R(q)\;
\frac{\sin qr}{qr}
.\end{aligned}$$ One faces now a well defined mathematical problem.
But one faces another problem: The integrand behaves like: $$\begin{aligned}
\frac{4\pi}{\beta_0\mathrm{ln}(q^2/\Lambda^2_{QCD})}\;
\frac{\lambda^2}{\lambda^2+q^2}\;
\frac{\sin qr}{qr}
.\end{aligned}$$ The large q behavior is dominated completely by the regulator and the rapid oscillations of the sine function. As compared to them, the logarithm is hyper slowly varying. A mathematician would treat the problem by replacing the slowly varying term by a constant $\alpha'_c$ and take it out of the integral. (He would call this step the saddle point approximation.) He thus would get $$\begin{aligned}
V(r) &=& -
\frac{2\alpha'_c}{\pi}\int_{0}^{\infty}\!d q\;R(q)\;
\frac{\sin qr}{qr}
\\ &=& \displaystyle
\alpha'_c\Big[\frac{\mathrm{e}^{-\lambda r}}{r} -
\frac{1}{r}\Big]
\,.\label{eq:9}\end{aligned}$$ Here are the Coulomb’s and Yukawas which the advocate refers to in his point \[Brodsky:Yukawa\]. He is right, of course, in stating that they are identical at short distances $r \to 0$.
The notion ‘identical’ is misleading: The leading singularities cancel, but still $V(r\to 0)$ is non-zero. The systematic expansion for short distances rather gives $$\begin{aligned}
V(r) &=& -a + b\;r
\,,\label{eq:10}\end{aligned}$$ plus possible higher terms to which I shall come below.
The numerical values of $a$ and $b$ are less important for the argument. But for completeness, here they are: $$\begin{aligned}
a &=& \alpha'_c\;\lambda
\,,\\
b &=& \alpha'_c\;\lambda^2/2
\,.\end{aligned}$$ The considerations, which lead to Eq.(\[eq:10\]), also highlight another aspect. The only piece of non ambiguous experimental information is the coupling constant at very high momentum transfer: The ‘running’ constant at the $Z$ mass $M_Z\simeq 91.2\mbox{ GeV}$. Its experimental value is $\alpha_s(M_Z)\simeq 0.118$. The coupling constant $\alpha_s'=\alpha_s(q_0)$ can thus be taken at comparatively large values of $q_0$. The argument of $\alpha'_s=\alpha_s(Q=0)$ in [@Pau03a] is false, since $\alpha'_s=\alpha_s(q_0)$, but the preliminary value $$\begin{aligned}
\alpha'_s = 0.1695
\,,\end{aligned}$$ taken from [@Pau03a], is a useful first guess subject to later re adjustment. The numerical value of $\alpha'_s$ is comparatively small, much smaller in any case than the often quoted value $\alpha'_s = 0.7$. Setting the scale, like in [@Pau03a], $$\begin{aligned}
\lambda = 200\mbox{ MeV}
\,,\end{aligned}$$ determines all parameters.
The above argument shows that the asymptotic scale $\Lambda_{QCD}$ in Eq.(\[eq:1\]) is conceptually different from the regularization scale $\lambda$ in Eq.(\[eq:5\]). It is $\lambda$, not $\Lambda_{QCD}$, which sets the scale of the problem – contrary to point \[Brodsky:scale\] of the advocate. The confusion arises since both scales are often quoted with the same numerical value of about $200\mbox{ MeV}$, corresponding to a length scale of about $1\mbox{ fm}$.
The generalized regulator function
==================================
It should be emphasized that the limit $\lambda\to\infty$ may not be taken in Eq.(\[eq:9\]): For $\lambda\to\infty$, one is back to the undefined problem of Eq.(\[eq:3\]).
The same conclusion will also be reached with the renormalization group equations [@Pau03b]. The renormalization group equations [@Pau03b] require also that the regulator function has well defined derivatives with respect to $\lambda$. This excludes the step function of the sharp cut-off from the class of admitted regulator functions. The theta function is a distribution with only ill defined derivatives. Having understood these essentials, one can phrase things in a very simple way.
Do the above statements imply that the ‘soft’ regulator in Eq.(\[eq:5\]) is the only one admitted? — Of course not.
In fact, one has a large class of ‘generalized regulator functions’ [@Pau03a]: $$\begin{aligned}
R(q) = \left[1+\sum_{n=1}^{N}
(-1)^n s_n \lambda^n\frac{\partial^n}{\partial\lambda^n}\right]
\frac{\lambda^2}{\lambda^2+q^2}
\,.\label{eq:35}\end{aligned}$$ The partials $\lambda^n\,\partial^n/\partial\lambda^n$ are dimensionless and independent of a change in $\lambda$. The arbitrarily many coefficients $s_1,\dots,s_N$ are renormalization group invariants and, as such, subject to be determined by experiment.
The generalized regulator yields immediately the generalized Coulomb potential $$\begin{aligned}
V(r) &=& -\frac{\alpha_c}{r} \Big[1+\sum_{n=1}^{N}
(-1)^n s_n \lambda^n\frac{\partial^n}{\partial\lambda^n}\Big]
\Big(1-\mathrm{e}^{-\lambda r}\Big)
\nonumber\\ &=& \phantom{-}
\frac{\alpha_c}{r}\Big[
-1+ \mathrm{e}^{-{\lambda r}}\sum_{n=0}^{N}s_n({r\lambda})^n \Big] ,
\label{eq:36}\end{aligned}$$ with $s_0\equiv1$. This result illustrates an other important point: The power series in front of the exponential are nothing but a spelled out version of Laguerre polynomials. Laguerre polynomials are a complete set of functions. Eq.(\[eq:36\]) is thus potentially able to reproduce an arbitrary function of $r$. The description in terms of a generating function, as in Eqs.(\[eq:35\]) or (\[eq:36\]), is therefore *complete*.
**The large number of parameter** in Eq.(\[eq:36\]) can be controlled by the following construction: The coefficients $s_n$ in Eq.(\[eq:36\]) are expressed in terms of only three parameters $a$, $b$, and $c$, by $$\begin{aligned}
s_n =
\frac{1}{n!} + \frac{a}{(n-1)!} + \frac{b}{(n-2)!} + \frac{c}{(n-3)!}
\,.\label{eq:37}\end{aligned}$$ The first few coefficients are then explicitly $$\begin{aligned}
\begin{array}{rc rc rc rc rc rc }
s_0 &=& 1 &,& \\
s_1 &=& 1 &+& a &,& \\
s_2 &=& \frac{1}{2} &+& a &+& b &,& \\
s_3 &=& \frac{1}{6} &+& \frac{a}{2} &+& b &+& c &,& \\
s_4 &=& \frac{1}{24} &+& \frac{a}{6} &+& \frac{b}{2} &+& c &.&
\end{array} \end{aligned}$$ As a consequence, the dimensionless Coulomb potential depends on $r$ only through the dimensionless combination $y=\lambda r$: $$\begin{aligned}
W_N(y;a,b,c) =
\frac{V(r)}{\alpha_c\lambda} =
\frac{1}{y}\Big(-1+ \mathrm{e}^{-y}\sum_{n=0}^{N}s_ny^n\Big)
\,.\label{eq:39}\end{aligned}$$ In the near zone, it is a quadratic function of $y$, $$\begin{aligned}
W_N(y;a,b,c) = a + by + cy^2
\,,\label{eq:40}\end{aligned}$$ independent of the value of $N$. The remainder starts at most with power $y^{N+1}$. A value of $a=c=0$ and $b=1$ should yield a linear set of functions $W_N(y;0,1,0) \simeq y$ in the near zone. As shown in Fig. \[fig:lin\] with the original Mathematica plot, this happens to be true for surprisingly large values of $y$, *i.e.* not only for $y\ll 1$. The value of $N$ essentially controls the height of a barrier. Similarly, $W_N(y;0,0,1)=y^2$ generates a set of functions which are strictly quadratic in the near zone. Again, $N$ controls the height of a barrier.
**The physical picture** which develops is illustrated in Fig. \[fig:V(r)schem\]. In the far zone, for sufficiently large $r$, the potential energy coincides with the conventional Coulomb potential $-\frac{\alpha_c}{r}$. Since the potential is attractive, it can host bound states which are probably those realized in weak binding. In the near zone, for sufficiently small $r$, the potential behaves like a *power series* $c_0+c_1r+c_2r^2$ which potentially can host the bound states of strong coupling, provided the actual parameter values allow for that. In the intermediate zone, the actual potential must interpolate between these two extremes, since Eq.(\[eq:36\]) is an analytic function of $r$. Most likely this is done by developing a *barrier of finite height*, depending on the actual parameter values. The onset of the near and intermediate regimes must occur for relative distances of the quarks, which are comparable to the Compton wave length associated with their reduced mass. If the distance is smaller, one expects deviations from the classical regime by elementary considerations on quantum mechanics, indeed.
Conclusions and remaining mysteries
===================================
Once the arguments in Sec. \[sec:2\] are accepted for Quantum Chromodynamics (QCD), one must accept them also for Quantum Electrodynamics (QED). The integral in Eq.(\[eq:3\]) for QED diverges even stronger than for QCD. I thus arrive at the incredible conclusion that the potential energy between a muon and an electron is linear (or quadratic!) for sufficiently short distance.
Do I want to overthrow everything? — As a matter of fact: no! The scales are such that the hadronic length scale $R_H=\hbar/\lambda c\sim 1\mbox{ fm}$ is very much smaller than the Bohr scale $R_B=\hbar/m_ec\cdot 1/\alpha\sim 386\cdot 137\mbox{ fm}$. The electron simply ‘does not see’ the details at short distance except for a small perturbation. We know that from the hydrogen atom where the finite size of the proton gives only a small finite size correction.
The second important conclusion is that linear (or quadratic!) confinement can not go on forever. Confinement must be short distance phenomenon, in sharp contrast to the teleological or theological beliefs of the community. The hadronic potential allows for continuous spectra.
The only mystery remaining is that Lattice Gauge Theories and their representatives insist on a singularity at short distance, $$\begin{aligned}
V(r) &=& -\frac{\alpha'_c}{r} + b\;r \neq -a + b\;r
\,,\label{eq:12}\end{aligned}$$ although inherently to the method, Lattice Gauge Theory can not carry out calculations *at the singularity* $r=0$. The value of the constant $a$ can take arbitrarily large values, but it must be finite, by definition. I have insufficient working knowledge to comment any further on these questions.
“Subtle is the Lord, but nasty is he not.”
**Acknowledgement.** I thank my good friend and light cone mentor Stan Brodsky for an intense discussion by e-mail which popped up in early October 2003 and went over 10 rounds. I thank as well my friend Bogdan Povh for his continuous support over the years, and the encouragement to switch from nuclear physics to the light cone. I beg pardon to quote myself so often. I hate it that many authors refer to all their own papers ever written. But here, I have to do it by technical reasons. Sometimes it takes courage to be provincial, as Bogdan would say.
[99]{} H. C. Pauli, *Successful renormalization of a QCD-inspired Hamiltonian*. PiHPP, Trieste, 12 - 16 May 2003; arXiv:hep-ph/0310294. NUPP with CEBAF, Dubrovnik, 26 - 31 May, 2003. Light-cone meeting, Durham, 5 - 9 August, 2003. L. von Smekal, R. Alkofer and A. Hauck, *The infrared behavior of gluon and ghost propagators in Landau gauge QCD*, Phys. Rev. Lett. **79**, 3591 (1997). D. M. Howe and C. J. Maxwell, *All-orders infra-red freezing of R(e$^+$e$^-$) in perturbative QCD*, Phys. Lett. B **541**, 129 (2002); see also arXiv:hep-ph/0303163. A. C. Mattingly and P. M. Stevenson, *Optimization of R(e$^+$e$^-$) and ’freezing’ of the QCD couplant at low energies*, Phys. Rev. D **49**, 437 (1994). S. J. Brodsky, S. Menke, C. Merino and J. Rathsman, *On the behavior of the effective QCD coupling $\alpha(\tau)(s)$ at low scales,* Phys. Rev. D **67**, 055008 (2003). M. Baldicchi, G. M. Prosperi, *Infrared behavior of the running coupling constant and bound states in QCD*, Phys. Rev. D **66**, 074008 (2002). H. C. Pauli, *Possible solution to the problem of the non perturbative renormalization in a gauge theory (II)*, arXiv:hep-ph/0312xxx. H. C. Pauli, *On the fine and hyperfine interaction on the light cone (III)*, arXiv:hep-ph/0312xxx. H. C. Pauli, *A linear potential in a light cone QCD inspired model (IV)*, arXiv:hep-ph/0312xxx. H. C. Pauli, *A quadratic potential in a light cone QCD inspired model (V)*, arXiv:hep-ph/0312xxx.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'There exist a variety of star-galaxy classification techniques, each with their own strengths and weaknesses. In this paper, we present a novel meta-classification framework that combines and fully exploits different techniques to produce a more robust star-galaxy classification. To demonstrate this hybrid, ensemble approach, we combine a purely morphological classifier, a supervised machine learning method based on random forest, an unsupervised machine learning method based on self-organizing maps, and a hierarchical Bayesian template fitting method. Using data from the CFHTLenS survey, we consider different scenarios: when a high-quality training set is available with spectroscopic labels from DEEP2, SDSS, VIPERS, and VVDS, and when the demographics of sources in a low-quality training set do not match the demographics of objects in the test data set. We demonstrate that our Bayesian combination technique improves the overall performance over any individual classification method in these scenarios. Thus, strategies that combine the predictions of different classifiers may prove to be optimal in currently ongoing and forthcoming photometric surveys, such as the Dark Energy Survey and the Large Synoptic Survey Telescope.'
author:
- |
Edward J. Kim$^1$[^1], Robert J. Brunner$^{2,3,4}$, and Matias Carrasco Kind$^{2,4}$\
$^1$Department of Physics, University of Illinois, Urbana, IL 61801 USA\
$^2$Department of Astronomy, University of Illinois, Urbana, IL 61801 USA\
$^3$Department of Statistics, University of Illinois, Champaign, IL 61820 USA\
$^4$National Center for Supercomputing Applications, Urbana, IL 61801 USA
title: 'A Hybrid Ensemble Learning Approach to Star-Galaxy Classification'
---
\[firstpage\]
methods: data analysis – methods: statistical – surveys – stars: statistics – galaxies:statistics.
Introduction {#section:introduction}
============
The problem of source classification is fundamental to astronomy and goes as far back as @messier1781catalogue. A variety of different strategies have been developed to tackle this long-standing problem, and yet there is no consensus on the optimal star-galaxy classification strategy. The most commonly used method to classify stars and galaxies in large sky surveys is the morphological separation [@sebok1979optimal; @kron1980photometry; @valdes1982resolution; @yee1991faint; @vasconcellos2011decision; @henrion2011bayesian]. It relies on the assumption that stars appear as point sources while galaxies appear as resolved sources. However, currently ongoing and upcoming large photometric surveys, such as the Dark Energy Survey (DES[^2]) and the Large Synoptic Survey Telescope (LSST[^3]), will detect a vast number of unresolved galaxies at faint magnitudes. Near a survey’s limit, the photometric observations cannot reliably separate stars from unresolved galaxies by morphology alone without leading to incompleteness and contamination in the star and galaxy samples.
The contamination of unresolved galaxies can be mitigated by using training based algorithms. Machine learning methods have the advantage that it is easier to include extra information, such as concentration indices, shape information, or different model magnitudes. However, they are only reliable within the limits of the training data, and it can be difficult to extrapolate these algorithms outside the parameter range of the training data. These techniques can be further categorized into supervised and unsupervised learning approaches.
In supervised learning, the input attributes ([[e.g., ]{}]{}magnitudes or colors) are provided along with the truth labels ([[e.g., ]{}]{}star or galaxy). @odewahn1992automated pioneered the application of neural networks to the star-galaxy classification problem, and it has become a core part of the astronomical image processing software <span style="font-variant:small-caps;">SExtractor</span> [@bertin1996sextractor]. Other successfully implemented examples include decision trees [@weir1995automated; @suchkov2005census; @ball2006robust; @sevilla2015effect] and Support Vector Machines [@Fadely2012]. Unsupervised machine learning techniques are less common, as they do not utilize the truth labels during the training process, and only the input attributes are used.
Physically based template fitting methods have also been used for the star-galaxy classification problem [@robin2007stellar; @Fadely2012]. Template fitting approaches infer a source’s properties by finding the best match between the measured set of magnitudes (or colors) and the synthetic set of magnitudes (or colors) computed from a set of spectral templates. Although it is not necessary to obtain a high-quality spectroscopic training sample, these techniques do require a representative sample of theoretical or empirical templates that span the possible spectral energy distributions (SEDs) of stars and galaxies. Furthermore, they are not exempt from uncertainties due to measurement errors on the filter response curves, or from mismatches between the observed magnitudes and the template SEDs.
In this paper, we present a novel star-galaxy classification framework that combines and fully exploits different classification techniques to produce a more robust classification. In particular, we show that the combination of a morphological separation method, a template fitting technique, a supervised machine learning method, and an unsupervised machine learning algorithm can improve the overall performance over any individual method. In Section \[section:classification\_methods\], we describe each of the star-galaxy classification methods. In Section \[section:classification\_combination\_methods\], we describe different classification combination techniques. In Section \[section:data\], we describe the Canada-France Hawaii Telescope Lensing Survey (CFHTLenS) data set with which we test the algorithms. In Section \[section:results\_and\_discussion\], we compare the performance of our combination techniques to the performance of the individual classification techniques. Finally, we outline our conclusions in Section \[section:conclusions\].
Classification Methods {#section:classification_methods}
======================
In this section, we present four distinct star-galaxy classification techniques. The first method is a morphological separation method, which uses a hard cut in the half-light radius vs. magnitude plane. The second method is a supervised machine learning technique named TPC (Trees for Probabilistic Classification), which uses prediction trees and a random forest [@carrascokind2013tpz]. The third method is an unsupervised machine learning technique named SOMc, which uses self-organizing maps (SOMs) and a random atlas to provide a classification [@carrascokind2014somz]. The fourth method is a Hierarchical Bayesian (HB) template fitting technique based on the work by @Fadely2012, which fits SED templates from star and galaxy libraries to an observed set of measured flux values.
Collectively, these four methods represent the majority of all standard star-galaxy classification approaches published in the literature. It is very likely that any new classification technique would be functionally similar to one of these four methods. Therefore, any of these four methods could in principle be replaced by a similar method.
Morphological Separation
------------------------
The simplest and perhaps the most widely used approach to star-galaxy classification is to make a hard cut in the space of photometric attributes. As a first-order morphological selection of point sources, we adopt a technique that is popular among the weak lensing community [@Kaiser1995]. As Figure \[fig:morph\] shows, there is a distinct locus produced by point sources in the half-light radius (estimated by <span style="font-variant:small-caps;">SExtractor</span>’s FLUX\_RADIUS parameter) vs. the $i$-band magnitude plane. A rectangular cut in this size-magnitude plane separates point sources, which are presumed to be stars, from resolved sources, which are presumed to be galaxies. The boundaries of the selection box are determined by manually inspecting the size-magnitude diagram.
![Half-light radius vs. magnitude.[]{data-label="fig:morph"}](figures/morph.pdf){width="\columnwidth"}
One of the disadvantages of such cut-based methods is that it classifies every source with absolute certainty. It is difficult to justify such a decisive classification near a survey’s magnitude limits, where measurement uncertainties generally increase. A more informative approach is to provide probabilistic classifications. Although a recent work by @henrion2011bayesian implemented a probabilistic classification using a Bayesian approach on the morphological measurements alone, here we use a cut-based morphological separation to demonstrate the advantages of our combination techniques. In particular, we later show that the binary outputs ([[i.e., ]{}]{}0 or 1) of cut-based methods can be transformed into probability estimates by combining them with the probability outputs from other probabilistic classification techniques, such as TPC, SOMc, and HB.
Supervised Machine Learning: TPC
--------------------------------
TPC is a parallel, supervised machine learning algorithm that uses prediction trees and random forest techniques [@breiman1984classification; @breiman2001random] to produce a star-galaxy classification. TPC is a part of a publicly available software package called <span style="font-variant:small-caps;">MLZ</span>[^4] (Machine Learning for Photo-$z$). The full software package includes: <span style="font-variant:small-caps;">TPZ</span>, a supervised photometric redshift (photo-$z$) estimation technique [regression mode; @carrascokind2013tpz]; TPC, a supervised star-galaxy classification technique (classification mode); <span style="font-variant:small-caps;">SOM</span>$z$, an unsupervised photo-$z$ technique [regression mode; @carrascokind2014somz]; and SOMc, an unsupervised star-galaxy classification technique (classification mode).
TPC uses classification trees, a type of prediction trees that are designed to provide a classification or predict a discrete category. Prediction trees are built by asking a sequence of questions that recursively split the data into branches until a terminal leaf is created that meets a stopping criterion ([[e.g., ]{}]{}a minimum leaf size). The optimal split dimension is decided by choosing the attribute that maximizes the *Information Gain* ($I_G$), which is defined as
$$\label{eq:information_gain}
I_G \left(D_{\rmn{node}}, X\right)
= I_d \left( D_{\rmn{node}} \right)
- \sum_{x \in \rmn{values}(X)}
\frac{|D_{\rmn{node}, x}|}{|D_{\rmn{node}}|}
I_d \left( D_{\rmn{node}, x} \right),$$
where $D_{\rmn{node}}$ is the training data in a given node, $X$ is one of the possible dimensions ([[e.g., ]{}]{}magnitudes or colors) along which the node is split, and $x$ are the possible values of a specific dimension $X$. $|D_{\rmn{node}}|$ and $|D_{\rmn{node}, x}|$ are the size of the total training data and the number of objects in a given subset $x$ within the current node, respectively. $I_d$ is the impurity degree index, and TPC can calculate $I_d$ from any of the three standard different impurity indices: *information entropy*, *Gini impurity*, and *classification error*. In this work, we use the information entropy, which is defined similarly to the thermodynamic entropy:
$$I_d \left( D \right)
= - f_g \log_{2} f_g - \left(1 - f_g\right) \log_{2} \left(1 - f_g\right),$$
where $f_g$ is the fraction of galaxies in the training data. At each node in our tree, we scan all dimensions to identify the split point that maximizes the information gain as defined by Equation \[eq:information\_gain\], and select the attribute that maximizes the impurity index overall.
In a technique called random forest, we create bootstrap samples ([[i.e., ]{}]{}$N$ randomly selected objects with replacement) from the input training data by sampling repeatedly from the magnitudes and colors using their measurement errors. We use these bootstrap samples to construct multiple, uncorrelated prediction trees whose individual predictions are aggregated to produce a star-galaxy classification for each source.
We also use a cross-validation technique called Out-of-Bag [OOB; @breiman1984classification; @carrascokind2013tpz]. When a tree (or a map) is built in TPC (or SOMc), a fraction of the training data, usually one-third, is left out and not used in training the trees (or maps). After a tree is constructed using two-thirds of the training data, the final tree is applied to the remaining one-third to make a classification. This process is repeated for every tree, and the predictions from each tree are aggregated for each object to make the final star-galaxy classification. We emphasize that if an object is used for training a given tree, it is never used for subsequent prediction by that tree. Thus, the OOB data is an unbiased estimation of the errors and can be used as cross-validation data as long as the OOB data remain similar to the final test data set. The OOB technique can also provide extra information such as a ranking of the relative importance of the input attributes used in the prediction. The OOB technique can prove extremely valuable when calibrating the algorithm, when deciding which attributes to incorporate in the construction of the trees, and when combining this approach with other techniques.
Unsupervised Machine Learning: SOMc
-----------------------------------
A self-organizing map [@kohonen1990self; @kohonen2001self] is an unsupervised, artificial neural network algorithm that is capable of projecting high-dimensional input data onto a low-dimensional map through a process of competitive learning. In astronomical applications, the high-dimensional input data can be magnitudes, colors, or some other photometric attributes. The output map is usually chosen to be two-dimensional so that the resulting map can be used for visualizing various properties of the input data. The differences between a SOM and typical neural network algorithms are that a SOM is unsupervised, there are no hidden layers and therefore no extra parameters, and it produces a direct mapping between the training set and the output network. In fact, a SOM can be viewed as a non-linear generalization of a principal component analysis (PCA) algorithm [@yin2008self].
The key characteristic of SOM is that it retains the topology of the input training set, revealing correlations between input data that are not obvious. The method is unsupervised: the user is not required to specify the desired output during the creation of the lower-dimensional map, and the mapping of the components from the input vectors is a natural outcome of the competitive learning process.
During the construction of a SOM, each node on the two-dimensional map is represented by weight vectors of the same dimension as the number of attributes used to create the map itself. In an iterative process, each object in the input sample is individually used to correct these weight vectors. This correction is determined so that the specific neuron (or node), which at a given moment best represents the input source, is modified along with the weight vectors of that node’s neighboring neurons. As a result, this sector within the map becomes a better representation of the current input object. This process is repeated for every object in the training data, and the entire process is repeated for several iterations. Eventually, the SOM converges to its final form where the training data is separated into groups of similar features. Although the spectroscopic labels are not used at all in the learning process, they are used (only after the map has been constructed) to generate predictions for each cell in the resulting two-dimensional map.
In a similar approach to random forest in TPZ and TPC, SOM$z$ uses a technique called *random atlas* to provide photo-$z$ estimation [@carrascokind2014somz]. In random atlas, the prediction trees of random forest are replaced by maps, and each map is constructed from different bootstrap samples of the training data. Furthermore, we create random realizations of the training data by perturbing the magnitudes and colors by their measurement errors. For each map, we can either use all available attributes, or randomly select a subsample of the attribute space. This SOM implementation can also be applied to the classification problem, and we refer to it as SOMc in order to differentiate it from the photo-$z$ estimation problem (regression mode). We also use the random atlas approach in some of the classification combination approaches as discussed in Section \[section:classification\_combination\_methods\].
One of the most important parameter in SOMc is the topology of the two-dimensional SOM, which can be rectangular, hexagonal, or spherical. In our SOM implementation, it is also possible to use periodic boundary conditions for the non-spherical cases. The spherical topology is by definition periodic and is constructed by using <span style="font-variant:small-caps;">HEALPIX</span> [@gorski2005healpix]. Similar to TPC, we use the OOB technique to make an unbiased estimation of errors. We determine the optimal parameters by performing a grid search in the parameter space of different toplogies, as well as other SOM parameters, for the OOB data. We find that the spherical topology gives the best performance for the CFHTLenS data, likely due to its natural periodicity. Thus, we use a spherical topology to classify stars and galaxies in the CFHTLenS data. For a complete description of the SOM implementation and its application to the estimation of photo-$z$ probability density functions (photo-$z$ PDFs), we refer the reader to [@carrascokind2014somz].
Template fitting: Hierarchical Bayesian
---------------------------------------
One of the most common methods to classify a source based on its observed magnitudes is template fitting. Template fitting algorithms do not require a spectroscopic training sample; there is no need for additional knowledge outside the observed data and the template SEDs. However, any incompleteness in our knowledge of the template SEDs that fully span the possible SEDs of observed sources may lead to misclassification of sources.
Bayesian algorithms use Bayesian inference to quantify the relative probability that each template matches the input photometry and determine a probability estimate by computing the posterior that a source is a star or a galaxy. In this work, we have modified and parallelized a publicly available Hierarchical Bayesian (HB) template fitting algorithm by [@Fadely2012]. In this section, we provide a brief description of the HB template fitting technique; for the details of the underlying HB approach, we refer the reader to [@Fadely2012].
We write the posterior probability that a source is a star as
$$\label{eq:overall_posterior}
P \left( S | \bmath{x}, \bmath{\theta} \right)
= P \left( \bmath{x} | S, \bmath{\theta} \right)
P \left( S | \bmath{\theta} \right),$$
where $\bmath{x}$ represents a given set of observed magnitudes,. We have also introduced the *hyperparameter* $\bmath{\theta}$, a nuisance parameter that characterizes our uncertainty in the prior distribution. To compute the likelihood that a source is a star, we marginalize over all star and galaxy templates $\bmath{T}$. In a template-fitting approach, we marginalize by summing up the likelihood that a source has the set of magnitudes $\bmath{x}$ for a given star template as well as the likelihood for a given galaxy template:
$$\label{eq:marginalize_template}
P \left(\bmath{x} | S, \bmath{\theta} \right)
= \sum_{t \in \bmath{T}}
P \left(\bmath{x} | S, t, \bmath{\theta} \right)
P \left(t | S, \bmath{\theta} \right).$$
The likelihood of each template $P \left( \bmath{x} | S, \bmath{\theta} \right)$ is itself marginalized over the uncertainty in the template-fitting coefficient. Furthermore, for galaxy templates, we introduce another step that marginalizes the likelihood by redshifting a given galaxy template by a factor of $1 + z$.
Marginalization in Equation \[eq:marginalize\_template\] requires that we specify the prior probability $P \left(t | S, \bmath{\theta} \right)$ that a source has a spectral template $t$ (at a given redshift). Thus, the probability that a source is a star (or a galaxy) is either the posterior probability itself if a prior is used, or the likelihood itself if an uninformative prior is used. In a Bayesian analysis, it is preferable to use a prior, which can be directly computed either from physical assumptions, or from an empirical function calibrated by using a spectroscopic training sample. In an HB approach, the entire sample of sources is used to infer the prior probabilities for each individual source.
Since the templates are discrete in both SED shape and physical properties, we parametrize the prior probability of each template as a discrete set of weights such that
$$\sum_{t \in \bmath{T}}
P \left(t | S, \bmath{\theta} \right) = 1.$$
Similarly, we also parametrize the overall prior probability, $\left(S | \bmath{\theta}\right)$, in Equation \[eq:overall\_posterior\], as a weight. These weights correspond to the hyperparameters, which can be inferred by sampling the posterior probability distribution in the hyperparameter space. For the sampling, we use <span style="font-variant:small-caps;">emcee</span>, a Python implementation of the affine-invariant Markov Chain Monte Carlo (MCMC) ensemble sampler [@Foreman-Mackey2013].
As the goal of template fitting methods is to minimize the difference between observed and theoretical magnitudes, this approach heavily relies on both the use of SED templates and the accuracy of the transmission functions for the filters used for particular survey. For our stellar templates, we use the empirical SED library from @pickles1998stellar. The Pickles library consists of 131 stellar templates, which span all normal spectral types and luminosity classes at solar abundance, as well as metal-poor and metal-rich F–K dwarf and G–K giant and supergiant stars. We supplement the stellar library with 100 SEDs from @chabrier2000evolutionary, which include low mass stars and brown dwarfs with different $T_{\rmn{eff}}$ and surface gravities. We also include four white dwarf templates of @bohlin1995white, for a total of 235 templates in our final stellar library. For our galaxy templates, we use four CWW spectra from [-@coleman1980colors], which include an Elliptical, an Sba, an Sbb, and an Irregular galaxy template. When extending an analysis to higher redshifts, the CWW library is often augmented with two star bursting galaxy templates from [@kinney1996template]. From the six original CWW and Kinney spectra, intermediate templates are created by interpolation, for a total of 51 SEDs in our final galaxy library.
All of the above templates are convolved with the filter response curves to generate model magnitudes. These response curves consist of $u$, $g$, $r$, $i$, $z$ filter transmission functions for the observations taken by the Canada-France Hawaii Telescope (CFHT).
Classification combination methods {#section:classification_combination_methods}
==================================
Building on the work in the field of ensemble learning, we combine the predictions from individual star-galaxy classification techniques using four combination techniques. The main idea behind ensemble learning is to weight the predictions from individual models and combine them to obtain a prediction that outperforms every one of them individually [@rokach2010ensemble].
Unsupervised Binning {#section:random_atlas}
--------------------
Given the variety of star-galaxy classification methods we are using, we fully expect the relative performance of the individual techniques to vary across the parameter space spanned by the data. For example, it is reasonable to expect supervised techniques to outperform other techniques in areas of parameter space that are well-populated with training data. Similarly, we can expect unsupervised approaches such as SOM or template fitting approaches to generally perform better when a training sample is either sparse or unavailable.
We therefore adopt a binning strategy similar to [@carrascokind2014exhausting]. In this binning strategy, we allow different classifier combinations in different parts of parameter space by creating two-dimensional SOM representations of the full nine-dimensional magnitude-color space: $u$, $g$, $r$, $i$, $z$, $u-g$, $g-r$, $r-i$, and $i-z$. A SOM representation can be rectangular, hexagonal, or spherical; here we choose a 10$\times$10 rectangular topology to facilitate visualization as shown in Figure \[fig:som\_colors\]. We note that this choice is mainly for convenience and that the optimal topology and map size would likely depend on a number of factors, such as the number of objects and attributes. For all combination methods, we use only the OOB (cross-validation) data contained in each cell to compute the relative weights for the base classifiers. The weights within individual cells are then applied to the blind test data set to make the prediction.
Furthermore, we construct a collection of SOM representations and subsequently combine the predictions from each map into a meta-prediction. Given a training sample of $N$ sources, we generate $N_R$ random realizations of training data by perturbing the attributes with the measured uncertainty for each attribute. The uncertainties are assumed to be normally distributed. In this manner, we reduce the bias towards the data and introduce randomness in a systematic manner. For each random realization of a training sample, we create $N_M$ bootstrap samples of size $N$ to generate $N_M$ different maps.
After all maps are built, we have a total of $N_R \times N_M$ probabilistic outputs for each of the $N$ sources. To produce a single probability estimate for each source, we could take the mean, the median, or some other simple statistic. With a sufficient number of maps, we find that there is usually negligible difference between taking the mean and taking the median, and use the median in the following sections. We note that it is also possible to establish confidence intervals using the distribution of the probability estimates.
Weighted Average
----------------
The simplest approach to combine different combination techniques is to simply add the individual classifications from the base classifiers and renormalize the sum. In this case, the final probability is given by $$P\left(S | \bmath{x}, \bmath{M} \right)
= \sum_{i} P \left( S | \bmath{x}, M_{i} \right),$$
where $\bmath{M}$ is the set of models (TPC, SOMc, HB, and morphological separation in our work). We improve on this simple approach by using the binning strategy to calculate the weighted average of objects in each SOM cell separately for each map, and then combine the predictions from each map into a final prediction.
Bucket of Models (BoM)
----------------------
After the multi-dimensional input data have been binned, we can use the cross-validation data to choose the best model within each bin, and use only that model within that specific bin to make predictions for the test data. We use the mean squared error (MSE; also known as Brier score [@brier1950verification]) as a classification error metric. We define MSE as
$$\label{eq:mse}
\rmn{MSE} = \frac{1}{N} \sum^{N - 1}_{i = 0}
\left( y_i - \hat{y}_i \right)^2,$$
where $\hat{y}_i$ is the actual truth value ([[e.g., ]{}]{}0 or 1) of the $i^{\text{th}}$ data, and $y_{i}$ is the probability prediction made by the models. Thus, a model with the minimum MSE is chosen in each bin, and is assigned a weight of one, and zero for all other models. However, the chosen model is allowed to vary between different bins.
Stacking
--------
Instead of selecting a single model that performs best within each bin, we can train a learning algorithm to combine the output values of several other base classifiers in each bin. An ensemble learning method of using a meta-classifier to combine lower-level classifiers is known as *stacking* or *stacked generalization* [@wolpert1992stacked]. Although any arbitrary algorithm can theoretically be used as a meta-classifier, a logistic regression or a linear regression is often used in practice. In our work, we use a single-layer multi-response linear regression algorithm, which often shows the best performance [@breiman1996stacked; @ting1999issues]. This algorithm is a variant of the least-square regression algorithm, where a linear regression model is constructed for each class.
Bayesian Model Combination
--------------------------
We also use a model combination technique known as Bayesian Model Combination [BMC; @Monteith2011], which uses Bayesian principles to generate an ensemble combination of different classifiers. The posterior probability that a source is a star is given by
$$\label{eq:p_star_bmc}
P \left(S | \bmath{x}, \bmath{D}, \bmath{M}, \bmath{E} \right)
= \sum_{e \in \bmath{E}} P \left(S | \bmath{x}, \bmath{M}, e \right)
P \left(e | \bmath{D} \right),$$
where $\bmath{D}$ is the data set, and $e$ is an element in the ensemble space $\bmath{E}$ of possible model combinations. By Bayes’ Theorem, the posterior probability of $e$ given $\bmath{D}$ is given by
$$\label{eq:p_ensemble}
P \left(e | \bmath{D} \right)
= \frac{P \left(e \right)}{P \left(\bmath{D} \right)}
\prod_{d \in \bmath{D}} P \left( d | e \right)
\propto P \left(e\right) \prod_{d \in \bmath{D}} P \left(d | e \right).$$
Here, $P\left(e \right)$ is the prior probability of $e$, which we assume to be uniform. The product of $P\left(d | e\right)$ is over all individual data $d$ in the training data $\bmath{D}$, and $P\left(\bmath{D}\right)$ is merely a normalization factor and not important.
For binary classifiers whose output is either zero or one ([[e.g., ]{}]{}a cut-based morphological separation), we assume that each example is corrupted with an average error rate $\epsilon$. This means that $P\left(d|e\right) = 1-\epsilon$ if the combination $e$ correctly predicts class $\hat{y}_i$ for the $i^{\text{th}}$ object, and $P\left(d|e\right) = \epsilon$ if it predicts an incorrect class. The average rate $\epsilon$ can be estimated by the fraction $\left(M_g + M_s\right) / N$, where $M_g$ is the number of true galaxies classified as stars, $M_s$ is the number of true stars classified as galaxies, and $N$ is the total number of sources. Equation \[eq:p\_ensemble\] then becomes
$$\label{eq:p_ensemble_discrete}
P \left( e | \bmath{D} \right) \propto
P \left( e \right) \left(1 - \epsilon \right)^{N - M_s - M_g}
\left( \epsilon \right)^{M_s + M_g}.$$
For probabilistic classifiers, we can directly use the probabilistic predictions and write Equation \[eq:p\_ensemble\] as
$$\label{eq:p_ensemble_prob}
P \left( e | \bmath{D} \right) \propto
P \left( e \right) \prod_{i=0}^{N-1}
\hat{y}_i y_i +
\left(1 - \hat{y}_i\right) \left(1 - y_i\right).$$
Although the space $\bmath{E}$ of potential model combinations is in principle infinite, we can produce a reasonable finite set of potential model combinations by using sampling techniques. In our implementation, the weights of each combination of the base classifiers is obtained by sampling from a Dirichlet distribution. We first set all alpha values of a Dirichlet distribution to unity. We then sample this distribution $q$ times to obtain $q$ sets of weights. For each combination, we assume a uniform prior and calculate $P\left(e|\bmath{D}\right)$ using Equation \[eq:p\_ensemble\_discrete\] or \[eq:p\_ensemble\_prob\]. We select the combination with the highest $P\left(e|\bmath{D}\right)$, and update the alpha values by adding the weights of the most probable combination to the current alpha values. The next $q$ sets of weights are drawn using the updated alpha values.
We continue the sampling process until we reach a predefined number of combinations, and finally use Equation \[eq:p\_star\_bmc\] to compute the posterior probability that a source is a star (or a galaxy). In this paper, we use a $q$ value of three, and 1,000 model combinations are considered.
We also use a binned version of the BMC technique, where we use a SOM representation to apply different model combinations for different regions of the parameter space. We however note that introducing randomness though the construction of $N_R \times N_M$ different SOM representations does not show significant improvement over using only one single SOM representation. This similarity is likely due to the randomness that has already been introduced by sampling from the Dirichlet distribution. Thus, our BMC technique uses one SOM, while other base models (WA, BoM, and stacking) generate $N_R$ random realizations of $N_M$ maps.
Data {#section:data}
====
We use photometric data from the Canada-France-Hawaii Telescope Lensing Survey . This catalog consists of more than twenty five million objects with a limiting magnitude of $i_{\text{AB}} \approx 25.5$. It covers a total of 154 square degrees in the four fields (named W1, W2, W3, and W4) of CFHT Legacy Survey [CFHTLS; @gwyn2012canada] observed in the five photometric bands: $u$, $g$, $r$, $i$, and $z$.
We have cross-matched reliable spectroscopic galaxies from the Deep Extragalactic Evolutionary Probe Phase 2 [DEEP2; @davis2003science; @newman2013deep2], the Sloan Digital Sky Survey Data Release 10 [@Ahn2014 SDSS-DR10], the VIsible imaging Multi-Object Spectrograph (VIMOS) Very Large Telescope (VLT) Deep Survey [VVDS; @le2005vimos; @garilli2008vimos], and the VIMOS Public Extragalactic Redshift Survey [VIPERS; @garilli2014vimos]. We have selected only sources with very secure redshifts and no bad flags (quality flags -1, 3, and 4 for DEEP2; quality flag 0 for SDSS; quality flags 3, 4, 23, and 24 for VIPERS and VVDS). In the end, we have 8,545 stars and 57,843 galaxies available for the training and testing processes. We randomly select 13,278 objects for the blind testing set, and use the remainder for training and cross-validation. While HB uses only the magnitudes in the five bands, $u$, $g$, $r$, $i$, and $z$, TPC and SOMc are trained with a total of 9 attributes: the five magnitudes and their corresponding colors, $u-g$, $g-r$, $r-i$, and $i-z$. The morphological separation method uses <span style="font-variant:small-caps;">SExtractor</span>’s FLUX\_RADIUS parameter provided by the CFHTLenS catalog.
Our goal here is not to obtain the best classifier performance; for this we would have fine tuned individual base classifiers and chosen sophisticated models best suited to the particular properties of the CFHTLenS data. For example, [@hildebrandt2012cfhtlens] suggest that all objects with $i > 23$ in the CFHTLenS data set may be classified as galaxies without significant incompleteness and contamination in the galaxy sample. Although this approach works because the high Galactic latitude fields of the CFHTLS contain relatively few stars, it is very unlikely that such an approach will meet the science requirements for the quality of star-galaxy classification in lower-latitude, star-crowded fields. Rather, our goal for the CFHTLenS data set is to demonstrate the usefulness of combining different classifiers even when the base classifiers may be poor or trained on partial data. We also note that the relatively few number of stars in the CFHTLS fields might paint too positive a picture of completeness and purity, especially for the stars. Thus, we caution the reader that the specific completeness and purity values will likely vary in other surveys that observe large portions of the sky, and we emphasize once again that our aim is to highlight that there is a relative improvement in performance when we combine multiple star-galaxy classification techniques to generate a meta-classification.
Results and Discussion {#section:results_and_discussion}
======================
In this section, we present the classification performance of the four different combination techniques, as well as the individual star-galaxy classification techniques on the CFHTLenS test data.
Classification Metrics
----------------------
Probabilistic classification models can be considered as functions that output a probability estimate of each source to be in one of the classes ([[e.g., ]{}]{}a star or a galaxy). Although the probability estimate can be used as a weight in subsequent analyses to improve or enhance a particular measurement [@ross2011ameliorating], it can also be converted into a class label by using a threshold (a probability cut). The simplest way to choose the threshold is to set it to a fixed value, [[e.g., ]{}]{}$p_\rmn{cut} = 0.5$. This is, in fact, what is often done [[[e.g., ]{}]{} @henrion2011bayesian; @Fadely2012]. However, choosing $0.5$ as a threshold is not the best choice for an unbalanced data set, where galaxies outnumber stars. Furthermore, setting a fixed threshold ignores the operating condition ([[e.g., ]{}]{}science requirements, stellar distribution, misclassification costs) where the model will be applied.
### Receiver Operating Characteristic Curve
When we have no information about the operating condition when evaluating the performance of classifiers, there are effective tools such as the Receiver Operating Characteristic (ROC) curve [@swets2000better]. An ROC curve is a graphical plot that illustrates the true positive rate versus the false positive rate of a binary classifier as its classification threshold is varied. The Area Under the Curve (AUC) summarizes the curve information in a single number, and can be used as an assessment of the overall performance.
### Completeness and Purity
In astronomical applications, the operating condition usually translates to the completeness and purity requirements of the star or galaxy sample. We define the galaxy *completeness* $c_g$ (also known as recall or sensitivity) as the fraction of the number of true galaxies classified as galaxies out of the total number of true galaxies, $$c_g = \frac{N_g}{N_g + M_g},$$
where $N_g$ is the number of true galaxies classified as galaxies, and $M_g$ is the number of true galaxies classified as stars. We define the galaxy *purity* $p_g$ (also known as precision or positive predictive value) as the fraction of the number of true galaxies classified as galaxies out of the total number of objects classified as galaxies, $$p_g = \frac{N_g}{N_g + M_s},$$
where $M_s$ is the number of true stars classified as galaxies. Star completeness and purity are defined in a similar manner.
One of the advantages of a probabilistic classification is that the threshold can be adjusted to produce a more complete but less pure sample, or a less complete but more pure one. To compare the performance of probabilistic classification techniques with that of morphological separation, which has a fixed completeness ($c_g = 0.9964$, $c_s = 0.7145$) at a certain purity ($p_g = 0.9597$, $p_s = 0.9666$), we adjust the threshold of probabilistic classifiers until the galaxy completeness $c_g$ matches that of morphological separation to compute the galaxy purity $p_{g}$ at $c_g=0.9964$. Similarly, the star purity $p_{s}$ at $c_{s}=0.7145$ is computed by adjusting the threshold until the star completeness of each classifier is equal to that of morphological separation.
We can also compare the performance of different classification techniques by assuming an arbitrary operating condition. For example, weak lensing science measurements of the DES require $c_g > 0.960$ and $p_g > 0.778$ to control both the statistical and systematic errors on the cosmological parameters, and $c_s > 0.250$ and $p_s > 0.970$ for stellar Point Spread Function (PSF) calibration [@soumagnac2013star]. Although these values will likely be different for the science cases of the CFHTLenS data, we adopt these values to compare the classification performance at a reasonable operating condition. Thus, we compute $p_{g}$ at $c_g=0.960$ and $p_{s}$ at $c_s=0.250$. We also use the MSE defined in Equation \[eq:mse\] as a classification error metric.
Metric Meaning
-------------- ------------------------------------------
AUC Area under the Receiver Operating Curve
MSE Mean squared error
$c_g$ Galaxy completeness
$p_g$ Galaxy purity
$c_s$ Star completeness
$p_s$ Star purity
$p_g(c_g=x)$ Galaxy purity at $x$ galaxy completeness
$p_s(c_s=x)$ Star purity at $x$ star completeness
: The definition of the classification performance metrics.
\[table:metrics\]
Classifier Combination {#section:rich_training}
----------------------
\[table:metrics\_all\]
We present in Table \[table:metrics\_all\] the classification performance obtained by applying the four different combination techniques, as well as the individual star-galaxy classification techniques, on the CFHTLenS test data. The bold entries highlight the best technique for any particular metric. The first four rows show the performance of four individual star-galaxy classification techniques. Given a high-quality training data, it is not surprising that our supervised machine learning technique TPC outperforms other unsupervised techniques. TPC is thus shown in the first row as the benchmark.
The simplest of the combination techniques, WA and BoM, generally do not perform better than TPC. It is also interesting that, even with binning the parameter space and selecting the best model within each bin, BoM almost always chooses TPC as the best model in all bins, and therefore gives the same performance as TPC in the end. However, our BMC and stacking techniques have a similar performance and often outperform TPC. Although TPC shows the best performance as measured by the AUC, BMC shows the best performance in all other metrics.
![A two-dimensional 10$\times$10 SOM representation showing the mean $i$-band magnitude (top left), the fraction of true stars in each cell (top right), and the mean values of $u-g$ (bottom left) and $g-r$ (bottom right) for the cross-validation data.[]{data-label="fig:som_colors"}](figures/som_colors.pdf){width="\columnwidth"}
![A two-dimensional 10$\times$10 SOM representation showing the relative weights for the BMC combination technique applied to the four individual methods for the CFHTLenS data.[]{data-label="fig:weights"}](figures/weights.pdf){width="\columnwidth"}
In Figure \[fig:som\_colors\], we show in the top left panel the mean CFHTLenS $i$-band magnitude in each cell, and in the top right panel the fraction of stars in each cell. The bottom two panels show the mean $u-g$ and $g-r$ colors in each cell. These two-dimensional maps clearly show the ability of the SOM to preserve relationships between sources when it projects the full nine-dimensional space to the two-dimensional map. We note that these SOM maps should only be used to provide guidance, as the SOM mapping is a non-linear representation of all magnitudes and colors.
We can also use the same SOM from Figure \[fig:som\_colors\] to determine the relative weights for the four individual classification methods in each cell. We present the four weight maps for the BMC technique in Figure \[fig:weights\]. In these maps, a darker color indicates a higher weight, or equivalently that the corresponding classifier performs better in that region. These weight maps demonstrate the variation in the performance of the individual techniques across the two-dimensional parameter space defined by the SOM. Furthermore, since the maps in Figure \[fig:som\_colors\] and \[fig:weights\] are constructed using the same SOM, we can determine the region in the parameter space where each individual technique performs better or worse. Not surprisingly, the morphological separation performs best in the top left corner of the weight map in Figure \[fig:weights\], which corresponds to the brightest CFHTLenS magnitudes $i \la 20$ in the $i$-band magnitude map of Figure \[fig:som\_colors\]. It is also clear that the SOM cells where the morphological separation performs best have higher stellar fraction than the other cells. On the other hand, TPC seems to perform best in the region that corresponds to intermediate magnitudes $20\la i \la22.5$ and $1.5 \la u-g \la 3.0$. Our unsupervised learning method SOMc performs relatively better at fainter magnitudes $i \ga 21.5$ with $0 \la u-g \la 0.5$ and $0 \la g-r \la 0.5$. Although HB shows the worst performance when there exists a high-quality training data set, BMC still utilizes information from HB, especially at intermediate magnitudes $20\la i \la22$. Another interesting pattern is that the four techniques seem complementary, and they are weighted most strongly in different regions of the SOM representation.
In Figure \[fig:purity\_mag\], we compare the star and galaxy purity values for BMC, TPC, and morphological separation as functions of $i$-band magnitude. We use the kernel density estimation [KDE; @silverman1986density] with the Gaussian kernel to smooth the fluctuations in the distribution. Although morphological separation shows a slightly better performance in galaxy purity at bright magnitudes $i \la 20$, BMC outperforms both TPC and morphological separation at faint magnitudes $i \ga 21$. As the top panel shows, the number count distribution peaks at $i \sim 22$, and BMC therefore outperforms both TPC and morphological separation for the majority of objects. It is also clear that BMC outperforms TPC over all magnitudes. BMC can presumably accomplish this by combining information from all base classifiers, [[e.g., ]{}]{}giving more weight to the morphological separation method at bright magnitudes. The bottom panel shows that the star purity of morphological separation drops to $p_s < 0.8$ at fainter magnitudes $i > 21$. This is expected, as our crude morphological separation classifies every object as a galaxy beyond $i > 21$, and purity measures the number of true stars classified as stars. It is again clear that BMC outperforms both TPC and morphological separation in star purity values over all magnitudes.
In Figure \[fig:purity\_mag\_integrated\], we show the cumulative galaxy and star purity values as functions of magnitude. Although morphological separation performs better than TPC at bright magnitudes, its purity values decrease as the magnitudes become fainter, and TPC eventually outperforms morphological separation by 1–2% at $i > 21$. BMC clearly outperforms both TPC and morphological separation, and it maintains the overall galaxy purity of 0.980 up to $i \sim 24.5$.
![Purity as a function of the $i$-band magnitude as estimated by the kernel density estimation (KDE) method. The top panel shows the histogram with a bin size of 0.1 mag and the KDE for objects in the test set. The second panel shows the fraction of stars estimated by KDE as a function of magnitude. The bottom two panels compare the galaxy and star purity values for BMC, TPC, and morphological separation as functions of magnitude. Results for BMC, TPC, and morphological separation are in blue, green, and red, respectively. The $1 \sigma$ confidence bands are estimated by bootstrap sampling.[]{data-label="fig:purity_mag"}](figures/purity_mag.pdf){width="\columnwidth"}
![Cumulative purity as a function of the $i$-band magnitude. The upper panel compares the galaxy purity values for BMC (blue solid line), TPC (green dashed line), and morphological separation (red dashed line). The lower panel compares the star purity. The $1 \sigma$ error bars are computed following the method of @paterno2004calculating to avoid the unphysical errors of binomial or Poisson statistics.[]{data-label="fig:purity_mag_integrated"}](figures/purity_mag_integrated.pdf){width="\columnwidth"}
We also show the star and galaxy purity values as functions of photometric redshift estimate in Figure \[fig:purity\_z\]. Photo-$z$ is estimated with the <span style="font-variant:small-caps;">BPZ</span> algorithm [@benitez2000bayesian] and provided with the CFHTLenS photometric redshift catalogue [@hildebrandt2012cfhtlens]. The advantage of BMC over TPC and morphological separation is now more pronounced in Figure \[fig:purity\_z\]. Although the morphological separation method outperforms BMC at bright magnitudes in Figure \[fig:purity\_mag\], it is clear that BMC outperforms both TPC and morphological separation over all redshifts. We also present in Figure \[fig:purity\_g\_r\] how the star and galaxy purity values vary as a function of $g-r$ color. It is again clear that BMC outperforms both TPC and morphological separation over all $g-r$ colors.
![Similar to Figure \[fig:purity\_mag\] but as a function of photo-$z$. The bin size of histogram in the top panel is 0.02.[]{data-label="fig:purity_z"}](figures/purity_z.pdf){width="\columnwidth"}
![Similar to Figure \[fig:purity\_mag\] but as a function of $g-r$ color. The bin size of histogram in the top panel is 0.05.[]{data-label="fig:purity_g_r"}](figures/purity_g_r.pdf){width="\columnwidth"}
In Figure \[fig:p\_dist\_all\], we show the distribution of $P(S)$, the posterior probability that an object is a star, for BMC, TPC, and morphological separation. It is interesting that the BMC technique assigns a posterior star probability $P(S) \la 0.3$ to significantly more true galaxies than TPC, and a probability $P(S) \ga 0.8$ to significantly fewer true galaxies. By utilizing information from different types of classification techniques in different parts of the parameter space, BMC becomes more certain that an object is a star or a galaxy, resulting in improvement of overall performance.
![Histogram of the posterior probability that a source is a star for morphological separation (top), TPC (middle), and BMC (bottom) for a high-quality training data set. The true galaxies are in green, and true stars are in blue. The bin size is 0.05.[]{data-label="fig:p_dist_all"}](figures/p_dist.pdf){width="\columnwidth"}
Heterogeneous Training {#section:poor_training}
----------------------
It is very costly in terms of telescope time to obtain a large sample of spectroscopic observations down to the limiting magnitude of a photometric sample. Thus, we investigate the impact of training set quality by considering a more realistic case where the training data set is available only for a small number of objects with bright magnitudes. To emulate this scenario, we only use objects that have spectroscopic labels from the VVDS 0226-04 field (which is located within the CFHTLS W1 field) and impose a magnitude cut of $i < 22.0$ in the training data, leaving us a training set with only 1,365 objects. We apply the same four star-galaxy classification techniques and four combination methods, and measure the performance of each technique on the same test data set from Section \[section:rich\_training\]. As the top two panels of Figures \[fig:purity\_mag\_cut\], \[fig:purity\_z\_cut\], and \[fig:purity\_g\_r\_cut\] show, the demographics of objects in the training set are different from the distribution of sources in the test set. Thus, this also serves as a test of the efficacy of heterogeneous training.
\[table:metrics\_cut\]
We present in Table \[table:metrics\_cut\] the same six metrics for each method, and highlight the best method for each metric. Overall, the results obtained for the reduced data set are remarkable. With a smaller training set, our training based methods, TPC and SOMc, suffer a significant decrease in performance. The performance of morphological separation and HB is essentially unchanged from Table \[table:metrics\_all\] as they do not depend on the training data. Without sufficient training data, the advantage of combining the predictions of different classifiers is more obvious. Even WA, the simplest of combination techniques, outperforms all individual classification techniques in four metrics, AUC, $p_s$ at $c_s=0.7145$, $p_g$ at $c_g=0.9600$, and $p_s$ at $c_s=0.2500$. Although BoM always chooses TPC as the best model when we have a high-quality training set, it now chooses various methods in different bins and outperforms all base classifiers. While the performance of the stacking technique is only slightly worse than that of BMC when we have a high-quality training set, stacking now fails to outperform morphological separation. BMC shows an impressive performance and outperforms all other classification techniques in all six metrics. Overall, the improvements are small but still significant since these metrics are averaged over the full test data.
![Similar to Figure \[fig:som\_colors\] but for the reduced training data set. []{data-label="fig:som_colors_cut"}](figures/som_colors_cut.pdf){width="\columnwidth"}
![Similar to Figure \[fig:weights\] but for the reduced training data set. []{data-label="fig:weights_cut"}](figures/weights_cut.pdf){width="\columnwidth"}
In Figure \[fig:weights\_cut\], we again show the $10\times10$ two-dimensional weight map defined by the SOM. When the quality of training data is relatively poor, the performance of training based algorithms will decrease, while the performance of template fitting algorithms or morphological separation methods is independent of training data. Thus, when the weight maps of Figure \[fig:weights\] and Figure \[fig:weights\_cut\] are visually compared, it is clear that the BMC algorithm now uses more information from morphological separation and HB, while it uses considerably less information from our training based algorithms, TPC and SOMc. Not surprisingly, the morphological separation method performs best at bright magnitudes, and BMC assigns more weight to HB at fainter magnitudes.
We present the star and galaxy purity values as functions of $i$-band magnitude in Figure \[fig:purity\_mag\_cut\]. The normalized density distribution as a function of magnitude in the top panel and the stellar distribution in the second panel clearly show that the demographics of the training set and that of the test set are different. Since the training set is cut at $i < 22$, the density distribution falls off sharply around $i \sim 22$ and has a higher fraction of stars than the test set. Compared to the purity values in Figure \[fig:purity\_mag\], TPC now suffers a significant decrease in star and galaxy purity. However, the purity of BMC does not show such a significant drop and decreases by only 2–5%. As suggested by the weight maps in Figure \[fig:weights\_cut\], BMC can accomplish this by shifting the relative weights assigned to each base classifier in different SOM cells. As the quality of training set worsens, BMC assigns less weight to training based methods and more weight to HB and morphological separation.
In Figure \[fig:purity\_mag\_cut\_integrated\], we show the cumulative galaxy and star purity values as functions of magnitude. Compared to Figure \[fig:purity\_mag\_integrated\], the drop in the performance of TPC is clear. However, even when some classifiers have been trained on a significantly reduced training set, BMC maintains a galaxy purity of 0.970 and a star purity of 1.0 up to $i \sim 24.5$, and it sill outperforms morphological separation at fainter magnitudes $i \ga 21$.
![Purity as a function of the $i$-band magnitude for the reduced training data set. Top panel shows the histograms and KDEs for the number count distribution for the training (blue) and test (green) data set. The second panel shows the fraction of stars in the training and test data set in blue and green, respectively. The bottom two panels compare the galaxy and star purity values for BMC, TPC, and morphological separation as functions of $i$-band magnitude.[]{data-label="fig:purity_mag_cut"}](figures/purity_mag_cut.pdf){width="\columnwidth"}
![Similar to Figure \[fig:purity\_mag\_integrated\] but for the reduced training data set.[]{data-label="fig:purity_mag_cut_integrated"}](figures/purity_mag_cut_integrated.pdf){width="\columnwidth"}
We also show the star and galaxy purity values as functions of photo-$z$ in Figure \[fig:purity\_z\_cut\] and as functions of $g-r$ in Figure \[fig:purity\_g\_r\_cut\]. Compared to Figure \[fig:purity\_z\] and \[fig:purity\_g\_r\], the performance of BMC becomes worse in some photo-$z$ and $g-r$ bins. However, this drop in performance seems to be confined to only a small number of objects in particular regions of the parameter space, and BMC still outperforms both TPC and morphological separation for the majority of objects.
![Similar to Figure \[fig:purity\_mag\_cut\] but as a function of photo-$z$.[]{data-label="fig:purity_z_cut"}](figures/purity_z_cut.pdf){width="\columnwidth"}
![Similar to Figure \[fig:purity\_mag\_cut\] but as a function of $g-r$ color.[]{data-label="fig:purity_g_r_cut"}](figures/purity_g_r_cut.pdf){width="\columnwidth"}
Compared to Figure \[fig:p\_dist\_all\], the difference between the posterior star probability distribution of TPC and that of BMC is now more pronounced in Figure \[fig:p\_dist\_cut\]. The $P\left(S\right)$ distribution of BMC for true galaxies falls off sharply at $P\left(S\right)\approx0.95$, and BMC does not assign a star probability $P(S) \ga 0.95$ to any true galaxies, On the other hand, both TPC and morphological separation classify some true galaxies as stars with absolute certainty.
![Similar to Figure \[fig:p\_dist\_all\] but for the reduced training data set.[]{data-label="fig:p_dist_cut"}](figures/p_dist_cut.pdf){width="\columnwidth"}
The Quality of Training Data {#section:quality_training}
----------------------------
The combination techniques that we have demonstrated so far use two training based algorithms as base classifiers. Ideally, the training data should mirror the entire parameter space occupied by the data to be classified. Yet we have seen in Section \[section:poor\_training\] that the BMC technique does reliably extrapolate past the limits of the training data, even when some base classifiers are trained on a low-quality training data set. In this section, we further investigate if and where BMC begins to break down by imposing various magnitude, photo-$z$, and color cuts to change the size and composition of the training set.
In Figure \[fig:perform\_mag\_cut\], we present a visual comparison between different classification techniques, when various magnitude cuts are applied on the training data, and the performance is measured on the same test set from Section \[section:rich\_training\] and \[section:poor\_training\]. It is not surprising that the performance of TPC decreases as we decrease the size of training set by imposing more restrictive magnitude cuts, while the performance of HB and morphological separation is essentially unchanged. However, the effect of change in size and composition of the training set is significantly mitigated by the use of the BMC technique. BMC outperforms both HB and TPC in all four metrics, even when the training set is restricted to $i < 20.0$. BMC also consistently outperforms morphological separation until we impose a magnitude cut of $i < 20.0$ on the training data, beyond which point BMC finally performs worse than morphological separation. It is remarkable that BMC is able to reliably extrapolate past the training data to $i \sim 24.5$, the limiting magnitude of the test set, and outperform HB, TPC, and morphological separation in all performance metrics, even the demographics of training set do not accurately sample the data to be classified.
Similarly, we impose various spectroscopic redshift cuts on the training data in Figure \[fig:perform\_z\_cut\]. Since all stars have $z_{\rmn{spec}}$ values close to zero, we are effectively changing the demographics of training set by keeping all stars and gradually removing galaxies with high redshifts. BMC begins to perform worse than morphological separation when a conservative cut of $z_{\rmn{spec}} < 0.6$ is imposed. However, it is again clear that BMC is able to utilize information from HB and morphological separation to mitigate the drop in the performance of TPC.
In Figure \[fig:perform\_g\_r\_cut\], we decrease the size of training set by keeping red objects and gradually removing blue objects. A color cut seems to have a more pronounced effect on the performance of TPC and BMC, which perform worse than morphological separation when the training set is restricted to $g - r > 0.4$. The performance depends more strongly on the color distribution, because a significant fraction of blue objects consists of stars, while objects with fainter magnitudes and higher redshifts are mostly galaxies. We can verify this in Figure \[fig:som\_colors\], where the darker (higher stellar fraction) cells in the upper middle region of the stellar fraction map (top right panel) have bright magnitudes $i \la 20$ in the $i$-band magnitude map (top left panel) and blue colors $g-r \la 0.5$ in the $g-r$ color map (bottom right panel). On the other hand, the darker (fainter magnitude) cells in the right-hand side of the $i$-band magnitude map have almost no stars in them and are represented by bright (low stellar fraction) cells in the stellar fraction map. Thus, these results indicate that the performance of training based methods depends more strongly on the composition of training data than on the size, and it is necessary to have a sufficient number of the minority class in the training data set to ensure optimal performance.
![The classification performance metrics for BMC (blue), TPC (green), morphology (red), and HB (purple) as applied to the CFHTLenS data in the VVDS field with various magnitude cuts. The top panel shows the number of sources in the training set at corresponding magnitude cuts. We show only one of the four combination methods, BMC, which has the best overall performance.[]{data-label="fig:perform_mag_cut"}](figures/perform_mag_cut.pdf){width="\columnwidth"}
![Similar to Figure \[fig:perform\_mag\_cut\] but using $z_{\text{spec}}$ cuts.[]{data-label="fig:perform_z_cut"}](figures/perform_z_cut.pdf){width="\columnwidth"}
![Similar to Figure \[fig:perform\_mag\_cut\] but using $g - r$ color cuts.[]{data-label="fig:perform_g_r_cut"}](figures/perform_g_r_cut.pdf){width="\columnwidth"}
Conclusions {#section:conclusions}
===========
We have presented and analyzed a novel star-galaxy classification framework for combining star-galaxy classifiers using the CFHTLenS data. In particular, we use four independent classification techniques: a morphological separation method; TPC, a supervised machine learning technique based on prediction trees and a random forest; SOMc, an unsupervised machine learning approach based on self-organizing maps and a random atlas; and HB, a Hierarchical Bayesian template-fitting method that we have modified and parallelized. Both TPC and SOMc algorithms are currently available within a software package named <span style="font-variant:small-caps;">MLZ</span>[^5]. Our implementation of HB and BMC, as well as <span style="font-variant:small-caps;">IPython</span> notebooks that have been used to produce the results in this paper, are available at <https://github.com/EdwardJKim/astroclass>.
Given the variety of star-galaxy classification methods we are using, we fully expect the relative performance of the individual techniques to vary across the parameter space spanned by the data. We therefore adopt the binning strategy, where we allow different classifier combinations in different parts of parameter space by creating two-dimensional self-organizing maps of the full multi-dimensional magnitude-color space. We apply different star-galaxy classification techniques within each cell of this map, and find that the four techniques are weighted most strongly in different regions of the map.
Using data from the CFHTLenS survey, we have considered different scenarios: when an excellent training set is available with spectroscopic labels from DEEP2, SDSS, VIPERS, and VVDS, and when the demographics of sources in a low-quality training set do not match the demographics of objects in the test data set. We demonstrate that the Bayesian Model Combination (BMC) technique improves the overall performance over any individual classification method in both cases. We note that @carrascokind2014exhausting analyzed different techniques for combining photometric redshift probability density functions (photo-$z$ PDFs) and also found that BMC is in general the best photo-$z$ PDF combination technique.
The problem of star-galaxy classification is a rich area for future research. It is unclear if sufficient training data will be available in future ground-based surveys. Furthermore, in large sky surveys such as DES and LSST, photometric quality is not uniform across the sky, and a purely morphological classifier alone will not be sufficient, especially at faint magnitudes. Given the efficacy of our approach, classifier combination strategies are likely the optimal approach for currently ongoing and forthcoming photometric surveys. We therefore plan to apply the combination technique described in this paper to other surveys such as the DES. Our approach can also be extended more broadly to classify objects that are neither stars nor galaxies ([[e.g., ]{}]{}quasars). Finally, future studies could explore the use of multi-epoch data, which would be particularly useful for the next generation of synoptic surveys.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors thank the referee for a careful reading of the manuscript and comments that improved this work. We thank Ignacio Sevilla for helpful and insightful conversations. We acknowledge support from the National Science Foundation Grant No. AST-1313415. RJB acknowledges support as an Associate within the Center for Advanced Study at the University of Illinois.
This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1053575.
This work is based on observations obtained with MegaPrime/MegaCam, a joint project of CFHT and CEA/DAPNIA, at the Canada-France-Hawaii Telescope (CFHT) which is operated by the National Research Council (NRC) of Canada, the Institut National des Sciences de l’Univers of the Centre National de la Recherche Scientifique (CNRS) of France, and the University of Hawaii. This research used the facilities of the Canadian Astronomy Data Centre operated by the National Research Council of Canada with the support of the Canadian Space Agency. CFHTLenS data processing was made possible thanks to significant computing support from the NSERC Research Tools and Instruments grant program.
Funding for the DEEP2 survey has been provided by NSF grants AST-0071048, AST-0071198, AST-0507428, and AST-0507483.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
This paper uses data from the VIMOS Public Extragalactic Redshift Survey (VIPERS). VIPERS has been performed using the ESO Very Large Telescope, under the “Large Programme” 182.A-0886. The participating institutions and funding agencies are listed at http://vipers.inaf.it/.
This research uses data from the VIMOS VLT Deep Survey, obtained from the VVDS database operated by Cesam, Laboratoire d’Astrophysique de Marseille, France.
\[lastpage\]
[^1]: jkim575@illinois.edu
[^2]: http://www.darkenergysurvey.org/
[^3]: http://www.lsst.org/lsst/
[^4]: http://lcdm.astro.illinois.edu/code/mlz.html
[^5]: http://lcdm.astro.illinois.edu/code/mlz.html
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper describes an enumeration of all words having a combinatoric property called “rhythmic oddity property”named *rop-words*. This property was introduced by Simha Aron in the 1990s. The set of rop-words is not a subset of the set of Lyndon words, but is very closed. We show that there is a bijection between some necklaces and rop-words. This leads to a formula for counting the rop-words of a given length. <span style="font-variant:small-caps;">Keywords:</span> Combinatoric on words. Lyndon words. Rhythmic oddity. Music formalization'
author:
- Franck Jedrzejewski
title: Counting words satisfying the rhythmic oddity property
---
The *rhythmic oddity property* was discovered by ethnomusicologist Simha Aron [@Aro1991] in the study of Aka pygmies music. The rhythms satisfying this property are refinement of aksak rhythms described by C. Brăiloiu [@Bra1952] in 1952, and used also in Turkish and Bulgarian music. They have been studied by M. Chemillier and C. Truchet in [Che2003]{} and André Bouchet in [@Bou2010] who gave some nice characterizations. Related problems of asymmetric rhythms have been studied by Rachael Hall and Paul Klingsberg [@Hal2004; @Hal2006]. In this paper, we carry on these studies by showing a one to one correspondance between some necklaces and words satisfying the rhythmic oddity property.
The rhythmic oddity property
============================
Patterns with rhythmic oddity property are combinations of durations equal to 2 or 3 units, such as the famous Aka pygmies rhythm 32222322222, and such that when placing the sequence on a circle, “one cannot break the circle into two parts of equal length whatever the chosen breaking point.”In the langage of combinatorics of words, this property in terms of words $%
\omega $ over the alphabet $A =\{2 ,3\}$ is defined as follows. The height $%
h (\omega )$ of a word $\omega =\omega _{0} \omega _{1} \ldots \omega _{n
-1} $ of length $n$ is by definition the sum of its letters $h (\omega ) =
\sum _{j =0 ,\ldots ,n -1}\omega _{j}$. A word $\omega $ satisfies the rhythmic oddity property (rop) if $h (\omega )$ is even and no cyclic shift of $\omega $ can be factorized into two words $u v$ such that $h (u) =h (v)%
\text{.}$ For short, we call *rop*-word a word over the alphabet $\{2
,3\}$ satisfying the rhythmic oddity property. For instance, the word $%
32322 $ of height 12 is a rop-word, as well as all words of the form $32^{n}
32^{n +1}$ for all non-negative integers $n$, where the notation $2^{n}$ means the letter 2 is repeated $n$ times. The properties of rop-words have been outine in [@Che2003]. But contrary to what is sometimes written, the set of rop-words is not a subset of the set of Lyndon words. The words 222 and 233233233 are rop-words, but not Lyndon words. A Lyndon word is a string that is strictly smaller in lexicographic order than all of its rotations. Conversely, the set of Lyndon words is not included in the set of rop-words, since 2233 is a Lyndon word, but not a rop-word (the words 23 and 32 have the same height and 2332 is a rotation of 2233). We call a Lyndon rop-word a word of the monoid $\{2 ,3\}^{ \ast }$ that is both a Lyndon word and a rop-word. The aim of this paper is to count the number of Lyndon rop-words and the number of rop-words of length $n$.
A bijection between some necklaces and rop-words
================================================
André Bouchet gave some characterizations of rop-words in [@Bou2010]. We present the main results of his paper. Let $\varepsilon $ be the empty word and $\omega $ a word over $\{2 ,3\}$. The cycle $\delta $ of $%
\omega $ is defined by $\delta (\varepsilon ) =\varepsilon $ and $\delta (a
\omega ) =\omega a\text{,}$ for $a \in \{2 ,3\}$. The rotations of $\omega $ are the words $\delta ^{k} (\omega )\text{,}$ for all positive integer $k
>0$. In his paper, Bouchet shows the following lemma.
Let $\omega =\omega _{0} \omega _{1} \ldots \omega _{n -1}$ be a word over the alphabet $A =\{2 ,3\}$ of height $2 h$. The word $\omega $ is a rop-word if and only if the two conditions are satisfied:
\(i) The length of $\omega $ is odd, say $2\ell +1$.
\(ii) The height of the prefixes of length $\ell $ of the rotations of $%
\omega $ are equal to $h-2$ or $h-1$.
From this lemma, André Bouchet shows the following theorem. Let $\omega $ be a word of length $n$ and $d$ be an integer such that $0 <d \leq n/2$. A $%
d $-pairing of $\omega $ is a partition of the subset of indices $\{i :0
\leq i <n ,\omega _{i} =3\}$ into pairs of indices $\{j ,j +d\}$. Arithmetic operations on indices are to be understood $\mathop{\rm mod} ~n$.
Let $\omega $ be a word of even height. $\omega $ is a rop-word if and only if the two conditions are satisfied:
\(i) The length of $\omega $ is odd, say $2 \ell +1$.
\(ii) $\omega $ admits a $\ell $-pairing.
Let $n_{2}$ and $n_{3}$ be the number of symbols 2 and 3 in $\omega $ and $%
n =n_{2} +n_{3}$ be the length of $\omega \text{.}$ For a given $n_{2}$, we will use the $d$-pairing to show that there is a one to one correspondance between aperiodic necklaces of length $n$ with $n_{2}$ black beads (represented by letter 2) and $(n -n_{2})$ white beads (represented by letter 3) and Lyndon rop-words of length $n^{ \prime } =2 n -n_{2}$ with $%
n_{2}^{ \prime } =n_{2}$ letters 2 and $n_{3}^{ \prime } =2 (n -n_{2})$ letters 3. And also a one-to-one correspondance between necklaces (eventually periodic) of length $n$ with $n_{2}$ black beads (represented by letter 2) and $(n -n_{2})$ white beads (represented by letter 3) and rop-words. The correspondance is obtained by adding or removing the letters 3 coming from the pairing. Let us examine an example. (See fig. \[Cycle\])
Fix $n_{2}$, for instance $n_{2}=3$, and let $n$ be $n=5$. The word 2233233 is a (Lyndon) rop-word with odd length 7 ($n_{2}^{\prime }=3$, $%
n_{3}^{\prime }=4$) since it has a 3-pairing. Put the word on a circle, starting from the bottom and turn counterclockwise as shown on the figure \[Cycle\]. Now discard the second 3 of each pairing $(3,3)$ turning counterclockwise. Reading the remaining word clockwise starting from the bottom gives the word 22332, one of the two necklaces of length 5 with 3 letters 2. Conversely, starting from the word 22233, it is easy to add a 3-pairing by doubling each letter 3, with respect to the counterclockwise tour.
![Cyclic representation of rop-words[]{data-label="Cycle"}](rop01.jpg){width="7cm"}
Let $\omega =\omega _{0}\omega _{1}\ldots \omega _{n-1}$ be a word of $%
\{2,3\}^{\ast }$ and $p$ coprime with $n$. Denote by $\omega
^{(p)}=x_{0}x_{1}\ldots x_{n-1}$ the word obtained by reading all letters of $\omega $ by step $p$, starting from $\omega _{0}\text{.}$ Namely, each letter of $\omega ^{(p)}$ is $x_{j}=\omega _{k}$ with $k=jp~\mathop{\rm mod}
$ $p$, $0\leq j<p\text{.}$ For instance, the word $\omega =2233233$ depicted on fig. \[Cycle\] with $n=7$ and $\ell =3$ becomes $\omega ^{(2)}=233233.$ A. Bouchet shows
Let $\omega $ be a word of even height. $\omega $ is a rop-word if and only if the two conditions are satisfied:
\(i) The length of $\omega $ is odd, say $2 \ell +1$.
\(ii) $\omega ^{(\ell )}$ admits a $1$-pairing.
In other words, we can always transform a (resp. Lyndon) rop-word $\omega $ of length $2 \ell +1$ by a one-to-one map $\phi $ such that the letters 3 in $\omega ^{(\ell )}$ are always coupled by subwords 33. The bijection $%
\psi $ sending $2 \rightarrow 0$ and $33 \rightarrow 1$ maps $\omega
^{(\ell )}$ to a word $\omega ^{ \prime } \in \{0 ,1\}^{ \ast }$ corresponding to a (resp. aperiodic) necklace. $$\omega \overset{\phi }{ \longrightarrow }\omega ^{(\ell )}\overset{\psi }{
\longrightarrow }\omega ^{ \prime }$$ The table \[Iso\] shows the first Lyndon rop-words for $n_{2} =3$ and the corresponding aperiodic necklaces
Aperiodic Necklaces $n$ Lyndon Rop-words $n^{\prime }$
--------------------- ----- ------------------ ---------------
5
7
7
9
9
9
: Correspondance for $n_{2}=3$[]{data-label="Iso"}
Conversely, starting from the representing Lyndon word of a aperiodic necklace $\omega ^{ \prime }\text{,}$ we construct the word $\omega ^{(\ell
)}$ by the bijection $\psi ^{ -1}$ sending $0 \rightarrow 2$ and $1
\rightarrow 33\text{,}$ and the word $\omega $ by applying $\phi ^{ -1}\text{%
.}$ By construction, the height $h (\omega ^{(\ell )})$ is even and also $h
(\omega ) .\;$Moreover, $\omega $ has a $\ell $-pairing and then is a rop-word.
Enumeration of the rop-words
============================
The number of necklaces (see [@Aig2007] for details, and also [Cas2013]{} for applications to music theory) with $n_{2}$ black beads and $%
n_{3}/2$ white beads derives from the generating function of the action of the cyclic group $$Z(C_{n_{2}},x)=\frac{1}{n_{2}}\sum_{d|n_{2}}\varphi (d)x_{d}^{n_{2}/d}$$where the sum is over all divisors $d$ of $n_{2}$ and $\varphi $ is the Euler totient function, according to the substitution of $x_{j}$ by $\frac{1%
}{1-x^{j}}\text{.}$ The developpement gives the coefficients of $x_{j}$ which are precisely the number of necklaces with $n_{2}$ black beads and $j$ white beads. For example, if $n_{2}=p$ is prime, the developpement leads to the following equations: $$\begin{aligned}
Z(C_{p},x) &=&\frac{1}{p}\left( \varphi (1)x_{1}^{p}+\varphi (p)x_{p}\right)
\\
&=&\frac{1}{p}\frac{1}{(1-x)^{p}}+\frac{p-1}{p}\frac{1}{1-x^{p}} \notag \\
&=&\frac{1}{p}\left( 1+\overset{\infty }{\sum_{n=1}}\frac{p(p+1)\ldots
(p+n-1)}{n!}x^{n}\right) +\frac{p-1}{p}\left( \overset{\infty }{\sum_{n=0}}%
x^{np}\right) \notag \\
&=&1+\frac{1}{p}\overset{\infty }{\sum_{n=1}}\binom{p+n-1}{n}x^{n}+\frac{p-1%
}{p}\left( x^{p}+x^{2p}+x^{3p}+\ldots \right) \notag \\
&=&1+\overset{\infty }{\sum_{n=1}}a_{n}x^{n} \notag\end{aligned}$$with $$a_{n}=\left\{
\begin{array}{lll}
\binom{p+n-1}{n} & if & n%
%TCIMACRO{\TeXButton{\not }{\not}}%
%BeginExpansion
\not%
%EndExpansion
\equiv 0~\mathop{\rm mod}~p \\
\binom{p+n-1}{n}+p-1 & if & n\equiv 0~\mathop{\rm mod}~p%
\end{array}%
\right.$$The table \[Rn\] with $n_{2}$ on the horizontal axis and $n_{3}$ on the vertical axis shows the number of rop-words for $n_{2}$ and $n_{3}$ fixed. The number of rop-words of length $n$ is given by summing along the diagonal $n_{2}+n_{3}=n$. Each column of the table is obtained from the developpement of the generating function $Z(C_{n_{2}},x)$. For $n_{2}$ prime, the coefficients agree with the formula of $a_{n}$ given above.
1 3 5 7 9 11 13 15 17
---- --- ------ ------ ----- ------- ----- ------ -------- ------
2 1 1 1 1 1 1 1 1 1
4 1 2 3 4 5 6 7 8 9
6 1 *4* 7 12 *19* 26 35 *46* 57
8 1 5 14 30 55 91 140 204 285
10 1 7 *26* 66 143 273 476 *776* 1197
12 1 *10* 42 132 *335* 728 1428 *2586* 4389
: Number of rop-words $(n_{2},n_{3})$[]{data-label="Rn"}
In each column, we recover the number of binary necklaces with length $%
n_{2}+q$ and density $q=n_{3}/2$ given by the rhs of the next formula. From the bijection of the previous section, it follows that the number $%
R(n_{2},n_{3})$ of rop-words with $n_{2}$ symbols 2 and $n_{3}$ symbols 3 is the number of binary necklaces of length $n_{2}+n_{3}/2$ and density $%
n_{3}/2 $,
$$R(n_{2},2q)=\frac{1}{n_{2}+q}\sum_{d|\gcd (n_{2}+q,q)}\varphi (d)\left(
\begin{array}{c}
(n_{2}+q)/d \\
q/d%
\end{array}%
\right) ,\text{\quad \quad }q=1,2,3,\ldots$$
By computing Lyndon words on alphabet $\{2,3\}$ and deleting those which are not rop-words, we get the table \[Ln\] of the number of Lyndon rop-words for $n_{2}$ and $n_{3}$ fixed, with $n_{2}$ on the horizontal axis and $n_{3}$ on the vertical axis. The total number of Lyndon rop-words of length $n$ is obtained by summing along $n_{2}+n_{3}=n$. The differences between the tables \[Rn\] and \[Ln\] are in italics.
1 3 5 7 9 11 13 15 17
---- --- ----- ------ ----- ----- ----- ------ -------- ------
2 1 1 1 1 1 1 1 1 1
4 1 2 3 4 5 6 7 8 9
6 1 *3* 7 12 18 26 35 *45* 57
8 1 5 14 30 55 91 140 204 285
10 1 7 *25* 66 143 273 476 *775* 1197
12 1 *9* 42 132 333 728 1428 *2584* 4389
: Number of Lyndon rop-words $(n_{2},n_{3})$[]{data-label="Ln"}
In each column of table \[Ln\], we recover the number of fixed density Lyndon words given by the following formula, with $n_{3}=2q$. It follows from the previous section, that the number $L(n_{2},n_{3})$ of Lyndon rop-words with $n_{2}$ symbols 2 and $n_{3}$ symbols 3 is $$L(n_{2},2q)=\frac{1}{n_{2}+q}\sum_{d|\gcd (n_{2}+q,q)}\mu (d)\left(
\begin{array}{c}
(n_{2}+q)/d \\
q/d%
\end{array}%
\right) ,\text{\quad \quad }q=1,2,3,\ldots$$where $\mu $ is the Mobius function.
By summing these formulas along a diagonal $n=n_{2}+n_{3}$, we get the number $\mathcal{L}_{n}$ of Lyndon rop-words of length $n$ and the number $%
\mathcal{R}_{n}$ of rop-words of the length $n$:
$$\mathcal{L}_{n}=\sum_{n_{2}+n_{3}=n}L(n_{2},n_{3})=\overset{(n-3)/2}{%
\sum_{p=0}}~L(2p+1,n-2q-1)$$
and $$\mathcal{R}_{n}=\sum_{n_{2}+n_{3}=n}R(n_{2},n_{3})=1+\overset{(n-3)/2}{%
\sum_{p=0}}~R(2p+1,n-2p-1)$$These numbers are tabulated as follows:
------------------- --- --- --- ---- ---- ---- ---- ----- ----- ------ ------ ------ -------
$n$ 3 5 7 9 11 13 15 17 19 21 23 25 27
$\mathcal{R}_{n}$ 2 3 5 10 19 41 94 211 493 1170 2787 6713 16274
$\mathcal{L}_{n}$ 1 2 4 8 18 40 90 210 492 1164 2786 6710 16264
------------------- --- --- --- ---- ---- ---- ---- ----- ----- ------ ------ ------ -------
: Numbers of Lyndon rop-words and rop-words of length $n$[]{data-label="Tot"}
If $n$ is prime, the difference between the cardinal of the two sets is 1, since the word $2^{n}$ (where the letter 2 is repeated $n$ times) is a rop-word but not a Lyndon word. If $n$ is a product or a power of primes, some periodic words appear that are rop-words but not Lyndon words. This explains the differences between the set of rop-words and the set of *Lyndon* rop-words. For instance, if $n=9$, $(233)^{3}$ is a rop-word but not a Lyndon word. The same is true for the words $(22323)^{3}$, $(233)^{5}$ and $(23333)^{3}$ of length 15. For $n_{2}=9$ and $n_{3}=12$, there are 333 Lyndon rop-words and 335 rop-words. The two non Lyndon rop-words are: $%
(2233233)^{3}$ and $(2323233)^{3}$.
Generalization: from rop to rap words
=====================================
First generalization.
----------------------
Let $s$ be an integer $\geq 2.$ The rhythmic oddity property could be generalized in two ways. The first way is as follows.
A word $\omega \in \{2 ,3\}^{ \ast }$ is a $s$-rop word if
\(i) $h (\omega ) \equiv 0~\mathop{\rm mod} ~s$
\(ii) No cyclic shift of $\omega $ can be factorized into $s$ words $u_{1}
,u_{2} ,\ldots ,u_{s}$ such that $$h (u_{1}) =h (u_{2}) =\ldots =h (u_{s})$$
A Lyndon $s$-rop word is both a $s$-rop word and a Lyndon word. For instance, if $s=3$, the first 3-rop words are: 2223, (3333), 22233, 22323, (33333), 222333, 2222223, 2232333, 2233233, 2233323, 2323233, (3333333).Non Lyndon words are given in parenthesis. A computation of the number of the first 3-rop words of length $n$ is given in table \[T5\].
------------------------- --- --- --- --- ---- --- ---- ---- ---- ----- ----- ----- ----- ------
$n$ 4 5 6 7 8 9 10 11 12 13 14 15 16 17
$\mathcal{R}_{n}^{(3)}$ 2 3 1 6 11 6 25 46 41 117 232 278 631 1237
$\mathcal{L}_{n}^{(3)}$ 1 2 1 5 9 6 22 45 40 116 226 278 620 1236
------------------------- --- --- --- --- ---- --- ---- ---- ---- ----- ----- ----- ----- ------
: Number of 3-rop words[]{data-label="T5"}
For $s=4$, the first 4-rop words are: 22233, 22323$^{\ast }$, (222222), 223333, 232333, (233233), 2222233, 2222323, 2223223$^{\ast }$, 2333333$%
^{\ast }$. Non Lyndon words are in parenthesis. The star indicates 2-rop words. A computation of the number of the first 4-rop words of length $n$ is given in table \[T6\]
------------------------- --- --- --- --- ---- ---- ---- ---- ----- ----- ----- ----- ------
$n$ 5 6 7 8 9 10 11 12 13 14 15 16 17
$\mathcal{R}_{n}^{(4)}$ 2 4 4 5 13 27 47 50 131 284 479 685 1450
$\mathcal{L}_{n}^{(4)}$ 2 2 4 5 12 24 47 50 131 279 473 683 1440
------------------------- --- --- --- --- ---- ---- ---- ---- ----- ----- ----- ----- ------
: Number of 4-rop words[]{data-label="T6"}
Second generalization.
----------------------
The second way of generalization is to change the alphabet and to consider words over $\mathcal{A} =\{1 ,2 ,\ldots ,s\}$.
A word $\omega \in \{1 ,2 ,\ldots ,s\}^{ \ast }$ has the rhythmic arity property (rap) of order s if
\(i) $h (\omega ) \equiv 0~\mathop{\rm mod} ~s$
\(ii) No cyclic shift of $\omega $ can be factorized into $s$ non-empty words $u_{1} ,u_{2} ,\ldots ,u_{s}$ such that $$h (u_{1}) =h (u_{2}) =\ldots =h (u_{s})$$
For short, we call *s-rap-word* a word with the rhythmic arity property of order $s$. For example, on the alphabet {1,2,3}, the words 111 and 333 are not 3-rap-words, but 123 and 132 are. The word 11133 is not a 3-rap-word since the subwords $u_{1}=111$, $u_{2}=3$ and $u_{3}=3$ have the same height, but the word 11313 is a rap-word. A *Lyndon s-rap-word* is both a $s$-rap-word and a Lyndon word. For instance, 11133, 11313, 11322 are Lyndon 3-rap-words.
Acknowledgements {#acknowledgements .unnumbered}
================
I would like to thank Harald Fripertinger for valuable comments and remarks.
[9]{} Aigner M. *A Course in Enumeration,* Springer, 2007.
Arom S., *African Polyphony and Polyrhythm*, Cambridge: Cambridge University Press, 1991.
Bouchet A. *Imparité rythmique*, Paris: ENS, *Culturemath*, 2010.
Brăiloiu C., Le rythme aksak, *Revue de musicologie* 33 (1952) 71-108.
Case J., *Necklaces and bracelets: Enumeration, algebraic properties and their relationship to music theory,* Master Thesis, University of Maine, G. Markowsky, advisor, 2013.
Chemillier M., Truchet C. Computation of words satisfying the “rhythmic oddity property”, *Information Processing Letters* 86: 5, (2003), 255-261.
Hall R.W., Klingsberg P. Asymmetric Rhythms, Tiling Canons, and Burnside’s Lemma, *Bridges: Mathematical Connections in Art, Music, and Science*. R. Sarhangi and C. Sequin, eds., Winfield, Kansas (2004), 189-194.
Hall R.W., Klingsberg P. Asymetric Rhythms and Tiling Canons, *Amer. Math. Monthly* 113:10, (2006), 887-896.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The temporal autocorrelation (AC) function associated with monitoring order parameters characterizing conformational fluctuations of an enzyme is analyzed using a collection of surrogate models. The surrogates considered are phenomenological stochastic differential equation (SDE) models. It is demonstrated how an ensemble of such surrogate models, each surrogate being calibrated from a single trajectory, indirectly contains information about unresolved conformational degrees of freedom. This ensemble can be used to construct complex temporal ACs associated with a “non-Markovian" process. The ensemble of surrogates approach allows researchers to consider models more flexible than a mixture of exponentials to describe relaxation times and at the same time gain physical information about the system. The relevance of this type of analysis to matching single-molecule experiments to computer simulations and how more complex stochastic processes can emerge from a mixture of simpler processes is also discussed. The ideas are illustrated on a toy SDE model and on molecular dynamics simulations of the enzyme dihydrofolate reductase.'
author:
- 'Christopher P. Calderon $^\dagger$'
bibliography:
- 'dhfr.bib'
date:
title: |
A Random Force is a Force, of Course, of Coarse:\
Decomposing Complex Enzyme Kinetics with Surrogate Models
---
Introduction
============
When enzymes and other proteins are probed at the single-molecule level, it has been observed in both experiments [@xie_dynamicdis_98; @xie04; @Dobson_science06AdK; @natureFRET07] and simulation studies [@wolynes_crackingPNAS03; @AdK; @wolynesHFSP08; @arora09; @rief_09; @Portman_CaMcracking09] that conformational fluctuations at several disparate timescales have physically significant influence on both large scale structure and biochemical function. In this article, a method for using a collection of Marovian surrogate models [@SPA1; @SPA2; @SPAJCTC] to predict kinetics that would often be considered non-Markovian is presented [@klafter08; @kou_08]. The ideas in [@SPAJCTC] are extended to treat a system with more complex kinetics. The aim of the approach is to obtain a better quantitative understanding the factors contributing to complex time autocorrelations (ACs) associated with quantities modulated by slowly evolving conformational degrees of freedom. The focus is on systems where certain thermodynamically important conformational degrees of freedom evolve over an effective free energy surface with relatively low-barriers; this situation is often relevant to molecules undergoing a “population shift" or “selected-fit" mechanism [@AdK; @arora09] and the connection to “dynamic disorder" [@xie_dynamicdis_98] is also discussed. The particular enzyme studied is dihydrofolate reductase (DHFR) because of its biological relevance to therapeutics and also due to the rich kinetics observed in certain order parameters [@arora09].
Surrogate models are used to describe the short time dynamics. These surrogates are fairly simple phenomenological parametric stochastic differential equation (SDE) models. Specifically the Ornstein-Ulhlenbeck process and an overdamped Langevin equation with a position dependent diffusion function [@hummer_posdep; @SPA1; @wangPNAS07; @SPAJCTC] are considered as the candidate surrogate models. Position dependent diffusion is often observed when a few observables (or order parameters) are used to describe an underlying complex system such as a protein. Position dependent noise models allow one to consider ACs having a different functional form than an exponential decay and it is demonstrated that this added flexibility can be of assistance in both understanding short and long timescale kinetics. Maximum likelihood type estimates utilizing transition densities, exact and approximate [@aitECO; @ozaki; @chen09], are used to fit our surrogate SDE models. The fitting method does not require one to discretize [@hummer_posdep] state space (the surrogates assume a continuum of states). The temporal AC is not used directly as a fitting criterion [@kou_08; @stuart_09], but the surrogate models are able to accurately predict the AC after the model parameters are fit. Maximum likelihood based approaches employing accurate transition density approximations and a parametric structure posses several advantages in this type of application [@llglassy]. An accurate transition density of a parametric SDE facilitates goodness-of-fit tests appropriate for both stationary [@aitFAN05; @chen08] and nonstationary time series [@hong]. The latter is particularly relevant to many systems (like the one considered here) where the diffusion coefficient is modulated by factors not directly monitored [@klafter08; @SPAJCTC] and the hypothesis of a fixed/stationary surrogate model describing the modeled time series is questionable. Statistically testing the validity of various assumptions explicitly or implicitly behind a candidate surrogate model, such as Markovian dynamics, state dependent noise and/or regime switching is helpful in both experimental and simulation data settings [@SPAfilter; @SPAdsDNA].
The type of modeling approach presented is attractive from a physical standpoint for a variety of reasons. The items that follow are discussed further in the Results and Discussion:
- In situations where the magnitude of the local fluctuations depend significantly on the instantaneous value of the order parameter monitored, simple exponential (or a finite mixture of exponentials [@socci96]) can be inadequate to describe the relaxation and/or AC function [@SPAJCTC]. The surrogates proposed can account for this situation when overdamped diffusion models can be used; additionally the estimated model parameters can be physically interpreted.
- It has been observed that even in single-molecule trajectories that “dynamic disorder" can be observed due to ignoring certain conformational degrees of freedom [@xie_dynamicdis_98; @xie04; @klafter08]; the methods proposed here can be used to account for this type of variability and show promise for comparing frequently sampled single-molecule experimental time series to computer simulations where dynamic disorder is relevant.
- Changes in conformational fluctuation magnitudes have been suggested to lead to physically interesting phenomena, so possessing a means for quantitatively describing an ensemble of dynamical responses can help one in better understanding the complex dynamics of enzymes, e.g. [@wolynes_crackingPNAS03; @AdK; @wolynesHFSP08; @Portman_CaMcracking09].
- There is general interest in showing how more complex stochastic processes arise from a collection of simpler parts [@granger80a; @granger80b; @cox91; @klafter09]. We discuss how, within a single trajectory, a continuous type of regime switching of Markovian surrogate models gives rise to an AC that would be considered non-Markovian.
The remainder of this article is organized as follows: Section \[sec:bacmeth\] reviews the background and presents the models considered. Section \[sec:acmash\] introduces the new modeling procedure used to approximate the AC function of a molecule experiencing multiple types of fluctuations. Section \[sec:results\] presents the Results and Discussion, Section \[sec:conc\] provides the Summary and Conclusions and this is followed by an Appendix.
Background and Methods {#sec:bacmeth}
======================
Effective Dynamics and Statistical Inference
--------------------------------------------
The trajectory generated by a detailed molecular dynamics (MD) simulation will be denoted by $\{ z_i \}_{i=1}^N$. The dynamics of the order parameter monitored is assumed to be complex (nonlinear, modulated by unobserved factors, etc.) even at the relatively short $O(ns)$ time intervals the order parameter time series is observed over. However, over short $\approx 50-100 ps$ time intervals a continuous stochastic differential equation (SDE) having the form $$\begin{aligned}
dz_t= & \mu(z_t; \theta,\Gamma)dt+ \sqrt{2}\sigma(z_t; \theta,
\Gamma)dB_t,
\label{eq:SDEgeneric}\end{aligned}$$ can often approximate the effective stochastic dynamics of the order parameter [@SPA1; @SPAJCTC]. In the above $\mu(\cdot)$ and $\sigma^2(\cdot) $ are the nonlinear deterministic drift and diffusion functions (respectively) and $B_t$ represents the standard Brownian motion [@raoDIFF]. The finite dimensional parameter vector is denoted by $\theta$ and $\Gamma$ is used to represent unresolved lurking degrees of freedom that slowly modulate the dynamics [@gAIoan].
The surrogate SDE models are formed by first dividing each trajectory into $L$ temporal partitions. Each estimated parameter vector is denoted by $\theta_\ell$ using the sequence $\{ z_i\}_{i=T_{\ell-1}}^{T_{\ell}}$ and an assumed model , where $\ell$ is an index of a partition, $1=:T_0< \ldots < T_\ell < \ldots < T_L := N$, used to divide a time series into $L$ disjoint local temporal windows. Within each of these windows, the data and the assumed model structure is used to compute $\theta_\ell $ using maximum likelihood type methods (exact [@chen09] and approximate [@aitECO] depending on the model). The parametric structures considered are presented in the next section. It is to be stressed that we estimate a collection of models, $\{\theta_\ell\}_{\ell=1}^L$ for each trajectory. The differences in the estimated parameters are due in part to random slowly evolving forces modulate the dynamics and also in part to unavoidable estimation uncertainty associated with a finite time series. It is demonstrated that a collection of surrogate model parameter vectors is needed to summarize conformational fluctuations inherent to many complex biomolecules. This procedure is repeated for each observed MD trajectory / time series.
The term “local diffusion coefficient" $\equiv
\widetilde{D}(z;\Gamma):= \sigma^2(z;\theta, \Gamma)$ is introduced in order to distinguish the coefficient in the Eqn. \[eq:SDEgeneric\] from the diffusion coefficient usually implied in the physical sciences: we estimate the former from observed data. The term “diffusion coefficient" used in statistical physics [@zwanzig] is not necessarily the same as $\widetilde{D}(z;\Gamma)$. If $\Gamma$ does not modulate the dynamics, the two definitions are effectively identical. However one theme of this paper is that some traditional dynamical summaries of statistical physics, such as diffusion coefficient and ensemble based AC, can be modified or made less coarse by using a collection of surrogate models. Such a procedure may help in intpretting/understanding single-molecule time series.
Candidate Surrogate SDE Models
------------------------------
Two local parametric SDE models considered. MATLAB scripts illustrating how to obtain parameter estimates of both models from discretely observed data are available online [www.caam.rice.edu/tech\_reports/2008\_abstracts.html\#TR08-25](www.caam.rice.edu/tech_reports/2008_abstracts.html#TR08-25). The first is a linear, constant additive noise process: $$dz_t= B(A-z_t)dt + \sqrt{2}CdB_t.$$ The above SDE has a rich history in both the physical sciences [@gardiner] where it is usually referred to as the Ornstein-Uhlenbeck (OU) process and in econometrics where it is sometimes referred to as the Vasicek process [@chen09]. The parameter vector to estimate in this model is $\theta \equiv (A,B,C)$. This model is appealing for a variety of reasons, one being that the exact transition density and maximum likelihood parameter vector for a discretely sampled process [^1] can be written in closed-form, i.e. a numerical optimization is not needed to find the parameter vector because the parameter estimate can be written explicitly in terms of $\theta$ and the observed data [@chen09].
The second is a nonlinear, position dependent overdamped (PDOD) Langevin type SDE [@schultenJCP04; @SPA1; @SPAJCTC]: $$\begin{aligned}
\nonumber dz_t= & \beta \Big(C+D(z_t-\psi_0)\Big)^2 \Big(A+B(z_t-\psi_0)\Big)dt \\
& + \sqrt{2}\Big(C+D(z_t-\psi_0)\Big)dB_t.
\label{eq:PDOD}
\end{aligned}$$ The variable $\beta\equiv 1/(k_BT)$ is the inverse of the product of Boltzmann’s constant and the system temperature. $\psi_0$ represents a free parameter; in this article it coincides with the umbrella sampling point specified in the simulation. The parameter vector to estimate in this model is $\theta \equiv (A,B,C,D)$. Each parameter is estimated using the observed data and the transition density expansions [@aitECO] associated with Eqn. \[eq:PDOD\] is used to construct a log likelihood cost function. A Nelder-Mead search is then used to find the $\theta$ maximizing the associated cost function. The effective force in the above model is assumed to be linear in $z$, e.g. $F(z):= A+B(z-\psi_0)$ whereas the diffusion function $\equiv \widetilde{D}(z;\Gamma):= \Big(C+D(z-\psi_0)\Big)^2$ is quadratic in $z$. The overdamped appellation comes from multiplying the effective force by the effective friction (as determined by the Einstein relation [@schultenJCP04]) corresponding to this diffusion function.
In this article, all stochastic integrals used are Itô integrals. When a complex high-dimensional system with multiple timescales is approximated with a low dimensional SDE possessing position dependent noise the choice of the Itô or Stratonovich integral influences the interpretation of the drift function and the issue of which interpretation is “physically correct" is a nontrivial problem [@razstuart04; @dima]. A related item is the so-called “noise-induced drift" [@risken; @arnold00]. Such a term is sometimes explicitly added to the drift [@arnold00], one thermodynamic motivation for this is discussed further in Section \[sec:standardcomp\].
An appealing feature of the data-driven modeling procedure presented here and elsewhere [@SPA1; @molsim; @SPAfilter; @SPAJCTC; @gAIoan] is that various SDE models, of an explicitly specified form, can be considered, estimated, and tested using observed trajectories. Statistical hypothesis tests making use of the conditional distribution (not just moments) of the assumed surrogate model can then be used to test if the model assumptions are justified for the observed data. Tools from mathematical statistics [@hong; @aitFAN05] facilitate quantitatively and rigorously testing if certain features are required to adequately describe the stochastic dynamics. Many features, e.g. position-dependent noise, would be hard to statistically check using AC based heuristic methods. Such heuristic checks are traditionally used in statistical physics, e.g. [@bussi_08; @dima].
The data-driven models are used to approximate the stochastic evolution of black-box data and the estimated parameters do have a loose physical phenomenological interpretation. If one desires to compute unambiguous physical quantities from the estimated coefficients using a particular definition from statistical physics, the models can also be used to generate data for this purpose. For example, surrogate models can generate nonequilibrium (surrogate) work [@SPA1; @SPA2; @gAIoan] and, under various assumptions, a well-defined thermodynamic potential of mean force (PMF) can be derived from such data [@hummerPNAS01; @schultenJCP04]. This contrasts the case where one starts with a high dimensional stochastic process (of known functional form) and then uses stochastic analysis to reduce the dimensionality of the system by first appealing to asymptotic arguments [@razstuart04] and then possibly modifying the resulting equations to achieve a desired physical constraint [@arnold00]. In both analytical and data-driven cases, the goal is often to construct a single limiting low-dimensional evolution equation that can be used to predict statistical properties of the complex system valid over longer time-scales [@schultenJCP04; @razstuart04; @dima; @hummer_posdep]. It is not quantitatively clear at what timescale such an approximation (if any such useful approximation exists at all) is valid over. Furthermore it is usually difficult to determine if an equilibrium concept such as a PMF connects simply to trajectory-wise kinetics in small complex systems experiencing fluctuations. Again, an appealing feature of the approach advocated here is that various statistical hypothesis tests [@hong; @aitFAN05; @chen08] can be used to quantitatively assess the validity of proposed (reduced) evolution equations to see if physically convenient models are consistent with the observed data. For example, such tests can be used to determine the time one needs to wait before inertia can be neglected [@SPAJCTC].
Regarding the validity of using a *single* surrogate SDE to approximate “long term" $> O(ns)$ trends, a main underlying theme of this paper and others [@gAIoan; @SPA2; @SPAJCTC; @SPAfilter] is that the presence of a lurking slowly evolving degree of freedom, $\Gamma$, can significantly complicate using a single equation and that methods for quantitatively accounting for this sort of variation are underdeveloped. Information in these types of models have proven useful in both theoretical chemistry computations [@gAIoan; @SPA2] and in characterizing nanoscale experimental data [@SPAdsDNA; @SPAfilter; @SPAfric]. Throughout this article, it is shown how the *collection* of surrogate models can be linked with the ideas of “dynamic-disorder" [@xie_dynamicdis_98] to make quantitative statements about systems observed at the single-molecule level.
A Method for Computing the AC Function of Complex Systems {#sec:acmash}
=========================================================
Before providing the algorithmic details of the method, the basic idea and motivation for the approach is sketched in words. It is assumed that a $\Gamma$ type coordinate slowly evolves (diffusively) over a relatively flat region of an effective free energy surface. This evolution modulates the stochastic dynamics of the order parameter modeled, e.g. it changes the diffusion coefficient function [@SPAJCTC; @klafter08]. However, due to the almost continuous nature of the parameter change, a sudden or sharp change in the process dynamics is assumed difficult to detect in short segments of the time series (sudden regime changes or barrier crossings are not readily apparent in the data). Over longer time intervals, the changes become significant and the validity of a simple SDE model like the ones considered here to describe the global dynamics become suspect. However if the evolution rules are updated as time progresses in the spirit of a dynamic disorder description [@xie_dynamicdis_98], then there is hope for using a collection of these models to summarize the dynamics. Even if the data is truly stationary, some fluctuations due to a $\Gamma$ type coordinate may take a long time to be “forgotten" [@xie04; @kou_08]. The idea proposed here is to essentially to use the estimated model for a time commensurate with time interval length used for estimation/hypothesis testing and then suddenly switch model parameters. By doing this, one can take a collection of fairly simple stochastic models and construct another stochastic process possessing a more complex AC function.
One advantage of such a procedure is that an ensemble of elementary or phenomenological pieces can be constructed to gain a better understanding of how variation induced by slowly evolving fluctuations affects some system statistics and this information may help in better quantitatively understanding some recently proposed enzyme mechanisms [@wolynes_crackingPNAS03; @AdK; @wolynesHFSP08; @arora09; @rief_09; @Portman_CaMcracking09]. This method is in line with the single-molecule philosophy that dynamical details should not be obscured by bulk averaging artifacts when possible. It is demonstrated how using a traditional AC summary of the data would obscure information of this sort on a toy example. Since the timescales at which simulations and single-molecule experiments span are rapidly converging, this new type of dynamical summary can also be used to help in matching the kinetics of simulations and experiments and/or can be used to understand how more complex dynamics emerge from simpler evolution rules [@granger80a; @granger80b; @cox91; @klafter09].
Now for the algorithm details. Recall that for a single trajectory coming from a high dimensional system, the time series data is divided into partitions and within each partition the parameters of both candidate models are estimated by methods discussed in the previous section. This results in a collection $\{\theta_\ell\}_{\ell=1}^L $ for a each trajectory observed. The Euler-Maruyama [@kp] scheme is used here to simulate $N_{MC}$ trajectories for each $\theta \in \{\theta_\ell\}_{\ell=1}^L $. The surrogate SDEs are recorded every $\delta t$ and denote simulated order parameter time series by $\big\{ {\{ z^{s,(j)}_{i} \}^{T_{\ell}}_{i=T_{\ell-1}}} \big\}_{\ell=1}^{L}$ with $j=1,\ldots N_{MC}$. To construct a new time series using the $N_{MC}$ trajectories generated, set $x^s=\{\},\ell=1$ and for t=1 to $n\times L$ repeat the following:
1. Draw Uniform Integer u $\in [1,N_{MC}]$.
2. Set $x^s=\Big\{x^s \ , \ {\{ z^{s,(u)} \}^{T_{\ell}}_{i=T_{\ell-1}}} \Big\}$
3. Update counter $\ell=\mathrm{mod}(t,L)+1$
The procedure described results in a new time series $\{x^s_i\}_{i=1}^{N^\prime}$ where $N^\prime\equiv n\times N$. Note that the time ordering of the original data is maintained and the last step forces the series $\{x^s_i\}_{i=1}^{N^\prime}$ to be periodic so time lags $>N\delta t$ cannot be resolved with this method. If the integer $n>1$, the series $\{x^s_i\}_{i=1}^{N^\prime}$ contains more temporal samples than the original series. A larger sample size reduces the statistical uncertainty in an empirically determined AC. The issue of reducing uncertainty is subtle and is discussed in detail using the toy model presented in the next section. If the time ordering is believed irrelevant, the first step can be modified to drawing 2 random integers. The other random integer can be used to randomize the $\ell$ index. [^2].
This procedure can then be repeated for each trajectory coming from a high dimensional system. It is to be stressed that suddenly and relatively infrequently regime switches (“barrier hopping") cannot be described with this method. If the simulation or experiment is associated with a system possessing a jagged/rough free energy surface with many small barriers and if a single trajectory can frequently sample the hops, then there is hope for using this method. However note that the method is designed to treat relatively smooth regime changes (i.e. regime changes hard to identify by simple visual inspection). A discussion on how the surrogate models can be potentially used in more complex situations is briefly discussed later.
Results and Discussion {#sec:results}
======================
Toy Model
---------
In order to demonstrate the AC method on a simple example and illustrate some points in a controlled setting we use the following SDE model: $$\begin{aligned}
dy^{\mathrm{I}}_t= & \kappa^0(\alpha^0-y^\mathrm{I}_t)dt + \eta^0 dB^1_t \\
\nonumber dy^{\mathrm{II}}_t= & \kappa^0(\alpha_t-y^\mathrm{II}_t)dt + \eta^0 dB^1_t \\
\nonumber dy^\mathrm{III}_t= & \kappa_t(\alpha_t-y^\mathrm{III}_t)dt + \eta_t dB^1_t \\
\nonumber d\alpha_t= & 1/\tau_0(\alpha^0-\alpha_t)dt + \sigma^{\alpha} dB^2_t \\
\nonumber d\kappa_t= & 1/\tau_0(\kappa^0-\kappa_t)dt + \sigma^{\kappa} dB^3_t \\
\nonumber d\eta_t= & 1/\tau_0(\eta^0-\eta_t)dt + \sigma^{\eta} dB^4_t ,\end{aligned}$$
where the constants $\alpha^0,\kappa^0,\eta^0$ are meant to play the role of the surrogate parameters $(A,B,C)$ in the OU model. In the above expressions, superscripts are used simply to distinguish different constants or processes and do not represent exponentiation. Superscripts on the $dB_t$ terms are used to distinguish separate independent standard Brownian motions. The Roman numeral superscripts distinguish three cases: I) the standard OU model; II) an OU type model where the mean level, $\alpha$ evolves stochastically; and III) an OU type model where all parameters evolve stochastically. The parameter $\tau_0$ dictates the time scale at which the OU parameters stochastically evolve. The evolution studied here is made to be slow relative to that dictated by $\kappa^0$. The (assumed unobserved) processes $\alpha_t$ , $\kappa_t$ ,$\eta_t$ are meant to mimic a dynamic disorder [@xie04] type situation.
\[ht\] ![ The AC of 4 realizations from 4 toy processes discussed in the text (left). The standard deviation of the AC curves measured in a trajectorywise fashion from a population of 100 trajectories. The computed standard deviation reflects the variance in a population of 100 empirically determined ACs. []{data-label="fig:1"}](fig1a.eps "fig:"){width=".235\textwidth"} ![ The AC of 4 realizations from 4 toy processes discussed in the text (left). The standard deviation of the AC curves measured in a trajectorywise fashion from a population of 100 trajectories. The computed standard deviation reflects the variance in a population of 100 empirically determined ACs. []{data-label="fig:1"}](fig1b.eps "fig:"){width=".235\textwidth"}
In addition, a fourth process referred to as “III (Proxy)" will be evolved to demonstrate the AC method of Section \[sec:acmash\]. This process is constructed by simply setting the parameters $\alpha^0,\kappa^0,\eta^0$ equal to the corresponding parameters of process III at time $t$ and then evolving this process like a standard OU model until the time index hits $t+\mathcal{T}$ when the parameters are updated to those of process III at the same time. This procedure is then iterated. Randomizing $\mathcal{T}$ had little influence on the accuracy here, but can be entertained. The processes above are simulated using the Euler-Maruyama scheme with time step size $\delta s$ and the process is observed discretely every $\delta t$ time unit. The remaining parameters are tuned to provide a parameter distribution consistent with those observed in some DHFR studies and are reported in the Appendix.
The toy model is used to investigate how variation induced by slowly evolving $\Gamma$ type factors influence the computed empirical AC on a controlled example where the assumptions behind the method introduced are satisfied. The features discussed are relevant to the DHFR system studied later and are also likely relevant to oher single-molecules studies. The example is also used to highlight issues relevant to nonergodic sampling [@klafter08], i.e. when temporal averages are not equivalent to ensemble averages. In this type of situation, single-molecule data is particularly helpful. Use of the same Brownian motion term to drive three separate processes facilitates studying contributions to variance in these types of studies. In addition, estimation is not carried out to keep the discussion simple and to remove an additional source of uncertainty.
The left panel of Fig. \[fig:1\] plots the empirical AC computed by sampling 4 realizations from this process using 5000 observations uniformly spaced by $\delta t=0.15 ps$. These time series lengths are commensurate with those used in typical MD applications [@socci96; @aroraAdK07; @arora09]. One observes that the slowly evolving parameters do influence the AC measured. The fairly simple method of periodically updating the evolution parameters is able to mimic the AC associated with $y^\mathrm{III}$ for both short and long time scales. Furthermore, the variation induced by the relaxation and noise level (modulated by $\kappa_t$ and $\eta_t$) influences both the short time and longer time responses. The stochastic response of a dynamic disorder type process is clearly richer than a single exponential. An advantage the surrogate approach offers over popular existing methods for treating this situation [@socci96] is that other kinetic schemes, e.g. those associated with overdamped models with position dependent diffusion, can be entertained. In enzymes associated with complex dynamics, other kinetics schemes may be needed to accurately reflect the stochastic dynamics of the order parameter monitored. For example it is demonstrated in Fig. \[fig:3\] that the PDOD surrogate is needed accurately captured relaxation kinetics even at short $O(ps)$ timescales. Over timescales relevant to experimentally accessible order parameters characterizing conformational fluctuations, one may need to account for dynamical responses much richer than a mixture of exponentials [@SPAJCTC; @karplus08; @kou_08]. The procedure presented demonstrated how “elementary" pieces could be patched together to characterize relaxations/fluctuations occurring over longer timescales. This is attractive to both computer simulations and experimental data sets. In what follows the attention is shifted to focusing on limitations of using a single AC to describe single-molecule time series.
The right panel of Fig. \[fig:1\] plots the standard deviation of the AC function associated with a trajectory population. For each observed trajectory, 100 different SDE trajectories were used to compute 100 empirical ACs from the time series associated with the trajectories. The pointwise standard deviation measured over the 100 ACs is plotted. The curve shadings distinguish different time series sample sizes. The three cases studied consisted of $(0.5,2,8)\times 10^4$ discrete temporal observations; each time series was uniformly sampled with $0.15 ps$ between observations. Note that the influence of the evolving parameters on the measured AC is substantial. Recall all $y$ processes used common Brownian paths (so computer generated random numbers do not contribute to the differences observed). In addition, observe that the difference between $y^{\mathrm{II}}$ and $y^{\mathrm{III}}$ persists for a fairly long time and the length of time this difference is measurably noticeable depends on the temporal sample size used to compute the AC. In some applications, the variation induced by conformational fluctuations is important in computations [@gAIoan] or to characterize a system [@AdK; @aroraAdK07]. The standard deviation in the measured AC here contains contributions coming from factors meant to mimic the influence unresolved conformational fluctuations whose influence persists for a fairly long time. In the AC computed with longer time series, i.e. spanning a larger time since the time between observations is fixed, the process has more time to “mix" and hence the difference between temporal and ensemble averages is reduced. Said differently, the influence of the initial conformation, or “memory", diminishes. By using a single long time series trajectory and only reporting one AC computed from this “mixed" series, these types of physically relevant fluctuations can get washed out by using a single AC function. This goes against the spirit of single-molecule experiments.
The example considered here is admittedly simple and was constructed to illustrate the types of assumption behind the method introduced. If the dynamic disorder induced by large kinetic barriers or a complicated interaction with the surroundings, then one would need to construct more sophisticated processes for determining how and when the parameters regime switch. Combining the surrogate models with efforts along these lines, e.g. [@schuetteMMS09], may be able to help these more exotic situations. Exploring the various routes by which complex and/or heavy tailed ACs [@xie04; @kou_08] can emerge from simpler dynamical rules can help in a fundamental understanding of the governing physics [@granger80a; @granger80b; @cox91; @klafter09]. However, if the ensemble average decay rate is deemed the only quantity of physical relevance then the collection of surrogate models can still potentially be used to help in roughly predicting the rate of decay of more complex ACs. This is particularly relevant to simulations where obtaining long enough trajectories to reliably calibrate models possessing complex AC exhibiting long range dependence from observed data is problematic [@Kaulakys_06; @SPAJCTC; @kou_08]. Even in cases where one only requires the asymptotic time decay of an ensemble of conformations for a physical computation [@deshaw] and can simulate for a long enough time to directly monitor kinetics, an understanding of the distribution of surrogate models estimated will likely be of help in linking computer simulation force fields to single-molecule experimental time series. The remaining results use simulations of DHFR to illustrate some of these points.
Dihydrofolate Reductase(DHFR)
-----------------------------
\[ht\] ![ Average local diffusion function estimated (left axis) and free energy (right axis) as a function of order parameter. The population average of the estimated $C$ is plotted to give a feel for the position dependence of this quantity; it is stressed that the average alone is not adequate the describe the dynamics here due to a type of “dynamic disorder" phenomena [@xie_dynamicdis_98]. The free energy was computed by K. Arora using methods described in Ref. [@arora09]. []{data-label="fig:2"}](fig2.eps "fig:"){width=".45\textwidth"}
### DHFR Simulation Details
The detailed computational details are reported in Ref. [@arora09]. Briefly, an order parameter denoted by $\Delta D_{\mathrm{rmsd}}$, was defined using the root mean square distance between two crystal structures [@arora09]. This order parameter provides an indication of the proximity to the “closed" and “occluded" enzyme state and is reported in units of $\mathrm{\AA}$ throughout. The initial path between the closed and occluded conformations of DHFR was generated using the Nudged Elastic Band (NEB) method [@chu03]. Subsequently, $\approx $50 configurations obtained from NEB path optimization were subjected to US simulations. During these US simulations, production dynamics of 1.2$ps$ at 300K was performed after equilibration using a weak harmonic restraint.
### DHFR Results
Figure \[fig:2\] plots the average local diffusion coefficient of the surrogate SDE models using two different observation frequencies on the left axis and on the right axis the free energy computed in Ref. [@arora09] is plotted. Each surrogate model was estimated using 400 time series observations with either $\delta=0.15$ or $\delta=0.30 \ ps$ separating adjacent observations corresponding to $L=20$ or $L=10$ (respectively). The average local diffusion coefficient demonstrates a relatively smooth increasing trend for a majority of the order parameter values explored, but then suddenly changes abruptly around $\Delta D_{\mathrm{rmsd}} \approx 3 \mathrm{\AA}$. It has been observed that an interesting interplay between free energy, fluctuations and stiffness, exists in some enzyme systems [@wolynes_crackingPNAS03; @AdK; @wolynesHFSP08; @Portman_CaMcracking09] and this plot suggests that future works investigating some of the finer structural factors leading to these change may be worthwhile, though this direction is left to future work because it is outside the scope of this study. It is to be stressed that the mean of each US window is not adequate to summarize the dynamics. That is, a single fixed parameter surrogate SDE like the ones considered here cannot mimic the longer time statistics of the process. This is why the AC procedure introduced in Section \[sec:acmash\] is needed. Figure \[fig:3\] demonstrates that the individual PDOD models do capture features simpler surrogates cannot. This is due part to the position dependence of the local diffusion coefficient. The PDOD surrogate model combined with the procedure of Section \[sec:acmash\] can accurately summarize the long time dynamics. These points are explained further in the discussion associated with Figs. \[fig:3\]-\[fig:5\].
\[ht\] ![ Local AC Function of MD data and that predicted by surrogates. The total time series of length 1.2 ns was divided into separate blocks, each containing 400 observations spaced by 0.15 ps. This data was used to estimate a collection of surrogate models and a collection of MD ACs corresponding to US target points of $\Delta D_{\mathrm{rmsd}}\approx -1.8, 0.9, 1.9$ and $3.1 \mathrm{\AA}$. The AC corresponding to the estimate surrogate models is also plotted where the initial lag is normalized to unity to facilitate comparison (the raw units are shown in Fig. \[fig:3\]). []{data-label="fig:3"}](fig3a.eps "fig:"){width=".225\textwidth"} ![ Local AC Function of MD data and that predicted by surrogates. The total time series of length 1.2 ns was divided into separate blocks, each containing 400 observations spaced by 0.15 ps. This data was used to estimate a collection of surrogate models and a collection of MD ACs corresponding to US target points of $\Delta D_{\mathrm{rmsd}}\approx -1.8, 0.9, 1.9$ and $3.1 \mathrm{\AA}$. The AC corresponding to the estimate surrogate models is also plotted where the initial lag is normalized to unity to facilitate comparison (the raw units are shown in Fig. \[fig:3\]). []{data-label="fig:3"}](fig3b.eps "fig:"){width=".225\textwidth"} ![ Local AC Function of MD data and that predicted by surrogates. The total time series of length 1.2 ns was divided into separate blocks, each containing 400 observations spaced by 0.15 ps. This data was used to estimate a collection of surrogate models and a collection of MD ACs corresponding to US target points of $\Delta D_{\mathrm{rmsd}}\approx -1.8, 0.9, 1.9$ and $3.1 \mathrm{\AA}$. The AC corresponding to the estimate surrogate models is also plotted where the initial lag is normalized to unity to facilitate comparison (the raw units are shown in Fig. \[fig:3\]). []{data-label="fig:3"}](fig3c.eps "fig:"){width=".225\textwidth"} ![ Local AC Function of MD data and that predicted by surrogates. The total time series of length 1.2 ns was divided into separate blocks, each containing 400 observations spaced by 0.15 ps. This data was used to estimate a collection of surrogate models and a collection of MD ACs corresponding to US target points of $\Delta D_{\mathrm{rmsd}}\approx -1.8, 0.9, 1.9$ and $3.1 \mathrm{\AA}$. The AC corresponding to the estimate surrogate models is also plotted where the initial lag is normalized to unity to facilitate comparison (the raw units are shown in Fig. \[fig:3\]). []{data-label="fig:3"}](fig3d.eps "fig:"){width=".225\textwidth"}
\[ht\] ![ The global AC prediction. The procedure introduced here is used along with PDOD data to predict the long time AC. The relaxation predicted by the individual surrogate (data shown in Fig. \[fig:2\] is also shown to stress that the SDE parameters are not fixed, but evolving and any single surrogate cannot capture the richer long time dynamics. []{data-label="fig:4"}](fig4a.eps "fig:"){width=".225\textwidth"} ![ The global AC prediction. The procedure introduced here is used along with PDOD data to predict the long time AC. The relaxation predicted by the individual surrogate (data shown in Fig. \[fig:2\] is also shown to stress that the SDE parameters are not fixed, but evolving and any single surrogate cannot capture the richer long time dynamics. []{data-label="fig:4"}](fig4b.eps "fig:"){width=".225\textwidth"} ![ The global AC prediction. The procedure introduced here is used along with PDOD data to predict the long time AC. The relaxation predicted by the individual surrogate (data shown in Fig. \[fig:2\] is also shown to stress that the SDE parameters are not fixed, but evolving and any single surrogate cannot capture the richer long time dynamics. []{data-label="fig:4"}](fig4c.eps "fig:"){width=".225\textwidth"} ![ The global AC prediction. The procedure introduced here is used along with PDOD data to predict the long time AC. The relaxation predicted by the individual surrogate (data shown in Fig. \[fig:2\] is also shown to stress that the SDE parameters are not fixed, but evolving and any single surrogate cannot capture the richer long time dynamics. []{data-label="fig:4"}](fig4d.eps "fig:"){width=".225\textwidth"}
The ability of the PDOD model to capture features that a single exponential (e.g. the AC associated with an OU process) cannot is demonstrated in Fig. \[fig:3\]. Results from four different US points, each possessing a different degrees of position dependence on the noise are shown. Here the results obtained using both the OU and PDOD surrogates calibrated using the $\delta t=0.15
ps$ with 400 temporal observations and the corresponding AC predictions are shown in the plot. The empirical ACs computed using the short segments of MD data used for surrogate model parameter estimation are also reported. Results with 400 blocks possessing observations spaced by $\delta t=0.30 ps$ were similar in their AC prediction, but hypothesis tests strongly rejected the assumption of a fixed local diffusion (see Fig \[fig:5\]). The 400 $\delta t=0.15 ps$ samples allowed the OU model structure to provide a better fit (as measured the fraction rejected) because the local diffusion function had less time to evolve/change value. For cases where the position dependence is moderate, the PDOD and OU surrogate models predict qualitatively similar AC functions. However, the PDOD model captures the short time relaxation dynamics better than the OU for cases where the position dependence of the local diffusion is more substantial and hence for clarity we focus on the PDOD models in the remaining kinetic studies.
\[ht\] ![ Goodness-of-fit tests. The test of Ref. [@hong] was computed given the parameter estimate and the observed data. The median of each umbrella sampling window is reported. []{data-label="fig:5"}](fig5.eps "fig:"){width=".45\textwidth"}
Figure \[fig:4\] plots the empirically determined AC obtained from different MD production simulation data. The case labelled “in-sample" was the one used for estimation of the local models reported in Fig. \[fig:3\] and that labeled “out-of-sample" was computed by running a longer 3.6 ns simulation and computing the AC from the last 1.2 ns of this time series. The PDOD version of these models were used along with the procedure outlined in Section \[sec:acmash\] using blocks of size 800 and randomizing the time index. The 400 blocks results were similar. Respecting the time ordering of the surrogate models only improved results marginally. Note also that the general trends of the long time decay of the MD data is captured with the procedure and that there is substantial difference between the “in-sample" and “out-of-sample" MD trajectories [^3]. The physical relevance of such variation was previously discussed and will be expanded on when results of stationary DHFR density prediction are shown. The primary observation is that a *collection* of PDOD surrogate models were able to capture the basic relaxation trends of the enzyme that a single surrogate could not. Recall that even at short timescales a single exponential decay was inadequate to fit the data. Similar trends were observed for all 51 US windows explored. However, it is to be stressed that the procedure shown here is to *decompose* kinetics in the longest contiguous block of discrete time series observed. If complex dynamics occur over longer timescales and data is not available that directly samples these scales, then the method cannot be used to predict the long time behavior that was not sampled.
\[ht\] ![ Stationary density/histogram prediction. The bars denote 1.2 $ns$ MD data, the solid thick line to the mixture PDOD method (see text), the solid dotted line to the average PDOD model, and the thin lines to the 4 local equilibrium densities used in constructing the mixture density. []{data-label="fig:6"}](fig6.eps "fig:"){width=".45\textwidth"}
The goodness-of-fit of the surrogate models using the two candidate SDEs is shown in Fig. \[fig:5\] for various US windows. The median of the Q-test statistic introduced by [@hong] is reported. This test statistic under the null is asymptotically normally distributed with mean zero and unit variance, but has also been proven to be useful in small samples [@hong; @molsim; @SPAJCTC; @gAIoan]. Recall that each MD time series (at each umbrella sampling window) was divided into small pieces. In the portions near the edges (larger $|\Delta_{\mathrm{rmsd}}|$ values), where the position dependence of the noise is greatest, one observes that the OU model population has a median that would typically indicate a collection of poor dynamical models. If conformational fluctuations slowly modulate the dynamics, the longer the time series one has, the likelihood of departing from any simple surrogate model increases [^4]. Goodness-of-fit tests, like the one presented here, can be used to quantitatively approximate when simple models begin departing from various assumptions.
The stationary density predicted by the surrogate OU models in a case where position-dependence was shown to be marginal for the time interval data was monitored is plotted in Fig. \[fig:6\]. Here the mixture method discussed in Ref. [@SPAJCTC] is reported due to its relevance to a collection of surrogates and dynamic disorder. The histogram of the $1.2 ns$ MD data is also plotted as well as the stationary density predicted by a the average of the surrogate models taken at the US window near $\Delta D_{\mathrm{rmsd}}\approx 2$. Using a single model obtained by aggregating all time series together in hopes of reducing surrogate parameter uncertainty due to the resulting smaller time series sample sizes actually worsens the results. The mixture of OU models calibrated using 4 sets of noncontiguously spaced $60 ps$ data (i.e. $ 400$ entries spaced $0.15 ps$) were sampled every 300 ps from the MD process and this was used to compute 4 surrogate OU model parameters. The goodness-of-fit tests indicated that the local surrogates given the data were reasonable dynamical models. So portions where the “local equilibrium" density, i.e. the stationary density predicted by a surrogate with estimated parameters, possessing significant probability mass can be though of the regions of phase space sampled due to fast-scale motion for a relatively fixed (and unobserved) value of $\Gamma$ [@SPAJCTC; @SPAfilter]. If variation in the conformational coordinate is important to thermodynamic averages, as the data here suggests to be the case in DHFR, then one needs to use a collection of “local equilibrium" densities [@SPAJCTC]. The advantage of such an approach is that short bursts of simulations started from different initial conditions can be run, then surrogate models can be calibrated and tested. If the surrogate is found suitable, it can then be used to make predictions on the local equilibrium density, and the variation in the local equilibrium densities can be used to partially quantify the degree to which a slow conformational degree of freedom modulates the dynamics. This treatment is appealing when data on other physically relevant order parameters is unknown are not easy to access.
Summary and Conclusions {#sec:conc}
=======================
Single-molecule experiments and simulations offer the potential for a detailed fundamental understanding of complex biomolecules without artifacts of bulk measurements obscuring the results. However, one must deal with complex multiscale fluctuations at this level of resolution and the factors contributing to the noise often contain physically relevant information such as quantitative information about conformational degrees of freedom [@SPAfilter]. The abundance of data available to researchers and recent advances in computational and statistical methods are allowing researchers to entertain new methods of summarizing information relevant to modeling systems at the nanoscale [@SPAfilter; @SPAfric; @gAIoan].
By applying surrogate models to the data coming from biased MD simulations of DHFR, it was demonstrated that a collection of stochastic dynamical models can be used to better understand the factors contributing to the shape of the autocorrelation function associated with fluctuations coming from multiple time scales. The surrogate models were estimated by appealing to maximum likelihood type methods [@aitECO; @ozaki; @chen09] and were tested using goodness-of-fit tests which utilized the transition density of the assumed surrogate and were appropriate for the data. For example, the time series data was not assumed to be stationary; the stationarity assumption is often suspect in simulation data. The tests used [@hong] indicated that taking the position dependence of the noise into account was required to provide a statistically acceptable model in many regions of phase space explored. For short timescales, the individual surrogate models (taking position dependence noise into account) were capable of predicting quantities outside of the fitting criterion, e.g. a parametric likelihood function was fit but the models were able to predict short timescale autocorrelation functions and these physically based models were able to fairly accurately summarize/model relaxation kinetics that a simple exponential relaxation could not. Other enzymes systems have exhibited this type of behavior [@SPAJCTC] and it is likely that future single-molecule experiments will yield data possessing this feature.
Perhaps more importantly, we demonstrated that a population of surrogate models was required to represent the complex dynamical system because an unobserved conformational degree freedom modulated the dynamical response and this “random force" had to be accounted for in order to predict autocorrelations valid for longer temporal trajectories. A method using parametric surrogate models calibrated over short timescales while at the same time respecting the variability induced by unresolved coordinates evolving over longer timescale was presented. The DHFR system was another instance where aggregating a collection of simpler dynamical models gave rise to a more complex stochastic process [@granger80a; @granger80b; @cox91; @SPAJCTC; @klafter09]. The basic idea is applicable to situations where a hidden slowly evolving degree of freedom modulates the dynamics and this coordinate evolves on an effective free energy surface possessing relatively low barriers [@SPAJCTC]. Issues associated with extensions were briefly discussed.
Even if a coarse system description, such as a single autocorrelation function, can be used to adequately approximate the physically relevant statistical properties of all experimentally accessible observables, the approach presented still has appeal. One circumstance where this is particularly relevant is when computer simulation trajectories are compared to frequently sampled experimental single-molecule time series [@SPAfilter]. In experimental time series, many conformational coordinates cannot typically be resolved [@SPAfilter; @SPAJCTC], so constructing a simulation that matches all relevant degrees of freedom is highly problematic. Quantitative knowledge of how the variability induced by such hidden degrees of freedom is reflected in the surrogate model parameters distribution may help in refining force fields to match kinetic properties at multiple timescales. If the force fields are believed valid, then turning to the simulations for details of the structural dynamics can help us in understanding complex molecular machines [@schultenscireview]. This type of extra detail may also assist (or lead to new) methods for computing transition rates [@chandler00]. Furthermore, as nanotechnology demands higher resolution at smaller length and timescales, one may want to avoid using a single autocorrelation function constructed by aggregating many meso or microscopic states each possessing different dynamical features because doing so may unnecessarily wash out physically relevant information. The phenomenologically motivated simple bottom-up strategy presented was one contribution in this direction.
Acknowledgements
================
The author thanks Karunesh Arora and Charles Brooks III for sharing the DHFR data.
Appendix
=========
Toy Model Parameters
--------------------
$\delta t=0.15, \delta s=\delta t/50, \mathcal{T}=200\times\delta t, \tau_0=120, (\alpha^0,\kappa^0,\eta^0)=(4,0.2,0.5), (\sigma^{\alpha},\sigma^{\kappa},\sigma^{\eta})=(6.5\times 10^{-2},6.5\times 10^{-3},1.9\times 10^{-2})$. The last set of parameters were selected to give the evolving OU parameters a stationary distribution characterized by three independent normals each having mean $(\alpha^0,\kappa^0,\eta^0)$ and standard deviation (1/2,1/20, 3/20). The initial condition of each $y$ process was set to $\alpha^0$ and the OU parameters were all set to $(\alpha^0,\kappa^0,\eta^0)$. 100 batches of 4 independent Brownian motion processes were used to evolve the system.
Predicting Quantities with Surrogate Models {#sec:standardcomp}
-------------------------------------------
The OU process is attractive for a variety of reasons. Its conditional and stationary density are both known analytically and it can be readily estimated from discrete data. For parameters possessing a stationary distribution, these can all be written explicitly in terms of Normal densities. Another appealing feature is that the AC function, denote this function by $AC(t)$, [@socci96] associated with a stationary process can readily be computed after parameter estimates are in hand, namely $AC(t)= \exp{-B t}$; recall the drift of the OU process is given by $B(A-z)$.
Unfortunately, these types of statistical summaries are more difficult to obtain with other SDEs. Position dependence of the diffusion function and nonlinear models severely complicate obtaining analytic expressions for the autocorrelation function. Note that once a single SDE models is estimated, a new large collection of sample paths can be simulated and quantities like the autocorrelation function associated with a given SDE model and $\theta$ can be empirically determined (the computational cost of simulating a scalar SDE is typically marginal in relation to a MD simulation). This can be repeated for each surrogate SDE estimated from each MD path.
However, a stationary density, under mild regularity conditions, of a scalar SDE can often be expressed in closed-form using only information contained in the estimated SDE coefficient functions via the relation [@kutoyants; @risken] $$p^{SD}(z;\Gamma)=\mathcal{Z}/\sigma(z)^2 \exp \Big( \int^z_{z_{\mathrm{REF}}} \mu(z')/\sigma(z')^2 dz' \Big),
\label{eq:invardens}$$ where in the above the SDE functions’ dependence on $\theta$ and $\Gamma$ has been suppressed to streamline the notation. $\mathcal{Z}$ represents a constant to ensure that the density integrates to unity and $z_{\mathrm{REF}}$ represents an arbitrary fixed reference point. When evaluating $p^{SD}(\cdot)$, one can encounter technical difficulties if the diffusion coefficient is allowed to take a zero or negative value (this is relevant to the PSOD model). Some heuristic computational approaches to dealing with this are discussed in Refs. [@molsim; @SPAJCTC].
Sometimes a thermodynamic motivation exists for expressing the stationary density of the high-dimensional molecular system in terms of some potential, denoted here by $V(z)$, that does not explicitly depend on the diffusion function [@risken; @dima]. In time-homogeneous scalar overdamped Brownian dynamics, where the forces of interest acting on $z$ are believed related to the gradient of $V(z)$, a “noise-induced drift" term [@risken] can be added to the drift function and this addition cancels out the contribution coming from the $1/\sigma(z)^2$ term outside the exponential. The stationary density of the modified SDE can then be expressed as being proportional to $\exp{\big(-V(z)\big)}$ in such a situation. This type of modification has a thermodynamic appeal when $z$ is the only important variable of the system and the fast-scale noise has been appropriately dealt with [@razstuart04]. The utility of such an approach in describing the *pathwise* kinetics of trajectories is another issue and single-molecule studies are one area where the distinction may be important (one may not care as much about the stationary *ensemble* distribution) .
However when there are slowly evolving lurking variables like $\Gamma$ modulating the dynamics (as is the case in many biomolecular systems) using simple expression like Eq. \[eq:invardens\] to approximate the stationary density of the high-dimensional system (with or without “noise-induced drift" corrections) is highly problematic. Note that the $\Gamma$ variable has been retained in the left hand side of Eq. \[eq:invardens\]; the stationary density estimate is only meant to be valid for a fixed estimated SDE surrogate corresponding to one value $\theta$. In this paper and others, it is assumed that for a short time interval both $\theta$ and $\Gamma$ are effectively frozen. Given a model and short time data, this can be tested using goodness-of-fit tests. However over longer timescales, $\Gamma$ evolves and modulates the dynamics so the estimated $\theta$ evolves in time (this is why the situation can be though of as a type of dynamic disorder [@xie_dynamicdis_98]). For this long time evolution, it is assumed that the form of a stochastic process depending only on $z$ is completely unknown to the researcher. Furthermore it was assumed that another order parameter (i.e. system observable) is unavailable or is unknown [@SPAJCTC; @SPAfilter; @karplus08]. Hence to approximate the stationary distribution of the high-dimensional molecular system one would require a collection of $p^{SD}$’s (each with different $\Gamma$’s) to approximate this quantity. This procedure is presented in Ref. [@SPAJCTC].
[^1]: We assume here that the initial condition is not random and hence does not contribute to the log likelihood function.
[^2]: Using a randomized $\ell$ index turns out to be adequate for DHFR (the accuracy gain of respecting the time ordering provided only marginal improvements), though we present the version respecting time ordering in the algorithm to simplify the exposition.
[^3]: The large differences persist even if the time series length is increased by a factor of 3.
[^4]: As always, when the amount of evidence increases, the likelihood of rejecting any over simplified model increases. However the tests developed in Ref. [@hong] are associated with “diagnostics" which can help one in determining if a rejected model might still contain useful information nonetheless.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The presence of nuclear medium and collective phenomena which involve several nucleons modify the parton distribution functions of nuclei (nPDFs) compared to those of a free nucleon. These modifications have been investigated by different groups using global analyses of high energy nuclear reaction world data resulting in modern nPDF parametrizations with error estimates, such as EPS09(s), HKN07 and nDS. These phenomenological nPDF sets roughly agree within their uncertainty bands, but have antiquarks for large-$x$ and gluons for the whole $x$-range poorly constrained by the available data. In the kinematics accessible at the LHC this has negative impact on the interpretation of the heavy-ion collision data, especially for the $p + A$ benchmarking runs. The EMC region is also sensitive to the proper definition of $x$, where the nuclear binding effects have to be taken into account, and for heavy nuclei one also needs to take into account that a fraction of the nucleus momentum is carried by the equivalent photons which modifies the momentum sum rule. We study how these effects affect the predictions for the nuclear modification ratios at the LHC kinematics using a model where we combine theoretical input for the leading twist nuclear shadowing (the FGS model) and the EKS98s/EPS09s nPDF set where the spatial dependence is formulated as a power series of the nuclear thickness functions $T_A$.\
[**Keywords:**]{} [*nuclear parton distribution function; LHC; impact parameter; EMC region*]{}
---
****
Modeling nuclear parton distribution functions
\
[$^a$*The Pennsylvania State University, 104 Davey Lab, University Park, PA 16802, USA* ]{}\
[$^b$*National Research Center ”Kurchatov Institute”, Petersburg Nuclear Physics Institute (PNPI), Gatchina, 188300, Russia*]{}
Proper definition of $x$
========================
Nuclear parton distribution functions (nPDFs) are usually defined for each parton flavor in terms of nuclear modifications $R_i^A(x,Q^2)$ and the corresponding free proton PDF $f_i^p(x,Q^2)$ such that $$f_i^A(x,Q^2)\equiv R_i^A(x,Q^2)\,f_i^p(x,Q^2) \,,$$ where the Bjorken $x =A\,Q^2/(2 q\cdot p_A)$, with $0\le x\le A$. In the collider frame, $x$ is simply the fraction of the nucleus momentum scaled by the factor $A$. However, since these nPDF sets are built “on top” of proton PDF sets, the tail $1\le x\le A$ is usually ignored.
Phenomenological parametrizations for nuclear modifications, such as EKS98 [@Eskola:1998df], EPS09 [@Eskola:2009uj], HKN07 [@Hirai:2007sx], DSZS [@deFlorian:2003qf], nCTEQ [@Kovarik:2010uv] etc. are largely based on deep inelastic scattering (DIS) data, which is given as a function of $x_p=Q^2/(2 q_0 m_p)$, which is independent of the target mass. The difference between $x_p$ and $x$ thus originates from the nuclear binding (see Refs. [@Frankfurt:2011cs; @Frankfurt:2012qs][^1] $$x_p= x\,(1\,+r^A_x) \,,
\label{xp}$$ where $$r^A_x = \frac{1}{m_p}\left((m_n-m_p)N/A-\epsilon_A\right)<0 \,.
\label{rx}$$ As an example, the nuclear binding energy $\epsilon_A\approx 7.88\, (7.68)$ MeV for Pb (C). In addition to the nuclear binding energy, the fraction of nucleus momentum carried by equivalent photons has to be taken into account in high energy heavy nuclear collisions. The fraction of the nucleus momentum carried by the photons is found to be [@Frankfurt:2012qs] $$\eta_{\gamma}({}^{12}{\rm C})=0.11\%,\,\,\eta_{\gamma}({}^{208}{\rm Pb})=0.7\%.$$ Since the gluon nPDFs are least constrained by the DIS and DY data, presence of the photons in the momentum sum rule mostly affects the overall momentum carried by gluons. The effect of the equivalent photon field can be taken into account by rescaling the gluons after Eq.(\[xp\]) has been applied to satisfy $$\sum_i\int_0^1 \mathrm{d}x x f^A_{i}(x,Q^2)=1-\eta_{\gamma}(A).
\label{rescale}$$ To apply the “conventional” nPDFs given as a function of $x_p$ for the calculation of the nuclear effects in the ultra-relativistic heavy ion collisions where $x$ is used, one has to translate to the “conventional” nPDFs, given as a function of $x_p$, by taking into account the difference between $x_p$ and $x$ which from now on we explicitly call $x_{\rm shift}$, as follows $$x_{\rm shift}f_i^A(x_{\rm shift},Q^2)=\begin{cases} \frac{x_p}{1+r^A_x} f_i^A(\frac{x_p}{1+r^A_x},Q^2),\,\, i=q,\bar q \\
g_{\rm scale}\frac{x_p}{1+r^A_x} f_i^A(\frac{x_p}{1+r^A_x},Q^2), \,\, i=g \end{cases},
%\nonumber\\
%&=&\begin{cases} \frac{\frac{x_p}{1+r^A_x} f_i^A(\frac{x_p}{1+r^A_x},Q^2)}
%{x_pf^A_i(x_p,Q^2)}{x_pf^A_i(x_p,Q^2)}, i=q,\bar q \\
% \frac{g_{\rm scale}\frac{x_p}{1+r^A_x} f_i^A(\frac{x_p}{1+r^A_x},Q^2)}{x_pf^A_%i(x_p,Q^2)}{x_pf^A_i(x_p,Q^2)}, i=g \end{cases}
\label{xshift}$$ where the scaling factor for gluons, $g_{\rm scale}$, is determined via Eq.(\[rescale\]). Note that for a free proton, $x=x_p=x_{\rm shift}$. It follows from Eq. (6) that while the fraction of the nucleus momentum carried by quarks is invariant with respect to the $x_p \to x_{\rm shift}$ conversion, the amount of the nucleus momentum carried by gluons, $\eta_g \equiv \int^{1}_{0} dx_{\rm shift} x_{\rm shift}
g_A(x_{\rm shift},Q^2)$, decreases by the factor of $g_{\rm
scale}=\frac{\eta_g-\eta_{\gamma}(A)}{\eta_g}$ with the rescaling.
Theoretically motivated nPDF model
==================================
In this work we combine a small-$x$ theoretical model for the leading twist nuclear shadowing, the FGS model [@Frankfurt:2011cs], with the phenomenological EKS98/EPS09 nPDF set. The FGS model is based on the generalization of the Gribov-Glauber multiple scattering formalism and QCD factorization theorems. Using the picture of high energy scattering in the laboratory frame [^2] and the notion of cross section fluctuations of energetic projectiles, multiple interactions are modeled using the effective $x$-dependent and flavor-dependent rescattering cross section $\sigma^i_{\rm soft}(x,Q^2)$, which controls the strength of the resulting nuclear shadowing. In [@Frankfurt:2011cs], based on the phenomenological analysis of cross section fluctuations in virtual photons, two models were suggested: model 1 (here referred to as FGS1) and model 2 (FGS2) corresponding to the upper and lower bounds on the predicted nuclear shadowing, respectively. Both of the models were built on top of CTEQ5 PDFs [@Lai:1999wy] (given as a function of $x_p$), and we will use this set for the combined model as well. In this paper, we will work in LO.
In the following the initial scale sea quark and gluon ($Q_0^2=2.5$ GeV$^2$) nuclear modifications for the region $10^{-4}\le x\le 0.01$, where data practically do not constrain nPDFs, are taken from the FGS1 and FGS2 parametrizations; for $0.03\le x\le 1.0$, the nuclear modifications are taken from the EKS98 parametrization [@Eskola:1998df] (which in this region is very similar to the newer set EPS09 [@Eskola:2009uj]). For the valence quarks, the nuclear modifications are taken from EKS98 for the whole $x$-range. The two parametrizations are combined by performing polynomial interpolation between them, and (after being corrected for the difference in the argument according to Eq. (\[xp\])) the gluons are rescaled as in Eq. (\[xshift\]). The resulting scaling factor at $Q_0^2=2.5$ GeV$^2$ for the gluons is $g_{\rm scale}\sim 0.984$ ($g_{\rm scale}\sim 0.997$) for Pb (C) nucleus. (The change in the scaling factor is $~0.4\%$ at $Q^2=10000^2$ GeV$^2$, so in practice a uniform scaling factor can be used for any scale.) Consequently the amount of the momentum carried by the gluons decreases by 0.88% (0.14%) for Pb (C). The fraction of the momentum carried by the gluons for each model is listed in Table 1 (41.80% for the proton in CTEQ5L).
\[momtab1e\]
[cccc]{} & EKS98 & FGS1+EKS98 & FGS2+EKS98\
&[Pb C]{} & [Pb C]{} & [Pb C]{}\
$x_p$: & [43.98 42.61]{} & [44.20 42.58]{} & [44.30 42.62]{}\
$x_{\rm shift}$: & [43.58 42.54 ]{}& [43.81 42.52]{} & [43.91 42.56]{}\
In Fig. \[gluon\_ratio\_p\_play.eps\] we show how changing the definition of $x$ affects the EKS98 gluon modification ratio ${\rm Pb}/p$ at the initial scale $Q_0^2=2.5$ GeV$^2$ and how this effect evolves up to the higher scale $Q^2=100^2$ GeV$^2$. In this work we use the QCDNUM DGLAP evolution code [@Botje:2010ay]. At the initial scale the original EKS98 gluon modification ratio $x_pG^{\rm Pb}(x_p)/x_pG^p(x_p)$ (solid line) is first modified setting $g_{\rm scale}=1$ in Eq. (\[xshift\]) (dotted-dashed line). As a result, the gluon modification ratio is only essentially modified at the EMC-region, $x_p>0.5$, where the (n)PDFs are decreasing rapidly. For the full conversion with $g_{\rm scale}\sim 0.98$ (dotted line), the gluon nPDF is naturally scaled down over the whole $x_p$-range. When evolved up to $Q^2=100^2$ GeV$^2$, the differences persist and spread towards smaller values of $x_p$. Using $x_{\rm shift}$ instead of $x_p$ obviously affects gluons for the whole $x_p$-range due to the rescaling (and at higher scales also the sea quarks via the DGLAP evolution), but for all the parton flavors the most prominent effect sets in at the EMC-region, where the parton distribution functions change quickly. As seen from above, the gluon ratio ${{\rm Pb}(x_{\rm shift})}/{\rm Pb}(x_p)$ (dashed line) is $<1$ for the whole $x_p$-range.
Note in passing that even for the valence quark distributions, the experimental information on the EMC effect in the region where the leading twist contribution dominates is very limited as the higher twist effects give a large (dominant ?) contribution to the $eA$ scattering cross section for $x\ge 0.5$ in the SLAC and JLab kinematics. Hence, to date, practically no data on the quark modification for $x\ge 0.5$ in the scaling region is available for heavy nuclei such as, e.g., lead. In Fig. \[gluon\_ratio\_d.eps\] we show the gluon ratio (${\rm Pb}/d$) for the combined models FGS1+EKS98 (dotted-dashed) and FGS2+EKS98 (dotted), together with the EKS98 (solid) gluon modification. At the initial scale above $x_p>0.01$, the difference between the models originates only from the different definition of $x$; below $x_p\le 0.01$, the two FGS nuclear shadowing models span a region considerably smaller than the error band of EPS09 gluons (see [@Eskola:2009uj] for details). When evolved up to $Q^2=10000^2$ GeV$^2$, a relevant scale in the LHC kinematics, it is evident that processes which are sensitive to the EMC-region are also sensitive to the proper definition of $x$.
![Gluon ratio for ${\rm Pb}(x_{\rm shift})/{\rm Pb}(x_p)$ at $Q_0^2=2.5$ GeV$^2$ and for ${\rm Pb}/p$ at $Q_0^2=2.5$ GeV$^2$ and $Q^2=100^2$ GeV$^2$ for EKS98. []{data-label="gluon_ratio_p_play.eps"}](Pb_gluon_ratio_to_p_play.eps){width="9.5cm"}
![Gluon ratios for ${\rm Pb}/d$ at $Q_0^2=2.5$, 100$^2$ and 10000$^2$ GeV$^2$ for EKS98, and the combination model FGS1(2)+EKS98.[]{data-label="gluon_ratio_d.eps"}](Pb_model_gluon_ratio_to_d.eps){width="9.5cm"}
Consequences for the LHC
========================
In order to understand the sensitivity of the LHC kinematics to the EMC effect, we study inclusive $\pi^0$ production in $p+{\rm Pb}$ collisions, which schematically can be expressed as $$\begin{aligned}
&& \sigma^{p+{\rm Pb} \to \pi^0 + X} = \nonumber \\
&& \sum_{i,j,k=q,\bar{q},g} {f_i^p}(x_1,Q^2) \otimes {f_j^{\rm Pb}}(x_2,Q^2) \otimes \hat{\sigma}^{ij\to k + X}(x_1,x_2,Q^2) \otimes D_{k \to \pi^0}(z,\mu_F^2),\end{aligned}$$ where the factorization and renormalization scales have been set equal (see e.g. [@Eskola:2002kv] for the formulae and details). In this work, we choose $\mu_F=p_T$ (the outgoing pion transverse momentum) and $Q=q_T$ (partonic transverse momentum).
In Fig. \[Pb\_eta\_0.0\_xbin.eps\] we show the LO invariant cross section $E d^3\sigma/dp^3$ for ${p+{\rm Pb}\to\pi^0+X}$ at $p_T=3.0,\, 10.0$ and 100.0 GeV as a function of $x_2$ (the momentum fraction carried by the parton in Pb, without conversion given in Eq. (\[xshift\])). The results have been computed with the EPS09 nuclear modifications [@Eskola:2009uj], CTEQ6L PDFs [@Pumplin:2002vw] and DSS fragmentation functions [@deFlorian:2007aj]. Working in LO the overall normalization of the spectra is not fixed, but the $x_2$-distribution and the relative normalization are not affected by this. The upper panel shows the mid-rapidity and the lower panel – the forward rapidity (at the LHC, the Pb rapidity is positive) $\pi^0$ production. For each $p_T$ value studied, the mid-rapidity production peaks at an order of magnitude smaller values of $x_2$ than the forward rapidity results, and remains significant over a wider range of $x_2$. In the forward direction the pion production is concentrated on a rather narrow $x_2$-range, making it a more sensitive probe of nuclear effects. In particular, at high-$p_T$, the pions are produced exclusively from the EMC-region, making them also sensitive to the definition of $x_2$.
Figures \[pPb\_per\_pp\_mb\_res.eps\] and \[pPb\_per\_pp\_mb\_res\_eta\_3.5.eps\] show the FGS1(2)+EKS98 results for the minimum bias nuclear modification ratio, $$R^{\pi^0}_{p{\rm Pb}}(p_T,\eta)=\frac{d^3\sigma^{p{\rm Pb}}/dp^3}{d^3\sigma^{pp}/dp^3} \,,$$ at the LHC at $\sqrt{s}=5500$ GeV. For comparison, the EKS98 results with the CTEQ5L PDFs are also shown. In the mid-rapidity (Fig. \[pPb\_per\_pp\_mb\_res.eps\]) the difference between the EKS98 and FGS+EKS98 results remain within a few-%. As can be seen from Fig. \[Pb\_eta\_0.0\_xbin.eps\] (upper panel), up to $p_T\sim 100$ GeV, the pion spectra mostly originate from the gluon distribution dominated $x$-range below the EMC-region. Therefore the differences between the models are caused both by the different assumptions about shadowing (see the lower panel where the small-$p_T$ part is shown on a logarithmic scale) and the scaling of the gluon distribution with $g_{\rm scale}<1$. of $x$. The difference between the models remain moderate even above $p_T>100$ GeV.
As seen from Fig. \[pPb\_per\_pp\_mb\_res\_eta\_3.5.eps\], the situation is drastically different in the forward rapidity pion production. Above $p_T\sim 10$ GeV, the two FGS-models are indistinguishable, but start to deviate from the EKS98 result as $p_T$ increases. This sizable effect is caused by the correction to the $x$-definition alone.
Until now we have discussed the minimum bias results, where the impact parameter dependence of the nuclear effects has been spatially averaged. In the FGS model the transverse position ${\bf s}$ dependence is naturally built in as functions of $T_A({\bf s})$ since the nuclear shadowing is first calculated for fixed ${\bf s}$ and next the integral over ${\bf s}$ is taken. In the EKS98s/EPS09s model [@Helenius:2012wd], the EKS98 and EPS09 nPDF parametrizations were also assumed to have spatial dependence as a power series of $T_A({\bf s})$. With the centrality classes modeled using the optical Glauber model, the EPS09s results were found to be consistent with the mid-rapidity PHENIX $R_{d{\rm Au}}^{\pi^0}$ centrality systematics [@Adler:2006wg]. However, as already seen in Fig. \[pPb\_per\_pp\_mb\_res\_eta\_3.5.eps\], the proper definition of $x$ has a major effect on the LHC predictions at forward rapidity. In Fig. \[pPb\_per\_pp.eps\] we applied the procedure described in [@Helenius:2012wd] to $R_{p{\rm Pb}}^{\pi^0}$ at $\eta=3.5$ for a selection of different centrality classes, and compared the EKS98s results (with CTEQ6L PDFs) with and without the $x$-corrections. Irrespective of the centrality, the correction causes a clear, measurable effect.
For an impact parameter dependent theoretically motivated nPDF model it is also important to pay special attention to the EMC region for another reason. In [@Weinstein:2010rt] the magnitude of the EMC effect was shown to be linearly related to the short range correlations (SRC) scale factor measured from electron inclusive scattering at $x\ge 1$. Consequently the impact parameter dependence of the EMC effect should be proportional to the local density [@Frankfurt:1988nt] and the EMC effect thus should be strongest in the center of the nucleus. We will address this issue in [@upcoming], where a full NLO impact parameter dependent nPDF set will be released.
This work was supported in part by US DOE Contract Number DE-FG02-93ER40771.
[0]{}
K. J. Eskola, V. J. Kolhinen and C. A. Salgado, Eur. Phys. J. C [**9**]{}, 61 (1999) \[hep-ph/9807297\].
K. J. Eskola, H. Paukkunen and C. A. Salgado, JHEP [**0904**]{}, 065 (2009) \[arXiv:0902.4154 \[hep-ph\]\].
M. Hirai, S. Kumano and T. -H. Nagai, Phys. Rev. C [**76**]{}, 065207 (2007) \[arXiv:0709.3038 \[hep-ph\]\].
D. de Florian and R. Sassot, Phys. Rev. D [**69**]{}, 074028 (2004) \[hep-ph/0311227\].
K. Kovarik, I. Schienbein, F. I. Olness, J. Y. Yu, C. Keppel, J. G. Morfin, J. F. Owens and T. Stavreva, Phys. Rev. Lett. [**106**]{}, 122301 (2011) \[arXiv:1012.0286 \[hep-ph\]\].
L. Frankfurt, V. Guzey and M. Strikman, Phys. Rept. [**512**]{}, 255 (2012) \[arXiv:1106.2091 \[hep-ph\]\].
L. Frankfurt and M. Strikman, Int. J. Mod. Phys. E [**21**]{}, 1230002 (2012) \[arXiv:1203.5278 \[hep-ph\]\].
H. L. Lai [*et al.*]{} \[CTEQ Collaboration\], Eur. Phys. J. C [**12**]{}, 375 (2000) \[hep-ph/9903282\]. M. Botje, Comput. Phys. Commun. [**182**]{}, 490 (2011) \[arXiv:1005.1481 \[hep-ph\]\].
K. J. Eskola and H. Honkanen, Nucl. Phys. A [**713**]{}, 167 (2003) \[hep-ph/0205048\].
J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. M. Nadolsky and W. K. Tung, JHEP [**0207**]{}, 012 (2002) \[hep-ph/0201195\].
D. de Florian, R. Sassot and M. Stratmann, Phys. Rev. D [**75**]{}, 114010 (2007) \[hep-ph/0703242 \[HEP-PH\]\].
S. S. Adler [*et al.*]{} \[PHENIX Collaboration\], Phys. Rev. Lett. [**98**]{}, 172302 (2007) \[nucl-ex/0610036\]. I. Helenius, K. J. Eskola, H. Honkanen and C. A. Salgado, JHEP [**1207**]{}, 073 (2012) \[arXiv:1205.5359 \[hep-ph\]\].
L. B. Weinstein, E. Piasetzky, D. W. Higinbotham, J. Gomez, O. Hen and R. Shneor, Phys. Rev. Lett. [**106**]{}, 052301 (2011) \[arXiv:1009.5666 \[hep-ph\]\]. L. L. Frankfurt and M. I. Strikman, Phys. Rept. [**160**]{}, 235 (1988).
H. Honkanen, M. Strikman, V. Guzey, [*Work in progress*]{}.
![The $x$-distribution for the minimum bias $\pi^0$ production at the LHC at $\eta=0$ and $\eta=3.5$ for different $p_T$.[]{data-label="Pb_eta_0.0_xbin.eps"}](Pb_xbin.eps){width="12cm"}
![Minimum bias FGS1(2)+EKS98 results for $R^{\pi^0}_{p{\rm Pb}}(p_T)$ for the LHC at $\sqrt{s}=5500$ GeV and $\eta=0.0$. For comparison, the EKS98 grid result with CTEQ5L pdfs is also shown. Upper panel on linear scale, lower panel on logarithmic scale.[]{data-label="pPb_per_pp_mb_res.eps"}](pPb_per_pp_mb_res.eps "fig:"){width="12cm"} ![Minimum bias FGS1(2)+EKS98 results for $R^{\pi^0}_{p{\rm Pb}}(p_T)$ for the LHC at $\sqrt{s}=5500$ GeV and $\eta=0.0$. For comparison, the EKS98 grid result with CTEQ5L pdfs is also shown. Upper panel on linear scale, lower panel on logarithmic scale.[]{data-label="pPb_per_pp_mb_res.eps"}](pPb_per_pp_mb_res_log.eps "fig:"){width="12cm"}
![Minimum bias FGS1(2)+EKS98 results for$R^{\pi^0}_{p{\rm Pb}}$ for the LHC at $\sqrt{s}=5500$ GeV and $\eta=3.5$. For comparison, the EKS98 grid result with CTEQ5L pdfs is also shown. The FGS1+EKS98 and FGS2+EKS98 curves are indistinguishable.[]{data-label="pPb_per_pp_mb_res_eta_3.5.eps"}](pPb_per_pp_mb_res_eta_3.5.eps){width="12cm"}
![$R^{\pi^0}_{p{\rm Pb}}$ for the LHC at $\sqrt{s}=5500$ GeV and $\eta=3.5$ for the EKS98 (with CTEQ6L1) results computed with $x_p$ and $x_{\rm shift}$. The different centrality classed are computed as in [@Helenius:2012wd]. The FGS1+EKS98 and FGS2+EKS98 curves are indistinguishable.[]{data-label="pPb_per_pp.eps"}](pPb_per_pp_eta_3.5.eps){width="16cm"}
[^1]: In Ref [@Frankfurt:2012qs] there is a sign error in the corresponding formula, Eq.(20). For detailed discussion, see the revised version of \[7\].
[^2]: The equivalent picture can be formulated in the nucleus fast frame [@Frankfurt:2011cs].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We introduce a notion of metric on a Lie groupoid, compatible with multiplication, and we study its properties. We show that many families of Lie groupoids admit such metrics, including the important class of proper Lie groupoids. The exponential map of these metrics allow us to establish a Linearization Theorem for Riemannian groupoids, obtaining both a simpler proof and a stronger version of the Weinstein-Zung Linearization Theorem for proper Lie groupoids. This new notion of metric has a simplicial nature which will be explored in future papers of this series.'
address:
- 'Instituto de Matemática Pura e Aplicada, Estrada Dona Castorina 110, Rio de Janeiro, 22460-320, Brazil'
- 'Department of Mathematics, University of Illinois at Urbana-Champaign, 1409 W. †Green Street, Urbana, IL 61801, USA'
author:
- Matias del Hoyo
- Rui Loja Fernandes
title: Riemannian metrics on Lie groupoids
---
[^1]
Introduction
============
This is the first of a series of papers dedicated to the study of Riemannian metrics on Lie groupoids. There have been several attempts to propose a good notion for metrics on Lie groupoids (see, e.g., [@gghr; @glick; @hepworth; @ppt]), but to our knowledge, all of them fail to take into account the groupoid multiplication. The compatibility with the multiplication may not be required in certain particular cases, such as étale groupoids, describing orbifolds. But for general Lie groupoids, describing more general singular spaces, taking into account the groupoid multiplication becomes crucial. In this paper we introduce a condition of compatibility between the metric and the groupoid multiplication, which is automatically satisfied for étale groupoids, but yields a novel concept for more general Lie groupoids.
The basic idea underlying our approach can be easily described using the simplicial model of the Lie groupoid $G{\rightrightarrows}M$, say its [*nerve*]{} $NG$: $$\xymatrix@1{ \dots \ar@<0.60pc>[r]\ar@<0.30pc>[r] \ar[r] \ar@<-0.30pc>[r]\ar@<-0.60pc>[r]& {G^{(n)}} \ar@<0.45pc>[r]\ar@<0.15pc>[r]\ar@<-0.15pc>[r]\ar@<-0.45pc>[r] &\cdots \ar@<0.45pc>[r]\ar@<0.15pc>[r]\ar@<-0.15pc>[r]\ar@<-0.45pc>[r]&{G^{(2)}} \ar@<0.30pc>[r] \ar[r] \ar@<-0.30pc>[r]& {G^{(1)}} \ar@<0.2pc>[r]\ar@<-0.2pc>[r] & {G^{(0)}}[\ar[r]& {G^{(0)}}/{G^{(1)}}]}.$$
The manifold ${G^{(n)}}$ consists of chains of $n$ composable arrows, or equivalently, commutative $n$-simplices on $G$. This viewpoint clarifies the definition of the face maps ${G^{(n)}}\to{G^{(n-1)}}$, as well as the existence of an action of the symmetric group $S_{n+1}{\curvearrowright}{G^{(n)}}$ by permuting the vertices of a simplex. We define an **$n$-metric on the Lie groupoid $G{\rightrightarrows}M$** to be a Riemannian metric ${\eta^{(n)}}$ on ${G^{(n)}}$ invariant under $S_{n+1}$ and for which the fibers of one (and therefore every) face map ${G^{(n)}}\to{G^{(n-1)}}$ are locally equidistant. For small values of $n$ one finds:
(i) If $n=0$ we recover the notion of a *transversely invariant metric* on the space of units $M={G^{(0)}}$, i.e., a metric transversely invariant under the action of $G$ on its units. Such metrics and their properties have been studied in [@ppt].
(ii) If $n=1$ we recover the notion of Riemannian groupoid introduced first in [@gghr]. This is a metric in the space arrows ${G^{(1)}}=G$ invariant under inversion and for which the source and target fibers are equidistant.
(iii) If $n=2$ we obtain a new notion of metric on a Lie groupoid, a metric in the space of composable arrows ${G^{(2)}}$ which is $S_3$-invariant and transverse to the multiplication map. We will stress the advantages of this new concept.
An $n$-metric on ${G^{(n)}}$ induces a $k$-metric on each ${G^{(k)}}$ for $0\le k<n$, for which all face maps ${G^{(k+1)}}\to{G^{(k)}}$ are Riemannian submersions. Since the nerve of a Lie groupoid is completely determined once one reaches ${G^{(2)}}$, it is enough for many purposes to consider only 2-metrics. We study 2-metrics in this paper and reserve a discussion on general $n$-metrics for the next papers in this series [@fdh1; @fdh2]. Still, it maybe worth to mention here the following fact, to be proved in [@fdh1]: there is at most one 3-metric inducing a given 2-metric in $G{\rightrightarrows}M$, and if such a 3-metric exists, it has a unique extension to an $n$-metric for every $n$. On the other hand, we will give later examples of groupoids which admit an $n$-metric, but do not admit an $n+1$-metric, for $n=0,1,2$. Also, uniqueness fails in these low degrees, so one can have, e.g., two different 2-metrics on ${G^{(2)}}$ inducing the same metric on ${G^{(1)}}$.
A 2-metric ${\eta^{(2)}}$ on a Lie groupoid $G{\rightrightarrows}M$ can be used to great profit. We will call the pair $(G{\rightrightarrows}M,{\eta^{(2)}})$ a **Riemannian groupoid**. As we will see later, there are plenty of classes of groupoids admitting 2-metrics: Lie groups, étale groupoids, transitive Lie groupoids, locally trivial Lie group bundles, etc. More important, our first main result shows that:
\[thmz:main:1\] Any Hausdorff proper Lie groupoid $G{\rightrightarrows}M$ admits a 2-metric.
In fact, a Hausdorff proper groupoid admits an $n$-metric, for any $n$, but we choose to ignore this for now, since our focus in this paper is on 2-metrics.
Our first major application of the existence of 2-metrics is to the *linearization problem* for Lie groupoids. Given a Lie groupoid $G{\rightrightarrows}M$ and an embedded saturated submanifold $S\subset M$ (i.e. $S$ is a collection of orbits of $G$), one has a *linear local model* for $G$ around $S$: one can show that the normal bundle to the restriction $G_S{\rightrightarrows}S$ is a Lie groupoid $\nu(G_S){\rightrightarrows}\nu(S)$, which only depends on the first jet of $G_S$. Then the linearization problem asks if there exists a groupoid isomorphism from a neighborhood of $G_S$ in $G$ to a neighborhood of $G_S$ in $\nu(G_S)$. Special cases of this problem, where the answer is affirmative, include some classical results in differential geometry, such as:
- Ehresmann’s Theorem for submersions,
- Local Reeb stability for foliations,
- The Bochner Linearization Theorem and, more generally, the Tube Theorem for proper Lie group actions.
In recent years, this problem has been intensively studied by several authors focusing on the case of proper groupoids [@cs; @ppt; @wein1; @wein2; @zung]. Using 2-metrics and their exponential maps, we will give a simple, geometric proof of the following much more general result:
\[thmz:main:2\] Let $(G{\rightrightarrows}M,{\eta^{(2)}})$ be a Riemannian Lie groupoid and let $S\subset M$ be a saturated embedded submanifold. Then the exponential map defines a linearization of $G$ around $S$.
This theorem yields all known linearization theorems for Lie groupoids. In particular, from Theorems \[thmz:main:1\] and \[thmz:main:2\] one can easily obtain the linearization theorem for proper Lie groupoids, where the groupoid neighborhood can be taken to be a full groupoid neighborhood, and the invariant linearization theorem for source proper Lie groupoids, where the groupoid neighborhood can be taken to be saturated (see Section \[subsec:linearization\] for details).
Summarizing, our metric approach to the linearization problem has the following advantages over the approaches one can find in the literature: (i) it is valid for any Riemannian groupoid, which include proper groupoids but also many other classes, enlarging the range of application, (ii) it holds around any saturated submanifold $S$ and not only around orbits $O$, (iii) it gives some control over the linearization map, and (iv) it has a simple conceptual proof that enlight the subject considerably.
The fact that the linearization is valid around any saturated submanifold $S$, instead of just an orbit $O$, has a nice interpretation from a *stacky* viewpoint: we will prove in [@fdh2] that our notion of metric is Morita invariant, hence it leads to a notion of metrics on smooth stacks, an essential tool in the study of their geometry. From this perspective, our linearization around an orbit $O$ yields normal coordinates around a point in the stack, while linearization around a saturated set $S$ yields an exponential tubular neighborhood around a substack. We will develop these and other aspects of metrics on Lie groupoids in the forthcoming papers [@fdh1; @fdh2].
[**Acknowledgments.** ]{} We thank IST-Lisbon, IMPA and U. Utrecht for hosting us at several stages of this project. We also thank H. Bursztyn, M. Crainic, I. Marcut, D. Martinez-Torres, H. Posthuma and I. Struchiner for many fruitful discussions.
Transversely invariant metrics
==============================
In this section we discuss metrics which are invariant under groupoid actions which are relevant for our study of metrics on groupoids. Since a groupoid action *is not by diffeomorphisms* of the space where the groupoid acts, this notion has some subtleties that one needs to address.
Background on Riemannian submersions
------------------------------------
Given $E$, $B$ smooth manifolds endowed with metrics $\eta^E$, $\eta^B$, a [**Riemannian submersion**]{} $p:E\to B$ is a submersion whose fibers are equidistant. This condition amounts to requiring that for every point $e\in E$ the map ${{\rm d}}_ep:T_eF^\bot\to T_bB$ is an isometry, where $b=p(e)$ and $T_eF^\bot$ denotes the subspace of vectors orthogonal to the fiber $F=p^{-1}(b)$.
Given an inner product $\eta$ on a finite dimensional vector space $V$, we denote by $\eta^*$ the [**dual inner product**]{} on the dual space $V^*$, which results from the identification $V\cong V^*$ defined by $v\mapsto \eta(v,\cdot)$. In terms of the dual inner products, the condition for a Riemannian submersion can be rephrase as follows: for all $e\in E$ the map $({{\rm d}}_ep)^*:T_b^*B\to T_eF^\circ$ is an isometry, where $T_eF^\circ$ denotes the annihilator of the vectors tangent to the fiber.
For a Riemannian submersion $p:E\to B$ the metric on $B$ is completely determined by the metric on $E$. In fact, given any submersion $p:E\to B$ and $\eta^E$ a metric on $E$, let us call $\eta^E$ [**$p$-transverse**]{} if for all $e,e'$ belonging to the same fiber $F$ the composition $T_eF^\circ {\leftarrow}T_b^*B \to T_{e'}F^\circ$ is an isometry. Of course, this can also be rephrased in terms of the orthogonal spaces to the fibers.
When $\eta^E$ is $p$-transverse we can endow $B$ with a [**push-forward metric**]{} $p_*\eta^E$, defined as the unique metric $\eta^B$ which makes the maps $T_b^*B \to T_{e}F^\circ$ isometries. This gives a smooth well-defined metric on $B$, and hence we can deconstruct the notion of Riemannian submersion as follows: $$p:E\to B \text{ is Riemannian}
\quad\iff\quad \begin{cases} \eta^E \text{ is $p$-transverse, and}\\p_*\eta^E=\eta^B.\end{cases}$$
Just like any manifold can be made into a Riemannian manifold, [*every submersion can be made Riemannian*]{}:
\[every.submersion.is.Riemannian\] For any submersion $p:E\to B$ there exists a $p$-transverse metric $\eta^E$ on $E$.
It is enough to set an Ehresmann connection, declare it orthogonal to the fibers, and construct the metric on the fibers arbitrary and on the connection as a pullback of a fixed one in $B$.
The correspondence $\eta\mapsto \eta^*$ between metrics on $V$ and $V^*$ is not linear, and this makes the two points of view on Riemannian submersions, via the annihilators of the fibers or via the normal spaces to the fibers, somewhat different. In fact, we advocate that it is more advantageous to take the cotangent space point of view in the study of Riemannian submersions, as illustrated by the following proposition:
Let $p:E\to B$ be a submersion and let $\{\eta_1,\dots,\eta_k\}$ be $p$-transverse metrics. Then their [**cotangent average**]{} $\eta$ defined by $$\eta:=\frac{1}{k}\left(\sum_{i=1}^k \eta_i^*\right)^*,$$ is also $p$-transverse.
In contrast, the tangent average is usually not $p$-transverse, as shown by the following simple example:
\[cotangent.principle\] Let $\eta$ be the canonical metric on ${\mathbb R}^2$ and let $\tilde \eta$ be the metric on ${\mathbb R}^2$ with orthonormal frame $\{\partial_x + f(x,y) \partial_y,\partial_y\}$. If $p:{\mathbb R}^2\to{\mathbb R}$ is the projection in the first factor, both metrics are $p$-transverse, with push-forward the canonical metric. The same also holds for their *cotangent average* $\frac 1 2(\eta^* +\tilde\eta^*)^*$. However, the *tangent average* metric $\frac 1 2(\eta+\tilde\eta)$ is not $p$-trasnverse, for the vector $\partial_x+ \frac{1} 2 f(x,y)\partial_y$ is normal to $\partial_y$, its projection does not depend on $y$, but its norm does.
The philosophy that one must work on the cotangent bundle, rather than on the tangent bundle, will occur frequently in the sequel. For example, for proper groupoids we will use averaging methods to construct suitable metrics, and these methods will only work when considering the dual metrics on the cotangent bundle.
We now establish a simple fact involving [**composition of Riemannian submersions**]{} that will be very useful later.
\[triangle\] Let $q:\tilde E\to E$ be a surjective submersion, let $\eta$ be a $q$-transverse metric on $\tilde E$, and consider another surjective submersion $p:E\to B$. $$\xymatrix@1{\tilde E \ar[r]^q & E \ar[r]^p & B}$$ Then $\eta$ is $pq$-transverse if and only if the push-forward metric $q_*\eta$ is $p$-transverse. In that case, we have $p_*(q_*(\eta))=(pq)_*(\eta)$.
Given $\tilde e\in \tilde E$ denote $e=q(\tilde e)$ and $b=p(e)$. Let us denote by $F_x$ the fiber corresponding to the map $x$. We have the following commutative diagram, where the horizontal arrows are linear isomorphisms: $$\newdir{ (}{{}*!/-5pt/@^{(}}
\xymatrix@R=10pt{
T_{\tilde e}F_{pq}^\circ \ar@<2pt>@{ (->}[d] &
T_eF_p^\circ \ar[l] \ar@<2pt>@{ (->}[d] &
T_b^*B \ar[l]\\
T_{\tilde e}F_{q}^\circ \ar@<2pt>@{ (->}[d] &
T_e^*E \ar[l] & & \\
T_{\tilde e}^*\tilde E}$$ Since $\eta$ is $q$-transverse we can push it forward so as the map $T_{\tilde e}F_q^\circ{\leftarrow}T_e^*E$ is an isometry, and hence also its restriction $T_{\tilde e}F_{qp}^\circ{\leftarrow}T_eF_p^\circ$. It becomes now clear that $T_eF_p^\circ{\leftarrow}T_b^*B$ is an isometry if and only if $T_{\tilde e}F_{qp}^\circ{\leftarrow}T_b^*B$ is, from where the result easily follows.
\[pullback metric\] One final observation concerning Riemannian submersions: Given $f_1:E_1\to B$ and $f_2:E_2\to B$ Riemannian submersions, the fibred product $E_1\times_B E_2$ has a natural [**pullback metric**]{} $\eta^{E_1\times_B E_2}$. It is defined by $$\eta^{E_1\times_B E_2}((v,w),(v',w'))=\eta^{E_1}(v,v')+\eta^{E_2}(w,w')-\eta^B(u,u')$$ where $u={{\rm d}}f_1(v)={{\rm d}}f_2(w)$ and $u'={{\rm d}}f_1(v')={{\rm d}}f_2(w')$. This metric is smooth and has the property that all maps in the commutative diagram $$\xymatrix{
E_1\times_B E_2\ar[d]\ar[r] & E_2\ar[d]^{f_2}\\
E_1\ar[r]_{f_1} & B}$$ are Riemannian submersions.
Transversely invariant metrics for groupoid actions {#sub:inv.metrics}
---------------------------------------------------
Suppose a Lie groupoid $G{\rightrightarrows}M$ acts on a space $E\to M$. What does it mean for a metric in $E$ to be invariant under that action? An arrow $g$ in $G$ determines only a diffeomorphism $E_{s(g)}\to E_{t(g)}$ so the notion of invariant metric is not a straightforward generalization of the case of Lie group actions, which we start by recalling.
Given $\theta:G{\curvearrowright}E$ a smooth action of a Lie group $G$ on a manifold $E$, for each $g\in G$ we let $\theta_g:E\to E$, $e\mapsto g\cdot e$ denote the translation by $g$. If $\eta^E$ is a Riemannian metric on $E$ then we say that
(a) $\eta^E$ is [**$\theta$-invariant**]{} if for each $g\in G$ the map $\theta_g:E\to E$ is an isometry.
(b) $\eta^E$ is [**transversely $\theta$-invariant**]{} if for each orbit $O$ and each $g\in G$ the map $\bar \theta_g:TE/TO\to TE/TO$ is an isometry, or equivalently, if $(\theta_g)^*:TO^\circ\to TO^\circ$ is an isometry.
Clearly (a) implies (b), which is weaker in general. For instance, any metric on ${\mathbb R}$ is obviously transversely invariant under translations ${\mathbb R}{\curvearrowright}{\mathbb R}$, but it may not be invariant.
When moving from groups to groupoids, note that the notion of invariance as in (a) does not make sense. If $G{\rightrightarrows}M$ is a Lie groupoid and $G{\curvearrowright}E$ is an action, there is not a natural lift to a tangent action $G{\curvearrowright}TE$. Nevertheless, we can still define transversely invariant metrics as in (b) by using the so-called normal representation, which we now recall.
Given $G{\rightrightarrows}M$ a Lie groupoid and $O\subset M$ an orbit, the [**normal representation**]{} $\lambda:G_O{\curvearrowright}\nu(O)$ is a linear action of the restriction groupoid $G_O{\rightrightarrows}O$ over the normal bundle $\nu(O)\to O$. It encodes the linear infinitesimal data around the orbit and plays a fundamental role in the theory (see eg. [@survey]). The action is defined by the composition $$G_O\times_O \nu(O) {\xrightarrow}\cong \nu(G_O){\xrightarrow}{\overline{dt}} \nu(O)$$ where the first map is the natural identification between the normal bundle of $G_O\subset G$ and the pullback of $\nu(O)$ through the source map . Unraveling this construction, for each arrow in $G_O$ the map $\lambda_g:\nu_x(O)\to \nu_y(O)$ can be geometrically described as follows: if $\gamma$ is a curve on $M$ whose velocity at 0 represents $v\in \nu_x(O)$, and $\tilde\gamma$ is a curve on $G$ such that $\tilde\gamma(0)=g$ and $s\tilde\gamma=\gamma$, then $\lambda_g(v)\in \nu_y(O)$ is defined by the velocity at 0 of $t\tilde\gamma$.
The [**conormal representation**]{} $\lambda^*:G_O{\curvearrowright}\nu^*(O)$ is defined by dualizing the previous one. Explicitly, we use the natural identification $\nu^*(O)\cong TO^\circ$ and an arrow in $G_O$, acts on the conormal space at $x$ by $$\lambda^*_g:T_x O^\circ\to T_{y}O^\circ, \quad \phi\mapsto \phi\circ \lambda_{g^{-1}}.$$
\[rmk:naturality.normal-representation\] The normal and the conormal representations are natural. More precisely, let $\Phi:G\to G'$ be a Lie groupoid map covering $\phi:M\to M'$. Given an arrow in the orbit $O\subset M$, write $x'=\phi(x)$, etc. and consider the following commutative diagram: $$\xymatrix{
T_xO^\circ \ar[r]^{ds^ *} & T_g G_O^\circ & T_yO^\circ \ar[l]_{dt^ *}\\
T_{x'}O'^\circ \ar[r]^{ds'^*} \ar[u]^{{{\rm d}}\phi^*} & T_{g'} G_{O'}^\circ \ar[u]^{{{\rm d}}\Phi^*} & T_{y'}O'^\circ \ar[u]^{{{\rm d}}\phi^*} \ar[l]_{dt'^ *}
}$$ The horizontal arrows are all isomorphisms, the upper composition gives the conormal action corresponding to $g$, and the bottom composition the corresponding to $g'=\Phi(g)$, showing the naturality of the conormal representation. The argument for the normal representation is analog.
If we have a metric $\eta^E$ on the manifold $E$ and $O\subset E$ is a submanifold, then on the normal bundle $\nu(O)$ we consider the metric coming from the identification $\nu(O)\cong TO^\perp\subset TE$ and on the conormal bundle $\nu^*(O)=(TO)^\circ\subset T^*E$ we consider the restriction of the metric.
\[invariant-metric\] Let $\theta:G{\curvearrowright}E$ be a Lie groupoid action. We say that a metric $\eta^E$ on $E$ is [**transversely $\theta$-invariant**]{} if the normal representation of the action groupoid $G\times_M E{\rightrightarrows}E$ is by isometries, or equivalently, if the conormal representation is by isometries.
Transversely invariant metrics and quotients
--------------------------------------------
Given $G$ a Lie groupoid, there is a correspondence between free and proper actions $G{\curvearrowright}P$ and principal $G$-bundles $P\to B$: the quotient space $P/G$ of a free proper action naturally inherits a smooth manifold structure, for which the projection $\pi:P\to P/G$ is a submersion (see e.g. [@survey]). This extends the better known case of free and proper actions of Lie groups and principal group bundles (see e.g. [@sharpe], Appendix E). The next result is a Riemannian version of this correspondence.
\[quotient.metrics\] Let $\theta:G{\curvearrowright}P$ be a free and proper groupoid action with quotient $\pi_P:P\to P/G$. A metric $\eta^P$ on $P$ is transversely $\theta$-invariant if and only if it is $\pi_P$-transverse. In such a case, $P/G$ inherits a well-defined [**quotient metric**]{} $\bar\eta=\pi_*\eta$, for which the projection is a Riemannian submersion.
The orbits $O$ of $\theta$ coincide with the fibers of $\pi_P:P\to P/G$, so $\eta^P$ is $\pi_P$-transverse if and only if for all $e,e'\in O$ the maps: $$(T_e O)^\circ {\xrightarrow}{\cong} (T_{e'}O)^\circ,\quad e,e'\in O,$$ are isometries. But this map coincides with the conormal representation along the unique arrow . Therefore, being transversely $\theta$-invariant amounts to be the same as being $\pi_P$-transverse. The quotient metric is just the push-forward metric as defined in a previous paragraph.
\[ex:left:right:action\] A Lie groupoid acts on its arrows by left translations. This actions is principal with projection the source map $s:G\to M$. Hence a metric $\eta$ on $G$ is transversely left invariant if and only if it is $s$-transverse. Analogously, a metric is transversely right invariant if and only if it is $t$-transverse.
Many of the metrics that we are going to deal with are quotient metrics for free and proper actions, as in Lemma \[quotient.metrics\]. Then the following technical result becomes an important tool for constructing examples.
\[quotient.metrics:maps\] Let $(E,\eta^E)$, $(E',\eta^{E'})$ be Riemannian manifolds endowed with free proper groupoid actions $G{\curvearrowright}E$, $G{\curvearrowright}E'$, for which the metrics are transversely invariant. If $p:(E,\eta^E)\to (E',\eta^{E'})$ is an equivariant Riemannian submersion then the induced map between the quotients $\bar p:(E/G,\bar\eta^E)\to(E'/G,\bar\eta^{E'})$ is a Riemannian submersion as well.
Call $\pi_E:E\to E/G$ and $\pi_{E'}:E'\to E'/G$ the corresponding projections, which are Riemannian submersions by construction (see Lemma \[quotient.metrics\]): $$\xymatrix{ E \ar[d]_{\pi_E} \ar[r]^p & E' \ar[d]^{\pi_{E'}} \\ E/G \ar[r]^{\bar p} & E'/G}$$ Since $\pi_{E'}$ and $p$ are Riemannian submersions then so does their composition $\pi_{E'}p=\bar p\pi_E$, and since $\pi_E$ is a Riemannian submersion we conclude that $\bar p$ also is (cf. Lemma \[triangle\]).
Riemannian groupoids {#sec:Riemannian:groupoids}
====================
Given $G{\rightrightarrows}M$ a Lie groupoid, we let ${G^{(n)}}\subset G^n$ be the embedded submanifold consisting of chains of $n$ composable arrows in $G$. Thus ${G^{(0)}}=M$ are the objects, ${G^{(1)}}=G$ are the arrows, and ${G^{(2)}}=\{(h,g)|s(h)=t(g)\}\subset G\times G$ are the pairs of composable arrows, or equivalently the commutative triangles in $G$. We now discuss appropriate notions of metrics on each of these spaces. We emphasize that we will be mostly interested in metrics in ${G^{(2)}}$, which we consider to be the right notion of metric on a Lie groupoid. Nevertheless, the metrics in ${G^{(1)}}$ and in ${G^{(0)}}$, which have been studied before (see [@ppt; @gghr]), also play a role in our approach.
Metrics on the space of units: 0-metrics
----------------------------------------
We restate the definition of [@ppt] in our language:
A [**0-metric**]{} on a Lie groupoid $G{\rightrightarrows}M$ is a Riemannian metric ${\eta^{(0)}}$ on the manifold ${G^{(0)}}=M$ which is transversely invariant for the canonical action $G{\curvearrowright}M$, $g\cdot s(g)=t(g)$.
In [@ppt] the authors develop some aspects of the theory of 0-metrics. One of their main results states that every proper Lie groupoid admits a 0-metric [@ppt Prop. 3.14], that for this metric the orbit foliation is a singular Riemannian foliation [@ppt Prop. 6.4], and that the corresponding distance function induces a topology on the orbit space that agrees with the quotient topology [@ppt Thm. 6.1]. Some of these results will be generalized later.
For now, let us observe that obviously not every Lie groupoid admits a 0-metric. Next we give a counter-example inspired in a foliation with infinite holonomy.
\[no.example\] The action ${\mathbb R}{\curvearrowright}S^1\times{\mathbb R}$ defined by $\lambda\cdot(z,t)=(e^{2\pi i\lambda}z,e^\lambda t)$), leads to a Lie groupoid $G{\rightrightarrows}M$ with a single compact orbit $S^1\times 0$. For $x=(1,0)$ the isotropy group $G_x\cong{\mathbb Z}$ is generated by $g=(1,1,0)$, and its normal representation is given by multiplication by $e$, say $g\cdot[\partial_t]=e[\partial_t]$. Hence, there cannot be a 0-metric on $G{\rightrightarrows}M$.
Nevertheless, there are plenty of examples of Lie groupoids endowed with 0-metrics. In some cases the invariance condition becomes vacuous, hence a 0-metric amounts to be the same as a metric on the units.
The unit groupoid $M{\rightrightarrows}M$ of a manifold $M$, a Lie group $G{\rightrightarrows}\ast$ (viewed as a groupoid with a single object), and more generally any transitive Lie groupoid $G{\rightrightarrows}M$, any metric on $M$ is a 0-metric: there is just one orbit, so the condition is vacuous.
Similarly, for a [**Lie group bundle**]{} $G{\rightrightarrows}M$, that is, a Lie groupoid on which the source and target maps agree, any metric on $M$ is a 0-metric: the orbits are just the points of $M$ and the normal representations are trivial.
Let us turn to less trivial examples.
If $p:M\to N$ is a submersion, then a 0-metric on the corresponding submersion groupoid $M\times_N M{\rightrightarrows}M$ is the same thing as a $p$-transverse metric on $M$.
The notion of a 0-metric arose first in the context of *étale groupoids*. Recall that an [**étale groupoid**]{} $G{\rightrightarrows}M$ is one on which both $G$ and $M$ have the same dimension, and therefore all the structural maps $s,t,m,u,i$ are étale, i.e., local diffeomorphisms, and the orbits are 0-dimensional. In an étale Lie groupoid every arrow $g$ induces exactly one (germ of a) bisection. The corresponding pseudogroup of transformations of $M$ is called the [**effect**]{} of $G$. The following lemma is immediate.
\[0-metrics.on.etal\] A 0-metric on an étale Lie groupoid $G{\rightrightarrows}M$ is the same as a metric on $M$ invariant under the effect of $G$, that is, a metric for which the bisections act by isometries.
The orbit space of a proper effective étale groupoid has an [*orbifold structure*]{}. Conversely, every (effective) orbifold can be obtained in this way, setting a 1-1 correspondence between Morita classes of such groupoids and isomorphism classes of effective orbifolds (see [@mm]). Under this correspondence, it is easy to see that a 0-metric on the groupoid amounts to be the same thing as a metric on the orbifold, as defined for instance in loc. cit. (see also [@hepworth Prop 5.5]).
In some works (see, e.g., [@glick]) a [*Riemannian groupoid*]{} $G{\rightrightarrows}M$ is defined as a Lie groupoid endowed with a metric $\eta$ on the units for which any bisection acts by isometries. For étale groupoids this is just the notion of a 0-metric, but for general groupoids this is not a very useful definition: if a Lie groupoid admits such a metric then all its orbits must be 0-dimensional. In fact, if $O\subset M$ is an orbit with positive dimension and is an arrow on $O$, then for arbitrary non-zero vectors $w\in T_yO$ and $v\in T_xO$ we can always find $u\in T_gG$ with ${{\rm d}}t(u)=w$ and ${{\rm d}}s(u)=v$ (see [@survey Prop. 3.5.1]). A subspace $S\subset T_gG$ complementing both $\ker_g {{\rm d}}s$ and $\ker_g {{\rm d}}t$ and containing $u$ then leads to a bisection relating $w$ and $v$, from where they should have the same norm, contradicting that they are arbitrary.
Metrics on the space of arrows: 1-metrics
-----------------------------------------
We recall now the definition of a metric on a groupoid that one can find in [@gghr], which we reformulate in our language:
A [**1-metric**]{} on $G{\rightrightarrows}M$ is a metric ${\eta^{(1)}}$ on the manifold ${G^{(1)}}=G$ which is transversely left invariant (see Example \[ex:left:right:action\]) and for which the inversion $i$ is an isometry.
As we have seen, transversely left invariant is the same as $s$-transverse, and assuming that the inversion $i:G\to G$ is an isometry, this is equivalent to $t$-transverse, by Lemma \[triangle\]. This gives many equivalent formulations for previous definition. Note that if a metric $\eta$ on $G$ is both $s$-transverse and $t$-transverse, it may not be invariant under inversion, but it leads to a groupoid metric by considering the cotangent average ${\eta^{(1)}}=(\frac 1 2(\eta^* + i^*\eta^*))^*$.
Given $G{\rightrightarrows}M$ a Lie groupoid and ${\eta^{(1)}}$ a 1-metric, since the inversion is an isometry, by Lemma \[triangle\] the push-forward metrics $s_*{\eta^{(1)}}$ and $t_*{\eta^{(1)}}$ on $M$ agree. We denote this common induced metric on $M$ by ${\eta^{(0)}}$. Note that ${\eta^{(0)}}$ differs in general from the restriction of ${\eta^{(1)}}$ along the unit map, as is observed in [@gghr] and we will see in many examples. The proof of the first two items in the following proposition can also be found in [@gghr]:
\[metric.on.the.units\] Given $G{\rightrightarrows}M$ a Lie groupoid, ${\eta^{(1)}}$ a 1-metric and ${\eta^{(0)}}$ the induced metric on $M$. Then:
(i) The source and target maps are Riemannian submersions, and the units $u(M)\subset G$ form a totally geodesic submanifold;
(ii) ${\eta^{(0)}}$ makes the orbit foliation $F_M$ of $M$ into a singular Riemannian foliation;
(iii) ${\eta^{(0)}}$ is transversely invariant for the action $G{\curvearrowright}M$, namely it is a 0-metric.
The units form a totally geodesic submanifold because they form the fixed point set of the inversion map, which is an isometry. That the source and target are Riemannian submersions is immediate. It implies in particular that the foliations on $G$ given by the source and target fibers are Riemannian, namely a geodesic which is orthogonal to some leaf remains orthogonal to every leaf it meets. Now, since for every $g\in G_O$ we have $T_g G_O=\ker {{\rm d}}_g s + \ker {{\rm d}}_g t$, it follows that the foliation $F_G$ on $G$ given by the orbits of the groupoid is also Riemannian. This easily implies that the foliation $F_M=s_*(F_G)$ is also Riemannian, for we can locally lift geodesics along $s$.
Regarding the last statement, given an arrow on $G$, its action under the normal representation consists of the composition $\overline{{{\rm d}}t} \circ\overline{{{\rm d}}s}^{-1}$. $$N_xO {\xleftarrow}{\overline{ds}} N_gG_O {\xrightarrow}{\overline{dt}} N_yO$$ Since the source and target map are Riemannian submersions, and since $T_gG_O=ds^{-1}(T_xO)=dt^{-1}(T_yO)$, the two maps above are isometries and we are done.
If ${\eta^{(1)}}$ is a 1-metric inducing ${\eta^{(0)}}$ we will say that ${\eta^{(1)}}$ is an [**extension**]{} of ${\eta^{(0)}}$. Let us start by giving an example of a 0-metric than cannot be extended to a 1-metric.
Consider the vector field $X=x^2\partial_x$ in ${\mathbb R}$, whose integral curve with initial condition $x\neq 0$ is given by $\gamma_x(\epsilon)=\frac{x}{1-x\epsilon}$. Denote by $G{\rightrightarrows}{\mathbb R}$ the corresponding [*flow groupoid*]{}: the arrows in $G$ are pairs $(\epsilon,x)$ for which $\gamma_x(\epsilon)$ is defined. For this groupoid, we have $s(\epsilon,x)=x$ and $t(\epsilon,x)=\gamma_x(\epsilon)$, while multiplication is defined in the obvious way.
Now any metric ${\eta^{(0)}}$ on ${\mathbb R}$ is a 0-metric: if $x\neq 0$ then the normal space to the orbit is trivial, so there is no requirement on ${\eta^{(0)}}$. If $x=0$, the normal space to the orbit is ${\mathbb R}$, but the normal representation of $G$ on ${\mathbb R}$ is trivial because the linear part of the vector field $X$ is zero, so again there is no condition on the metric. On the other hand, there are no 1-metrics on $G$: this can be seen using either (i) or (ii) in Proposition \[metric.on.the.units\]: the orbit foliation $F_M$ of $M$ has orbits ${\mathbb R}_{<0}$, $\{0\}$ and ${\mathbb R}_{>0}$, so cannot be a Riemannian foliation; also, there is no metric ${\eta^{(1)}}$ in $G$ for which *both* the s-fibres and the t-fibres are Riemannian foliations.
On the other hand, there are cases on which the extension is unique, e.g. étale groupoids, and still other cases on which the same 0-metric admits many extensions.
\[pair.groupoid\] Given ${\eta^{(0)}}$ any metric on the manifold $M$, then ${\eta^{(1)}}:={\eta^{(0)}}\oplus{\eta^{(0)}}$ yields a 1-metric on the pair groupoid $M\times M{\rightrightarrows}M$. However, not every extension of ${\eta^{(0)}}$ arises in this way. Consider, for instance, $M={\mathbb R}$, $a\neq 0$, then we have extensions ${\eta^{(1)}}$ of the canonical metric ${\eta^{(0)}}$ on ${\mathbb R}$ of the form: $$({\eta^{(1)}})^*=\begin{bmatrix}1 & a\\ a & 1\end{bmatrix}.$$ relative to the canonical basis for ${\mathbb R}^2$.
There are classes of groupoids where a 0-metric can always be extended to a 1-metric (although maybe not in unique way, like in the previous example). Recall that a [**foliation groupoid**]{} $G{\rightrightarrows}M$ is a Lie groupoid for which every isotropy group is discrete. Examples of foliation groupoids include étale groupoids and submersion groupoids. Alternative characterizations of foliations groupoids are: (i) a Lie groupoid which is Morita equivalent to an étale groupoid, or (ii) a Lie groupoid whose Lie algebroid has injective anchor map (see [@cm]).
\[prop:foliation.groupoids\] If $G{\rightrightarrows}M$ is a foliation groupoid and ${\eta^{(0)}}$ is a 0-metric, then there exists a 1-metric extending it.
Call $q$ the codimension of an orbit on $M$. Given on $G$, we can decompose $T_xM$ and $T_yM$ as orthogonal sums between the tangent spaces to the orbit $O$ and their complements. The differential of the anchor map $${{\rm d}}_g (s,t):T_gG\to T_yM\times T_x M= T_yO\oplus N_yO \oplus T_xO\oplus N_xO$$ is injective, hence we can identify $T_gG$ with its image, which are the vectors $(v_1,v_2,v_3,v_4)$ satisfying $\rho_g [v_4]=[v_2]$, where $\rho_g$ denotes the normal representation. By hypothesis, the norm of $v_2$ equals that of $v_4$. We define (the square of) the norm of $(v_1,v_2,v_3,v_4)$ as that of $(v_1,v_2,v_3,0)$. In other words, we preserve the metric along the orbit and correct it on the normal direction. It is immediate that the metric ${\eta^{(1)}}$ defined this way is both $s$-fibred and $t$-fibred, and that is sitting over $\eta$. Since $(s,t)\circ i = \tau \circ (s,t)$ it easily follows that our metric is invariant for the inversion, which completes the proof.
Note that the construction in this proof does not depend on any choices, so it leads to a completely determined metric ${\eta^{(1)}}$. Still, the example of the pair groupoid shows that there may be other extensions.
Given a regular foliation ${\mathcal F}$ on a manifold $M$, we have its [**monodromy groupoid**]{} $\operatorname{Mon}({\mathcal F}){\rightrightarrows}M$ and its [**holonomy groupoid**]{} $\operatorname{Hol}(F){\rightrightarrows}M$. The first consists of homotopy classes of paths within a leaf, and the second is a quotient of the previous one, where two paths are identified if they induce the same holonomy. These examples of foliation groupoids may fail to be Hausdorff.
If $M$ is endowed with a metric for which the leaves stay equidistant, namely every geodesic orthogonal to a leaf remains orthogonal to every leaf it meets, then ${\mathcal F}$ is a [**Riemannian foliation**]{}. For a Riemannian foliation holonomy and linear holonomy coincide, so the holonomy groupoid acts faithfully by isometries on the normal space to the foliation, and hence it is always Hausdorff.
A 0-metric on $\operatorname{Hol}(F){\rightrightarrows}M$ is the same as a metric on $M$ making the foliation ${\mathcal F}$ Riemannian. Proposition \[prop:foliation.groupoids\] recovers the main result of [@gghr], namely that the holonomy groupoid of a regular Riemannian foliation has a 1-metric extending the original one.
Metrics on the space of composable arrows: 2-metrics {#sec:2:metrics}
----------------------------------------------------
The notions of 0-metrics and 1-metrics, discussed in the previous paragraphs, do not take into account the groupoid multiplication $m:{G^{(2)}}\to G$. Hence, they ignore a fundamental ingredient of the concept of a Lie groupoid. This is fixed by considering metrics on the space of composable arrows ${G^{(2)}}$. In order to do that we need first to discuss certain symmetries of the manifold ${G^{(2)}}$, which are formal consequences of the combinatorics underlying its structure.
Each point in ${G^{(2)}}$ represents a pair of composable arrows, or equivalently, a commutative triangle of arrows. Hence we have a natural action of the symmetric group $S_3{\curvearrowright}{G^{(2)}}$, by permuting the objects of a given triangle. For example, in the diagram below, the permutation $(xy)(z)$ sends the point $(h,g)\in{G^{(2)}}$ to $(hg,g^{-1})$: $$\begin{matrix}
\xymatrix{ &y \ar@/_/[dl]_{h} & \\ z& & x \ar@/_/[ul]_{g} \ar@/^/[ll]^{h\cdot g}}
\end{matrix}
\qquad\mapsto \qquad
\begin{matrix}
\xymatrix{ &x \ar@/_/[dl]_{h\cdot g} \ar@/^/[dr]^{g}& \\ z& & y \ar@/^/[ll]^{h}}
\end{matrix}$$ We leave it to the reader to find the other transformations for this action in terms of the pair notation in ${G^{(2)}}$.
A precise formulation can be given as follows. Denote by $[2]$ the pair groupoid on a set with three objects $\{0,1,2\}$. Then $S_3$ naturally identifies with the groupoid automorphisms $[2]\to[2]$, and ${G^{(2)}}$ with the set of functors $[2]\to G$. Then the action $S_3{\curvearrowright}{G^{(2)}}$ is just given by pre-composition.
There are also three commuting left groupoid actions $\theta^1,\theta^2,\theta^3:G{\curvearrowright}{G^{(2)}}$, defined as follows. $$\theta^1(k)(g,h)=(kg,h)
\qquad
\theta^2(k)(g,h)=(gk^{-1},kh)
\qquad
\theta^3(k)(g,h)=(g,hk^{-1})$$ They are free and proper, and their orbits agree with the fibers of the maps $\pi_2,m,\pi_1:{G^{(2)}}\to G$, respectively, hence yielding three principal $G$-bundles. Observe that the action $S_3{\curvearrowright}{G^{(2)}}$ interchanges these principal actions $\theta^i$.
A [**2-metric**]{} on $G{\rightrightarrows}M$ is a metric ${\eta^{(2)}}$ on the manifold ${G^{(2)}}$ which is transversely invariant for the action $\theta^1:G{\curvearrowright}{G^{(2)}}$, $k\cdot (g,h)=(kg,h)$, and for which the group $S_3$ acts by isometries. The pair $(G{\rightrightarrows}M,{\eta^{(2)}})$ is called a **Riemannian groupoid**.
Unfortunately, the term *Riemannian groupoid* has been used in the literature with different meanings, as in [@gghr] and [@glick], which we have already discussed. Other approach that may seem natural is to consider metrics that are multiplicative tensors, in the sense of [@bc], but a simple computation shows that being positive definite implies that the groupoid must be 0-dimensional. We hope to convince the reader that a 2-metric is the right geometric structure that one should add to a Lie groupoid in order to call it *Riemannian*.
It should be clear that the definition of a 2-metric can be reformulated in many equivalent ways. For example, we can say:
- ${\eta^{(2)}}$ is transversely invariant for (any of) the actions $\theta^i:G{\curvearrowright}{G^{(2)}}$ and the group $S_3$ acts by isometries, or equivalently,
- ${\eta^{(2)}}$ is transverse to (any of) the maps $\pi_2,m,\pi_1:{G^{(2)}}\to G$ and the group $S_3$ acts by isometries.
In the next section we will construct many examples of 2-metrics. For now, let us relate this notion with those introduced before:
\[2-metric.induces.1-metric\] Let $G{\rightrightarrows}M$ be a Lie groupoid. A 2-metric ${\eta^{(2)}}$ on ${G^{(2)}}$ induces a 1-metric ${\eta^{(1)}}$ on ${G^{(1)}}$, and hence also a 0-metric on ${G^{(0)}}$.
Given a 2-metric on ${G^{(2)}}$, since ${\eta^{(2)}}$ is invariant under the action of $S_3$, which interchanges the maps $m,\pi_1,\pi_2$, it easily follows that the push-forward metrics $m_*{\eta^{(2)}},(\pi_1)_*{\eta^{(2)}},(\pi_2)_*{\eta^{(2)}}$ all agree (see Lemma \[triangle\]). Denote the resulting metric on the manifold $G$ by ${\eta^{(1)}}$.
We have $i m = m \phi$ for some $\phi\in S_3$, which implies that the inversion map $i$ preserves the metric ${\eta^{(1)}}$ (again Lemma \[triangle\]). It remains to show that ${\eta^{(1)}}$ is transverse to the source map $s:G\to M$, or equivalently, that ${\eta^{(2)}}$ is transverse to the map $s\pi_2=sm:{G^{(2)}}\to M$ (once more, Lemma \[triangle\]).
The principal actions $\theta^1,\theta^2:G{\curvearrowright}{G^{(2)}}$ can be combined to give a simultaneous action $$\theta^{12}:\tilde G{\curvearrowright}{G^{(2)}} \qquad (g',g)\cdot [h_1,h_2,h_3]=[gh_1,g'h_2,h_3]$$ where $\tilde G{\rightrightarrows}G$ is the so-called [*arrow groupoid*]{} of $G{\rightrightarrows}M$, whose objects are the arrows of $G$ and whose arrows are the commutative squares. The new action $\theta^{12}$ is again free and proper, and its orbits are exactly the fibers of $s\pi_2$. We will show that ${\eta^{(2)}}$ is transversely $\theta^{12}$-invariant (Proposition \[quotient.metrics\]).
By hypothesis, ${\eta^{(2)}}$ is both transversely $\theta^1$ and $\theta^2$ invariant. And the action groupoids of $\theta^1$ and $\theta^2$ embed into the action groupoid of $\theta^{12}$, via the inclusions $$(g',[h_1,h_2,h_3])\mapsto ((g',{{\rm id}}),[h_1,h_2,h_3]) \qquad (g,[h_1,h_2,h_3])\mapsto (({{\rm id}},g),[h_1,h_2,h_3]).$$ Then, we can use the naturality of the conormal representation with respect to these inclusions (cf. \[rmk:naturality.normal-representation\]), and the fact that every arrow $(g',g)$ in $\tilde G$ factors as $(g',{{\rm id}})({{\rm id}},g)$, to conclude the result.
Thus, given a 2-metric ${\eta^{(2)}}$ in a Lie groupoid $G{\rightrightarrows}M$, there are induced metrics ${\eta^{(1)}}$ on $G$ and ${\eta^{(0)}}$ on $M$ and the following five maps are Riemannian submersions. $$\xymatrix@1{ {G^{(2)}} \ar@<0.5pc>[r]^{\pi_1} \ar[r]|m \ar@<-0.5pc>[r]_{\pi_2} & G \ar@<0.25pc>[r]^s \ar@<-0.25pc>[r]_t & M}$$ Moreover, the metrics ${\eta^{(2)}}$ and ${\eta^{(1)}}$ are preserved by the natural actions of the symmetric groups $S_3{\curvearrowright}{G^{(2)}}$ and $S_2{\curvearrowright}{G^{(1)}}$.
Given $G{\rightrightarrows}M$ a Lie groupoid and ${\eta^{(1)}}$ a 1-metric, it is natural to ask if there exists a 2-metric ${\eta^{(2)}}$ inducing ${\eta^{(1)}}$. If that is the case we will say that ${\eta^{(2)}}$ is an [**extension**]{} of ${\eta^{(1)}}$. This [*extension problem*]{}, which consists on deciding whether there is an extension and if such an extension is unique, admit several different answers, depending on the groupoid. In the next section we will show with examples that one can have any of the following situations:
(i) there may exists a unique extension,
(ii) there may be more than one extensions, or
(iii) there may be no extension at all.
Examples of Riemannian groupoids
================================
First examples
--------------
For some Lie groupoids $G{\rightrightarrows}M$ the extension problem from 1-metrics to 2-metrics has exactly a unique solution. Hence for these Lie groupoids our notion of Riemannian groupoid is essentially the same as the one introduced in [@gghr].
In an étale Lie groupoid $G{\rightrightarrows}M$ the manifolds $M,G,{G^{(2)}}$ all have the same dimension. If ${\eta^{(1)}}$ is a 1-metric on it, then it induces a 0-metric ${\eta^{(0)}}$, which is the same as a metric on $M$ invariant under the effect of $G$ (see Lemma \[0-metrics.on.etal\]). For such a metric ${\eta^{(1)}}$ it is not hard to see that the pullback metrics $m^*{\eta^{(1)}},\pi_1^*{\eta^{(1)}},\pi_2^*{\eta^{(1)}}$ agree and give the only possible extension ${\eta^{(2)}}$.
More generally, as we will show in the forthcoming paper of this series [@fdh1], when working with foliation groupoids there exists exactly one extension of a $k$-metric ${\eta^{(k)}}$ to a $(k+1)$-metric ${\eta^{(k+1)}}$, for any $k\geq 1$.
Next we construct an example of a Lie group, viewed as a groupoid with a single object, on which there are many possible ways to extend the 1-metric.
Let $G{\rightrightarrows}M$ be the abelian group of vectors in the plane, so $G={\mathbb R}^2$ and $M=\ast$. The canonical metric ${\eta^{(1)}}$ on ${\mathbb R}^2$ is a 1-metric for our groupoid. Given a metric ${\eta^{(2)}}$ on ${G^{(2)}}={\mathbb R}^2\times{\mathbb R}^2$, let us write the matrix of its dual metric on the canonical basis at a point $a=(x_1,x_2,y_1,y_2)\in{G^{(2)}}={\mathbb R}^2\times{\mathbb R}^2$ by $$\begin{bmatrix}
A & B^t \\
B & C
\end{bmatrix}.$$ Using coordinates $(z_1,z_2)$ on ${G^{(1)}}={\mathbb R}^2$, the cotangent linear maps associated to $\pi_2,m,\pi_1:{G^{(2)}}\to{G^{(1)}}$ are given respectively by $$d(z_i)\overset{d_a\pi_2}\mapsto d(y_i) \qquad
d(z_i)\overset{d_a m}\mapsto d(x_i)+d(y_i) \qquad
d(z_i)\overset{d_a\pi_1}\mapsto d(x_i)$$ Thus, if ${\eta^{(2)}}$ is such that $\pi_2,m,\pi_1$ are Riemannian submersions, then the following equations should hold. $$C=I \qquad A+B+B^t+C=I \qquad A=I$$ This leaves one degree of freedom, namely $$B=
\begin{bmatrix}
-1/2 & -\beta \\
\beta & -1/2
\end{bmatrix}$$ for some smooth function $\beta:{\mathbb R}^4\to{\mathbb R}$. In order to have a 2-metric on ${\mathbb R}^2$ we need further to require that $S_3{\curvearrowright}{G^{(2)}}$ is an isometric action, which easily translates into the following functional equations on $\beta$: $$\begin{cases}
\beta(x_1,x_2,y_1,y_2) &= -\beta(-y_1,-y_2,-x_1,-x_2)\\
\beta(x_1,x_2,y_1,y_2) &= \beta(y_1,y_2,-x_1-y_1,-x_2-y_2).
\end{cases}$$ Lastly, different 2-metrics ${\eta^{(2)}}$ extending our original ${\eta^{(1)}}$ arise from instance from $\beta\equiv 0$ and $\beta\equiv x_1y_1(x_1+y_1)$. Although we do not consider here 3-metrics, it is easy to see that for most choices of $\beta$, the metric ${\eta^{(2)}}$ cannot be extended to a 3-metric.
Roughly speaking, a 1-metric on a Lie groupoid leaves some freedom along the isotropies, but it completely describes the Riemannian structure on the longitudinal and the transversal directions. This phenomenon can be made more precise in the context of stacks, namely any two different extensions of a 1-metric induce the same metric on the quotient stack. We will pursue this point of view in part III of this series of papers [@fdh2].
We now give an example of a Lie groupoid with a 1-metric which is not extendable to a 2-metric.
Let $G{\rightrightarrows}M$ be a Lie group bundle, i.e., a groupoid where the source and target maps agree. We can think of such a groupoid as a smooth family of Lie groups, parametrized by the base manifold $M$. The orbits are just points and the normal representations are trivial. A 1-metric on a Lie group bundle $G{\rightrightarrows}M$ is the same thing as a metric on $G$ transverse to the projection $s=t$ and invariant under the inversion map, so it is easy to construct such 1-metrics on any Lie group bundle (see Lemma \[every.submersion.is.Riemannian\] and Proposition \[cotangent.principle\]). However, such a 1-metric may not be extendable, and moreover, there may not exists a 2-metric at all.
For a concrete example, let $(G{\rightrightarrows}M)=({\mathbb R}^3{\rightrightarrows}{\mathbb R})$ be the Lie group bundle settled by $$s(\lambda,x,\epsilon)=t(\lambda,x,\epsilon)=\epsilon
\qquad
m((\lambda,x,\epsilon),(\lambda',x',\epsilon))=(\lambda+\lambda',x+e^{\lambda \epsilon}x',\epsilon).$$ This can be seen as a 1-parameter family of Lie group structures ${\mathbb R}^2_\epsilon=({\mathbb R}^2,\cdot_\epsilon)$, where $(\lambda,x)\cdot_\epsilon(\lambda',x')=(\lambda+\lambda',x+e^{\lambda \epsilon}x')$. In other words, the group ${\mathbb R}^2_\epsilon$ is the semi-direct product ${\mathbb R}\ltimes{\mathbb R}$ under the action $\phi:{\mathbb R}{\curvearrowright}{\mathbb R}$, $\phi_\lambda(x)=e^{\lambda \epsilon}x$.
As any Lie group bundle, $G$ admits a 1-metric, but it turns out that $G$ cannot be endowed with a 2-metric. We give now a direct proof of this, and we will provide a more conceptual explanation later.
Assume that ${\eta^{(2)}}$ is a 2-metric, and endow $G,M$ with the induced metrics. Write $\pi:G\to M$ and $\pi':{G^{(2)}}\to M$ for the time projections, which are Riemannian submersions. For each $(\lambda,x)\in G_0$ let $(a(\lambda,x),b(\lambda,x),1)\in T_{(\lambda,x,0)}G$ be the orthogonal lift of $\partial_\epsilon\in T_0M$ along $\pi$, and let $w\in T_{\lambda,x,\lambda',x',0}{G^{(2)}}$ be the orthogonal lift of $\partial_\epsilon\in T_0M$ along $\pi'$.
From the equations $\pi\pi_1=\pi'$ and $\pi\pi_2=\pi'$ it follows that $$w=(a(\lambda,x),b(\lambda,x),a(\lambda',x'),b(\lambda',x'),1).$$ Now, the differential of the multiplication on the canonical basis at a central point has the form $$d m_{(\lambda,x,\lambda',x',0)}=\begin{bmatrix}
1 & 0 & 1 & 0 & 0 \\
0 & 1 & 0 & 1 & x'\lambda\\
0 & 0 & 0 & 0 & 1
\end{bmatrix}$$ so from the equation $\pi m=\pi'$ the next conditions on $w$ arise: $$\begin{cases}
\begin{matrix}
a(\lambda,x)+a(\lambda',x') & = & a(\lambda+\lambda',x+x')\\
b(\lambda,x)+b(\lambda',x')+\lambda x' & = & b(\lambda+\lambda', x+x').
\end{matrix}
\end{cases}$$ These equations should hold for any choice of $\lambda,x,\lambda',x'$, but it is easy to see that there are no such functions $a,b$. Note for instance that the right hand sides are symmetric and the left hand sides are not.
In fact, we will show later in Example \[ex:lie.alg.bundle\], as a consequence of the Linearization Theorem, that if a bundle of Lie groups admits a 2-metric, then its associated bundle of Lie algebras is locally trivial (see also Example \[ex:local:trivial:bundles\] for a partial converse). In the example above, the Lie algebra of ${\mathbb R}^2_\epsilon$ has generators $a,b$ and bracket $[a,b]_\epsilon=\epsilon b$, so it is a non-trivial deformation of the abelian Lie algebra of dimension 2, hence a 2-metric cannot exist.
The gauge trick
---------------
Every Lie group admits a 2-metric, hence it can be regarded as a Riemannian Lie groupoid. However, since 2-metrics are concerned with the transverse geometry of a Lie groupoid, this is not surprising. There is however one thing to be learned: the same recipe that allow us one to construct 2-metrics on Lie groups also works in several other situations.
\[ex:Lie:group\] Let $G$ be a Lie group, and let $\eta$ be a right invariant metric on $G$. If $\<,\>$ denontes the value of $\eta$ at the origin, then for any $g\in G$ and any $\phi,\psi\in T_g^*G$ we have $$\eta^*_g(\alpha,\beta)=\<R_g\alpha,R_g\beta\> \qquad \alpha,\beta\in T_g^*G$$ The product metric $\eta\times\eta\times\eta$ on $G\times G\times G$ is a 2-metric on the pair groupod $G\times G{\rightrightarrows}G$, and moreover, it is transversely invariant under the action of $G$ by right multiplication. Hence, it induces a 2-metric on the quotient groupoid, which is actually isomorphic to the group $G$ (cf. Lemma \[quotient.metrics\] and Proposition \[quotient.metrics:maps\]).
Note that the induced metric ${\eta^{(1)}}$ does not agree in general with the original metric $\eta$. A tedious computation shows that the resulting metrics ${\eta^{(2)}}$ and ${\eta^{(1)}}$ are given by: $$\begin{aligned}
({\eta^{(2)}})^*(({\alpha}_1,{\alpha}_2),({\beta}_1,{\beta}_2))_{(g,h)} =&
\langle R_{g}^*{\alpha}_1, R_{g}^*{\beta}_1\rangle + \langle R_{h}^*{\alpha}_2, R_{h}^*{\beta}_2\rangle -\langle L_{g}^*{\alpha}_1, R_{h}^*{\beta}_2\rangle -\\
&-\langle R_{h}^*{\alpha}_2, L_{g}^*{\beta}_1\rangle +\langle L_{g}^*{\alpha}_1, L_{g}^*{\beta}_1\rangle + \langle L_{h}^*{\alpha}_2, L_{h}^*{\beta}_2\rangle. \\
({\eta^{(1)}})^*({\alpha},{\beta})_{g}=&
\langle R_{g}^*{\alpha}, R_{g}^*{\beta}\rangle + \langle L_{g}^*{\alpha}, L_{g}^*{\beta}\rangle. \end{aligned}$$
\[ex:local:trivial:bundles\] We can use the existence of 2-metrics on Lie groups to show that any locally trivial bundle of Lie groups admits a 2-metric. First, for a trivial bundle $M\times G\to M$ we can construct a 2-metric on the space of composable arrows $M\times G\times G\to M$: we chose a 2-metric $\eta^G$ in the Lie group $G$ and a metric $\eta^M$ in $M$, and we form the product metric $\eta^M\oplus\eta^G$. Then, for a locally trivial bundle $G{\rightrightarrows}M$, we cover $M$ by trivializing open sets $U_{\alpha}$ and 2-metrics ${\eta^{(2)}}_{\alpha}$ on the restrictions $G|_{U_{\alpha}}$. If $\{\rho_{\alpha}\}$ is a partition of unity subordinated to the cover, then we obtain a 2-metric ${\eta^{(2)}}$ in $G{\rightrightarrows}M$ by setting: $${\eta^{(2)}}:=\left(\sum_{{\alpha}}\rho_{\alpha}({\eta^{(2)}}_{\alpha})^*\right)^*.$$
The 2-metric that we have constructed on a Lie group was obtained as the quotient metric of a suitable one in $G^3$. This can be generalized for any Lie groupoid, now considering the manifold $G^{[3]}\subset G^3$ of triples of arrows with the same source, and the map $\pi^{(2)}:G^{[3]}\to{G^{(2)}}$ given by $\pi^{(2)}(h_1,h_2,h_3)= (h_1h_2^{-1},h_2h_3^{-1})$. The fibers of $\pi^{(2)}$ coincide with the orbits of the right-multiplication action, $$\begin{matrix}
G^{[3]}{\curvearrowleft}G\\
(h_1,h_2,h_3)\cdot k = (h_1k,h_2k,h_3k).
\end{matrix}
\qquad \qquad
\begin{matrix}
\xymatrix@R=10pt{\bullet & & \\ \bullet & \bullet \ar[lu]_{h_1} \ar[l]|{h_2} \ar[ld]^{h_3} & \bullet \ar[l]^k \\ \bullet }\end{matrix}$$ and this action is free and proper, hence defining a principal $G$-bundle. The general estrategy will be to define a nice enough metric on $G^{[3]}$ in a way such that it can be pushed forward along $\pi^{(2)}$, and that the resulting metric is a 2-metric.
Notice that the group $S_3$ acts on the manifold $G^{[3]}$ by permuting its coordinates, and there are three left groupoid actions $G{\curvearrowright}G^{[3]}$, each consisting in left multiplication on a given coordinate. $$\begin{matrix}
\xymatrix@R=10pt{\bullet & \bullet \ar[l]^k & \\ & \bullet & \bullet \ar[lu]_{h_1} \ar[l]|{h_2} \ar[ld]^{h_3} \\ & \bullet }\end{matrix}
\qquad\qquad
\begin{matrix}
\xymatrix@R=10pt{ & \bullet & \\\bullet & \bullet \ar[l]|k & \bullet \ar[lu]_{h_1} \ar[l]|{h_2} \ar[ld]^{h_3} \\ & \bullet }\end{matrix}
\qquad\qquad
\begin{matrix}
\xymatrix@R=10pt{ & \bullet & \\ & \bullet & \bullet \ar[lu]_{h_1} \ar[l]|{h_2} \ar[ld]^{h_3} \\\bullet & \bullet \ar[l]_k &}\end{matrix}$$ These four actions commute with the above right action, and cover the actions $S_3{\curvearrowright}{G^{(2)}}$, $G{\curvearrowright}{G^{(2)}}$ used when defining 2-metrics.
The map $\pi^{(2)}:G^{[3]}\to{G^{(2)}}$ plays a key role for the groupoid. In fact, one can generalize this construction to any $k\geq 0$ in an obvious way, obtaining principal $G$-bundles $\pi^{(k)}:G^{[k+1]}\to{G^{(k)}}$, which are $S_{k}$-invariant. They together form the simplicial model for the universal principal $G$-bundle $EG\to BG$. We will come back to this when studying $n$-metrics on Lie groupoids in [@fdh1].
Emmulating what we have done for Lie groups, we can construct 2-metrics for more general Lie groupoids. The following is a first generalization.
\[gauge.trick.1\] Let $G$ be a Lie group acting on the right on the manifolds $P,N$ and let $q:P\to N$ be an equivariant submersion. Assume that $G{\curvearrowright}P$ is free and $G{\curvearrowright}N$ is proper. Then the [**gauge groupoid**]{} $(P\times_N P)/G{\rightrightarrows}P/G$ admits a 2-metric.
We repeatedly make use of the fact that on a manifold with a proper action an equivariant metric can be constructed by means of a classic averaging argument. Thus, by setting a preliminary equivariant metric on $P$, which exists for this action is also proper, we can construct an Ehresmann connection $H$ for $q:P\to N$ which is $G$-invariant: take just the orthogonal to the fibers. This way we have $TP=H\oplus F$.
Now, we will modify the preliminary metric by pullbacking to $H$ an equivariant metric on $N$. The resulting metric $\eta^P$ is both $q$-fibred and $G$-invariant
The induced metric ${\eta^{(2)}}$ on $P\times_N P\times_N P$ make the submersion groupoid $P\times_N P{\rightrightarrows}P$ a Riemannian groupoid, and since ${\eta^{(2)}}$ is $G$-invariant, it induces a quotient metric which yields the desired structure on the gauge groupoid (cf. Lemma \[quotient.metrics\] and Proposition \[quotient.metrics:maps\]).
Let us derive now some immediate corollaries.
Every transitive groupoid can be obtained as the gauge groupoid over some free proper action $G{\curvearrowright}P$, or equivalently, over some principal $G$-bundle $P\to M$. Hence, by letting $N=\ast$ in the previous proposition, we recover a recipe to construct 2-metrics on any transitive groupoid.
\[ex:proper-actions\] If $G{\curvearrowright}M$ is a Lie group acting over a manifold, then it is easy to see that the corresponding action groupoid $G\ltimes M{\rightrightarrows}M$ is isomorphic to the gauge construction over the projection $G\times M\to M$, where $G$ acts over the product $G\times M$ by $g\cdot(k,x)=(kg^{-1},gx)$. Thus, by the proposition, we may conclude that every Lie groupoid arising from a proper Lie group action admits a 2-metric. Moreover, we can take as the auxiliary metric $\eta^P$ appearing in \[gauge.trick.1\] the product of a right invariant metric on $G$ and a $G$-equivariant metric on $M$.
We will refer to the construction behind Proposition \[gauge.trick.1\] as the [*gauge trick*]{}.
2-metrics on proper groupoids
-----------------------------
Recall that a groupoid $G{\rightrightarrows}M$ is proper if the anchor map $\rho:G\to M\times M$, $g\mapsto (t(g),s(g))$ is a proper map. In this section we will prove that every Hausdorff proper groupoid admits a 2-metric, which is a groupoid version of the well-known fact that every manifold admits a metric, or that every orbifold admits a metric. We will construct such a 2-metric by adapting the *gauge trick*, introduced in last section. The delicate point here is that a groupoid action does not lift to an action on the tangent/cotangent bundle. There is however a tangent/cotangent [*quasi-action*]{}, which will allow us to apply averaging methods to produce 2-metrics. We have collected the relevant material about quasi-actions, tangent lifts and averaging in the Appendix \[appendix:averaging\].
Let $G{\rightrightarrows}M$ be a Lie groupoid and $\sigma$ a connection on it. We consider an action $G{\curvearrowright}E$ and its cotangent lift $(G\ltimes E)\tilde{\curvearrowright}T^*E$. In order to simplify the notation we will write just $ge=\theta_g(e)$, $gv=T_\sigma\theta_{(g,e)}(v)$ and so on.
Given a metric $\eta$ on $E$, we can view it as a section of the second symmetric power, say $\eta\in \Gamma(E,S^2(T^*E))$, and analogously $\eta^*\in\Gamma(E,S^2(TE))$. Then the metric $\eta$ is transversely invariant if and only if for each orbit $O\subset E$ the section $\eta^*|_{TO^\circ\times TO^\circ}\in\Gamma(O,S^2(TE/TO))$ is an invariant section for the corresponding lifted action (cf. Proposition \[tan.lift.nor.representation\]). Now suppose that $G{\rightrightarrows}M$ is proper, so we can fix $\mu$ a Haar density on it, and consider its associated averaging operators (cf. Definition \[def:aver.oper\]).
Let $\theta:G{\curvearrowright}E$ be a groupoid action, and $\eta$ a metric on $E$. Its [**cotangent average**]{} $\tilde\eta\in \Gamma(E,S^2(T^*E))$ is defined by averaging its dual, say $$(\tilde\eta)^*_e(\alpha,\beta):=I_\theta(\eta^*)_e(\alpha,\beta)=
\int_{G(-,x)} \eta^*_{ge}(g\alpha,g\beta) \mu^x(g),$$ where $I_\theta$ is the operator on $\Gamma(E,S^2(TE))$, $x=q(e)$, , and $\alpha,\beta\in T_e^*E$.
The following proposition plays a key role in the paper:
\[cotangent.average\] Given $\eta$ a metric on $E$, then:
(i) Its cotangent average $\tilde\eta$ is a transversely $G$-invariant metric on $E$.
(ii) If $\eta$ is already transversely $G$-invariant, then $\eta$ and $\tilde\eta$ agree in the directions normal to the orbits.
It is easy to see that the section $\tilde\eta$ is positive definite, hence a metric. To see that it is transversely $G$-invariant, let $O\subset E$ be an orbit, and consider the following vector bundle map: $$S^2(TE|_O)\to S^2(TE|_O/TO) \qquad \eta^* \mapsto \eta^*|_{TO^\circ\times TO^\circ}.$$ This is surjective and equivariant, where we endow $S^2(TE|_O)$ with the quasi-action induced by the tangent lift, and $S^2(TE|_O/TO)$ with the conormal representation (cf. Proposition \[tan.lift.nor.representation\]). By Proposition \[prop:aver.oper\](iv) the metric $\tilde\eta|_{TO^\circ\times TO^\circ}$ agrees with the averaging of the restriction $\eta^*|_{TO^\circ\times TO^\circ}$ with respect to the conormal representation. This yields an invariant section by \[prop:aver.oper\](i), showing that $\tilde\eta$ is transversely $G$-invariant. The last statement readily follows from \[prop:aver.oper\](ii).
Notice that if the action $\theta:G{\curvearrowright}E$ is free and proper, hence defining a principal $G$-bundle, then the push-forward metric $q_*\tilde\eta$ of the cotangent average along $q:E\to E/G$ is a well-defined metric in the quotient, and if $\eta$ is already transversely $G$-invariant, then we have $q_*\tilde\eta=q_*\eta$.
Let us show now that the cotangent average behaves well with respect to Riemannian equivariant submersions.
\[prop:cot.ave.&.submersions\] Let $p:(E,\eta^E)\to (B,\eta^B)$ be a Riemannian submersion which is equivariant for actions $\theta^E:G{\curvearrowright}E$ and $\theta^B:G{\curvearrowright}B$. If $\tilde\eta^E,\tilde\eta^B$ denote the cotangent averages of the metrics, then $p:(E,\tilde\eta^E)\to(B,\tilde\eta^B)$ is also a Riemannian submersion.
Given $e\in E$ and $\alpha,\beta\in T_b^*B$, where $b=p(e)$, by using first that $p$ is Riemannian and then that is equivariant, we get the following chain of equalities, $$\begin{aligned}
(\tilde\eta^{B}_b)^*(\alpha,\beta)&=\int_{G(-,x)} (\eta_{gb}^{B})^*(g\alpha,g\beta)\mu^x(g) \\
&=\int_{G(-,x)} (\eta_{ge}^{E})^*(({{\rm d}}_{ge} p)^*g\alpha,({{\rm d}}_{ge} p)^*g\beta)\mu^x(g)\\
&=\int_{G(-,x)} (\eta_{ge}^{E})^*(g({{\rm d}}_e p)^*\alpha,g({{\rm d}}_e p)^*\beta)\mu^x(g)\\
&=(\tilde\eta_e^{E})^*(({{\rm d}}_e p)^*\alpha,({{\rm d}}_e p)^*\beta),\end{aligned}$$ from which we conclude that $p$ is also Riemannian for the averaged metrics.
Now we have collected all the preliminaries needed to establish our first fundamental theorem.
\[thm:main:1\] Every Hausdorff proper groupoid $G{\rightrightarrows}M$ admits a 2-metric ${\eta^{(2)}}$.
Endow the manifold $G$ with a Riemannian structure $\eta^{[1]}$ transverse to the source map $G\to M$. For each $k=1,2,3,\dots$ endow the $k$-fold pullback along the source map $G^{[k]}$ with the corresponding pullback metric $\eta^{[k]}$ (cf. \[pullback metric\]). Then every face map of the submersion groupoid arising from the source is a Riemannian submersion. $$\xymatrix@1{
\cdots \,
G^{[3]} \ar@<0.5pc>[r]^{\pi_1} \ar[r]|{m} \ar@<-0.5pc>[r]_{\pi_2} &
G^{[2]} \ar@<0.25pc>[r]^{s} \ar@<-0.25pc>[r]_{t} &
G^{[1]}}$$ Moreover, each of these Riemannian submersions is equivariant for the right action $G^{[k]}{\curvearrowleft}G$. Thus, after replacing each metric $\eta^{[k]}$ by its cotangent average $\tilde\eta^{[k]}$ we still have that every face map is a Riemannian submersion (cf. Proposition \[prop:cot.ave.&.submersions\]). But now we can push-forward each metric $\eta^{[k+1]}$ through the quotient map $G^{[k+1]}\to{G^{(k)}}$ and obtain a (fully extendable) $k$-metric on $G{\rightrightarrows}M$. In particular, for $k=2$, we obtain the desired 2-metric.
Notice that the proof shows that the 2-metric is a [*simplicial metric*]{}, i.e., it extends to an $n$-metric, for all $n\in{\mathbb N}$. We will come back to this notion of metrics in the forthcoming paper [@fdh1].
Linearization of Riemannian Groupoids {#section.linearization}
=====================================
The linearization problem {#subsec:linearization}
-------------------------
Let $G{\rightrightarrows}M$ be a Lie groupoid, and let $S\subset M$ be an embedded saturated submanifold of codimension $q$, i.e., $S$ is a submanifold which is a union of orbits of $G$. We denote by $G_S\subset G$ the set of arrows whose source and target belong to $S$: $$G_S=s^{-1}(S)=t^{-1}(S).$$ Note that $G_S$ is an embedded submanifold of the same codimension $q$ as $S$. We denote by $G_S^{(k)}\subset G^{(k)}$ the set of $k$-tuples of composable arrows in $G_S$, which is again an embedded submanifold of codimension $q$.
For any saturated submanifold of $G{\rightrightarrows}M$, there is a [**local linear model**]{} for $G$ around $S$. It can be defined the groupoid-theoretic normal bundle: $$\xymatrix{ \nu(G_S) \ar@<0.25pc>[r] \ar@<-0.25pc>[r] \ar[d] & \nu(S) \ar[d] \\
G_S \ar@<0.25pc>[r] \ar@<-0.25pc>[r] & S}$$ Its objects and arrows are given by the total spaces of the normal bundles $\nu(S)=T_SM/TS$ and $\nu(G_S)=T_{G_S}G/TG_S$, respectively. The structural maps $s,t,m,u,i$ are induced by the total differential of those of $G{\rightrightarrows}M$. Notice that this local linear model depends only on the linear infinitesimal data around $S$.
\[rmk:core-zero\] The structure maps of the local linear model $\nu(G_S){\rightrightarrows}\nu(S)$ are vector bundle maps and fiberwise isomorphisms. One consequence is that $\nu(G_S)$ identifies with the pullback $G_S \times_S \nu(S)$ along the source map, yielding a representation $$G_S \times_S \nu(S)\cong \nu(G_S) {\xrightarrow}t \nu(S).$$ When $S$ is an orbit this recovers the normal representation recalled in Section \[sub:inv.metrics\]. Another consequence, that will play a key role later, is that the pairs of composable arrows of the local linear model $\nu(G_S)^{(2)}$ canonically identifies with the normal bundle , for it can be seen as the top square of a cartesian cube.
The linearization problem consists on establishing an isomorphism between the local model and the original groupoid on suitable neighborhoods. There are however several possibilities for this choice of neighborhoods, as we now explain.
Let $G{\rightrightarrows}M$ be a Lie groupoid and let $S\subset M$ be a saturated submanifold. A [**groupoid neighborhood**]{} of $G_S{\rightrightarrows}S$ in $G{\rightrightarrows}M$ is a pair of open subsets $S\subset U\subset M$ and $G_S\subset \tilde{U} \subset G$ such that $\tilde{U}{\rightrightarrows}U$ is a subgroupoid of $G{\rightrightarrows}M$. Such a groupoid neighborhood is called [**full**]{} if: $$\tilde{U}=G_U=s^{-1}(U)\cap t^{-1}(U).$$
We note that for proper Lie groupoids we have:
\[lem:full:neighborhood\] For a proper Lie groupoid, a groupoid neighborhood always contains a full neighborhood.
Let $G{\rightrightarrows}M$ be a proper groupoid, let $S\subset M$ be a saturated submanifold, and let $\tilde{U}{\rightrightarrows}U$ be a groupoid neighborhood of $G_S{\rightrightarrows}S$.
We start by observing that each $x\in S$ has a neighborhood $V_x$ in $M$ such that $G_{V_x}\subset \tilde{U}$. This is because $G_x$ is the fiber over $(x,x)$ of the proper map $\rho=(t,s):G\to M\times M$, and therefore, the open $\tilde{U}$ that contains $G_x$ must also contain an open tube $(t,s)^{-1}(V_x\times V_x)=G_{V_x}$. Since $S$ is second countable, we can find a countable family $\{V_n\}$ with $G_{V_n}\subset \tilde{U}$ and such that $S\subset \bigcup_{n=1}^\infty V_n$. We can further assume that $\overline{V}_n$ is compact for all $n$.
The naive tentative would be to take the full neighborhood given by $\bigcup_{n=1}^\infty V_n$, but as a matter of fact, there may be arrows from some $V_i$ to some $V_j$ not contained in $\tilde U$. To solve this we shrink each $V_n$ by defining $V'_n=V_n-C_n$, where $$\begin{aligned}
C_n:&=\pi_1((M\times \cup_{i=1}^{n-1}\overline{V}_i)\cap \rho(G-\tilde{U}))\\
&=\{y\in M: \exists g\in G-\tilde{U}\text{ with }t(g)=y\text{ and }s(g)\in \cup_{i=1}^{n-1}\overline{V}_i\}.\end{aligned}$$ Since both $\rho:G\to M\times M$ and the projection $\pi_1:M\times \cup_{i=1}^{n-1}\overline{V}_i\to M$ are closed maps, we have that $C_n$ is closed, and hence $V'_n$ is open.
Since $G_S\subset \tilde{U}$, it follows that $S\cap C_n=\emptyset$, then $S\cap V_n=S\cap V'_n$ and the collection $\{V'_n\}$ still covers $S$. We finally define $V:=\bigcup_{n=1}^\infty V'_n$, it is easy to check that $G_V\subset \tilde{U}$, and the lemma follows.
However, for a general Lie groupoid, a groupoid neighborhood may fail to contain a full neighborhood: for example, for the Lie groupoid associated with the flow of a vector field, if $S=\{x_0\}$ is a non-degenerate sink of the vector field, then small enough groupoid neighborhoods do not contain any full neighborhoods.
Using the notion of groupoid neighborhood we can formulate the various versions of the linearization problem:
Let $G{\rightrightarrows}M$ be a Lie groupoid and let $S\subset M$ be a saturated submanifold. Then we say that:
(a) $G$ is [**weakly linearizable**]{} at $S$ if there are groupoid neighborhoods $\tilde{U}{\rightrightarrows}U$ of $G_S{\rightrightarrows}S$ in $G{\rightrightarrows}M$ and $\tilde{V}{\rightrightarrows}V$ of $G_S\to S$ in the local model $\nu(G_S){\rightrightarrows}\nu(S)$, and an isomorphism of Lie groupoids: $$(\tilde{U}{\rightrightarrows}U)\overset\phi\cong (\tilde{V}{\rightrightarrows}V)$$ which is the identity on $G_S$.
(b) $G$ is [**linearizable**]{} at $S$ if both $\tilde{U}$ and $\tilde{V}$ can be chosen to be full neighborhoods, so that there is an isomorphism of Lie groupoids $$(G_U{\rightrightarrows}U)\overset\phi\cong (\nu(G_S)_V{\rightrightarrows}V),$$ which is the identity on $G_S$.
(c) $G$ is called [**invariantly linearizable**]{} at $S$ if it is linearizable and both $U$ and $V$ can be taken to be saturated.
The linearization problem has been intensively studied in the last decade in the case of *proper* Lie groupoids. See, e.g., [@cs] and references therein for the most updated account of the linearization problem in the proper case.
Let ${\mathbb R}\times {\mathbb Z}{\rightrightarrows}{\mathbb R}$ be the trivial bundle of Lie groups with fiber ${\mathbb Z}$ and let $G{\rightrightarrows}{\mathbb R}$ be the subgroupoid with fibers $G_t={\mathbb Z}$ for $t\neq 0$ and $G_0=\ast$, and let $S=\{0\}$. It is easy to see that $G{\rightrightarrows}{\mathbb R}$ is weakly linearizable at $S$ but it is not linearizable at $S$.
If $G=M\times_N M{\rightrightarrows}M$ is the Lie groupoid arising from a submersion $p:M\to N$ and $S\subset M$ is the preimage of any embedded submanifold, then we will see below that $G{\rightrightarrows}M$ is always linearizable around $S$. However, it is invariantly linearizable if and only if the submersion is locally trivial at the points of $S$.
A Lie groupoid arising from a proper action of a Lie group is invariantly linearizable around an orbit. This is in fact a way to rephrase the Tube Theorem, see e.g. [@dk].
Exponential neighborhoods
-------------------------
Let $(M,\eta)$ be a Riemannian manifold. Denote by $\mathcal E_M\subset TM$ the domain of the [**exponential map**]{}, that is, the open set consisting of tangent vectors $v\in TM$ for which the corresponding geodesic $\gamma_v(t)$ is defined for all $0\le t \le 1$. Then the exponential map $$\exp:\mathcal E_M\to M \qquad \exp(v)=\gamma_v(1)$$ is smooth, it is the identity over $M$ (viewed as the zero section), and its differential at points of $M$ has the form: $$T_{0_x}(TM)\cong T_xM\times T_xM\to TM,\quad (v,w)\mapsto v+w.$$
Let $S\subset (M,\eta)$ be an embedded submanifold. We identify the abstract normal bundle $\nu(S)$ with the orthogonal bundle $(TS)^\bot$. An open subset $S\subset U\subset \mathcal E_M\cap \nu(S)$ is called an [**admissible neighborhood**]{} if the exponential map is injective and étale over $U$, hence an open embedding. Then we call the image $\exp(U)\subset M$ an [**exponential neighborhood**]{} of $S$. Of course, this is the standard way tubular neighborhoods are constructed for $S\subset M$ out of $\eta$. When $S$ consist of a single point this construction yields normal coordinates around the point (see e.g. [@lee]).
The existence of admissible opens is well-known:
Any submanifold $S$ of a Riemannian manifold $(M,\eta)$ has an admissible open neighborhood $S\subset U\subset \mathcal E_M\cap \nu(S)$.
We are interested in the construction of exponential neighborhoods related to a Riemannian submersion $p:(E,\eta^{E})\to(B,\eta^B)$. Let us denote by $H\subset TE$ the horizontal vector bundle, consisting of vectors orthogonal to the fibers. A curve $\tilde\gamma$ on $E$ is [**horizontal**]{} if its tangent vectors $\dot{\tilde{\gamma}}(t)$ belong to $H$. Since $H$ is an example of an Ehresmann connection, one has local lifting of curves, i.e., for any $e\in E$ and any curve $\gamma$ in $B$ with $p(e)=\gamma(0)$ there exists a unique horizontal curve $\tilde\gamma$ such that $\tilde\gamma(0)=e$ and $\gamma(t)=p(\tilde\gamma(t))$.
Since $p:(E,\eta^{E})\to(B,\eta^B)$ is a Riemannian submersion, a horizontal curve is a geodesic if and only if its projection is a geodesic. In particular, if a geodesic on $E$ is normal to a fiber, then it is normal to every fiber it meets.
\[prop:commuting.exponential\] Let $p:(E,\eta^{E})\to(B,\eta^B)$ be a Riemannian submersion. If $S\subset B$ is an embedded submanifold and $\tilde S=p^{-1}(S)$, then for any open subsets $\tilde S\subset \tilde U\subset \mathcal E_E\cap \nu(\tilde S)$ and $S\subset U\subset\mathcal E_B\cap \nu(S)$ such that ${{\rm d}}p(\tilde U)\subset U$, the following square commutes: $$\xymatrix{
\tilde U \ar[r]^{\exp} \ar[d]_{{{\rm d}}p} & E \ar[d]^p\\
U \ar[r]^{\exp} & B
}$$ Moreover, if $U$ is admissible then $\tilde U$ is also admissible.
Any vector normal to $\tilde S$ is, in particular, normal to the corresponding fiber. Hence, it gives rise to a horizontal geodesic whose projection is also a geodesic. Therefore, the diagram in the statement of the proposition commutes.
Assume now that $U$ is admissible. To show that $\tilde U$ is also admissible we need to show that $\exp:\tilde U\to E$ is both injective and étale:
- To prove that $\exp:\tilde U\to E$ is injective, lets assume that $\exp(v)=\exp(v')$ and denote by $\gamma$ and $\gamma'$ the geodesics arising from $v$ and $v'$, respectively. Their projections $p(\gamma)$ and $p(\gamma')$ are geodesics in $U$ arising from ${{\rm d}}p(v))$ and ${{\rm d}}p(v')$. Since $\exp$ is injective over $U$ we conclude that ${{\rm d}}p(v)={{\rm d}}p(v')$. Thus we see that $\gamma$ and $\gamma'$ are two horizontal lifts of the same curve which end at the same point. By the uniqueness of lifting we conclude that they are the same curve. This establishes injectivity.
- To prove that $\exp:\tilde U\to E$ is étale, let $v\in\tilde U$ and set $e=\exp(v)$. The above commutative diagram gives the following map of short exact sequences: $$\xymatrix{
0 \ar[r] & \ker{{\rm d}}_v ({{\rm d}}p)|_{\tilde U} \ar[r] \ar[d] & T_{v}\tilde U \ar[d]_{{{\rm d}}_v\exp} \ar[r] & T_{{{\rm d}}p(v)}U \ar[d] \ar[r] & 0\\
0 \ar[r] & \ker{{\rm d}}_e p \ar[r] & T_e E \ar[r] & T_{p(e)} B \ar[r] & 0
}$$ The last vertical arrow is an isomorphism because $U$ is admissible. The first vertical arrow identifies the tangent spaces to the fibers of $p$ and $dp$. It follows that the middle arrow is also an isomorphism, so $\exp:\tilde U\to E$ is a local diffeomorphism, as claimed.
\[Ehresmann.Theorem\] The previous proposition establishes the existence of exponential neighborhoods adapted to a Riemannian submersion. When $S=\{y\}$ consists of a single point, it follows that a submersion looks like a projection around a fiber, since in this case the normal bundle $\nu(\tilde S)\cong p^*T_yB\cong p^{-1}(y)\times {\mathbb R}^n$ is trivial. This statement can be thought of as a structure theorem for a submersion. Ehresmann’s Theorem, asserting that a proper submersion is locally trivial, can be easily obtained from this statement.
The Linearization Theorem for Riemannian groupoids
--------------------------------------------------
We have now everything in place to state and prove one of our main results.
\[thm:main:2\] Let $G{\rightrightarrows}M$ be a Lie groupoid endowed with a 2-metric $\eta^{(2)}$, and let $S\subset M$ be a saturated embedded submanifold. Then the exponential map defines a weak linearization of $G$ around $S$.
Let $S\subset V\subset \nu(S)$ be an admissible neighborhood for ${\eta^{(0)}}$, and let $$\tilde V = ({{\rm d}}s)^{-1}(V)\cap ({{\rm d}}t)^{-1}(V) \cap \mathcal E_G \cap \nu(G_S)$$ be the arrows of $\nu(G_S){\rightrightarrows}\nu(S)$, with source and target in $V$, belonging to the domain of the exponential map $\exp^{(1)}$ of the metric ${\eta^{(1)}}$.
Let us show that $\tilde V{\rightrightarrows}V$ is a groupoid neighborhood of $G_S{\rightrightarrows}S$ in $\nu(G_S){\rightrightarrows}\nu(S)$. The compatibility with the structural maps $s,t,i,u$ follows from Proposition \[metric.on.the.units\]. To see that $\tilde V$ is closed under multiplication, let $({\tilde w},{\tilde v})\in\tilde V\times_V\tilde V\subset \nu(G_S)^{(2)}$, and identify $\nu(G_S)^{(2)}\cong \nu(G_S^{(2)})$ in the canonical way (cf. Remark \[rmk:core-zero\]).
We claim that the geodesic $\gamma_{(\tilde w,\tilde v)}$ in ${G^{(2)}}$ with initial condition $({\tilde w},{\tilde v})$ is the curve $(\gamma_{\tilde w}(t),\gamma_{\tilde v}(t))$. In fact, note that $\gamma_{(\tilde w,\tilde v)}(t)$ is perpendicular to the fibers of $\pi_1$ and $\pi_2$ at $t=0$. Since $\pi_1,\pi_2:{G^{(2)}}\to G$ are both Riemannian submersions, $\gamma_{(\tilde w,\tilde v)}(t)$ stays perpendicular to those fibers, and we conclude that $\pi_1(\gamma_{(\tilde w,\tilde v)}(t))$ and $\pi_2(\gamma_{(\tilde w,\tilde v)}(t))$ are both geodesics in $G$ with initial conditions ${\tilde w}$ and ${\tilde v}$. Hence, we must have: $$\gamma_{(\tilde w,\tilde v)}(t)=(\pi_1(\gamma_{(\tilde w,\tilde v)}(t)),\pi_2(\gamma_{(\tilde w,\tilde v)}(t))=(\gamma_{\tilde w}(t),\gamma_{\tilde v}(t)),$$ as claimed. We have proven that $\tilde V\times_V \tilde V$ is included in the domain of $\exp^{(2)}$.
Now, since $m$ is also Riemannian and $({\tilde w},{\tilde v})$ is perpendicular to an $m$-fiber, it follows also that $m(\gamma_{\tilde w}(t),\gamma_{\tilde v}(t))$ is a geodesic in $G$: it is actually the geodesic with initial condition ${{\rm d}}m({\tilde w},{\tilde v})$. We conclude that ${{\rm d}}m({\tilde w},{\tilde v})$ belongs to the domain of $\exp^{(1)}$, hence in $\tilde V$, proving finally that $\tilde V{\rightrightarrows}V$ is a groupoid neighborhood.
Proposition \[prop:commuting.exponential\] applied to the Riemannian submersions $s,t,m,\pi_1,\pi_2$ shows that we have a commutative diagram: $$\xymatrix@R=15pt{
\tilde V\times_V \tilde V \ar[rr]^{\exp^{(2)}}\ar@<0.5pc>[d] \ar[d] \ar@<-0.5pc>[d]& & {G^{(2)}} \ar@<0.5pc>[d] \ar[d] \ar@<-0.5pc>[d] \\
\tilde V \ar[rr]^{\exp^{(1)}} \ar@<0.25pc>[d] \ar@<-0.25pc>[d] & & G\ar@<0.25pc>[d] \ar@<-0.25pc>[d]\\
V \ar[rr]^{\exp^{(0)}} & &M}$$ where $V$, $\tilde V$ and $\tilde V\times_V \tilde V$ are all admissible. We conclude that the exponential maps of ${\eta^{(1)}}$ and ${\eta^{(0)}}$ give the desired weak linearization: $$(\nu(G_S){\rightrightarrows}\nu(S))\supset(\tilde V{\rightrightarrows}V)\overset{\exp}\cong (\exp(\tilde V){\rightrightarrows}\exp(V))\subset (G{\rightrightarrows}M).$$
The main step in the proof of Theorem \[thm:main:2\] was showing the following property of the metric ${\eta^{(1)}}$: if $\tilde w,\tilde v\in \nu(G_S)$ are such that ${{\rm d}}s(\tilde w)={{\rm d}}t(\tilde v)$, then $(\gamma_{\tilde w}(t),\gamma_{\tilde v}(t))$ belongs to ${G^{(2)}}$ for all $t$ and $m(\gamma_{\tilde w}(t),\gamma_{\tilde v}(t))$ is the geodesic in $G$ with initial condition ${{\rm d}}m(\tilde w,\tilde v)$. If one has a metric on $G$ with this property, then the proof shows that one can linearize $G$ around $S$. Notice that this condition on the metric involves multiplication and does not require *a priori* any metric on ${G^{(2)}}$. One is tempted to call a groupoid with a metric satisfying this property for any invariant submanifold $S$ a *Riemannian groupoid*. However, we don’t know of any method to produce such metrics apart from the metrics associated with 2-metrics.
We can easily deduce from Theorem \[thm:main:2\] the main results on linearization of proper groupoids that one can find in the literature [@cs; @ppt; @wein1; @wein2; @zung]:
\[linearization.proper\] If $G{\rightrightarrows}M$ is a Hausdorff proper groupoid and $S\subset M$ is a saturated embedded manifold, then $G$ is linearizable around $S$.
By Theorem \[thm:main:1\], we can endow our groupoid with a 2-metric. By Theorem \[thm:main:2\], we obtain a groupoid neighborhood $\tilde V{\rightrightarrows}V$ of $G_S{\rightrightarrows}S$ in the local model which can be embedded into $G{\rightrightarrows}M$. The proof is completed by observing that in a proper groupoid every groupoid neighborhood contains a full groupoid neighborhood (cf. Lemma \[lem:full:neighborhood\]).
If $G{\rightrightarrows}M$ is a Hausdorff groupoid whose source map is proper and $S\subset M$ is a saturated embedded manifold, then $G$ is invariantly linearizable around $S$.
Every $s$-proper groupoid is proper, and its orbits are stable, namely every neighborhood $U$ of a saturated embedded manifold $S$, contains a saturated neighborhood of $S$ (cf. [@survey Prop 5.3.3]). The proof now is clear.
Invariant linearization of $s$-proper groupoids covers a large number of related classical results, on fibrations, group actions and foliations. But as explained in [@cs], it does not imply the Tube Theorem for proper actions (cf. [@dk Thm. 2.4.1]), where invariant linearization holds without requiring s-properness. Theorem \[thm:main:2\] also does not yield the Tube Theorem, but our strategy of proof should still work: if we use the metric described in Example \[ex:proper-actions\], then the admissible neighborhoods can be taken to be invariant. Nevertheless, a statement for Lie groupoids generalizing the Tube Theorem is still lacking. The one conjectured in [@survey] is that invariant linearization holds for any $s$-locally trivial proper groupoid.
We can also formulate an infinitesimal version of the linearization theorem, which gives a criterion to conclude that a given Lie algebroid does not admit a proper integration.
Given $G{\rightrightarrows}M$ a Lie groupoid and $S\subset M$ a saturated embedded submanifold, we can define the [**infinitesimal local linear model**]{} as the Lie algebroid of the local linear model, say $\operatorname{Lie}(\nu(G_S){\rightrightarrows}S)=A_{\nu(G_S)}\to S$. We will say that the groupoid is [**infinitesimally linearizable**]{} around $S$ if there are opens $S\subset U\subset M$ and $S\subset V\subset \nu(S)$ and a Lie algebroid isomorphism $$A_G|_U\cong A_{\nu(G_S)|_V}$$
A Riemannian groupoid is infinitesimally linearizable around any saturated submanifold $S$.
Notice that at the infinitesimal level weak linearization and linearization agree. More precisely, given $\tilde U{\rightrightarrows}U$ an open subgroupoid of $G{\rightrightarrows}M$, the inclusion $$(\tilde U {\rightrightarrows}U)\to (G_U{\rightrightarrows}U)$$ defines an isomorphism between the corresponding Lie algebroids. The result now follows from Theorem \[thm:main:2\].
\[ex:lie.alg.bundle\] Let $G{\rightrightarrows}M$ be a Lie group bundle. In this case the linear local model around a point $x\in M$ can be identified with the product $G_x\times T_xM{\rightrightarrows}T_xM$. Hence, if there exists a 2-metric in $G{\rightrightarrows}M$ then the underlying bundle of Lie algebras $A_G\to M$ must be locally trivial.
Some technical background {#appendix:averaging}
=========================
In this appendix, we recall the concept of quasi-action, with focus on the tangent lift of an action. We then introduce averaging operators for quasi-actions. This is a crucial technique that we use in the paper to construct 2-metrics on proper groupoids.
The tangent lift of an action
-----------------------------
Unlike the group case, an action of a Lie groupoid on a manifold does not induce a tangent action on the tangent bundle. We do have natural actions in the normal directions of the underline foliation, the so-called normal representations that we have already discussed. Still, sometimes it is necessary to put them all in a common framework. This can be done with the help of a connection on the groupoid, which allow us to define a quasi-action on the tangent bundle.
Let $G{\rightrightarrows}M$ be a Lie groupoid, and let $E$ be a manifold. A [**quasi-action**]{} $\theta:G\tilde{\curvearrowright}E$ with [**moment map**]{} $q:E\to M$ consists of a smooth map $$\theta:G\times_M E\to E \qquad (g,e)\mapsto \theta_g(e)$$ satisfying $q(\theta_g(e))=t(g)$ for all $(g,e)\in G\times_ME=\{(g,e)\in G\times E: s(g)=q(e)\}$. In other words, a quasi-action associates to each arrow in $G$ a smooth map $\theta_g:E_x\to E_y$. The quasi-action is called:
(i) [**unital**]{} if $\theta_{1_x}={{\rm id}}_{E_x}$ for all $x\in M$;
(ii) [**flat**]{} if $\theta_{g_1}\theta_{g_2}=\theta_{g_1g_2}$ for all $g_1,g_2\in {G^{(2)}}$.
(iii) [**linear**]{} if $q:E\to M$ is a vector bundle and $\theta_g:E_x\to E_y$ is linear for all $g$.
Thus, with these definitions, an *action* is the same as a unital flat quasi-action and a *representation* is the same as a linear action.
An action $\theta:G{\curvearrowright}E$ can be lifted to a quasi-action of the action groupoid $G\ltimes E$ over the tangent bundle $TE$ with the help of a connection on the groupoid. By a [**connection**]{} $\sigma$ on the Lie groupoid $G{\rightrightarrows}M$ we mean a vector bundle map $\sigma:s^*TM\to TG$ such that ${{\rm d}}s\cdot \sigma={{\rm id}}_{s^*TM}$ and $\sigma|_M={{\rm d}}u$. Hence, a connection yields a splitting for the following sequence of vector bundles over $G$: $$\xymatrix{\ 0\ \ar[r] &\ t^* A \ \ar[r] &\ TG \ \ar[r]^{s_*} & \ s^*TM \ \ar@/^/[l]^{\sigma} \ar[r]& \ 0\ }.$$ A connection $\sigma$ is [**multiplicative**]{} if its image is a subgroupoid of $TG{\rightrightarrows}TM$. Using a partition of the unity, one can show that every Lie groupoid admits a connection (see e.g. [@ac]), however a groupoid may not have a multiplicative connection. For instance, a multiplicative connection for the pair groupoid $M\times M{\rightrightarrows}M$ is the same thing as a trivialization of the tangent bundle $TM\to M$ which, of course, does not exists in general.
Given $\theta:G{\curvearrowright}E$ an action and $\sigma$ a connection on $G$, the [**tangent lift**]{} of $\theta$ is the quasi-action $T_\sigma\theta:(G\ltimes E)\tilde{\curvearrowright}TE$ which has moment map the projection $TE\to E$ and is defined by $$T_\sigma\theta:G\times_ M TE \to TE \qquad
(T_\sigma\theta)_{(g,e)}(v)={{\rm d}}\theta(\sigma_g({{\rm d}}_e q(v)),v).$$ By transposition, we define the [**cotangent lift**]{} $T_\sigma\theta:(G\ltimes E)\tilde{\curvearrowright}T^*E$: $$T_\sigma^*\theta:G\times_M T^*E\to T^*E \qquad
\<(T_\sigma^*\theta)_{g,e}(\alpha),v\> = \< \alpha, (T_\sigma \theta)_{(g^{-1},ge)}(v) \>.$$
We will often denote $(T_\sigma\theta)_{(g,e)}(v)$ just by $gv$ and similar for the cotangent lift. With these notations we have $\<g\alpha,v\>=\<\alpha,g^{-1}v\>$. The tangent lift $T_\sigma\theta$ and the cotangent lift $T_\sigma^*\theta$ are both unital, but rarely flat. In fact, the tangent and cotangent lift are flat if and only if the connection is multiplicative. As we saw above in the example of the pair groupoid, this may be a quite restrictive condition, and that is why we need to consider quasi-actions to work with general groupoids.
When $G{\rightrightarrows}M$ is an étale groupoid the map $s_*$ is an isomorphism and there exists a unique connection, namely $\sigma=s_*^{-1}$. Moreover, this connection is multiplicative. Therefore, when working with étale groupoids (and orbifolds) the tangent and cotangent lift are canonically defined, and they are actual actions, which greatly simplifies the whole theory.
Although the tangent and cotangent lift depend on the choice of a connection, their action along the directions transversal to the orbits is intrinsic:
\[tan.lift.nor.representation\] Let $G{\rightrightarrows}M$ be a Lie groupoid, $\sigma$ a connection, $\theta:G{\curvearrowright}E$ an action and $O\subset E$ an orbit. Then $TO^\circ\subset T^*E$ is invariant for the cotangent quasi-action $T^*_\sigma\theta$, and the restriction $(T^*_\sigma\theta)|_{TO^\circ}$ agrees with the conormal representation of the action groupoid. Hence, it is an action which does not depend on $\sigma$.
The connection $\sigma$ consists in choosing for each $(g,e)\in G\times_M E$ a retraction for the linear map $d_{(g,e)}s^*:T_e^*E\to T_{(g,e)}^*(G\times_M E)$ in a smooth way. The value of such a retraction over the image $d_{(g,e)}s^*(T_e^*E)$ is totally settled. Writing $\tilde O=(G\times_M E)_O$, the result follows by noting that $T\tilde O^\circ\subset d_{(g,e)}s^*(T_e^*E)$.
We end this subsection by stating the following naturality properties of the tagent lift, whose proof are straightforward.
\[naturality.tangent.lift\] Let $G{\rightrightarrows}M$ be a Lie groupoid and fix a connection $\sigma$ on it. Then:
(i) If $\theta^E:G{\curvearrowright}E$ and $\theta^F:G{\curvearrowright}F$ are two groupoid actions with moment maps $q^E, q^F$, respectively, then for any equivariant map $p:E\to F$ the differential ${{\rm d}}p:TE\to TF$ is also equivariant for the tangent lifts $T_\sigma\theta^E$ and $T_\sigma\theta^F$;
(ii) If $\theta^1:G{\curvearrowright}E$ and $\theta^2:G{\curvearrowright}E$ are two commuting actions with moment maps $q_1,q_2:E\to M$, then the tangent lifts $T_\sigma\theta^1$ and $T_\sigma\theta^2$ also commute.
Haar systems and averaging methods
----------------------------------
Haar systems on Lie groupoids generalize Haar systems on Lie groups, they always exist for proper groupoids, and allow some averaging arguments on functions and sections of equivariant vector bundles. We show that this can even be extended so as to include vector bundles endowed with quasi-actions, and apply in this way averaging arguments to metrics.
Recall that a *smooth density* on a vector bundle $E\to M$ of rank $r$ is a nowhere vanishing smooth section $\mu$ of the trivial line bundle $(\wedge^r E)\otimes (\wedge^r E)$. For instance, when $E$ is orientable, any volume form in $E$, i.e., a nowhere vanishing section $\omega$ of $\wedge^r E$, determines a density $\omega\otimes\omega$. Let $G{\rightrightarrows}M$ be a Lie groupoid with associated algebroid $A\to M$. Given a smooth density $\mu$ on the underlying vector bundle, we denote by $\mu^x$ the pullback density on $G(-,x)=s^{-1}(x)$ through the target map: $$\begin{matrix}
\xymatrix{TG(-,x) \ar[d] \ar[r]^(.6){\phi} & A \ar[d] \\ G(-,x) \ar[r]^(.6)t & M}
\end{matrix}
\qquad \phi(g,v)=(t(g), {{\rm d}}R_{g^{-1}}(v)).$$ The family of densities $\{\mu^x\}_{x\in M}$ satisfies the following two properties:
(i) (Smoothness) The function $$x\mapsto \int_{G(-,x)}f(g)\mu^x(g)$$ is smooth for all $f\in C^\infty(G)$.
(ii) (Right-invariance) For any arrow and $f\in C^\infty(G(-,x))$ one has: $$\int_{G(-,y)} f(gh)\mu^y(g)= \int_{G(-,x)}f(g)\mu^x(g).$$ In other words, we have $\mu^y = R_h^*(\mu^x)$, where $R_h:G(-,y)\to G(-,x)$ denotes right multiplication.
We say that $\mu$ is a [**normalized Haar density**]{} if the family $\{\mu^x\}_{x\in M}$ also satisfies the following property:
(iii) (Normalization) The support ${\rm supp}(\mu^x)$ is compact and for all $x\in M$: $$\int_{G(-,x)}\mu^x(g)=1.$$
We have the following important fact:
A proper groupoid $G{\rightrightarrows}M$ admits a normalized Haar density.
For a proof we refer to [@crainic; @tu]. The basic idea is that one can construct such a density $\mu$ as the product $c\tilde\mu$ of a nowhere vanishing density $\tilde\mu$ and a cut-off function $c$. Here by a [**cut-off**]{} function $c:M\to{\mathbb R}$ we mean a function whose support intersects the saturation of any compact set in a compact set, or equivalently, such that $s:\operatorname{supp}(c\circ t)\to{\mathbb R}$ is proper, plus the normalization condition: $$\int_{G(-,x)}c(t(g))\mu^x(g)=1.$$ Now let $G{\rightrightarrows}M$ be a Lie groupoid and let $\theta:G{\curvearrowright}E$ be an action with moment map $q:E\to M$. We say that a function $f\in C^\infty(E)$ is [**$\theta$-invariant**]{} if it is constant along the orbits, namely $f(\theta_g e)=f(e)$ for all $g, e$ for which the action is defined. A normalized Haar density allow us to construct for any $f\in C^\infty(E)=\Gamma(E,{\mathbb R}_E)$ a $\theta$-invariant function $I_\theta(f)$ by averaging over the orbits.
In the same fashion it is possible to average sections of more general vector bundles $\Gamma(E,V)$. More precisely, let $V\to E$ be a vector bundle, let $\theta^E:G{\curvearrowright}E$ be an action and let $\theta^V:(G\ltimes E)\tilde{\curvearrowright}V$ be a linear quasi-action. The main examples to keep in mind are the tangent and cotangent lifts of an action. Writing $ge=\theta^E_g(e)$ and $gv=\theta^V_{(g,e)}(v)$, we say that a section $f\in\Gamma(E,V)$ is [**$\theta$-invariant**]{} if $f(g e)=g f(e)$ for all $g,e$ for which the action is defined.
\[def:aver.oper\] Given $G{\rightrightarrows}M$ a Lie groupoid with normalized density $\mu$, $\theta^E:G{\curvearrowright}E$ an action and $\theta^V:(G\ltimes E){\curvearrowright}V$ a linear quasi-action, the associated [**averaging operator**]{} is defined by $$I_\theta:\Gamma(E,V)\to \Gamma(E,V) \qquad
I_\theta(f)(e):=\int_{G(-,x)} g^{-1}f(g(e))\mu^{x}(g) \qquad x=q(e).$$
Note that $I_\theta(f)(e)$ only depends on the restriction of $f$ to the orbit of $e$. The main properties of this averaging operator are summarized in the following proposition. The proof is straightforward.
\[prop:aver.oper\] With the above notations, the following hold:
(i) If $\theta^V$ is flat then $I_\theta(f)$ is $\theta$-invariant for any $f$.
(ii) If $f$ is already $\theta$-invariant then $I_\theta(f)=f$.
(iii) If $\theta^1,\theta^2:G{\curvearrowright}E$ are two commuting actions then $I_{\theta^1}I_{\theta^2}(f)=I_{\theta^2}I_{\theta^1}(f)$.
(iv) For any equivariant map $\phi:V_1\to V_2$ over vector bundles endowed with linear quasi-actions of $G\ltimes E$, the averaging operators commute with $\phi$.
[xxx]{}
C. Abad and M. Crainic; Representations up to homotopy and Bott’s spectral sequence for Lie groupoids; *Adv. Math.* **248** (2013), 416–452. Preprint *arXiv:0911.2859*.
K. Behrend; On the de Rham cohomology of differential and algebraic stacks; *Adv. Math.* **198** (2005), 583-622.
H. Bursztyn and A. Cabrera; Multiplicative forms at the infinitesimal level; *Math. Ann.* **353** (2012), 663-705.
M. Crainic; Differentiable and algebroid cohomology, van Est isomorphisms, and characteristic classes; *Comment. Math. Helv.* **78** (2003), 681-721.
M. Crainic and R.L. Fernandes; Lectures on integrability of Lie brackets; *Geom. Topol. Monogr.* **17** (2011), 1-107.
M. Crainic and I. Moerdijk; Foliation Groupoids and their Cyclic Homology; *Adv. in Math.* **157** (2001), 177-197.
M. Crainic and I. Struchiner; On the Linearization Theorem for proper Lie groupoids; *Ann. Scient. Éc. Norm. Sup.* $4^e$ série, **46** (2013), 723-746.
M. del Hoyo and R.L. Fernandes; Simplicial metrics on Lie groupoids; *Work in progress*.
M. del Hoyo and R.L. Fernandes; A stacky version of Ehresmann’s Theorem; *Work in progress*.
M. del Hoyo; Lie groupoids and their orbispaces; *Portugal. Math.* **70** (2013), 161-209.
J. Duistermaat and J. Kolk; *Lie groups*; Universitext, Springer-Verlag, Berlin 2000.
E. Gallego, L. Gualandri, G. Hector and A. Reventos; Groupoïdes Riemanniens; *Publ. Mat.* **33** (1989), no. 3, 417–422.
D. Glickenstein; Riemannian Groupoids and Solitons for Three-Dimensional Homogeneous Ricci and Cross-Curvature Flows; *Int. Math. Res. Notices* Vol. **2008** (2008).
A. Gracia-Saz and R. Mehta; VB-groupoids and representations up to homotopy of Lie groupoids; Preprint *arXiv:1007.3658*.
R. Hepworth; Morse inequalities for Orbifold Cohomology; *Algebraic & Geometric Topology* **9** (2009), 1105-1175.
J. Lee; *Riemannian manifolds: An introduction to curvature*; Graduate Texts in Mathematics, Vol. 176, Springer-Verlag, Berlin, 1997.
I. Moerdijk and J. Mrcun; *Introduction to foliations and Lie groupoids*; Cambridge Stud. Adv. Math. 91, Cambridge University Press, Cambridge 2003.
M. Pflaum, H. Posthuma and X. Tang; Geometry of orbit spaces of proper Lie groupoids; [*J. Reine Angew. Math.*]{} [**694**]{} (2014), 49–84. Preprint *arXiv:1101.0180*.
R. Sharpe; *Differential geometry. Cartan’s generalization of Klein’s Erlangen program*; Graduate Texts in Mathematics, 166. Springer-Verlag, New York, 1997.
J.L. Tu; La conjecture de Novikov pour les feuilletages hyperboliques; *K-Theory* **16** (1999), no. 2, 129-184.
A. Weinstein; Linearization problems for Lie algebroids and Lie groupoids, *Lett. Math. Phys.* **52** (2000), 93-102.
A. Weinstein; Linearization of regular proper groupoids, *J. Inst. Math. Jussieu.* **1** (2002), 493-511.
N.T. Zung; Proper groupoids and momentum maps: linearization, affinity and convexity, *Ann. Scient. Éc. Norm. Sup.* $4^e$ série, **39** (2006), 841-869.
[^1]: MdH was partially supported by the ERC Starting Grant No. 279729. RLF was partially supported by NSF grant DMS 1308472 and FCT/Portugal. Both authors acknowledge the support of the *Ciências Sem Fronteiras* grant 401817/2013-0.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let $k$ be a finite field. Wintenberger used the field of norms to give an equivalence between a category whose objects are totally ramified abelian $p$-adic Lie extensions $E/F$, where $F$ is a local field with residue field $k$, and a category whose objects are pairs $(K,A)$, where $K\cong k((T))$ and $A$ is an abelian $p$-adic Lie subgroup of ${\mathrm{Aut}}_k(K)$. In this paper we extend this equivalence to allow ${\mathrm{Gal}}(E/F)$ and $A$ to be arbitrary abelian pro-$p$ groups.'
author:
- |
Kevin Keating\
Department of Mathematics\
University of Florida\
Gainesville, FL 32611\
USA\
[keating@math.ufl.edu]{}
title: 'Wintenberger’s Functor for Abelian Extensions'
---
Introduction
============
Let $q=p^f$ and let $k$ be a finite field with $q$ elements. We define a category ${\mathcal{A}}$ whose objects are totally ramified abelian extensions $E/F$, where $F$ is a local field with residue field $k$, and $[E:F]$ is infinite if $F$ has characteristic 0. An ${\mathcal{A}}$-morphism from $E/F$ to $E'/F'$ is defined to be a continuous embedding $\rho:E{\rightarrow}E'$ such that
1. $\rho$ induces the identity on $k$.
2. $E'$ is a finite separable extension of $\rho(E)$.
3. $F'$ is a finite separable extension of $\rho(F)$.
Let $\rho^*:{\mathrm{Gal}}(E'/F'){\rightarrow}{\mathrm{Gal}}(E/F)$ be the map induced by $\rho$. It follows from conditions 2 and 3 that $\rho^*$ has finite kernel and finite cokernel.
We also define a category ${\mathcal{B}}$ whose objects are pairs $(K,A)$, where $K$ is a local field of characteristic $p$ with residue field $k$ and $A$ is a closed abelian subgroup of ${\mathrm{Aut}}_k(K)$. A ${\mathcal{B}}$-morphism from $(K,A)$ to $(K',A')$ is defined to be a continuous embedding $\sigma:K{\rightarrow}K'$ such that
1. $\sigma$ induces the identity on $k$.
2. $K'$ is a finite separable extension of $\sigma(K)$.
3. $A'$ stabilizes $\sigma(K)$, and the image of $A'$ in ${\mathrm{Aut}}_k(K)$ is an open subgroup of $A$.
Let $\sigma^*:A'{\rightarrow}A$ be the map induced by $\sigma$. It follows from conditions 2 and 3 that $\sigma^*$ has finite kernel and finite cokernel.
The field of norms construction [@cn] gives a functor ${\mathcal{F}}:{\mathcal{A}}{\rightarrow}{\mathcal{B}}$ defined by $$\label{functor}
{\mathcal{F}}(E/F)=(X_F(E),{\mathrm{Gal}}(E/F)).$$ We wish to prove the following:
\[main\] ${\mathcal{F}}$ is an equivalence of categories.
Wintenberger ([@WZp; @Wab]; see also [@lie]) has shown that ${\mathcal{F}}$ induces an equivalence between the full subcategory ${\mathcal{A}}_L$ of ${\mathcal{A}}$ consisting of extensions $E/F$ such that ${\mathrm{Gal}}(E/F)$ is an abelian $p$-adic Lie group, and the full subcategory ${\mathcal{B}}_L$ of ${\mathcal{B}}$ consisting of pairs $(K,A)$ such that $A$ is an abelian $p$-adic Lie group. (Contrary to [@lie; @Wab] we consider finite groups to be $p$-adic Lie groups. The equivalence of categories proved in [@lie; @WZp; @Wab] extends trivially to include the case of finite groups.) The proof of Theorem \[main\] is based on reducing to the equivalence between ${\mathcal{A}}_L$ and ${\mathcal{B}}_L$.
The following result, proved by Laubie [@ram], is a consequence of Theorem \[main\]:
\[laubie\] Let $(K,A)\in{\mathcal{B}}$. Then there is $E/F\in{\mathcal{A}}$ such that $A$ is isomorphic to $G={\mathrm{Gal}}(E/F)$ as a filtered group. That is, there exists an isomorphism $i:A{\rightarrow}G$ such that $i(A[x])=G[x]$ for all $x\ge0$, where $A[x]$, $G[x]$ denote the ramification subgroups of $A$, $G$ with respect to the lower numbering.
The finite field $k\cong{\mathbb F}_q$ is fixed throughout the paper, as is the field $K=k((T))$ of formal Laurent series over $k$. We work with complete discretely valued fields $F$ whose residue field is identified with $k$, and with totally ramified abelian extensions of such fields. The ring of integers of $F$ is denoted by ${\mathcal{O}}_F$ and the maximal ideal of ${\mathcal{O}}_F$ is denoted by ${\mathcal{M}}_F$. We let $v_F$ denote the valuation on the separable closure $F^{sep}$ of $F$ which is normalized so that $v_F(F^{\times})={{\mathbb Z}}$, and we let $v_p$ denote the $p$-adic valuation on ${{\mathbb Z}}$. We say that the profinite group $G$ is finitely generated if there is a finite set $S\subset G$ such that $\langle S\rangle$ is dense in $G$.
Ramification theory and the field of norms
==========================================
In this section we recall some facts from ramification theory, and summarize the construction of the field of norms for extensions in ${\mathcal{A}}$.
Let $E/F\in{\mathcal{A}}$. Then $G={\mathrm{Gal}}(E/F)$ has a decreasing filtration by the upper ramification subgroups $G(x)$, defined for nonnegative real $x$. (See for instance [@cl IV].) We say that $u$ is an upper ramification break of $G$ if $G(u+\epsilon)\subsetneqq G(u)$ for every $\epsilon>0$. Since $G$ is abelian, by the Hasse-Arf Theorem [@cl V§7, Th.1] every upper ramification break of $G$ is an integer. In addition, since $E/F$ is a totally ramified abelian extension, it follows from class field theory that $E/F$ is arithmetically profinite (APF) in the sense of [@cn §1]. This means that for every $x\ge0$ the upper ramification subgroup $G(x)$ has finite index in $G=G(0)$. This allows us to define the Hasse-Herbrand functions $$\psi_{E/F}(x)=\int_0^x|G(0):G(t)|\,dt$$ and $\phi_{E/F}(x)=\psi_{E/F}^{-1}(x)$. It follows that the ramification subgroups of $G$ with the lower numbering can be defined by $G[x]=G(\phi_{E/F}(x))$ for $x\ge0$. We say that $l$ is a lower ramification break for $G$ if $G[u+\epsilon]\subsetneqq G[u]$ for every $\epsilon>0$. It is clear from the definitions that $l$ is a lower ramification break if and only if $\phi_{E/F}(l)$ is an upper ramification break.
When $(K,A)\in{\mathcal{B}}$ the abelian subgroup $A$ of ${\mathrm{Aut}}_k(K)$ also has a ramification filtration. The lower ramification subgroups of $A$ are defined by $$A[x]=\{\sigma\in A: v_K(\sigma(T)-T)\ge x+1\}$$ for $x\ge0$. Since $A[x]$ has finite index in $A=A[0]$ for every $x\ge0$, the function $$\phi_A(x)=\int_0^x\frac{dt}{|A[0]:A[t]|}$$ is strictly increasing. Hence we can define the ramification subgroups of $A$ with the upper numbering by $A(x)=A[\psi_A(x)]$, where $\psi_A(x)=\phi_A^{-1}(x)$. The upper and lower ramification breaks of $A$ are defined in the same way as the upper and lower ramification breaks of ${\mathrm{Gal}}(E/F)$. The lower ramification breaks of $A$ are certainly integers, and Laubie’s result (Corollary \[laubie\]) together with the Hasse-Arf theorem imply that the upper ramification breaks of $A$ are also integers.
For $E/F\in{\mathcal{A}}$ let $i(E/F)$ denote the smallest (upper or lower) ramification break of the extension $E/F$. The following basic result from ramification theory is presumably well-known (cf. [@cn 3.2.5.5]).
\[breaks\] Let $M/F\in{\mathcal{A}}$, let $L\in{\mathcal{E}}_{M/F}$, and let $L'/L$ be a finite totally ramified abelian extension which is linearly disjoint from $M/L$. Assume that $M'=ML'$ has residue field $k$, so that $M'/F'\in{\mathcal{A}}$. Then $i(M'/F')\le\psi_{F'/F}(i(M/F))$, with equality if the largest upper ramification break $u$ of $F'/F$ is less than $i(M/F)$.
[[*Proof:* ]{}]{}Set $G={\mathrm{Gal}}(M'/F)$, $H={\mathrm{Gal}}(M'/M)$, and $N={\mathrm{Gal}}(M'/F')$. Then $G=HN\cong H\times N$. Let $y=\phi_{F'/F}(i(M'/F'))$. Then $$N=N(i(M'/F'))=N(\psi_{F'/F}(y))=G(y)\cap N.$$ It follows that $G(y)\supset N$, and hence that $G/H=G(y)H/H=(G/H)(y)$. Therefore $y\le i(M/F)$, which implies $i(M'/F')\le\psi_{F'/F}(i(M/F))$.
If $u<i(M/F)$ then the group $$(G/N)(i(M/F))=G(i(M/F))N/N$$ is trivial. It follows that $G(i(M/F))\subset N$, and hence that $$N(\psi_{F'/F}(i(M/F)))=G(i(M/F))\cap N=G(i(M/F)).$$ The restriction map from ${\mathrm{Gal}}(M'/F')=N$ to ${\mathrm{Gal}}(M/F)\cong
G/H$ carries $G(i(M/F))$ onto $$G(i(M/F))H/H=(G/H)(i(M/F))=G/H.$$ Thus $N(\psi_{F'/F}(i(M/F))=N$, so we have $i(M'/F')\ge\psi_{F'/F}(i(M/F))$. Combining this with the inequality proved above we get $i(M'/F')=\psi_{F'/F}(i(M/F))$. [$\square$]{}
Let $E/F\in{\mathcal{A}}$. Since $E/F$ is an APF extension, the field of norms of $E/F$ is defined: Let ${\mathcal{E}}_{E/F}$ denote the set of finite subextensions of $E/F$, and for $L',L\in{\mathcal{E}}_{E/F}$ such that $L'\supset L$ let ${\mbox{N}}_{L'/L}:L'{\rightarrow}L$ denote the norm map. The field of norms $X_F(E)$ of $E/F$ is defined to be the inverse limit of $L\in{\mathcal{E}}_{E/F}$ with respect to the norms. We denote an element of $X_F(E)$ by $\alpha_{E/F}=(\alpha_L)_{L\in{\mathcal{E}}_{E/F}}$. Multiplication in $X_F(E)$ is defined componentwise, and addition is defined by the rule $\alpha_{E/F}+\beta_{E/F}=\gamma_{E/F}$, where $$\gamma_L=\lim_{L'\in{\mathcal{E}}_{E/L}}{\mbox{N}}_{L'/L}(\alpha_{L'}+\beta_{L'})$$ for $L\in{\mathcal{E}}_{E/F}$. We embed $k$ into $X_F(E)$ as follows: Let $F_0/F$ be the maximum tamely ramified subextension of $E/F$, and for $\zeta\in k$ let $\tilde{\zeta}_{F_0}$ be the Teichmüller lift of $\zeta$ in ${\mathcal{O}}_{F_0}$. Note that for any $L\in{\mathcal{E}}_{E/F_0}$ the degree of the extension $L/F_0$ is a power of $p$. Therefore there is a unique $\tilde{\zeta}_L\in L$ such that $\tilde{\zeta}_L^{[L:F_0]}=\tilde{\zeta}_{F_0}$. Define $f_{E/F}(\zeta)$ to be the unique element of $X_F(E)$ whose $L$ component is $\tilde{\zeta}_L$ for every $L\in{\mathcal{E}}_{E/F_0}$. Then the map $f_{E/F}:k{\rightarrow}X_F(E)$ is a field embedding. By choosing a uniformizer for $X_F(E)$ we get a $k$-isomorphism $X_F(E)\cong k((T))$.
The ring of integers ${\mathcal{O}}_{X_F(E)}$ consists of those $\alpha_{E/F}\in X_F(E)$ such that $\alpha_L\in{\mathcal{O}}_L$ for all $L\in{\mathcal{E}}_{E/F}$ (or equivalently, for any $L\in{\mathcal{E}}_{E/F}$). A uniformizer $\pi_{E/F}=(\pi_L)_{L\in{\mathcal{E}}_{E/F}}$ for $X_F(E)$ gives a uniformizer $\pi_L$ for each finite subextension $L/F$ of $E/F$, and also a uniformizer $\pi_{M/F}=(\pi_L)_{L\in{\mathcal{E}}_{M/F}}$ of $X_F(M)$ for each infinite subextension $M/F$ of $E/F$. The action of ${\mathrm{Gal}}(E/F)$ on the fields $L\in{\mathcal{E}}_{E/F}$ induces a $k$-action of ${\mathrm{Gal}}(E/F)$ on $X_F(E)$. By identifying ${\mathrm{Gal}}(E/F)$ with the subgroup of ${\mathrm{Aut}}_k(X_F(E))$ which it induces, we get the functor ${\mathcal{F}}(E/F)=(X_F(E),{\mathrm{Gal}}(E/F))$ which was mentioned in (\[functor\]).
Let $E'$ be a finite extension of $E$ such that $E'/F\in{\mathcal{A}}$. Then there is $M\in{\mathcal{E}}_{E/F}$ and a finite extension $M'$ of $M$ such that $E'=EM'$ and $E$, $M'$ are linearly disjoint over $M$. We define an embedding $j:X_F(E){\rightarrow}X_F(E')$ as follows. For $\alpha_{E/F}\in X_F(E)$ set $j(\alpha_{E/F})=\beta_{E'/F}$, where $\beta_{E'/F}$ is the unique element of $X_F(E')$ such that $\beta_{LM'}=\alpha_L$ for all $L\in{\mathcal{E}}_{E/M}$ [@cn 3.1.1]. The embedding $j$ makes $X_F(E')$ into a finite separable extension of $X_F(E)$ of degree $[E':E]$; in this setting we denote $X_F(E')$ by $X_{E/F}(E')$. If $E''\supset E'\supset E$ are finite extensions such that $E''/F\in{\mathcal{A}}$ then $X_{E/F}(E')/X_F(E)$ is a subextension of $X_{E/F}(E'')/X_F(E)$. Let $D/E$ be an infinite extension such that $D/F\in{\mathcal{A}}$. Then $X_{E/F}(D)$ is defined to be the union of $X_{E/F}(E')$ as $E'$ ranges over the finite subextensions of $D/E$.
For $E/F\in{\mathcal{A}}$ set $r(E/F)=\lceil\frac{p-1}{p}\cdot i(E/F)\rceil$. The proof of Theorem \[main\] depends on the following two results, the first of which was proved by Wintenberger:
\[quotient\] Let $E/F\in{\mathcal{A}}$, let $L\in{\mathcal{E}}_{E/F_0}$, and define $\xi_L:{\mathcal{O}}_{X_F(E)}{\rightarrow}{\mathcal{O}}_L/{\mathcal{M}}_L^{r(E/L)}$ by $\xi_L(\alpha_{E/F})=\alpha_L\pmod{{\mathcal{M}}_L^{r(E/L)}}$. Then\
(a) $\xi_L$ is a surjective ring homomorphism.\
(b) $\xi_L$ induces the map $\zeta\mapsto\zeta^{p^{-a}}$ on $k$, where $a=v_p([L:F])$.
[[*Proof:* ]{}]{}This is proved in Propositions 2.2.1 and 2.3.1 of [@cn]. [$\square$]{}
\[commute\] Let $E/F\in{\mathcal{A}}$, let $L\in{\mathcal{E}}_{E/F}$, and let $L'/L$ be a finite totally ramified abelian extension which is linearly disjoint from $E/L$. Assume that $E'=EL'$ has residue field $k$, so that $E'/L'\in{\mathcal{A}}$. Then the following diagram commutes, where the bottom horizontal map is induced by the inclusion ${\mathcal{O}}_L\hookrightarrow {\mathcal{O}}_{L'}$: $$\label{diagram}
\begin{array}{ccc}
{\mathcal{O}}_{X_F(E)}&\overset{j}{\longrightarrow}&{\mathcal{O}}_{X_{E/F}(E')} \\[.3cm]
\!\!\!\!\!\!\xi_L\downarrow&&\downarrow\xi_{L'}\!\!\!\!\!\! \\[.3cm]
{\mathcal{O}}_L/{\mathcal{M}}_L^{r(E/L)}&\longrightarrow&
{\mathcal{O}}_{L'}/{\mathcal{M}}_{L'}^{r(E'/L')}
\end{array}$$ Furthermore, for $\zeta\in k$ we have $j\circ f_{E/F}(\zeta)=f_{E'/F}(\zeta^{p^b})$, where $b=v_p([L':L])$.
[[*Proof:* ]{}]{}Using Lemma \[breaks\] we get $$i(E'/L')\le\psi_{L'/L}(i(E/L))\le[L':L]i(E/L).$$ Thus $r(E'/L')\le[L':L]r(E/L)$, so the bottom horizontal map in the diagram is well-defined. Let $\alpha_{E/F}=(\alpha_M)_{M\in{\mathcal{E}}_{E/F}}$ be an element of ${\mathcal{O}}_{X_F(E)}$. Then $j(\alpha_{E/F})$ is the unique element of ${\mathcal{O}}_{X_{E/F}(E')}$ whose $ML'$-component is $\alpha_M$ for every $M\in{\mathcal{E}}_{E/L}$. In particular, the $L'$-component of $j(\alpha_{E/F})$ is $\alpha_L$. Hence $\xi_L(\alpha_{E/F})$ and $\xi_{L'}(j(\alpha_{E/F}))$ are both congruent to $\alpha_L$ modulo ${\mathcal{M}}_{L'}^{r(E'/L')}$, which proves the commutativity of (\[diagram\]). The second claim follows from (\[diagram\]) and Proposition \[quotient\](b). [$\square$]{}
Proof of Theorem \[main\] {#proof}
=========================
In this section we prove Theorem \[main\]. To do this, we must show that the functor ${\mathcal{F}}:{\mathcal{A}}{\rightarrow}{\mathcal{B}}$ is essentially surjective and fully faithful.
We begin by showing that ${\mathcal{F}}$ is essentially surjective. Let $A$ be a closed abelian subgroup of ${\mathrm{Aut}}_k(K)$, where $K=k((T))$. Then $A$ is a $p$-adic Lie group if and only if $A$ is finitely generated. Since ${\mathcal{F}}$ induces an equivalence between the categories ${\mathcal{A}}_L$ and ${\mathcal{B}}_L$, it suffices to prove that $(K,A)$ lies in the essential image of ${\mathcal{F}}$ in the case where $A$ is [*not*]{} finitely generated.
Let $F\cong k((T))$, let $E/F$ be a finite totally ramified abelian extension, and let $\pi$ be a uniformizer of $E$. Then for each $\sigma\in{\mathrm{Gal}}(E/F)$ there is a unique $f_{\sigma}\in k[[T]]$ such that $\sigma(\pi)=f_{\sigma}(\pi)$. Let $a=v_p([E:F])$ and define $$G(E/F,\pi)=\{\gamma\in{\mathrm{Aut}}_k(K):\gamma(T)=f_{\sigma}^{p^a}(T)
\mbox{ for some }\sigma\in{\mathrm{Gal}}(E/F)\},$$ where $f_{\sigma}^{p^a}(T)$ is the power series obtained from $f_{\sigma}(T)$ by replacing the coefficients by their $p^a$ powers. Then $G(E/F,\pi)$ is a subgroup of ${\mathrm{Aut}}_k(K)$ which is isomorphic to ${\mathrm{Gal}}(E/F)$.
Let $l_0<l_1<l_2<\ldots$ denote the positive lower ramification breaks of $A$. For $n\ge0$ set $r_n=\lceil\frac{p-1}{p}\cdot l_n\rceil$ and let ${\overline{\Gamma}}_n$ denote the quotient of $\Gamma={\mathrm{Aut}}_k(K)$ by the lower ramification subgroup $$\Gamma[r_n-1]=
\{\sigma\in\Gamma:\sigma(T)\equiv T\pmod{{\mathcal{M}}_K^{r_n}}\}.$$ Let ${\mathcal{S}}_n$ denote the set of pairs $(E,\pi)$ such that
1. $E/F$ is a totally ramified abelian subextension of $F^{sep}/F$ such that ${\mathrm{Gal}}(E/F)[l_n]$ is trivial. (Such an extension is necessarily finite.)
2. $\pi$ is a uniformizer of $E$ such that the image of $G(E/F,\pi)$ in ${\overline{\Gamma}}_n$ is equal to the image of $A$ in ${\overline{\Gamma}}_n$.
Since there are only finitely many extensions $E/F$ satisfying condition 1, and condition 2 depends only on the class of $\pi$ modulo ${\mathcal{M}}_E^{r_n}$, the set ${\mathcal{S}}_n$ is compact. Using the following lemma we get a map from ${\mathcal{S}}_n$ to ${\mathcal{S}}_{n-1}$.
\[nu\] Let $n\ge1$, let $(E,\pi)\in{\mathcal{S}}_n$, and let $\tilde{E}$ denote the fixed field of ${\mathrm{Gal}}(E/F)[l_{n-1}]$. Then $(\tilde{E},{\mbox{N}}_{E/\tilde{E}}(\pi))\in{\mathcal{S}}_{n-1}$.
[[*Proof:* ]{}]{}It follows from the definitions that $\tilde{E}/F$ is a totally ramified abelian extension and that ${\mathrm{Gal}}(\tilde{E}/F)[l_{n-1}]$ is trivial. Set $\tilde{\pi}={\mbox{N}}_{E/\tilde{E}}(\pi)$, choose $\sigma\in{\mathrm{Gal}}(E/F)$, and let $\tilde{\sigma}$ denote the restriction of $\sigma$ to $\tilde{E}$. By [@cn Prop.2.2.1] the norm ${\mbox{N}}_{E/\tilde{E}}$ induces a ring homomorphism from ${\mathcal{O}}_E$ to ${\mathcal{O}}_{\tilde{E}}/{\mathcal{M}}_{\tilde{E}}^{r_{n-1}}$. Therefore $$\label{sigtilde}
\tilde{\sigma}(\tilde{\pi})={\mbox{N}}_{E/\tilde{E}}(\sigma(\pi))=
{\mbox{N}}_{E/\tilde{E}}(f_{\sigma}(\pi))\equiv
f_{\sigma}^{p^b}({\mbox{N}}_{E/\tilde{E}}(\pi))\pmod{{\mathcal{M}}_{\tilde{E}}^{r_{n-1}}},$$ where $b=v_p([E:\tilde{E}])$. Let $\tilde{a}=v_p([\tilde{E}:F])$ and let $f_{\tilde{\sigma}}\in k[[T]]$ be such that $\tilde{\sigma}(\tilde{\pi})=f_{\tilde{\sigma}}(\tilde{\pi})$. Then by (\[sigtilde\]) we have $$\begin{aligned}
{2}
f_{\tilde{\sigma}}(T)&\equiv
f_{\sigma}^{p^{b}}(T)&&\pmod{T^{r_{n-1}}} \\
f_{\tilde{\sigma}}^{p^{\tilde{a}}}(T)&\equiv
f_{\sigma}^{p^{a}}(T)&&\pmod{T^{r_{n-1}}}.\end{aligned}$$ It follows that $G(\tilde{E}/F,\tilde{\pi})$ and $G(E/F,\pi)$ have the same image in ${\overline{\Gamma}}_{n-1}$, and hence that $G(\tilde{E}/F,\tilde{\pi})$ and $A$ have the same image in ${\overline{\Gamma}}_{n-1}$. [$\square$]{}
Since each $A/A[l_n]$ is finite there is a sequence $A_0\le A_1\le A_2\le\dots$ of finitely generated closed subgroups of $A$ such that $A_nA[l_n]=A$ for all $n\ge0$. Recall that ${\mathcal{F}}$ induces an equivalence of categories between ${\mathcal{A}}_L$ and ${\mathcal{B}}_L$. Since $(K,A_n)\in{\mathcal{B}}_L$, for $n\ge0$ there exists $L_n/F_n\in{\mathcal{A}}_L$ such that ${\mathcal{F}}(L_n/F_n)$ is ${\mathcal{B}}$-isomorphic to $(K,A_n)$. Since $A_n$ is a normal subgroup of $A$, the action of $A$ on $K$ gives a ${\mathcal{B}}$-action of $A$ on the pair $(K,A_n)$. Since ${\mathcal{F}}(L_n/F_n)\cong(K,A_n)$ and ${\mathcal{F}}$ induces an equivalence between ${\mathcal{A}}_L$ and ${\mathcal{B}}_L$, this action is induced by a faithful ${\mathcal{A}}$-action of $A$ on $L_n/F_n$. Since ${\mathrm{Gal}}(L_n/F_n)\cong A_n$ is finitely generated, and $A$ is not finitely generated, this implies that ${\mathrm{Aut}}_k(F_n)$ is not finitely generated. Hence $F_n$ has characteristic $p$. Therefore we may assume $F_n=F$ and $L_n\subset F^{sep}$ with $F\cong k((T))$ fixed.
For $n\ge0$ let $$i_n:(K,A_n){\longrightarrow}(X_{F}(L_n),{\mathrm{Gal}}(L_n/F))$$ be an ${\mathcal{A}}$-isomorphism, and set $\pi_{L_n/F}=i_n(T)$. Let $E_n\subset L_n$ be the fixed field of ${\mathrm{Gal}}(L_n/F)[l_n]$. Then either $E_n=L_n$, or $i(L_n/E_n)\ge l_n$, in which case $r(L_n/E_n)\ge r_n$. Therefore by Proposition \[quotient\](a) we have $(E_n,\pi_{E_n})\in{\mathcal{S}}_n$, so ${\mathcal{S}}_n\not=\varnothing$. For $n\ge1$ let $\nu_n:{\mathcal{S}}_n{\rightarrow}{\mathcal{S}}_{n-1}$ be the map defined by Lemma \[nu\]. Since each ${\mathcal{S}}_n$ is compact, by Tychonoff’s theorem there exists a sequence of pairs $(E_n,\pi_{E_n})\in{\mathcal{S}}_n$ such that $\nu_n(E_n,\pi_{E_n})=(E_{n-1},\pi_{E_{n-1}})$ for $n\ge1$. It follows in particular that $F\subset E_0\subset E_1\subset E_2\subset\dots$. Let $E_{\infty}=\cup_{n\ge 0}E_n$. Then $E_{\infty}$ is a totally ramified abelian extension of $F$, and the uniformizers $\pi_{E_n}$ for $E_n$ induce a uniformizer $\pi_{E_{\infty}/F}$ for $X_F(E_{\infty})$. Let $\tau$ denote the unique $k$-isomorphism from $K=k((T))$ to $X_F(E_{\infty})$ such that $\tau(T)=\pi_{E_{\infty}/F}$. It follows from our construction that $\tau$ induces a ${\mathcal{B}}$-isomorphism from $(K,A)$ to $${\mathcal{F}}(E_{\infty}/F)=(X_F(E_{\infty}),{\mathrm{Gal}}(E_{\infty}/F)).$$ Thus $(K,A)$ lies in the essential image of ${\mathcal{F}}$, so ${\mathcal{F}}$ is essentially surjective.
We now show that ${\mathcal{F}}$ is faithful. Let $E/F$ and $E'/F'$ be elements of ${\mathcal{A}}$, and set $G={\mathrm{Gal}}(E/F)$ and $G'={\mathrm{Gal}}(E'/F')$. We need to show that the map $$\label{Hom}
\Psi:{\mathrm{Hom}}_{{\mathcal{A}}}(E/F,E'/F'){\longrightarrow}{\mathrm{Hom}}_{{\mathcal{B}}}((X_F(E),G),(X_{F'}(E'),G'))$$ induced by the field of norms functor is one-to-one. Suppose $\rho_1,\rho_2\in{\mathrm{Hom}}_{{\mathcal{A}}}(E/F,E'/F')$ satisfy $\Psi(\rho_1)=\Psi(\rho_2)$. Let $\pi_{E/F}=(\pi_L)_{L\in{\mathcal{E}}_{E/F}}$ be a uniformizer for $X_F(E)$. Then $\Psi(\rho_1)(\pi_{E/F})=\Psi(\rho_2)(\pi_{E/F})$, and hence $(\rho_1(\pi_L))_{L\in{\mathcal{E}}_{E/F}}=(\rho_2(\pi_L))_{L\in{\mathcal{E}}_{E/F}}$. It follows that $\rho_1(\pi_L)=\rho_2(\pi_L)$ for every $L\in{\mathcal{E}}_{E/F}$. Since $\rho_1$ and $\rho_2$ are $k$-algebra homomorphisms, this implies that $\rho_1=\rho_2$.
It remains to show that ${\mathcal{F}}$ is full, i.e., that $\Psi$ is onto. It follows from the arguments given in the proof of [@Wab Th.2.1] that the codomain of $\Psi$ is empty if ${\mbox{char}}(F)\not={\mbox{char}}(F')$, and that $\Psi$ is onto if $G$ and $G'$ are finitely generated. In particular, $\Psi$ is onto if ${\mbox{char}}(F)=0$ or ${\mbox{char}}(F')=0$. If one of $G$, $G'$ is finitely generated and the other is not then the codomain of $\Psi$ is empty. Hence it suffices to prove that $\Psi$ is onto in the case where ${\mbox{char}}(F)={\mbox{char}}(F')=p$ and neither of $G$, $G'$ is finitely generated.
We first show that every isomorphism lies in the image of $\Psi$. Let $$\tau:(X_F(E),G){\longrightarrow}(X_{F'}(E'),G')$$ be a ${\mathcal{B}}$-isomorphism. For $n\ge1$ let $F_n$ denote the fixed field of $G[l_n]=G(u_n)$. If ${\displaystyle}\lim_{n{\rightarrow}\infty}l_n/[F_n:F]=\infty$ then an argument similar to that used in [@WZp §2] shows that $\tau$ is induced by an ${\mathcal{A}}$-isomorphism from $E/F$ to $E'/F'$. This limit condition holds for instance if ${\mbox{char}}(F)=p$ and ${\mathrm{Gal}}(E/F)$ is finitely generated, but it can fail if ${\mathrm{Gal}}(E/F)$ is not finitely generated. Therefore we use a different method to prove that $\tau$ lies in the image of $\Psi$, based on a characterization of $F_n/F$ in terms of $(X_F(E),G)$.
Let $d$ denote the $F_n$-valuation of the different of $F_n/F$, and let $c$ be an integer such that $c>\phi_{F_n/F}(\frac{p}{p-1}(l_{n-1}+d))$. Since $G/G(c)$ is finite there exists a finitely generated closed subgroup $H$ of $G$ such that $HG(c)=G$. Let $M\subset E$ be the fixed field of $H$ and set $M_n=F_nM$. Then $F_n/F$ and $M_n/M$ are finite abelian extensions. On the other hand, since $G$ is not finitely generated, ${\mathrm{Gal}}(M/F)\cong G/H$ is not finitely generated, and hence $M/F$ is an infinite abelian extension.
\[kisom\] Let $\pi_{E/F}$ be a uniformizer for $X_F(E)$, and recall that $\pi_{E/F}$ induces uniformizers $\pi_F$, $\pi_{F_n}$, $\pi_{M/F}$, and $\pi_{M_n/F}$ for $F$, $F_n$, $X_F(M)$, and $X_{M/F}(M_n)$. There exists a $k$-isomorphism $\zeta:X_{M/F}(M_n)/X_F(M){\rightarrow}F_n/F$ such that
1. $\zeta(\pi_{M/F})=\pi_F$;
2. $\zeta(\pi_{M_n/F})\equiv\pi_{F_n}
\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}}$;
3. $\gamma\cdot\zeta(\pi_{M_n/F})=\zeta(\gamma\cdot\pi_{M_n/F})$ for every $\gamma\in H$.
The proof of this proposition depends on the following lemma (cf. [@trunc p.88]).
\[near\] Let $F$ be a local field, let $g(T)\in{\mathcal{O}}_F[T]$ be a separable monic Eisenstein polynomial, and let $\alpha\in F^{sep}$ be a root of $g(T)$. Set $E=F(\alpha)$ and let $d=v_E(g'(\alpha))$ be the $E$-valuation of the different of the extension $E/F$. Choose $\pi\in F^{sep}$ such that $v_E(g(\pi))>d$. Then there is a root $\beta$ for $g(X)$ such that $v_E(\pi-\beta)\ge v_E(g(\pi))-d$.
[[*Proof:* ]{}]{}Let $\alpha_1,\alpha_2,\dots,\alpha_n$ be the roots of $g(T)$, and choose $1\leq j\leq n$ to maximize $w=v_E(\pi-\alpha_j)$. For $1\leq i\le n$ we have $$v_E(\pi-\alpha_i) \geq \min\{w,v_E(\alpha_j-\alpha_i)\},$$ with equality if $w>v_E(\alpha_j-\alpha_i)$. Since $w\geq
v_E(\pi-\alpha_i)$, this implies that for $i\not=j$ we have $v_E(\pi-\alpha_i)\leq v_E(\alpha_j-\alpha_i)$. Since $$g(\pi)=(\pi-\alpha_1)(\pi-\alpha_2)\dots(\pi-\alpha_n),$$ we get $$\label{m}
v_E(g(\pi))\leq
w+\sum_{\substack{1\le i\le n\\[.1cm]i\not=j}}
\,v_E(\alpha_j-\alpha_i)=w+d.$$ Setting $\beta=\alpha_j$ gives $v_E(\pi-\beta)=w\ge v_E(g(\pi))-d$. [$\square$]{}
[*Proof of Proposition \[kisom\]:*]{} Since $HG(c)=G$ we have $$i(M/F)\ge c>\phi_{F_n/F}(l_{n-1})=u_{n-1}.$$ Therefore by Lemma \[breaks\] we get $$i(M_n/F_n)=\psi_{F_n/F}(i(M/F))\ge\psi_{F_n/F}(c),$$ and hence $r(M_n/F_n)\ge s$, where $s=\lceil\frac{p-1}{p}
\cdot\psi_{F_n/F}(c)\rceil$. Let $g(T)$ be the minimum polynomial for $\pi_{M_n/F}$ over $X_F(M)$, and let $g_F(T)\in{\mathcal{O}}_F[T]$ be the polynomial obtained by applying the canonical map $\xi_F:X_F(M){\rightarrow}F$ to the coefficients of $g(T)$. Since $g(\pi_{M_n/F})=0$, it follows from Propositions \[quotient\](a) and \[commute\] that $v_{F_n}(g_F(\pi_{F_n}))\ge
r(M_n/F_n)\ge s$. On the other hand, let $\mu:X_F(M){\rightarrow}F$ be the unique $k$-algebra isomorphism such that $\mu(\pi_{M/F})=\pi_F$. Then by Proposition \[quotient\](a) we have $\mu(\alpha_{M/F})\equiv\alpha_F\pmod{{\mathcal{M}}_F^t}$ for all $\alpha_{M/F}\in {\mathcal{O}}_{X_F(M)}$, where $t=\lceil\frac{p-1}{p}\cdot c\rceil$. Let $g^{\mu}(T)\in{\mathcal{O}}_F[T]$ be the polynomial obtained by applying $\mu$ to the coefficients of $g(T)$. Then $g^{\mu}(T)\equiv g_F(T)\pmod{{\mathcal{M}}_F^t}$. Since $[F_n:F]\cdot c\ge
\psi_{F_n/F}(c)$ we have $[F_n:F]\cdot t\ge s$; therefore since $v_{F_n}(g_F(\pi_{F_n}))\ge s$ we have $v_{F_n}(g^{\mu}(\pi_{F_n}))\ge s>l_{n-1}+d$. It follows from Lemma \[near\] that there is a root $\beta$ of $g^{\mu}(T)$ such that $v_{F_n}(\pi_{F_n}-\beta)>l_{n-1}$. Therefore by Krasner’s Lemma we have $F(\beta)\supset F(\pi_{F_n})$. Since $$[F(\beta):F]=\deg(g)=[F(\pi_{F_n}):F]$$ we deduce that $F(\beta)=F(\pi_{F_n})=F_n$. Since $\pi_{M_n/F}$ is a root of $g(T)$, and $\beta$ is a root of $g^{\mu}(T)$, the isomorphism $\mu$ from $X_F(M)$ to $F$ extends uniquely to an isomorphism $\zeta$ from $X_{M/F}(M_n)/X_F(M)$ to $F_n/F$ such that $\zeta(\pi_{M_n/F})=\beta\equiv\pi_{F_n}
\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}}$.
We now show that $\zeta$ is $H$-equivariant. Let $\gamma\in H$ and define $\psi_{\gamma}\in k[[T]]$ by $$\psi_{\gamma}(\pi_{M_n/F})=\gamma\cdot\pi_{M_n/F}=
(\gamma\cdot\pi_L)_{L\in{\mathcal{E}}_{M_n/F}},$$ where we identify $k$ with a subfield of $X_F(M)$ using the map $f_{M/F}$. Using Propositions \[quotient\] and \[commute\] we get $$\gamma\cdot\pi_{F_n}\equiv \psi_{\gamma}(\pi_{F_n})
\pmod{{\mathcal{M}}_{F_n}^{r(M_n/F_n)}}.$$ Since $\zeta(\pi_{M_n/F})\equiv\pi_{F_n}\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}}$ and $r(M_n/F_n)\ge s\ge l_{n-1}+1$ this implies $$\begin{aligned}
{2}
\zeta(\gamma\cdot\pi_{M_n/F})&=\zeta(\psi_{\gamma}(\pi_{M_n/F})) \\
&=\psi_{\gamma}(\zeta(\pi_{M_n/F})) \\
&\equiv \psi_{\gamma}(\pi_{F_n})&&\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}} \\
&\equiv \gamma\cdot\pi_{F_n}&&\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}} \\
&\equiv\gamma\cdot\zeta(\pi_{M_n/F})&&
\pmod{{\mathcal{M}}_{F_n}^{l_{n-1}+1}}.\end{aligned}$$ Since $\zeta(\gamma\cdot\pi_{M_n/F})$ and $\gamma\cdot\zeta(\pi_{M_n/F})$ are both roots of $g^{\mu}(T)$ we deduce that $\gamma\cdot\zeta(\pi_{M_n/F})=
\zeta(\gamma\cdot\pi_{M_n/F})$. Since $\zeta$ is $k$-linear and $\gamma$ acts trivially on $k$, it follows that $\gamma\cdot\zeta(\alpha)=\zeta(\gamma\cdot\alpha)$ for all $\alpha\in X_{M/F}(M_n)$. [$\square$]{}
Since $\tau$ is an ${\mathcal{A}}$-isomorphism, $\tau^*:G'{\rightarrow}G$ is a group isomorphism. For $\gamma\in G$ set $\gamma'=(\tau^*)^{-1}(\gamma)$, and for $N\le G$ set $N'=(\tau^*)^{-1}(N)$. Then $\tau$ induces an isomorphism from $(X_F(E),N)$ to $(X_{F'}(E'),N')$. In particular, $\tau$ gives an isomorphism from $(X_F(E),H)$ to $(X_{F'}(E'),H')$. Using the isomorphism $X_{X_F(M)}(X_{M/F}(E))\cong X_F(E)$ from [@cn 3.4.1] we get an isomorphism $$\label{tauH}
\tau_H:(X_{X_F(M)}(X_{M/F}(E)),H){\longrightarrow}(X_{X_{F'}(M')}(X_{M'/F'}(E')),H'),$$ where $M'\subset E'$ is the fixed field of $H'$. Since $H$ is an abelian $p$-adic Lie group, it follows from [@lie; @WZp; @Wab] that $\tau_H$ is induced by an ${\mathcal{A}}$-isomorphism $$\rho:X_{M/F}(E)/X_F(M){\longrightarrow}X_{M'/F'}(E')/X_{F'}(M').$$ By restricting $\rho$ we get an isomorphism $$\tilde{\rho}:X_{M/F}(M_n)/X_F(M){\longrightarrow}X_{M'/F'}(M_n')/X_{F'}(M'),$$ where $M_n'=(M')_n=F_n'M'$ is the fixed field of $H'[l_n]=H[l_n]'$. Furthermore, for $\gamma\in H$ and $\alpha\in X_{M/F}(M_n)$ we have $\tilde{\rho}(\gamma(\alpha))=\gamma'(\tilde{\rho}(\alpha))$.
Let $\pi_{E/F}$ be a uniformizer for $X_F(E)$, set $\pi_{E'/F'}=\tau(\pi_{E/F})$, and let $$\begin{aligned}
\zeta&:X_{M/F}(M_n)/X_F(M){\longrightarrow}F_n/F \\
\zeta'&:X_{M'/F'}(M_n')/X_{F'}(M'){\longrightarrow}F_n'/F'\end{aligned}$$ be the isomorphisms given by Proposition \[kisom\]. Then $\omega_n=\zeta'\circ\tilde{\rho}\circ\zeta^{-1}$ gives a $k$-linear isomorphism from $F_n/F$ to $F_n'/F'$. It follows from Proposition \[kisom\] that $$\label{scong}
\omega_n(\pi_{F_n})\equiv\pi_{F_n'}\pmod{{\mathcal{M}}_{F_n'}^{l_{n-1}+1}},$$ and that for all $\gamma\in H$ we have $$\label{sequi}
\omega_n(\gamma(\pi_{F_n}))=\gamma'(\omega_n(\pi_{F_n})).$$ Since the restriction map from $H={\mathrm{Gal}}(E/M)$ to ${\mathrm{Gal}}(F_n/F)$ is onto, (\[sequi\]) is actually valid for all $\gamma\in G$.
Let ${\mathcal{T}}_n$ denote the set of $k$-isomorphisms $\omega_n:F_n/F{\rightarrow}F_n'/F'$ which satisfy (\[scong\]) and (\[sequi\]) for all $\gamma\in G$. Since $l_{n-1}$ is the only ramification break of $F_n'/F_{n-1}'$ we have $\psi_{F_n'/F_{n-1}'}(l_{n-1})=l_{n-1}$. Therefore by (\[scong\]) and [@cl V§6, Prop.8], for any $\omega_n\in{\mathcal{T}}_n$ we have $$\label{norm}
{\mbox{N}}_{L_n'/L_{n-1}'}(\omega_n(\pi_{F_n}))\equiv
{\mbox{N}}_{L_n'/L_{n-1}'}(\pi_{F_n'})\pmod{{\mathcal{M}}_{F_{n-1}'}^{l_{n-1}+1}}.$$ Since ${\mbox{N}}_{L_n/L_{n-1}}(\pi_{F_n})=\pi_{F_{n-1}}$ and ${\mbox{N}}_{L_n'/L_{n-1}'}(\pi_{F_n'})=\pi_{F_{n-1}'}$, it follows from (\[norm\]) and (\[sequi\]) that $$\omega_n(\pi_{F_{n-1}})\equiv\pi_{F_{n-1}'}
\pmod{{\mathcal{M}}_{F_{n-1}'}^{l_{n-1}+1}}.$$ Therefore restriction induces a map from ${\mathcal{T}}_n$ to ${\mathcal{T}}_{n-1}$.
Define a metric on ${\mathcal{T}}_n$ by setting $d(\omega_n,\tilde{\omega}_n)=2^{-a}$, where $a=v_{F_n'}(\omega_n(\pi_{F_n})-\tilde{\omega}_n(\pi_{F_n}))$. Then ${\mathcal{T}}_n$ is easily seen to be compact, and we showed above that ${\mathcal{T}}_n$ is nonempty. Therefore by Tychonoff’s theorem there is a sequence $(\omega_n)_{n\ge1}$ with $\omega_n\in{\mathcal{T}}_n$ and $\omega_n|_{F_{n-1}}=\omega_{n-1}$. Since $E=\cup_{n\ge1}F_n$ and $E'=\cup_{n\ge1}F_n'$ the isomorphisms $\omega_n:F_n/F{\rightarrow}F_n'/F'$ combine to give an ${\mathcal{A}}$-isomorphism $\Omega:E/F{\rightarrow}E'/F'$. Let $\theta=\Psi(\Omega)$ be the ${\mathcal{B}}$-isomorphism induced by $\Omega$ and let $m_n=\min\{l_{n-1}+1,r(E/F_n)\}$. It follows from (\[scong\]) and Proposition \[quotient\](a) that $$\theta(\pi_{E/F})\equiv\pi_{E'/F'}\pmod{{\mathcal{M}}_{X_{F'}(E')}^{m_n}}$$ for every $n\ge1$. Since $\lim_{n{\rightarrow}\infty}m_n=\infty$ we get $\theta(\pi_{E/F})=\pi_{E'/F'}=\tau(\pi_{E/F})$. Hence $\tau=\theta=\Psi(\Omega)$.
Now let $\sigma$ be an arbitrary element of ${\mathrm{Hom}}_{{\mathcal{B}}}((X_F(E),G),(X_{F'}(E'),G'))$. Since $X_{F'}(E')$ is a finite separable extension of $\sigma(X_F(E))$, by [@cn 3.2.2] there is a finite separable extension $\tilde{E}'$ of $E$ such that $\sigma$ extends to an isomorphism $$\tau:X_{E/F}(\tilde{E}'){\longrightarrow}X_{F'}(E').$$ It follows that each $\gamma'\in G'$ induces an automorphism $\tilde{\gamma}$ of $X_{E/F}(\tilde{E}')$ whose restriction to $X_F(E)$ is $\sigma^*(\gamma')\in G$. Since $X_{E/F}(F^{sep})$ is a separable closure of $X_F(E)$ [@cn Cor.3.2.3], $\tilde{\gamma}$ can be extended to an automorphism ${\overline{\gamma}}$ of $X_{E/F}(F^{sep})$. Since ${\overline{\gamma}}$ stabilizes $X_F(E)$, and ${\overline{\gamma}}|_{X_F(E)}=\sigma^*(\gamma')$ is induced by an element of $G={\mathrm{Gal}}(E/F)$, it follows from [@cn Rem.3.2.4] that ${\overline{\gamma}}$ is induced by an element of ${\mathrm{Gal}}(F^{sep}/F)$, which we also denote by ${\overline{\gamma}}$. Since ${\overline{\gamma}}$ stabilizes $X_{E/F}(\tilde{E}')$, it stabilizes $\tilde{E}'$ as well. Thus ${\overline{\gamma}}|_{\tilde{E}'}$ is an element of ${\mathrm{Aut}}_k(\tilde{E}')$ which is uniquely determined by $\gamma'$. Since ${\overline{\gamma}}|_{\tilde{E}'}$ induces the automorphism $\tilde{\gamma}$ of $X_{E/F}(\tilde{E}')$, we denote ${\overline{\gamma}}|_{\tilde{E}'}$ by $\tilde{\gamma}$ as well.
Let $\tilde{F}'$ denote the subfield of $\tilde{E}'$ which is fixed by the group $\tilde{G}'=\{\tilde{\gamma}:\gamma'\in G'\}$. Then $\tilde{F}'\supset F$, so $\tilde{E}'/\tilde{F}'$ is a Galois extension. Since the image of $\tilde{G}'\cong G'$ in $G$ is open, and $\tilde{E}'/E$ is a finite extension, it follows that $\tilde{F}'$ is a finite separable extension of $F$, and $\tilde{G}'={\mathrm{Gal}}(\tilde{E}'/\tilde{F}')$. In particular, $\tilde{F}'\cong k((T))$ is a local field with residue field $k$. Hence $(X_{\tilde{F}'}(\tilde{E}'),\tilde{G}')$ is an object in ${\mathcal{B}}$, and $\tau$ gives a ${\mathcal{B}}$-isomorphism from $(X_{\tilde{F}'}(\tilde{E}'),\tilde{G}')$ to $(X_{F'}(E'),G')$. By the arguments given above we see that $\tau$ is induced by an ${\mathcal{A}}$-isomorphism $\Omega:\tilde{E}'/\tilde{F}'{\rightarrow}E'/F'$. Furthermore, the embedding $E\hookrightarrow\tilde{E}'$ induces an ${\mathcal{A}}$-morphism $i:E/F{\rightarrow}\tilde{E}'/\tilde{F}'$. Let $$\alpha:(X_F(E),G){\longrightarrow}(X_{\tilde{F}'}(\tilde{E}'),\tilde{G}')$$ be the ${\mathcal{B}}$-morphism induced by $i$. Then $\sigma=\tau\circ\alpha=\Psi(\Omega\circ i)$.
[9]{}
K. Keating, Extensions of local fields and truncated power series, J. Number Theory [**116**]{} (2006), 69–101.
F. Laubie, Extensions de Lie et groupes d’automorphismes de corps locaux, Compositio Math. [**67**]{} (1988), 165–189.
F. Laubie, Ramification des groupes abéliens d’automorphismes de ${{\mathbb F}}_q((X))$, Canad. Math. Bull. [**50**]{} (2007), 594–597.
J.-P. Serre, [*Corps Locaux*]{}, Hermann, Paris (1962).
J.-P. Wintenberger, Automorphismes des corps locaux de charactéristique $p$, J. Théor. Nombres Bordeaux [**16**]{} (2004), 429–456.
J.-P. Wintenberger, Extensions abéliennes et groupes d’automorphismes de corps locaux, C. R. Acad. Sci. Paris Sér. A-B [**290**]{} (1980), A201–A203.
J.-P. Wintenberger, Le corps des normes de certaines extensions infinies de corps locaux; applications, Ann. Sci. École Norm. Sup. (4) [**16**]{} (1983), 59–89.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Stefano Boselli,'
- 'Ross Hunter,'
- Alexander Mitov
title: 'Prospects for the determination of the top-quark Yukawa coupling at future $\epem$ colliders'
---
Introduction {#sec:intro}
============
The high-precision determination of the Higgs boson couplings is one of the major tasks facing High Energy Physics [@Klute:2013cx; @Mariotti:2016owy]. Despite the wealth of information already delivered by the LHC [@deFlorian:2016spz], and expected from its future high-luminosity operation, a truly detailed study of the Higgs sector will likely be possible only at a future $\epem$ collider.
The top-quark Yukawa coupling $y_t$ dominates the renormalization group evolution of the Higgs potential at high energy scales (see, for instance, refs. [@Degrassi:2012ry; @Alekhin:2012py; @Buttazzo:2013uya]). Therefore $y_t$ is among the main drivers of SM predictions at very high energies and often dominates in, amongst others, studies of the self-consistency of the SM at GUT-scale energies and in searches for physics beyond the SM.
The ideal way for measuring $y_t$ would be through $t\bar t h$ final states since they allow cleaner interpretation of the measurement in terms of $y_t$. However, to produce such a final state a center-of-mass (c.m.) energy of at least $500$ GeV is required. From all proposed $\epem$ colliders - which we detail in sec. \[sec:approach\] below - only the Compact Linear Collider (CLIC) and the International Linear Collider (ILC) will be capable of achieving such c.m. energies. Preliminary studies concerning $t\bar t h$ final states suggest [@Abramowicz:2016zbo; @Barklow:2015tja; @Fujii:2015jha] that CLIC and ILC will be able to measure $y_t$ with a precision of about 4–5%. Given the importance of the top-quark Yukawa coupling such ultimate precision is not entirely satisfactory. Indeed, it can be contrasted to the expected 10% precision [@CMS:2013xfa] from the HL-LHC and the 1% precision expected at a future 100 TeV hadron collider [@Plehn:2015cta].
Given the central importance of the top-quark Yukawa coupling as well as the seemingly puzzling fact that future $\epem$ colliders may not be able to measure it better than a future hadron collider, in this work we set ourselves the goal of addressing the following question: [*what is the ultimate precision with which $y_t$ can be measured at future $\epem$ colliders*]{}?
To answer this question we explore a new approach for the determination of the top-quark Yukawa coupling, utilizing loop-induced Higgs production and decay processes, in a simple version of the so-called $\kappa$-framework [@LHCHiggsCrossSectionWorkingGroup:2012nn]. Loop-induced processes have an advantage in that they potentially allow for a precise determination of $y_t$ at colliders such as FCC-$ee$ or CEPC which are designed to operate at c.m. energies below the $t\bar t h$ threshold. Furthermore, even at colliders that can produce $t\bar t h$ final states, measurements at different c.m. energies could be combined in order to derive more precise determination of $y_t$ than from $t\bar t h$ final states alone. Such indirect approaches are already being pursued in, for example, the determination of the Higgs self-interaction at the LHC [@Degrassi:2016wml; @Maltoni:2017ims; @Maltoni:2018ttu; @Bizon:2016wgr; @DiVita:2017vrr; @DiVita:2017eyz].
This work is organized as follows: in sec. \[sec:approach\] we provide a general overview of future $\epem$ colliders and the Higgs processes considered in this paper. In sec. \[sec:fit\] we describe our approach for determining $y_t$ and detail the fitting procedure adopted in our analysis. Our numerical results are presented in sec. \[sec:results\]. In sec. \[sec:limitations\] we discuss the limitations of our study and possible future extensions. Our conclusions are summarized in sec. \[sec:conclusions\].
Higgs production and decay processes at $\epem$ colliders {#sec:approach}
=========================================================
A number of future lepton colliders have been proposed over the years: [^1]
- The Circular Electron Positron Collider (CEPC) in China would collect 5 $\abinv$ of integrated luminosity at 240 GeV [@CEPC-SPPCStudyGroup:2015csa; @CEPC-SPPCStudyGroup:2015esa; @Mo:2015mza]. A run at 350 GeV could also be envisioned;
- The Future Circular Collider with $\epem$ (FCC-$ee$) is a high-luminosity, high-precision circular collider envisioned in a new 80-100 km tunnel at CERN [@Gomez-Ceballos:2013zzn; @dEnterria:2016sca]. The FCC-$ee$ aims at collecting multi-$\abinv$ integrated luminosity at $\sqrt{s} =$ 90, 160, 240, and 350 GeV. In particular, 10 $\abinv$ of data would be collected at 240 GeV and 2.6 $\abinv$ at 350 GeV;
- The Compact Linear Collider (CLIC) at CERN would collect 100 $\fbinv$at the top threshold, 500 $\fbinv$ at 380 GeV, 1.5 $\abinv$ at 1.5 TeV, and 3 $\abinv$ at 3 TeV [@CLIC:2016zwp]. The study of Higgs measurements at CLIC [@Abramowicz:2016zbo] assumes a different scenario: 500 $\fbinv$ at 350 GeV, 1.5 $\abinv$ at 1.4 TeV and 2 $\abinv$ at 3 TeV. In the following we assume this scenario, for which an accurate estimate of the relevant experimental uncertainties is available;
- The International Linear Collider (ILC) is a proposed linear $\epem$ collider to run in the energy range between 200 GeV and 500 GeV [@Barklow:2015tja; @Fujii:2015jha]. Here we follow a scenario with three energy stages: 250, 350, and 500 GeV, and accumulated luminosity of, respectively, 2 $\abinv$, 200 $\fbinv$ and 4 $\abinv$. The 350 GeV stage of the ILC is mainly intended for the determination of the top quark mass from the $t\bar{t}$ threshold. This stage will not be considered in our study due to its low luminosity.
The dominant single-Higgs production mechanisms at the above future colliders are the $s$-channel Higgstrahlung process $\epem \to hZ$ and the $t$-channel charged vector-boson fusion (VBF) process resulting in $h \nu \bar{\nu}$ final states. The relative importance of these two processes depends on the c.m. energy; the Higgstrahlung process dominates around 240-250 GeV while the VBF cross section takes over around 500 GeV. At even higher energies, the neutral VBF process $\epem\to he^-e^+$ also becomes significant. One should keep in mind, however, that the separation of the various processes is not unambiguous once the $Z$ decays have been taken into account. In particular, Higgstrahlung with $Z$ decaying to neutrinos (electrons) yields the same final-state as the charged (neutral) vector boson fusion processes. This contamination is particularly relevant at 240 GeV where the inclusive $h \nu \bar{\nu}$ rate is dominated by Higgstrahlung.
For this reason in fig. \[fig:scan\] we show the c.m. dependence of the computed at leading order (LO) inclusive cross-section for the final states described above. The $h \nu \bar{\nu}$ and $h \epem$ channels include the Higgstrahlung contribution; the process labeled $hZ$ on the other hand includes all $Z$ decay modes. The same applies to the results in tables \[tab:xs1\],\[tab:xs2\] below. We note, however, that in table \[tab:1s\] the process labeled $hZ$ is not inclusive in the $Z$ decay and includes only certain $Z$ decay modes which are specific to the analyses referenced in that table.
In this study we are interested in processes which are sensitive to a non-vanishing anomalous contribution to $y_t$. For this reason, in fig. \[fig:scan\] we show two more Higgs production processes: $t\bar{t}h$, which is usually considered as the only available channel for the extraction of $y_t$, and the loop-induced process $h\gamma$ which is one of the inputs to our analysis.
To get an overall impression about the potential of the various Higgs production modes, in tables \[tab:xs1\] and \[tab:xs2\] we show the expected number of events for each run of the future colliders described above. The expected numbers of events are derived by multiplying the inclusive cross-sections with the corresponding luminosities shown in tables \[tab:xs1\] and \[tab:xs2\].
As far as Higgs decays are concerned, the loop-induced processes $h\to gg$ and $h\to\gamma\gamma$ are both sensitive to $y_t$ and will be considered in the following. The Higgs decay to gluons is generated by massive quarks in the loops with the top-quark being the dominant contribution. In the $m_t \to \infty$ limit this coupling is known with next-to-next-to-next to leading order (${\rm N}^3{\rm LO}$) QCD accuracy [@Baikov:2006ch]. In contrast, the Higgs decay to photons (as well as $h\gamma$ production) has a dominant contributions from loops involving gauge bosons which results in a reduced sensitivity to $y_t$ compared to $h\to gg$.
CEPC
-------------------------------------------------- --------------------- ---------------------- ---------------------
$\sqrt{s}$ (GeV) 240 350 240
$\mathcal{L}_\mathrm{int.}$ ($\mathrm{fb}^{-1}$) $1.0 \cdot 10^4$ $2.6 \cdot 10^3$ $5.0 \cdot 10^3$
$\sigma_{hZ}$ (fb) 240 130 240
$\mathcal{N}_{hZ}$ $2.4 \cdot 10^6$ $3.38 \cdot 10^5$ $1.2 \cdot 10^6$
$\sigma_{\nu\bar{\nu}h}$ (fb) 54.4 54.7 54.4
$\mathcal{N}_{\nu\bar{\nu}h}$ $5.44 \cdot 10^5$ $1.42 \cdot 10^5$ $2.72 \cdot 10^5$
$\sigma_{eeh}$ (fb) 7.9 7.13 7.9
$\mathcal{N}_{eeh}$ $7.9 \cdot 10^4 $ $1.85 \cdot 10^4$ $3.95 \cdot 10^4$
$\sigma_{h\gamma}$ (fb) $8.96\cdot 10^{-2}$ $3.18 \cdot 10^{-2}$ $8.96\cdot 10^{-2}$
$\mathcal{N}_{h\gamma}$ 896 82 448
: Inclusive LO cross-sections and numbers of expected events for the main Higgs production modes at FCC-$ee$ and CEPC. The process labeled $hZ$ includes all decays of the $Z$ boson.[]{data-label="tab:xs1"}
-------------------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- ----------------------
$\sqrt{s}$ (GeV) 350 1400 3000 250 500
$\mathcal{L}_\mathrm{int.}$ ($\mathrm{fb}^{-1}$) $5.0\cdot 10^2$ $1.5 \cdot 10^3$ $2.0 \cdot 10^3$ $2.0 \cdot 10^3$ $4.0 \cdot 10^3$
$\sigma_{hZ}$ (fb) 130 6.42 1.37 240 57.2
$\mathcal{N}_{hZ}$ $6.50 \cdot 10^4$ $9.6 \cdot 10^3$ $2.74 \cdot 10^3$ $4.80 \cdot 10^5$ $2.29 \cdot 10^5$
$\sigma_{\nu\bar{\nu}h}$ (fb) 54.4 293 498 55.0 85.2
$\mathcal{N}_{\nu\bar{\nu}h}$ $2.73 \cdot 10^4$ $4.39 \cdot 10^5$ $9.96 \cdot 10^5$ $1.10 \cdot 10^5$ $3.41 \cdot 10^5$
$\sigma_{eeh}$ (fb) 7.13 28.3 49.1 8.2 8.7
$\mathcal{N}_{eeh}$ $3.56 \cdot 10^3$ $4.24\cdot 10^4$ $9.82 \cdot 10^4$ $1.64 \cdot 10^4$ $3.48 \cdot 10^4$
$\sigma_{t\bar{t}h}$ (fb) - 1.33 0.41 - 0.27
$\mathcal{N}_{t\bar{t}h}$ - 1995 820 - $1.08 \cdot 10^3$
$\sigma_{h\gamma}$ (fb) $3.18 \cdot 10^{-2}$ $1.20 \cdot 10^{-2}$ $3.08 \cdot 10^{-3}$ $8.97 \cdot 10^{-2}$ $4.74 \cdot 10^{-2}$
$\mathcal{N}_{h\gamma}$ 16 18 6 179 189
-------------------------------------------------- ---------------------- ---------------------- ---------------------- ---------------------- ----------------------
: As in table \[tab:xs1\] but for CLIC and ILC.[]{data-label="tab:xs2"}
Our approach for determining $y_t$ {#sec:fit}
==================================
The problem of determining $y_t$ at a given run of a hypothetical future collider is formulated in terms of the new physics contribution $\Delta y_t$ to the top-quark Yukawa coupling $y_t$
$$y_t = y_t^\mathrm{SM} + \Delta y_t\,,
\label{eq:resc}$$
where we assume that this is the only source of deviation from the SM. This assumption is discussed in sec. \[sec:limitations\].
The main limitation of an analysis restricted to $t\bar th$ data is that it requires a c.m. energy of at least 500 GeV. To circumvent this limitation, we also consider $\epem$ observables which are [*indirectly*]{} sensitive to $y_t$ and, at the same time, have sufficiently large number of expected events. The $y_t$ dependence in such processes originates in the Higgs boson coupling to top quarks in loops, either in the production or in the decay of the Higgs boson [^2]. Thus, in addition to $t\bar t h$ production, we consider the Higgs decays to gluons and photons as well as Higgs production in association with a hard photon. The idea is to exploit the $y_t$ dependence of single Higgs processes with the added benefit that these processes are accessible at all c.m energies.
In order to constrain $\Delta y_t$ we define a global $\chi^2$ for each run of the future colliders described in the previous section $$\chi^2 (\Delta y_t) = \sum_{i=1}^{N_p}\sum_{j=1}^{N_d} \frac{\left[\mu_{ij}\left(\Delta y_t\right) - 1\right]^2}{\errors^2} \,,
\label{eqn:chi}$$ with $N_p$ and $N_d$ being, respectively, the number of available production and decay channels. The sums in eq. (\[eqn:chi\]) include only the processes for which $\errors$ values are explicitly shown in table \[tab:1s\].
The degrees of freedom of the $\chi^2$ in eq. (\[eqn:chi\]) are the signal-strengths $\mu_{ij}$ of all Higgs boson processes which are sensitive to a non-vanishing value of $\Delta y_t$ and which can be measured with a sufficient precision. The signal-strength $\mu_{ij}$ for a generic Higgs production mode $i$ and decay channel $j$ can be written in the narrow-width approximation as $$\mu_{ij} =
\left(\frac{\sigma_i}{\sigma_i^\mathrm{SM}}\right)
\left(\frac{\Gamma_j}{\Gamma_j^\mathrm{SM}}\right)
\left(\frac{\Gamma_h}{\Gamma_h^{\mathrm{SM}}}\right)^{-1}\,.
\label{eqn:muxy}$$ In eq. (\[eqn:muxy\]) $\Gamma_h$ is the total Higgs width and $\sigma_i$ and $\Gamma_j$ are the corresponding production cross-section and partial decay width. Due to the small number of expected $t\bar th$ and $h\gamma$ events, these two production channels are included with only the dominant $h\to b\bar b$ decay mode.
The one-sigma uncertainties $\errors$ appearing in eq. (\[eqn:chi\]) and listed in table \[tab:1s\] are taken from the literature [@Abramowicz:2016zbo; @Gomez-Ceballos:2013zzn; @CEPC-SPPCStudyGroup:2015csa; @Durieux:2017rsg]. Their values have been derived in a realistic framework that accounts for acceptance cuts, background contributions and detector simulation for the reconstruction of the final state. To the best of our knowledge such studies are not available for the $h\gamma$ process. For this reason the one-sigma uncertainties for this channel have been derived by calculating the Poissonian error and have to be considered as optimistic estimates of the total uncertainties.
------ ------ ------------------ ------ ----------------- ------ ----------------- ----------- -------------
$hZ$ $\nu\bar{\nu}h$ $hZ$ $\nu\bar{\nu}h$ $h\gamma$ $t\bar{t}h$
240 $1.0\cdot 10^4$ 1.4% - 3.0% - 4.4% -
350 $2.6\cdot 10^3$ 3.1% 4.7% 14% 21% 14% -
CEPC 240 $5.0\cdot 10^3$ 1.2% - 9.0% - 6.2% -
350 $5.0\cdot 10^2$ 6.1% 10% - - - -
1400 $1.5\cdot 10^3$ - 5.0% - 15% - 8.0%
3000 $2.0\cdot 10^3$ - 4.3% - 10% - 12.5%
250 $2.0\cdot 10^3$ 2.5% - 12% - 10% -
500 $4.0 \cdot 10^3$ 3.9% 1.4% 12% 6.7% 9.8% 9.9%
------ ------ ------------------ ------ ----------------- ------ ----------------- ----------- -------------
: The estimated one-sigma uncertainties $\errors$ used in eq. (\[eqn:chi\]). The ones for CLIC, CEPC and FCC-[*ee*]{} at $\sqrt{s}=240$ GeV are taken from refs. , respectively. The ones for ILC and FCC-[*ee*]{} at $\sqrt{s}=350$ GeV are from ref. [@Durieux:2017rsg]. The $\sqrt{s}=3$ TeV CLIC $t\bar{t}h$ result is derived by extrapolating the 1.4 TeV one with the corresponding number of events. The statistical uncertainties for $h\gamma$ production are derived from the expected number of events reported in tables \[tab:xs1\] and \[tab:xs2\]. The process labeled $hZ$ includes selected $Z$ decays, and their content is specific to each analysis referenced in this table.[]{data-label="tab:1s"}
The analytic expressions for the $h\gamma$ and $t\bar{t}h$ signal-strengths, as functions of $\Delta y_t$, read $$\begin{aligned}
\mu_{h\gamma}
\begin{pmatrix}
\sqrt{s} = 240 \, \mathrm{GeV} \\
\sqrt{s} = 250 \, \mathrm{GeV} \\
\sqrt{s} = 350\, \mathrm{GeV} \\
\sqrt{s} = 500\, \mathrm{GeV}
\end{pmatrix} &=&
\frac{\sigma_{h\gamma}}{\sigma^{\mathrm{SM}}_{h\gamma}} =
1 - \begin{pmatrix}
0.43 \\
0.45 \\
0.73 \\
0.13
\end{pmatrix} \Delta y_t \label{eq:hahg} \\
\mu_{t\bar{t}h}
\begin{pmatrix}
\sqrt{s} = 500 \, \mathrm{GeV} \\
\sqrt{s} = 1400\, \mathrm{GeV} \\
\sqrt{s} = 3000\, \mathrm{GeV}
\end{pmatrix} &=&
\frac{\sigma_{t\bar{t}h}}{\sigma^{\mathrm{SM}}_{t\bar{t}h}} =
1 + \begin{pmatrix}
1.99 \\
1.83 \\
1.71
\end{pmatrix} \Delta y_t,
\label{eq:ha}\end{aligned}$$
In the calculation of the above expressions we do not include corrections beyond LO. Such higher-order effects have been studied in ref. [@Abramowicz:2016zbo] for the 1.4 TeV run of CLIC. These corrections result in a relatively small shift in the corresponding coefficient in eq. (\[eq:ha\]) from 1.83 to 1.89. In turn, this slightly increases the $y_t$ precision in the $t\bar th$ channel.
For the two loop-induced Higgs decay processes we get $$\begin{aligned}
\mu_{h\to g g} &=&
\frac{\Gamma_{h\to gg}} {\Gamma_{h\to gg}^\mathrm{SM}} =
1 + 2 \Delta y_t\,, \label{eq:widthhgg}\\
\mu_{h\to \gamma \gamma} &=&
\frac{\Gamma_{h\to \gamma\gamma}}
{\Gamma_{h\to \gamma\gamma}^\mathrm{SM}} =
1 -0.56 \Delta y_t\,.
\label{eq:width}\end{aligned}$$ All computations in this work have been carried out in the $G_\mu$ input scheme with the help of the `Madgraph5_aMC@NLO_v2.6.1` code [@Alwall:2014hca]. Eqs. (\[eq:hahg\]–\[eq:width\]) have been derived in the following way: we first compute the corresponding cross-sections and decay widths for a number of different values of $\Delta y_t$ and then fit the resulting expressions for $\mu_{ij}$ with a parabola. Finally, we take its linear approximation for small values of $\Delta y_t$. In deriving $\mu_{h\to g g}$ the bottom quark contribution in the loop has been neglected.
Results {#sec:results}
=======
Our main results, namely, the 68% CL constraints following from eq. (\[eqn:chi\]), are displayed in table \[tab:yt\] and in fig. \[fig:fit\]. We report results for the following scenarios: the FCC-$ee$ runs at c.m. energies of $240$ GeV and $350$ GeV, the CEPC run at $240$ GeV, the three CLIC runs at $350$ GeV, $1.4$ TeV and $3$ TeV and the $250$ GeV and $500$ GeV runs of the ILC.
Collider $\sqrt{s}$ (GeV) $\mathcal{L}$ $\left(\mathrm{fb}^{-1}\right)$ $h \to g g$ $h \to \gamma \gamma$ $h\gamma$ $t\bar{t}h$ total
---------- ------------------ ----------------------------------------------- ------------- ----------------------- ----------- ------------- -------
240 $1.0\cdot 10^4$ 0.7% 5.3% 10% - 0.7%
350 $2.6\cdot 10^3$ 1.3% 21% 19% - 1.3%
CEPC 240 $5.0\cdot 10^3$ 0.6% 16% 14% - 0.6%
350 $5.0\cdot 10^2$ 2.6% - - - 2.6%
1400 $1.5\cdot 10^3$ 2.5% 27% - 4.4% 2.2%
3000 $2.0\cdot 10^3$ 2.2% 18% - 7.3% 2.1%
250 $2.0\cdot 10^3$ 1.2% 21% 23% - 1.2%
500 $4.0\cdot 10^3$ 0.7% 10% 75% 5.0% 0.7%
: 68% CL boundaries on $\Delta y_t$ for different runs and processes. In the last column we report the results of the global $\chi^2$ analysis described in sec. \[sec:results\].[]{data-label="tab:yt"}
From table \[tab:yt\] and fig. \[fig:fit\] we conclude that the decay process $h \to gg$ is a strong potential candidate for precise determination of $y_t$. Combining the high $y_t$ sensitivity of $h \to gg$ seen in eq. (\[eq:widthhgg\]) with the high luminosities at the 240 GeV FCC-$ee$ and CEPC runs and at the 500 GeV ILC run, one may potentially be able to determine $\Delta y_t$ with uncertainty of 0.6–0.7%.
The potential of $h \to \gamma\gamma$ for a precise determination of $y_t$ is much smaller than $h\to gg$. Despite the low cross-section of $\epem\to h\gamma$ (see fig. \[fig:scan\]), this loop-induced process allows access to $y_t$ at both FCC-$ee$ and CEPC. While not directly competitive with $h\to gg$, this additional $y_t$ sensitivity is on par with the one expected at HL-LHC and may be useful for disentangling Wilson coefficients in a more refined Effective Field Theory (EFT) approach (see sec. \[sec:limitations\] for further details and ref. [@Brivio:2017vri] for a recent review).
As far as CLIC is concerned, its 350 GeV run allows $y_t$ to be determined from purely loop-induced processes with precision of about 2.6%. At higher CLIC energies the precision in the $y_t$ determination from loop-induced processes is significantly larger than the one expected from the standard $t\bar t h$-based approach. Our estimates show that by combining the extraction of $y_t$ from $t\bar th$ with that from loop-induced final states one can reach $y_t$-precision of about 2.1–2.2% at both the $\sqrt{s}=1.4$ TeV and $\sqrt{s}=3.0$ TeV CLIC runs. This is 2-3 times better than the precision expected from purely $t\bar th$ final states.
From table \[tab:yt\] and fig. \[fig:fit\] we also conclude that for all collider runs considered by us, loop-induced processes (mostly $h\to gg$) could potentially lead to significantly more precise determination of $y_t$ compared to $t\bar th$ final states.
\[fig:mu\]
Limitations of the present study and possible further improvements {#sec:limitations}
==================================================================
The precision in the various $y_t$ determinations estimated in the previous section are based on a number of assumptions and approximations. We discuss them in turn.
For most processes and colliders we have used existing studies for Higgs production and decay. An exception is the process $\epem\to h\gamma$ [@Barroso:1985et; @Abbasabadi:1995rc; @Djouadi:1996ws; @Sang:2017vph]. In this work we have computed this process at LO in the SM with the help of `Madgraph5_aMC@NLO_v2.6.1`, i.e. accounting fully for the loop in the Higgs production process [^3]. No detector simulation, efficiency estimates or realistic estimate of backgrounds and systematic effects have been performed by us for this process. For this reason the purely statistical errors derived by us are, likely, optimistic.
In all our estimates we assume that the top-quark mass $m_t$ is perfectly known. While the future lepton colliders capable of reaching the $t\bar t$ threshold of $\sqrt{s}= 350$ GeV should be able to measure $m_t$ with excellent precision of about 50 MeV [@Beneke:2015lwa; @Beneke:2016kkb; @Beneke:2017rdn; @Penin:2014zaa; @Beneke:2015kwa; @Beneke:2016cbu], the timing of such measurements may be an issue. For example, a $t\bar t$ threshold scan may come only after the measurements at 240 GeV have been performed. In any case one can either reanalyze the lower energy measurement in light of new $m_t$ measurements or utilize then-available HL-LHC measurements. Indeed, the ${\cal O}(300~{\rm MeV})$ uncertainty on $m_t$ expected from HL-LHC will already be sufficiently precise to be a subdominant effect for even the most precise expected determination of $y_t$.
An important caveat for our study is the interpretation of the possible measurements in terms of uncertainty on $y_t$ alone. As we already mentioned in sec. \[sec:fit\], we work in a simple extension of the SM where $\Delta y_t$ is the only possible anomalous coupling. In effect this allows us to simply trade any uncertainty in the measurement for an uncertainty in $y_t$. In reality such an assumption is not well motivated since the assumption of non-vanishing $\Delta y_t$ implies physics beyond the SM, and once such an assumption is made then there is no good justification for assuming a single source of deviation from SM. In this sense, ideally, one would like to treat the problem of the $y_t$ determination within the $\kappa$-framework or a full-blown EFT approach (see refs. [@Elias-Miro:2013mua; @Jenkins:2013zja; @Pomarol:2013zra; @Jenkins:2013wua; @Alonso:2013hga; @Ellis:2014dva; @Falkowski:2015fla; @Butter:2016cvz; @Kobakhidze:2016mfx; @Ellis:2014jta; @Brivio:2017bnu; @deBlas:2018tjm; @Ellis:2018gqa; @Vryonidou:2018eyv] for global analyses in the LHC framework and refs. [@Durieux:2017rsg; @Ellis:2015sca; @Ellis:2017kfi] for EFT analyses in the context of future lepton colliders).
While the introduction of many EFT couplings will reduce the sensitivity to $y_t$ from the measurements we discuss, one should bear in mind that we have not assumed any prior knowledge on $y_t$ or any other EFT coupling. In reality the LHC, especially after its high luminosity phase, is expected to produce a significant set of constraints on both $y_t$ and the EFT couplings that will contribute to the extraction of $y_t$ at future lepton colliders (a recent LHC update with HL-LHC projections for relevant operators can be found in ref. [@Maltoni:2016yxb]). This will benefit the proposed extraction of $y_t$.
Our indirect approach for the extraction of $y_t$ could also be applied/contrasted with the HL-LHC. A naive translation of the latest projections from a $\kappa$-framework study on Higgs physics at the HL-LHC [@Cepeda:2019klc] into our framework using $\kappa_g^2 = \Gamma_{h\to gg}/\Gamma_{h\to gg}^\mathrm{SM}$ and eq. (\[eq:widthhgg\]), yields $\Delta y_t \approx 2.5\%$. This is better than the expected direct $y_t$ determination from $t\bar th$ final states, but still around a factor of four worse than what some future $\epem$ colliders can achieve.
Another tacit assumption made in our analysis is the perfect (or near perfect) knowledge of the SM predictions for the processes under consideration. Full NLO SM accuracy is now easily achievable for non-loop-induced process thanks to automated tools like `Madgraph5_aMC@NLO_v2.6.1`, `Sherpa` [@Gleisberg:2008ta] and `Whizard` [@Moretti:2001zz; @Kilian:2007gr]. Full NLO SM accuracy is also desired for the loop-induced processes discussed here. This requires the calculation of two-loop multiscale amplitudes. A lot of progress in this direction has recently been achieved at the LHC [@Kudashkin:2017skd; @Jones:2018hbb; @Borowka:2016ypz; @Borowka:2016ehy; @Davies:2018ood; @Borowka:2017idc; @Bogner:2017xhp; @Neumann:2018bsx; @Borowka:2018dsa] making this a doable, albeit non-trivial, problem.
Conclusions {#sec:conclusions}
===========
We estimate the ultimate precision with which the top-Yukawa coupling $y_t$ can be extracted at the various proposed high-energy $\epem$ colliders, utilising a simplified version of the $\kappa$-framework. Our motivation for embarking on this study stems from the recognition that the traditional approach for extracting $y_t$ from $t\bar t h$ final states may be too restrictive: such an approach can only be realised at a couple of proposed colliders (CLIC and ILC) and in both cases results in somewhat limited precision of about 4–5%. Such precision has to be viewed in the context of the high-luminosity LHC, which will precede any future $\epem$ machine, and where precision on $y_t$ of about 10%, or even better, is expected.
To increase the scope for precise extraction of $y_t$ at future $\epem$ colliders, in this work we consider an alternative set of final states that are [*indirectly*]{} sensitive to $y_t$. The main advantage in considering loop-induced processes is that due to their large expected event yields and sensitivity to $y_t$, such processes can significantly increase the range of lepton colliders at which $y_t$ can be precisely determined.
We find that potentially one could measure $y_t$ with precision of about 0.6% at the 240 GeV CEPC run. Similarly, the 240 GeV FCC-$ee$ and 500 GeV ILC runs have the potential for determining $y_t$ with precision of 0.7%. Such high $y_t$ precision is driven mainly by the Higgs decay to gluons. Furthermore, the inclusion of $h\to gg$ data significantly increases the sensitivity to $y_t$ at higher c.m. energies. For example, at the 1.4 TeV CLIC run one can get an improvement by a factor of about two compared to $t\bar th$-only data.
The loop-induced Higgs decay $h\to\gamma\gamma$ is not as sensitive to $y_t$ but still offers a decent, better than 6%, precision at the FCC-$ee$ collider. While not directly competitive with the $h\to gg$ decay, $h\to\gamma\gamma$ data could be useful for disentangling contributions from effective couplings and in some cases offers precision better than the one expected form the HL-LHC.
Finally, we have identified the loop-induced associated process $\epem\to h\gamma$, with $h\to b\bar b$, which does not rely on loop-induced Higgs decays and could allow $y_t$ precision of about 10%. Such a precision is comparable to the one expected from the HL-LHC.
Note added {#sec:noteadded}
==========
As mentioned in section \[sec:approach\], after our work was completed and made available on arXiv, updates to the baselines of FCC-$ee$, CEPC and CLIC appeared. The CEPC CDR [@CEPCStudyGroup:2018ghi] reports a modest $\sim 10\%$ increase in integrated luminosity. In contrast, CLIC now plans to take more than double the integrated luminosity than quoted here, whereas the FCC-$ee$ data-taking expectations [@Mangano:2018mur] have been roughly halved. Both have also made small changes to their planned centre-of-mass energies. Whilst these are large operational changes, none are predicted to make a significant effect on the $h\to gg$ process. The $h\to gg$ signal strength (eq. (\[eq:widthhgg\])) is not affected by $\sqrt{s}$, while the experimental error on extracting $\sigma \times \textrm{BR}(h\to gg)$ - as listed in table \[tab:1s\] - is not predicted by FCC-$ee$ or CEPC to significantly change with their new baselines. Since we find that this process dominates the extraction of $y_t$ at all future lepton colliders, the broad conclusions we have drawn here remain unchanged. This is illustrated by an approximate scaling of our results performed in light of CLIC’s new baseline, which modestly improves CLIC’s reach to below 2% on $\Delta y_t$ [@deBlas:2018mhx].
The authors are grateful to Ambresh Shivaji for useful discussions and for providing some independent checks. A.M. thanks the Department of Physics at Princeton University for hospitality during the completion of this work. The work of S.B. and A.M. is supported by the European Research Council Consolidator Grant NNLOforLHC2. The work of A.M. is also supported by the UK STFC grants ST/L002760/1 and ST/K004883/1.
[99]{}
M. Klute, R. Lafaye, T. Plehn, M. Rauch and D. Zerwas, EPL [**101**]{}, no. 5, 51001 (2013) \[arXiv:1301.1322 \[hep-ph\]\]. C. Mariotti and G. Passarino, Int. J. Mod. Phys. A [**32**]{}, no. 04, 1730003 (2017) \[arXiv:1612.00269 \[hep-ph\]\]. D. de Florian [*et al.*]{} \[LHC Higgs Cross Section Working Group\], arXiv:1610.07922 \[hep-ph\]. G. Degrassi, S. Di Vita, J. Elias-Miro, J. R. Espinosa, G. F. Giudice, G. Isidori and A. Strumia, JHEP [**1208**]{}, 098 (2012) \[arXiv:1205.6497 \[hep-ph\]\]. S. Alekhin, A. Djouadi and S. Moch, Phys. Lett. B [**716**]{}, 214 (2012) \[arXiv:1207.0980 \[hep-ph\]\]. D. Buttazzo, G. Degrassi, P. P. Giardino, G. F. Giudice, F. Sala, A. Salvio and A. Strumia, JHEP [**1312**]{}, 089 (2013) \[arXiv:1307.3536 \[hep-ph\]\]. H. Abramowicz [*et al.*]{}, Eur. Phys. J. C [**77**]{}, no. 7, 475 (2017) \[arXiv:1608.07538 \[hep-ex\]\]. T. Barklow, J. Brau, K. Fujii, J. Gao, J. List, N. Walker and K. Yokoya, arXiv:1506.07830 \[hep-ex\]. K. Fujii [*et al.*]{}, arXiv:1506.05992 \[hep-ex\]. \[CMS Collaboration\], arXiv:1307.7135 \[hep-ex\]. M. L. Mangano, T. Plehn, P. Reimitz, T. Schell and H. S. Shao, J. Phys. G [**43**]{}, no. 3, 035001 (2016) \[arXiv:1507.08169 \[hep-ph\]\]. A. David [*et al.*]{} \[LHC Higgs Cross Section Working Group\], arXiv:1209.0040 \[hep-ph\]. F. Maltoni, D. Pagani and X. Zhao, JHEP [**1807**]{}, 087 (2018) \[arXiv:1802.07616 \[hep-ph\]\]. G. Degrassi, P. P. Giardino, F. Maltoni and D. Pagani, JHEP [**1612**]{}, 080 (2016) \[arXiv:1607.04251 \[hep-ph\]\]. F. Maltoni, D. Pagani, A. Shivaji and X. Zhao, Eur. Phys. J. C [**77**]{}, no. 12, 887 (2017) \[arXiv:1709.08649 \[hep-ph\]\]. W. Bizon, M. Gorbahn, U. Haisch and G. Zanderighi, JHEP [**1707**]{}, 083 (2017) \[arXiv:1610.05771 \[hep-ph\]\]. S. Di Vita, G. Durieux, C. Grojean, J. Gu, Z. Liu, G. Panico, M. Riembau and T. Vantalon, JHEP [**1802**]{}, 178 (2018) \[arXiv:1711.03978 \[hep-ph\]\]. S. Di Vita, C. Grojean, G. Panico, M. Riembau and T. Vantalon, JHEP [**1709**]{}, 069 (2017) \[arXiv:1704.01953 \[hep-ph\]\]. CEPC-SPPC Study Group, Preliminary Conceptual Design Report. 1. Physics and Detector http://cepc.ihep.ac.cn/preCDR/volume.html (2018).
CEPC-SPPC Study Group, IHEP-CEPC-DR-2015-01, IHEP-AC-2015-01. X. Mo, G. Li, M. Q. Ruan and X. C. Lou, Chin. Phys. C [**40**]{}, no. 3, 033001 (2016) \[arXiv:1505.01008 \[hep-ex\]\]. M. Bicer [*et al.*]{} \[TLEP Design Study Working Group\], JHEP [**1401**]{}, 164 (2014) \[arXiv:1308.6176 \[hep-ex\]\]. D. d’Enterria, arXiv:1602.05043 \[hep-ex\]. M. J. Boland [*et al.*]{} \[CLIC and CLICdp Collaborations\], arXiv:1608.07537 \[physics.acc-ph\]. P. A. Baikov and K. G. Chetyrkin, Phys. Rev. Lett. [**97**]{}, 061803 (2006) \[hep-ph/0604194\]. M. Beneke, Y. Kiyo, A. Maier and J. Piclum, Comput. Phys. Commun. [**209**]{}, 96 (2016) \[arXiv:1605.03010 \[hep-ph\]\]. M. Beneke, A. Maier, T. Rauh and P. Ruiz-Femenia, JHEP [**1802**]{}, 125 (2018) \[arXiv:1711.10429 \[hep-ph\]\]. M. Beneke, A. Maier, J. Piclum and T. Rauh, Nucl. Phys. B [**899**]{}, 180 (2015) \[arXiv:1506.06865 \[hep-ph\]\]. G. Durieux, C. Grojean, J. Gu and K. Wang, JHEP [**1709**]{}, 014 (2017) \[arXiv:1704.02333 \[hep-ph\]\]. J. Alwall [*et al.*]{}, JHEP [**1407**]{}, 079 (2014) \[arXiv:1405.0301 \[hep-ph\]\]. I. Brivio and M. Trott, arXiv:1706.08945 \[hep-ph\]. A. Barroso, J. Pulido and J. C. Romao, Nucl. Phys. B [**267**]{}, 509 (1986). A. Abbasabadi, D. Bowser-Chao, D. A. Dicus and W. W. Repko, Phys. Rev. D [**52**]{}, 3919 (1995) \[hep-ph/9507463\]. A. Djouadi, V. Driesen, W. Hollik and J. Rosiek, Nucl. Phys. B [**491**]{}, 68 (1997) \[hep-ph/9609420\]. W. L. Sang, W. Chen, F. Feng, Y. Jia and Q. F. Sun, Phys. Lett. B [**775**]{}, 152 (2017) \[arXiv:1706.03572 \[hep-ph\]\]. M. Beneke, Y. Kiyo, P. Marquard, A. Penin, J. Piclum and M. Steinhauser, Phys. Rev. Lett. [**115**]{}, no. 19, 192001 (2015) \[arXiv:1506.06864 \[hep-ph\]\]. A. A. Penin and N. Zerf, JHEP [**1404**]{}, 120 (2014) \[arXiv:1401.7035 \[hep-ph\]\]. M. Beneke, P. Marquard, P. Nason and M. Steinhauser, Phys. Lett. B [**775**]{}, 63 (2017) \[arXiv:1605.03609 \[hep-ph\]\]. J. Elias-Miro, J. R. Espinosa, E. Masso and A. Pomarol, JHEP [**1311**]{}, 066 (2013) \[arXiv:1308.1879 \[hep-ph\]\]. E. E. Jenkins, A. V. Manohar and M. Trott, JHEP [**1310**]{}, 087 (2013) \[arXiv:1308.2627 \[hep-ph\]\]. A. Pomarol and F. Riva, JHEP [**1401**]{}, 151 (2014) \[arXiv:1308.2803 \[hep-ph\]\]. E. E. Jenkins, A. V. Manohar and M. Trott, JHEP [**1401**]{}, 035 (2014) \[arXiv:1310.4838 \[hep-ph\]\]. R. Alonso, E. E. Jenkins, A. V. Manohar and M. Trott, JHEP [**1404**]{}, 159 (2014) \[arXiv:1312.2014 \[hep-ph\]\]. J. Ellis, V. Sanz and T. You, JHEP [**1407**]{}, 036 (2014) \[arXiv:1404.3667 \[hep-ph\]\]. J. Ellis, V. Sanz and T. You, JHEP [**1503**]{}, 157 (2015) \[arXiv:1410.7703 \[hep-ph\]\]. A. Falkowski, Pramana [**87**]{}, no. 3, 39 (2016) \[arXiv:1505.00046 \[hep-ph\]\]. A. Butter, O. J. P. Éboli, J. Gonzalez-Fraile, M. C. Gonzalez-Garcia, T. Plehn and M. Rauch, JHEP [**1607**]{}, 152 (2016) \[arXiv:1604.03105 \[hep-ph\]\]. A. Kobakhidze, N. Liu, L. Wu and J. Yue, Phys. Rev. D [**95**]{}, no. 1, 015016 (2017) \[arXiv:1610.06676 \[hep-ph\]\]. I. Brivio and M. Trott, JHEP [**1707**]{}, 148 (2017) Addendum: \[JHEP [**1805**]{}, 136 (2018)\] \[arXiv:1701.06424 \[hep-ph\]\]. J. de Blas, O. Eberhardt and C. Krause, JHEP [**1807**]{}, 048 (2018) \[arXiv:1803.00939 \[hep-ph\]\]. J. Ellis, C. W. Murphy, V. Sanz and T. You, JHEP [**1806**]{}, 146 (2018) \[arXiv:1803.03252 \[hep-ph\]\]. E. Vryonidou and C. Zhang, JHEP [**1808**]{}, 036 (2018) \[arXiv:1804.09766 \[hep-ph\]\]. J. Ellis and T. You, JHEP [**1603**]{}, 089 (2016) \[arXiv:1510.04561 \[hep-ph\]\]. J. Ellis, P. Roloff, V. Sanz and T. You, JHEP [**1705**]{}, 096 (2017) \[arXiv:1701.04804 \[hep-ph\]\]. F. Maltoni, E. Vryonidou and C. Zhang, JHEP [**1610**]{}, 123 (2016) \[arXiv:1607.05330 \[hep-ph\]\]. M. Cepeda [*et al.*]{} \[Physics of the HL-LHC Working Group\], arXiv:1902.00134 \[hep-ph\]. T. Gleisberg, S. Hoeche, F. Krauss, M. Schonherr, S. Schumann, F. Siegert and J. Winter, JHEP [**0902**]{}, 007 (2009) \[arXiv:0811.4622 \[hep-ph\]\]. W. Kilian, T. Ohl and J. Reuter, Eur. Phys. J. C [**71**]{}, 1742 (2011) \[arXiv:0708.4233 \[hep-ph\]\]. M. Moretti, T. Ohl and J. Reuter, hep-ph/0102195. K. Kudashkin, K. Melnikov and C. Wever, JHEP [**1802**]{}, 135 (2018) \[arXiv:1712.06549 \[hep-ph\]\]. S. P. Jones, M. Kerner and G. Luisoni, Phys. Rev. Lett. [**120**]{}, no. 16, 162001 (2018) \[arXiv:1802.00349 \[hep-ph\]\]. S. Borowka, N. Greiner, G. Heinrich, S. P. Jones, M. Kerner, J. Schlenk and T. Zirke, JHEP [**1610**]{}, 107 (2016) \[arXiv:1608.04798 \[hep-ph\]\]. S. Borowka, N. Greiner, G. Heinrich, S. P. Jones, M. Kerner, J. Schlenk, U. Schubert and T. Zirke, Phys. Rev. Lett. [**117**]{}, no. 1, 012001 (2016); Erratum: \[Phys. Rev. Lett. [**117**]{}, no. 7, 079901 (2016)\] \[arXiv:1604.06447 \[hep-ph\]\]. J. Davies, G. Mishima, M. Steinhauser and D. Wellmann, JHEP [**1803**]{}, 048 (2018) \[arXiv:1801.09696 \[hep-ph\]\]. S. Borowka, G. Heinrich, S. Jahn, S. P. Jones, M. Kerner, J. Schlenk and T. Zirke, Comput. Phys. Commun. [**222**]{}, 313 (2018) \[arXiv:1703.09692 \[hep-ph\]\]. C. Bogner [*et al.*]{}, Comput. Phys. Commun. [**225**]{}, 1 (2018) \[arXiv:1709.01266 \[hep-ph\]\]. T. Neumann, J. Phys. Comm. [**2**]{}, 095017 (2018) \[arXiv:1802.02981 \[hep-ph\]\]. S. Borowka, T. Gehrmann and D. Hulme, JHEP [**1808**]{}, 111 (2018) \[arXiv:1804.06824 \[hep-ph\]\]. J. Guimaraes da Costa [*et al.*]{} \[CEPC Study Group\], arXiv:1811.10545 \[hep-ex\]. A. Abada [*et al.*]{} \[FCC Collaboration\], CERN-ACC-2018-0056. J. de Blas [*et al.*]{}, CERN Yellow Rep. Monogr. Vol. 3 (2018) \[arXiv:1812.02093 \[hep-ph\]\].
[^1]: Since this study was conducted and this publication prepared, FCC-$ee$, CEPC and CLIC have all produced updates to their operational baselines. The potential consequences for our analysis are discussed in sec. \[sec:noteadded\].
[^2]: We will not consider final states with top quarks but no Higgs. These final states are included in the $t\bar t$ threshold scan studies [@Beneke:2015lwa; @Beneke:2016kkb; @Beneke:2017rdn].
[^3]: The rational terms in the relevant `Madgraph5_aMC@NLO_v2.6.1` model file are written in terms of the top-quark mass. They need to be recast in terms of $y_t$ in order to obtain the correct $y_t$ dependence of the $h\gamma$ cross-section.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
bibliography:
- 'bibfile.bib'
---
[**Individual Differences in Dynamic Functional Brain Connectivity Across the Human Lifespan** ]{}\
Elizabeth N. Davison ^1,\*^, Benjamin O. Turner^2^, Kimberly J. Schlesinger ^3^, Michael B. Miller^2^, Scott T. Grafton^2^, Danielle S. Bassett^4,5^, Jean M. Carlson^3^,\
**1** Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, New Jersey, United States of America\
**2** Department of Psychological and Brain Sciences, University of California, Santa Barbara, Santa Barbara, California, United States of America\
**3** Department of Physics, University of California, Santa Barbara, Santa Barbara, California, United States of America\
**4** Department of Bioengineering, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America\
**5** Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, Pennsylvania, United States of America\
\* end@princeton.edu
Abstract {#abstract .unnumbered}
========
Individual differences in brain functional networks may be related to complex personal identifiers, including health, age, and ability. Understanding and quantifying these differences is a necessary first step towards developing predictive methods derived from network topology. Here, we present a method to quantify individual differences in brain functional dynamics by applying hypergraph analysis, a method from dynamic network theory. Using a summary metric derived from the hypergraph formalism—hypergraph cardinality—we investigate individual variations in two separate and complementary data sets. The first data set (“multi-task”) consists of 77 individuals engaging in four consecutive cognitive tasks. We observed that hypergraph cardinality exhibits variation across individuals while remaining consistent within individuals between tasks; moreover, one of the memory tasks evinced a marginally significant correspondence between hypergraph cardinality and age. This finding motivated a similar analysis of the second data set (“age-memory”), in which 95 individuals of varying ages performed a memory task with a similar structure to the multi-task memory task. With the increased age range in the age-memory data set, the correlation between hypergraph cardinality and age correspondence becomes significant. We discuss these results in the context of the well-known finding linking age with network structure, and suggest that age-related changes in brain function can be better understood by taking an integrative approach that incorporates information about the dynamics of functional interactions.
Author Summary {#author-summary .unnumbered}
==============
Complex patterns of activity in each individual human brain generates the unique range of thoughts and behaviors that person experiences. Individual differences in ability, age, state of mind, and other characteristics are tied to differences in brain activity, but determination of the exact nature of these relationships has been limited by the intrinsic complexity of the brain. Here, we apply dynamic network theory to quantify fundamental features of individual neural activity. We represent functional connections between brain regions as a time varying network, and then identify groups of these interactions that exhibit similar behavior over time. The result of this construction is referred to as a hypergraph, and each grouping within the hypergraph is called a hyperedge. We find that the number of these hyperedges in an individual’s hypergraph is a trait-like metric, with considerable variation across the population of subjects, but remarkable consistency within each subject as they perform different tasks. We find a significant correspondence between this metric and the subject’s age, indicating that the dynamics of functional brain activity in older individuals tends to be more dynamically segregated. This new insight into age-related changes in the dynamics of cognitive processing expands our knowledge of the effects of age on brain function and confirms our methods as promising for quantifying and examining individual differences.
Introduction {#introduction .unnumbered}
============
Functional connectivity (FC) analyses based on fMRI data are effective tools for quantifying and characterizing interactions between brain regions. Many approaches borrow methods from the field of graph theory, in which FC is used to build graphs that model the brain as a complex network, treating brain regions as nodes and using functional connections (pairs of nodes with significantly related BOLD signal dynamics) to determine the edge structure of the network [@bullmore2009complex; @friston2011functional]. Individual differences in both underlying FC and the complex network structure resulting from graph theory approaches have been investigated for a variety of task states, developmental stages, and clinical diagnoses [@tavor2016task; @zhang2010disease; @greicius2008resting].
Certain characteristics of FC have been found to vary consistently over the course of normal human aging. The loss of clear segmentation between neural systems is widely reported: many intrinsic functional connectivity networks in the brain tend to become less internally coherent with age, and the functional differences between these intrinsic networks generally become less pronounced [@dennis2014functional; @contreras2015structural; @sala2015reorganization]. These changes are most commonly reported in the default mode network (DMN) [@onoda2012decreased; @tomasi2012aging; @wang2012decoding; @song2014age; @ferreira2015aging; @geerligs2015brain; @ng_reduced_2016], although they have also been observed in other networks, including those associated with higher cognitive functions [@onoda2012decreased; @wang2012decoding; @chan2014decreased; @geerligs2015brain; @ng_reduced_2016]. In addition, inter-network connectivity between the DMN and other regions of the brain has been found to increase, diminishing the ability to discriminate between networks based on FC [@ferreira2015aging; @ng_reduced_2016]. There are some intrinsic functional networks, however, that show no changes or even increased intra-network conectivity with age, such as sensory networks[@tomasi2012aging; @song2014age; @geerligs2015brain].
The bulk of studies on age-related changes and other individual differences in FC, including those that use methods from complex networks and graph theory to represent FC patterns, are performed using static FC analysis, which represents the similarities of brain region activity (or some other measure of concordance) aggregated across an entire data set. Here, we build upon recent advances in network science to study individual differences in human brain activity and behavior from a dynamic network science perspective [@holme2012temporal]. Dynamic functional connectivity (DFC) extends FC to examine how functional organization evolves over time [@hutchison2013dynamic; @calhoun2014chronnectome], allowing investigation of the changes in FC during the course of a cognitive task or scanning session. Efforts to probe the dynamics of functional brain networks have revealed that functional structure reconfigures over time in response to task demands [@doron2012dynamic; @cohen2014quantifying; @monti2014estimating; @gonzalez-castillo_tracking_2015; @davison_brain_2015] and spontaneously at rest [@hutchison2013dynamic; @zalesky_time-resolved_2014]. DFC methods have also been used to inform understanding of individual differences related to aging. In particular, dynamic community structure was found to vary significantly with age [@Schlesinger2016age] and amplitude of low-frequency fluctuations of FC (ALFF-FC) was used to show age-dependent changes in the dynamics of interactions between networks [@qin2015predicting]. Both studies imply that functional dynamics should be considered when investigating how aging affects brain network organization.
Here, we use hypergraph analysis to examine individual differences in DFC network structure in fMRI data acquired as subjects perform cognitively demanding tasks. Compared to traditional graph theoretic techniques, hypergraphs address an existing methodological gap by characterizing not only activity, but also the co-evolution of activity over time. Hyperedges group connections that co-vary over both strong and weak interactions, thus enabling a more complete description of activity during both rest and cognitive tasks. Hypergraph methods extend standard graph methods to incorporate information about co-evolution of activity; whereas standard methods operate on the node-node connectivity matrix, hypergraph methods operate on the edge-edge connectivity matrix. The particular method we utilize here identifies groups of FC connections with statistically similar temporal profiles and links them into groups called hyperedges [@bassett_cross-linked_2014]. Standard FC characterizes interactions between pairs of brain regions and can be extended through DFC methods to capture the dynamics of those interactions. The groups of brain regions that comprise hyperedges are not necessarily strongly active or strongly interconnected brain regions. Rather, correlations in the dynamic connectivity of these regions are the defining characteristics that determine hyperedge structure. As a result, hypergraph analysis is able to identify groups of dynamic connections that change from strong to weak (or [*vice versa*]{}) cohesively together over time, providing complementary information to other DFC methods that focus on only the strongest node-node correlations, such as dynamic community detection [@bassett_robust_2013; @bassett2013task; @Schlesinger2016age].
In previous work, we demonstrated that hyperedges discriminate between diverse task states in a group-level analysis of an fMRI data set spanning four tasks, which we refer to as the “multi-task" data set [@davison_brain_2015]. We also observed notable variation in descriptive hypergraph measures across individuals. In this paper, we extend these results by developing and employing hypergraph measures that capture individual differences in functional brain dynamics to determine correspondences between dynamics and specific demographic and behavioral measures. In the multi-task data set, we find that hypergraph cardinality—the number of distinct hyperedges within a subject’s hypergraph—exhibits marked variation across individuals. At the same time, we find this measure is consistent within individuals, across overall hypergraphs and those associated with specific tasks.
To elucidate the drivers of this striking variation in hypergraph metrics observed across subjects, we explore systematic relationships between hypergraph cardinality and individual difference measures spanning distinct domains such as demographics, cognitive strategy, and personality. In the multi-task data set, we find a suggestive relationship between hypergraph cardinality and participant age. This relationship is confirmed with an independent analysis of a data set with participants who range in age from 18 to 75, which we refer to as the “age-memory” data set. We report a strong positive relationship between age and hypergraph cardinality: older participants are significantly more likely to have a larger number of distinct hyperedges in their hypergraph. This agrees with the widely reported phenomenon of the loss of cohesion within intrinsic functional brain systems, because an increase in the number of distinct hyperedges linking various brain regions points to interconnections between functional groups evolving in time [@ferreira2015aging; @ng_reduced_2016]. Thus, the hypergraph method agrees with previous descriptions of age-related brain changes, while capturing information about dynamics that adds a novel dimension to previous studies. This work further recommends the hypergraph as a useful tool in studying structure in dynamic functional connectivity.
Methods {#methods .unnumbered}
=======
Ethics Statement {#ethics-statement .unnumbered}
----------------
Informed written consent was obtained from each participant prior to experimental sessions for the multi-task and age-memory experiments. All procedures were approved by the University of California, Santa Barbara Human Participants Committee.
Background and Multi-Task Methods {#background-and-multi-task-methods .unnumbered}
---------------------------------
### Multi-Task Experimental Design {#multi-task-experimental-design .unnumbered}
Participants were scanned at rest (task-free) and while engaging in three distinct tasks designed to elicit distinct cognitive functions: an attention-demanding task, a memory task with lexical stimuli, and a memory task with face stimuli. Participants were instructed to lie still and look at a blank screen for the duration of the rest period. During the attention task, participants were instructed to attend to sequences of images on a screen and detect the presence or absence of a target stimulus in designated test displays. Prior to the test display, a cue arrow provided probabilistic information on whether and where the target stimulus might appear. The test display was flashed for approximately 50 ms, after which participants chose whether or not the target stimulus had been present. In both memory tasks, participants were presented with 180 previously examined stimuli and 180 novel stimuli and were asked to discriminate between the two. The memory tasks also included probabilistic cues indicating the probability that the stimulus was novel. For additional experimental details, see [@hermundstad_structural_2013] and [@aminoff_individual_2012].
After completing the scans described above, the following individual difference measures were obtained for study participants: self-reported demographic information, self-reported state of mind (including physical and mental comfort) information, results from the Beck Depression Inventory II [@beck1988inventory], tests for cognitive style (Santa Barbara Learning Style Questionnaire [@massa2006testing], Object Spatial Imagery Questionnaire[@blajenkova2006object], The Need for Cognition Questionnaire [@cacioppo1982need], Verbalizer-Visualizer Questionnaire [@richardson1977VVQ], Card Rotation and Paper Folding Tests [@ekstrom1976manual]), personality tests (Big Five Inventory [@john1999big] BIS/BAS scales[@carver1994behavioral], and PANAS mood assessment [@watson1988development]). More individual difference measures were also collected, but do not match the individual difference measures collected from subjects in the age-memory study.
### Image Acquisition and Processing {#image-acquisition-and-processing .unnumbered}
The MRI data were acquired from 116 participants at the UCSB Brain Imaging Center using a phased array 3T Siemens TIM Trio with a 12 channel head coil. In addition to functional data, a three dimensional high-resolution T1-weighted structural image of the whole brain was obtained for each participant. Functional MRI data were collected from 116 healthy adult participants over the four states described above. Due to various sources of attrition, only 77 participants completed the functional scan and accompanying survey of detailed in [@aminoff_individual_2012]. The sampling period (TR) was 2 s for the rest and attention tasks and 2.5 s for both memory tasks (TE = 30ms, FA = 90).
The functional data is parcellated into regions using a “hybrid" adaptation of the multi-resolution Lausanne2008 atlas registered to MNI space [@hagmann2008mapping] in order to apply the hypergraph analysis. This 194 region “hybrid" anatomical atlas minimizes variability in region size between subjects and brain regions [@davison_brain_2015].
The functional data are preprocessed using FSL [@jenkinson_fsl_2012], AFNI [@cox_afni:_1996] and Matlab [@MATLAB2015]. Preprocessing includes head motion correction with MCFLIRT [@jenkinson_improved_2002], non-brain removal and spatial smoothing with AFNI 3dAutomask/3dDespike, slice-timing correction with AFNI 3dTshift, and additional motion artifact correction with AFNI 3dDetrend. Additionally, each participant’s time-averaged fMRI image is aligned to their structural T1 scan using FSL’s FLIRT with boundary-based registration [@jenkinson_improved_2002; @greve_accurate_2009]. The inverse of this transformation is applied to all participants’ parcellation scales (generated in structural space) and parcellations are down-sampled into functional space with AFNI 3dfractionize. The mean signal across all voxels within a given brain region is calculated to produce a single representative time series. Time series for each task are concatenated to produce a single time series for each brain region.
### Construction of Temporal Networks {#construction-of-temporal-networks .unnumbered}
For each subject, we construct a dynamic network model of brain function that accounts for changes in connectivity over time. Each of the $N = 194$ brain regions is a node in the network. The BOLD signal time series from each brain region is bandpass filtered to obtain data in the 0.06-0.125 Hz frequency range that contains task-related brain activity [@sun2004measuring; @bassett2011dynamic; @lynall2010functional; @bassett2012altered]. This bandpassed time series is then windowed into one-minute sections. Node-node adjacency matrices of size $N\times N$ are constructed by taking Pearson’s correlations between each pair of the $N = 194$ nodes in each of the time windows. Given the lengths of each scan, this windowing yields four rest, 18 attention, 18 word memory, and 18 face memory node-node adjacency matrices. The set of node-node adjacency matrices, one for each windowed section of time, represents the dynamic functional connectivity network; each edge, or pairwise connection between nodes, has an edge weight time series describing its temporal evolution across time windows, as depicted in Figure \[fig1\].
![**Hypergraph construction:** Illustration of the method used to identify hyperedges. Edge weights are computed separately for each time window (A) and joined together to form edge weight time series (B). Significantly correlated edge time series are cross-linked to form a hyperedge (C). The group of hyperedges for an individual, with singletons removed, forms a hypergraph (D).[]{data-label="fig1"}](Fig1.pdf){width="0.9\linewidth"}
### Hypergraph Construction {#hypergraph-construction .unnumbered}
Since hyperedges link edges in this dynamic network that have related temporal profiles, hypergraph structure is determined from the correlations between time-evolving weights of network edges [@bassett_cross-linked_2014] (See Figure \[fig1\] for a schematic illustration of hypergraph construction). These are represented in an edge-edge adjacency matrix $\textbf{X}$, of size $E\times E$, where $E = N(N - 1)/2$ is the total number of possible edges in one time window of the DFC network. Each entry in $\textbf{X}$ is given by the Pearson correlation between the corresponding pair of edge weight time series in the DFC network. The $p$-values from these correlations are thresholded by a false discovery rate correction, which is more sensitive than other corrections for multiple comparisons and is thus effective for such neuroimaging network analyses [@genovese_thresholding_2002]. When the correlation between edges $i$ and $j$ is significant ($p <0.05$), we set $\xi_{ij} = X_{ij}$, to form the thresholded matrix $\xi$. All other elements of $\xi$ are set to zero. We binarize this thresholded matrix and obtain $\xi_{ij}^\prime$, where $$\begin{aligned}
\xi_{ij}^\prime = \begin{cases}
1, & \text{if} \;\;\xi_{ij} \neq 0 ;\\ 0, & \text{if} \;\;\xi_{ij} = 0.
\end{cases}\end{aligned}$$
Each connected component in the thresholded edge-edge correlation matrix $\xi^\prime$ – that is, each set of edges with correlations between any two edges in the set but no significant correlation with edges in any other set – forms a hyperedge. Taken together, all hyperedges in $\xi$ form a hypergraph. Since the edge weight time series are never thresholded and both high and low edge weights are preserved, hypergraphs provide information about edge dynamics without restricting the analysis to strong correlations in regional time series.
Our results are compared with a null model designed to ensure that hyperedges identified in our analysis can be attributed to system dynamics, rather than overall statistical properties of the data [@bassett_robust_2013]. To destroy temporal correspondences between edges but retain the mean and variance of each edge weight time series, the null model randomly reorders each edge time series individually and calculates correlations between the reordered edges.
Once hypergraphs are identified for each individual in the multi-task data set, hyperedges are classified according to whether the correlation in a cognitive state (i.e., rest or one of three cognitive tasks) is significant compared to a permutation null model over all states [@davison_brain_2015]. The hyperedges that satisfy these requirements are denoted as task-specific hyperedges, which we combine to form task-specific hypergraphs.
### Hypergraph Metrics {#hypergraph-metrics .unnumbered}
In this analysis, we examine several complementary measures on individual hypergraphs and focus on one method to extract meaningful information from the overall hyperedge distribution.
[*Hyperedge size:*]{} The size, $s(h)$, of a hyperedge $h$, is defined by $$\begin{aligned}
\label{eq:1}
s(h) = \sum_{i,j \in h} \xi^\prime_{i,j},\end{aligned}$$ where the sum is over the upper triangular elements of $\xi^\prime$, the binarized edge-edge adjacency matrix defined above. This is equivalent to the number of edges that are designated as part of this hyperedge.
[*Singletons:*]{} Singletons are hyperedges with $s(h) = 1$, edges with no significant correlation with any other edge in the network. We exclude singletons from the following analyses.
[*Hypergraph cardinality:*]{} The cardinality of an individual hypergraph is the number of non-singleton hyperedges present in the hypergraph.
[*Hyperedge node degree:*]{} The hyperedge degree of a node is the total number of hyperedges that contain that node.
[*Task-specific hyperedges:*]{} Hyperedges that exhibit a significantly higher correlation within one particular task are grouped into task-specific sets. The sets are calculated by using a permutation test to compare the correlation between edge time series for groups of edges in hypereges in a single task to the same correlation with edge time series data chosen randomly from all tasks. A Bonferroni correction for false positives due to multiple comparisons is employed to select task-specific hyperedges using the most stringent requirements [@hochberg_sharper_1988].
### Regression Procedure {#regression-procedure .unnumbered}
To investigate possible correlates of variability in individual hypergraph metrics, we perform a series of regression analyses. In each analysis, we use the hypergraph metric as the dependent variable and factors representing individual difference measures from the psychometric tests as the independent variables.
[*Behavioral data categorization:*]{} Behavioral and performance data for the multi-task study consist of 231 measures, while there are 115 measures for the age-memory study participants. There are 42 individual difference measures common to both studies, which we group into five categories, given in Table \[Table 1\]. These categories are comprised of differing numbers of individual difference measures, which are summarized in Table \[Table\_All\].
[**Category**]{} [**Factors**]{} [**Information Retained**]{}
------------------- ----------------- ------------------------------
Performance 2 91.41%
Demographics 2 92.62%
State of Mind 3 80.45%
Cognitive Factors 4 77.64%
Personality 6 77.79%
: **Information retained for multi-task study:** Categories, number of factors for each, and how much overall variance from the multi-task individual difference data was retained for each category. Each category represents a subset of the 42 individual difference measures and the factors represent a percentage of the variance contained in the category for the multi-task data.
\[Table 1\]
[*Singular value decomposition:*]{} Once the individual difference measures have been categorized, we demean all measures and perform a singular value decomposition (SVD) separately for each category. We choose the minimum number of factors from the SVD for each category that retain at least 75% of the variance across the category of measures from the multi-task study. Results from this process are presented in Table \[Table 1\].
[*$R^2$ change:*]{} The number of factors retained is not constant across categories, so we implement an adapted multivariate hierarchical regression [@raudenbush2002hierarchical; @miller2012individual] to establish the comparative informativeness of each category. To assess the explanatory power of a given category, all factors in that category are held out for a “control” regression, and the difference in model $R^2$ between this reduced model and the full model is denoted as the contribution for that category. This corresponds to repeatedly performing a hierarchical regression with each category computed last, which gives a conservative estimate for the amount of variance attributable to the category [@miller2012individual].
[*Significance test:*]{} To determine the significance of the regression coefficients, we use the $p$-values from $t$-tests on each multiple regression performed. The Bonferroni procedure for correcting for false positives due to multiple comparisons is used to adjust the $t$-test $p$-values over all regressions performed in this study [@hochberg_sharper_1988]. We employ the Bonferroni correction for multiple comparisons in all regression analyses because it is the most stringent test for significance.
Age-Memory Methods {#age-memory-methods .unnumbered}
------------------
The majority of the methods are identical to those discussed for the multi-task data set. Below, we point out aspects that differ between the two analyses.
### Age-Memory Experimental Design {#age-memory-experimental-design .unnumbered}
The word memory task in the age-memory study is constructed similarly to the word memory task in the multi-task data set. In addition to the memory task, participants completed a resting state scan and diffusion-tensor imaging, which we do not analyze further. Participants did not complete the face memory or attention tasks described in the first data set. The BOLD data were acquired while adult participants performed a recognition memory task with probabilistic cues. Prior to the scanning session, the participants studied 153 common English words, which were mixed with 153 novel lexical stimuli during the task. Participants were asked to determine whether the stimuli were studied or unstudied, with font color cues indicating whether the word had a 70% probability or a 30% probability of having been previously studied [@turner2015one].
### Image Acquisition and Processing {#image-acquisition-and-processing-1 .unnumbered}
Functional and structural data were collected from 126 healthy participants engaged in the word memory task. All functional data was acquired with a 3T Siemens TIM Trio MRI system with a 12-channel head coil. Scans consisted of T2\*-weighted single shot gradient echo, echo-planar sequences sensitive to BOLD contrast (TR = 1.6 s; TE = 30 ms; FA = 90) with generalized autocalibrating partially parallel acquisitions (GRAPPA). In additon to the functional scans, high-resolution anatomical scans were performed for each participant using an MPRAGE sequence (TR = 2.3 s; TE = 2.98 ms; FA = 9; 160 slices; 1.1 mm thickness). Study participants also underwent behavioral assessments and psychological testing. Functional data from 31 participants were excluded due to technical issues, metal screening issues, claustrophobia, attrition, or lack of a complete individual differences survey. The results presented here are from 95 participants with usable functional and individual difference data.
The functional data are preprocessed using FSL [@jenkinson_fsl_2012], AFNI [@cox_afni:_1996], and Matlab [@MATLAB2015]. Preprocessing includes head motion correction (MCFLIRT) [@jenkinson_improved_2002], non-brain removal (BET) [@smith_BET_2002], high-pass temporal filtering ($\sigma{} = 50\text{s}$), spatial smoothing, and grand mean intensity normalization (FEAT) [@wollrichBayesian2009]. Each voxel’s time series is further denoised using a nuisance regression. The nuisance regression includes regressors for the six motion correction terms returned by MCFLIRT, their temporal derivatives, and the mean signal time series from the cerebrospinal fluid. The denoised data is registered to MNI space using FLIRT [@jenkinsonGlobal2001; @jenkinsonOpt2002]. The T1 scan is first registered to the MNI template (12 df affine transformation), the functional data are registered with the T1 image (6 df affine transformation, trilinear interpolation), and the transformations are combined. As in the multi-task study, the mean BOLD signal across all voxels within a given brain region is calculated to produce a single representative time series.
### Construction of Temporal Networks {#construction-of-temporal-networks-1 .unnumbered}
Time series are demeaned and concatenated across the three functional runs of the word memory task to produce a single time series for each brain region. DFC networks are constructed here analogously to the multi-task study, with one key difference. In the age-memory analysis, we remove a single node-node adjacency matrix (i.e., a single time window) from the beginning and end of each functional run. This is to counteract edge effects from processing and ensure continuity across runs. We address this choice further in the section of the Supporting Information.
### Regression Procedure {#regression-procedure-1 .unnumbered}
The regression procedure is similar to the analysis performed on the multi-task data. The individual difference data is kept in the common format, where only the 42 measures common to both studies are used and the categories are the same. Furthermore, the $R^2$ change and significance tests are calculated as above.
[*Singular value decomposition:*]{} We demean all measures and perform a singular value decomposition (SVD) on the combined multi-task and age-memory data separately for each category. This differs from the multi-task analysis, where we only consider the variance retained over the multi-task data. We choose the minimum number of factors from each SVD that retain at least 75% of the variance across both studies. Results from this process are presented in Table \[Table 3\].
Category Factors Information Retained
------------------- --------- ----------------------
Performance 1 87.18%
Demographics 1 86.14 %
State of Mind 3 77.09%
Cognitive Factors 3 81.25%
Personality 4 78.56%
: **Factors common to the mutli-task and age-memory trials:** Categories, number of factors assigned to each, and how much of the overall variance was retained in each category. Each category represents a subset of the 42 individual difference measures and the factors represent a percentage of the variance contained in the category.
\[Table 3\]
Results {#results .unnumbered}
=======
As mentioned above, the hyperedge method has been applied to the multi-task data set in a previous study [@davison_brain_2015]. Here, we recapitulate the key findings from that investigation and provide results of exploratory analyses that motivate the followup analyses on the age-memory data set. We then present results from the age-memory analysis.
Summary of Prior Results {#summary-of-prior-results .unnumbered}
------------------------
A previous study of the multi-task data identified measures that capture significant differences in population-level hypergraph structure across tasks [@davison_brain_2015]. Furthermore, extensive variation was observed in several hypergraph measures, including hypergraph cardinality, across individuals. These results emphasize that hypergraph structure can be used to differentiate between task states and motivates our investigation of the correspondence between hypergraph structure and individual difference measures.
Figure \[fig2\] depicts the empirical cumulative hyperedge size distributions for all hyperedges found across all subjects in the multi-task data set. As a null test, we shuffle the data over time and find no hyperedges of size greater than one. There is a rough power law for the smaller sizes ($s < 100$), followed by a gap in the distribution from about 100 to 1000 and a sharp drop at the system size ($s = \binom{194}{2} = 18721$). The shape of the distribution is due to the consistent hypergraph structure across individuals; the majority of subjects in this study have a hypergraph composed of one large hyperedge and many small hyperedges. While this characteristic structure is common to most subjects in the study, the size of the largest hyperedge varies across individuals. This size is closely related to the hypergraph cardinality, defined as the number of hyperedges in a hypergraph, a measure which also exhibits large variation.
![**Multi-task cumulative size distribution:** The empirical cumulative distribution function of hyperedge sizes for all subjects in the multi-task study. Also shown is a trace for the empirical cumulative distribution functions of hyperedge sizes over all subjects for each of the four task-specific hypergraphs.[]{data-label="fig2"}](Fig2.pdf){width="0.9\linewidth"}
Figure \[fig2\] also depicts task-dependent differences in the cumulative size distributions of task-specific hyperedges. Memory-specific hyperedges tend to be more numerous than those specific to the rest and attention tasks. However, the total number of task-specific hyperedges for any task is at least ten times fewer than the total number of hyperedges. Our strict definition of task specificity includes only hyperedges specific to a single task and discards those associated with more than one task. This approach is conservative, and likely leaves some meaningfully task-related hyperedges unclassified. However, it reduces the complexity of the task-specific results, and provides greater confidence that any hyperedges classified as task-specific are indeed providing truly task-driven information due to coherence within that task alone, rather than coherence due to an unrelated driver that is common to several tasks.
There are significant differences in the spatial organization of task-specific hyperedges over all individuals that are visualized in Figure \[fig3\]. The plots depict task-specific hyperedge degree across the brain for each of the four tasks. In addition to the differences in magnitude between word memory and the other tasks, the locations of high hyperedge concentration vary with task.
![**Node degree spatial distribution:** Here, the number of hyperedges at each node over all individuals in the multi-task study is plotted on the brain. The scale is logarithmic, and higher values in a region indicate that there are more hyperedges that include the region.[]{data-label="fig3"}](Fig3.pdf){width="0.9\linewidth"}
These significant differences in hypergraph structure between the tasks confirm that hypergraph structure varies between task states. However, persistent variability in hypergraph measures across individuals indicates that the hypergraph method reflects innate differences beyond the current task state. The work presented here follows this line of inquiry, beginning with an analysis of individual differences in the multi-task data set.
Multi-Task Results: Individual Differences {#multi-task-results-individual-differences .unnumbered}
------------------------------------------
Here, we illustrate and quantify the wide variation in hypergraph measures across individuals in the multi-task data. In brief, we identify a particular measure, hypergraph cardinality, that demonstrates large variance across all individuals but is consistent within individuals. Following this, we investigate relationships between the variation in individual difference measures and the variation in hypergraph cardinality. The results from this study are not statistically significant due to the limited variation in individual difference measures and strict corrections for multiple comparisons. However, we report a marginally significant result relating demographics and word-memory hyperedge cardinality that motivates further analyses on the age-memory data set.
### Individual Variability and Consistency in Hypergraph Metrics {#individual-variability-and-consistency-in-hypergraph-metrics .unnumbered}
Although our previous study focused on group-level properties of hypergraphs across tasks, notable individual differences in functional dynamics were also seen [@davison_brain_2015]. Here, we confirm those preliminary observations by investigating the hypergraph cardinality measure and finding that it displays extreme variations across subjects in the multi-task data set, as shown in panel (A) of Figure \[fig4\]. These individual variations in hypergraph cardinality span several orders of magnitude.
Despite this large variation between participants, hypergraph cardinality follows a consistent pattern within each participant across tasks. Panel (B) of Figure \[fig4\] depicts individual measures of hypergraph cardinality for hyperedges specific to each task, with subjects sorted by rest hypergraph cardinality. Within participants, the task-specific hypergraph cardinality is consistent across task states and follows the distribution for rest-specific hyperedges, which further emphasizes the consistency of hypergraph cardinality within individuals.
Consistent hypergraph cardinality within participants over all tasks indicates that there are characteristics specific to individuals that drive hypergraph properties, even in designated task-specific hypergraphs. These patterns imply the existence of driving influences on hypergraph structure that are independent of performance on a specific task. To investigate this further, we examine how individual difference measures from demographic and behavioral data relate to hypergraph cardinality.
![**Individual variability:** Hypergraph cardinality for overall hyperedges (A), sorted from smallest to largest cardinality. The plot also includes task-specific cardinalities sorted by overall cardinality. Panel (B) depicts the cardinality for task-specific hyperedges, sorted by rest cardinality. The number of hyperedges across tasks is fairly consistent within individuals, in contrast to the range of hyperedge number across individuals.[]{data-label="fig4"}](Fig4.pdf){width="0.9\linewidth"}
### Drivers of Individual Variability {#drivers-of-individual-variability .unnumbered}
To investigate possible sources of the large variation in hypergraph cardinality seen above, as well as to quantify the extent of the consistency of hyperedge cardinality across tasks, we perform a series of multiple regression analyses on the multi-task data, as described in Methods.
First, using the cardinality of task-specific hypergraphs as the dependent variable, we perform a regression analysis for each non-resting task (attention, word memory, and face memory) that includes the cardinality of the rest-specific hypergraph and the factors shown in Table \[Table 1\] as independent variables. Table \[Table 2\] gives the $R^2$ change values and $p$-values associated with the rest predictor for each task-specific regression. In all three tasks, the rest predictor alone significantly explains the variance in task-specific hypergraph cardinality. This confirms and quantifies our observation in Figure \[fig4\] that hypergraph cardinality is consistent across each individual’s task-specific hypergraphs—i.e., it is trait-like. The individual difference measures used as independent variables are not significant after the Bonferroni correction for multiple comparisons over all tests. However, including the rest-specific hypergraph cardinality, which is closely linked to overall hypergraph cardinality, as an independent variable in the regression accounts for the variation across individuals that is consistent across tasks.
{width="0.9\linewidth"}
[**Attention**]{} [**Word Memory**]{} [**Face Memory**]{}
---------------------- ------------------- --------------------- ---------------------
[**$R^2$ change**]{} 0.72 0.58 0.68
[**$p$-value**]{} $<0.0001$ $<0.0001$ $<0.0001$
: **Rest regression $R^2$ values:** $R^2$ values for the regression between rest-specific hyperedge cardinality and hyperedge cardinality for each of the other three tasks.
\[Table 2\]
To identify possible drivers of this individual variation, we perform another regression analysis, using the individual difference measures from Table \[Table 1\] as independent variables and overall hypergraph cardinality as the dependent variable. Figure \[fig5\] depicts the $R^2$ changes from this analysis for each category of factors. The $t$-test identifies no factors with significant correspondence to hypergraph cardinality, but we observe that the demographics category has the largest $R^2$ change. The $t$-test $p$-value for one of the factors in the demographics category is $<0.05$ and is by far the lowest $p$-value in this stage of the analysis. However, due to our stringent requirements for correcting for multiple comparisons and the number of tests we performed, this correlation is not statistically significant. The marginally significant demographics factor has a loading of $-0.95$ for the age measure and $ -0.31$ for the years of education measure; the loading for sex and handedness demographic measures are comparatively negligible, with magnitudes $< 0.02$.
### Summary of Multi-task Results {#summary-of-multi-task-results .unnumbered}
On the basis of our previous results applying hyperedge analysis to this data set, which hints at substantial variability across individuals in hypergraph structure (Figure \[fig2\]), we carry out several regression analyses designed to identify individual drivers of this variability. There were two key results. The first result is that overall and task-specific hypergraph cardinality show notable variation between subjects, but remarkable consistency within subjects for all tasks (Figure \[fig4\]).
The second key result from this exploratory analysis is the finding of a marginally significant relationship between the demographics category and hyperedge cardinality. Limits to the explanatory power of the multi-task data set may be determined by limited variation in some demographic measures – particularly the small range (27–45) and variance (19) in subject age, which poorly represents the ages observed in the entire population. We thus extend our analysis to a complementary data set collected on a longer study of the word memory task with participants aged 18–75. In the next section, we report the results of our independent analysis of this age-memory data set, which confirm the relationship between age and hypergraph cardinality suggested by the multi-task results.
Age-Memory Results {#age-memory-results .unnumbered}
------------------
To supplement the findings from the multi-task data set, we perform a parallel set of analyses on the age-memory data set. The data set includes participants with ages ranging from 18 to 75, a range three times larger than the range of ages in the multi-task study. Furthermore, the age-memory study uses an almost identical task to the multi-task word-memory task. In this section, we combine hypergraph results for all participants in the age-memory data set and obtain a distribution of hyperedge size over all participants with similar features to the hyperedge size distribution from the word-memory task of the multi-task data. We then identify and test specific drivers of individual variation in hypergraph cardinality for the age-memory study participants. We find a strong correspondence between age and hypergraph cardinality that confirms the preliminary result from the multi-task study.
### Hypergraph Statistics {#hypergraph-statistics .unnumbered}
The cumulative size distribution of hyperedges for all individuals in the age-memory study is depicted in blue in Panel (A) of Figure \[fig6\]. To compare these age-memory hyperedges with the word memory portion of the multi-task study, we identify a new set of hyperedges using only the portion of the multi-task functional time series recorded during the word-memory task for each subject; the distribution of sizes for these hyperedges are plotted in pink. Note that these new word-memory hyperedges from the multi-task data are fundamentally different from the “word memory-specific” hyperedges depicted in Figure \[fig2\]. The “word memory-specific” hyperedges are those hyperedges computed over all tasks, but classified to be driven by correlations in the word memory task alone. In contrast, the new word-memory hyperedges in Figure \[fig6\] are found by using just the word-memory subset of the multi-task data, with no further classification applied.
The distributions of sizes are similar at smaller size scales, but differ somewhat at larger size scales. There are many more hyperedges close to the system size in the age-memory task, while the word-memory hyperedges from the multi-task data set tend to be smaller. The length of the multi-task word-memory time series is shorter than the age-memory time series, which may contribute to this effect [@birn2013effect]. To investigate the size distributions without the effect of full-brain hyperedges, we remove the largest hyperedge from each subject’s hypergraph and plot the resulting distribution in Panel B of Figure \[fig6\]. With this adjustment, the distribution of age-memory hyperedge sizes has a striking agreement with the size distribution of hyperedges constructed from the multi-task word memory data. In both distributions, there is power law behavior for small sizes, similar to that observed in Figure \[fig2\]. Furthermore, the distributions without the largest hyperedges are almost identical; the power of the fit to multi-task word memory data is $-2.21$ and the intercept is $7.91\times 10^{4}$, while the power of the fit to the age-memory data is $-2.37$ and the intercept is $1.46\times 10^{5}$.
We construct a null model, as detailed in the multi-task Methods section, by temporally shuffling the data and find no hyperedges with size greater than one, indicating that the hyperedges identified in the unshuffled data are capturing statistically significant aspects of brain dynamics. In addition, the close correspondence between these two distributions of word-memory hyperedges suggests that the analysis captures aspects of brain dynamics that are robust across imaging sessions and populations.
The inter-subject variability in multi-task hypergraph cardinality spanned several orders of magnitude and followed consistent patterns within subjects for differing cognitive states. We compare the individual hypergraph cardinality for the age-memory and multi-task word-only studies in Figure \[fig7\]. In the age-memory data, hypergraph cardinality ranges from 0 to 1817, which is a similar range of variability as that observed for the complete overall multi-task data set in Figure \[fig4\]. There are 79 subjects with nonzero hyperedge cardinality, indicating that significant non-singleton hyperedges are present in less than two thirds of the subjects. For the remaining analyses, we only consider the 79 subjects with nonzero hypergraph cardinality. For the overall hypergraphs, hypergraph cardinality ranges from 0 to 1832. The maximum hypergraph cardinality for the multi-task word-only data is 1408, which is markedly less than that observed for the age-memory data and may be a result of the shorter time series for the multi-task word task. The presence of near-system size hyperedges, which may also be due to the shorter multi-task word time series, affects hypergraph cardinality by resulting in hypergraphs with cardinality near one.
![**Sorted hypergraph cardinality:** Increasing hyperedge cardinality for individual multi-task word-only and age-memory hypergraphs. The variability for both studies is similar to the variability in multi-task overall hypergraph cardinality, depicted in Panel (A) of Figure \[fig4\].[]{data-label="fig7"}](Fig7.pdf){width="0.9\linewidth"}
### Age-Memory Hypergraph Correspondence With Age {#age-memory-hypergraph-correspondence-with-age .unnumbered}
Having confirmed that hypergraph composition is similar for the multi-task word study and the age-memory study, we investigate whether the individual variability in hypergraph cardinality seen in Figure \[fig7\] corresponds to individual difference factors for the age-memory study.
We perform a multiple regression on the 12 factors distributed across five categories in Table \[Table 3\]. Head motion has been found to induce correlations in FC analyses [@power2012spurious], and a previous study using this data found a significant correlation between age and amount of head motion during the experiment [@turner2015one]. To ensure that excessive head motion is not contributing to our result in any way, we include head motion (operationalized as the average relative movement as computed by MCFLIRT) as a predictor in this regression.
The overall $R^2$ value for the multiple regression analysis was 0.3452, indicating that the predictors explain about a third of the variance in the overall data. After a Bonferroni correction for multiple comparisons across all regression studies included in this paper [@hochberg_sharper_1988], the demographics factor is the only significant predictor of hyperedge cardinality. The normalized $R^2$ changes for hypergraph cardinality can be seen in Figure \[fig8\]; the demographics factor has the largest normalized $R^2$ change and the only significant $p$-value ($<0.005$) in the regression. These results correspond with the marginal result from the multi-task data set, where the demographics factor is a marginally significant predictor.
![**Age-memory $R^2$ changes:** Normalized $R^2$ changes with respect to hypergraph cardinality across individuals in the age-memory study. The largest normalized $R^2$ changes are from the demographics factor and head motion measure. The demographics factor is the only significant predictor of hypergraph cardinality, which we denote with a bold outline.[]{data-label="fig8"}](Fig8.pdf){width="0.9\linewidth"}
Much of the variation in the demographics factor (73.5%) is directly attributable to age. We attempt to isolate the specific relationship between age and hypergraph cardinality by performing a separate regression. In this regression, hypergraph cardinality is the dependent variable and the independent variables are age and head motion. The relationship between age and hypergraph cardinality is significant, with the $t$-test $p$-value well below the Bonferroni correction over all regression analyses presented in this work, at $p <0.001$.
This is a positive relationship, indicating that older individuals tend to have higher hypergraph cardinality, while younger participants tend towards lower hypergraph cardinality. An illustration of this correspondence between hypergraph cardinality and age is presented in Figure \[fig9\]. As age increases, the number of hyperedges in a participant’s hypergraph increases as well. We verify that this relationship holds beyond this particular study by reintroducing the word-memory data from the multi-task study and performing a correlation between hypergraph cardinality and age over both studies. Age and hypergraph cardinality have a Spearman correlation coefficient of $\rho = 0.32$, and the $p$-value for this correlation, $ p <10^{-5}$, is significant when we use the Bonferroni correction over all analyses presented in this paper.
![**Hypergraph cardinality and age:** Scatter plot of hypergraph cardinality as a function of age for the age-memory data set and word memory task of the multi-task data set. The correspondence between increasing age and larger hypergraph cardinality can be observed.[]{data-label="fig9"}](Fig9.pdf){width="0.9\linewidth"}
Discussion {#discussion .unnumbered}
==========
Improving our understanding of the drivers of individual differences in functional brain imaging data can give insight into the mechanisms that lead to individual behavior. Dynamic FC has been used over groups to explain changes in the brain attributed to individual differences in learning [@bassett2011dynamic; @bassett2013task; @mantzaris2013dynamic]. Hypergraphs in particular have been used to analyze how long-term learning impacts the functional network structure [@bassett2013task] and how the brain switches between cognitive states [@davison_brain_2015]. Here, we extend previous dynamic FC studies that showed common properties of the dynamics at the level of the group [@davison_brain_2015] to investigate the drivers of strong individual variations in certain hypergraph metrics.
Disparate Sources of Variability in Hypergraph Structure {#disparate-sources-of-variability-in-hypergraph-structure .unnumbered}
--------------------------------------------------------
As we showed in the Multi-Task Analysis, the hypergraph cardinality varies widely across individuals, but is consistent between task states. Previous work on the multi-task data set found that the probability for hypergraphs to appear in a particular network configuration over individuals was significantly different depending on task state [@davison_brain_2015]. Consistent spatial organization rules for each task existed at the level of the group. There were similarities in the spatial arrangement of hyperedges in the brain for differing tasks, but certain properties were found to vary significantly between tasks. Brain areas in the occipital lobe in particular were highly likely to participate in the hypergraph network across individuals and across tasks, likely due to the visual nature of most of the cognitive tasks studied.
Here, we study hypergraph cardinality, which displays high variability across individuals and consistency across tasks within individuals (Figure \[fig4\]). This indicates that hypergraph cardinality serves as an individual signature of a subject’s brain dynamics. The similarities across subjects in the spatial distributions of hypergraphs described in [@davison_brain_2015] capture information orthogonal to the information summarized by hypergraph cardinality. For example, there are some individuals for whom the visual brain regions are linked by many hyperedges, and some for whom those same regions are linked by relatively few hyperedges, but these regions are more likely than others to be included in hypergraphs in the majority of subjects. This suggests that, for some subjects, brain regions tend to be more dynamically integrated in general, with co-varying functional relationships across many brain circuits; in other subjects, connectivity dynamics are more fragmented across the brain.
The high degree of variability in hypergraph cardinality across subjects and consistency within subjects, combined with the significant differences in spatial hyperedge arrangement across tasks, indicate that hypergraphs are a useful analysis tool for investigating both individual and task-based differences in brain function in a variety of settings. At the same time, hypergraphs can provide a view of dynamic patterns that complements other commonly used DFC methods. For example, many FC methods exclusively investigate the structure of strong correlations in functional data [@hermundstad2014structurally; @bassett_robust_2013; @greicius2009resting; @bassett2008hierarchical]; hypergraph analysis captures information about both strongly and weakly correlated dynamics and how sets of brain regions transition between them [@bassett_cross-linked_2014].
Although they are highly informative, many of the hypergraph metrics we study here are representative measures that greatly reduce the complexity of the hypergraph and only reveal a small part of the information contained in its structure. Further development of methods to utilize more of the information that hypergraphs provide will allow characterization of the consistency of particular hyperedges and dynamic modes, an understanding of which are important for behavior, or influenced by demographics or disease. Future work is also needed to further quantify the spatial differences in hypergraph arrangement across both individuals and tasks, to clarify the extent of overlap between the two types of information, and to determine whether the individual variability in cardinality can be mapped to individual spatial differences in hypergraph structure.
Relationship Between Age and Changes in DFC Networks {#relationship-between-age-and-changes-in-dfc-networks .unnumbered}
----------------------------------------------------
FC studies have established clear trends associated with aging, including a decrease in connectivity within functional networks and an increase in connectivity across different functional networks in resting and task states [@geerligs_reduced_2014; @madden_adult_2010; @meunier_age-related_2009; @ng_reduced_2016; @betzel_changes_2014]. Many of these studies have considered resting-state FC, because the absence of task stimulus provides a simple and reliable setting for comparison between subjects [@ferreira_resting-state_2013], although recent studies have successfully used FC networks to study various cognitive proceses [@medaglia_cognitive_2015]. The default mode network (DMN) and similar resting-state analyses may miss functional changes evoked by task states; while the DMN FC decreases with age, task-related sensorimotor network FC has been shown to increase with age [@song2014age; @geerligs2015brain]. Similarly, FC in memory tasks shows increased segmentation with age [@sala-llonch_changes_2014]. Extending these analyses to incorporate the dynamics of functional interactions is a necessary step towards quantifying individual changes in functional brain dynamics associated with age.
Several efforts have been made to capture individual age-related differences with methods from dynamic FC. Dynamic community structure and amplitude of low-frequency fluctuation of FC were both found to be strongly correlated with age, illustrating that functional dynamics are closely linked with aging [@qin2015predicting; @Schlesinger2016age]. In the dynamic community detection analysis, functional communities were found to be more fragmented with age, which agrees with the hypergraph cardinality result presented here [@Schlesinger2016age]. A multi-scale community detection analysis uncovered similar fragmentation with age for small scales [@betzel2015functional]. Our finding that hypergraph cardinality also increases with age aligns with this result and provides further information based upon its ability to capture higher-order dynamic patterns across larger ensembles of brain regions. Not only do the functional similarities of communities of brain regions themselves become less distinct as humans age, but the temporal profiles of these functional similarities also become less integrated across brain regions. The agreement of this result with known age-related changes in FC [@dennis2014functional; @contreras2015structural; @sala2015reorganization; @ferreira2015aging; @ng_reduced_2016] demonstrates the ability of hypergraph methods to capture and quantify major brain changes. Moreover, since the hypergraph analysis is not limited to strong correlations, our analysis further suggests that age is related not only to the organization of functional activity in groups of brain regions with strongly coherent activity, but also to the coordination between groups of regions that transition from being strongly to weakly correlated over time (or [*vice versa*]{}).
The reported correspondence between age and hypergraph cardinality is significant in the age-memory data set, but our analysis did not include data that could verify this relationship for cognitive tasks other than the word memory task. Although memory is a cognitive ability known to decline with age in many individuals, it is unlikely that the specific task studied in the age-memory data set drives this result. Rather, the consistency of hypergraph cardinality across tasks seen in the multi-task data set in Figure \[fig4\](B) suggests that similar hypergraph cardinalities may be found during other tasks in data sets with higher age variability, and that the relationship between age and cardinality is unlikely to depend primarily on the behavioral task. Further investigation is needed to determine whether individual differences in hyperedge structure have any significant relationship to behavioral or cognitive performance on any particular task.
Conclusion {#conclusion .unnumbered}
==========
Here, we have shown that the considerable differences in functional connectivity dynamics across individuals are closely linked with age. The hypergraph method is presented as a complex analysis tool that captures information about group-level similarities that differ between task states as well as individual differences that are consistent within individuals, across tasks. Further investigation into a single hypergraph metric (hypergraph cardinality) that varies across individuals uncovers a significant relationship between hypergraph cardinality and age. Specifically, there are a greater number of hyperedges in older individuals’ hypergraphs, suggesting that there are more small groups of regions with cohesively evolving dynamics and indicating a loss of coherence across larger, spatially distributed intrinsic functional connectivity networks. This complements widely reported relationships between FC and human aging by providing new insight into how FC activity and the co-evolution of FC activity are altered with increasing age, including the loss of large groups of co-evolving brain regions in older individuals. The correspondence with and extension of classic FC results to new dynamic regimes, along with the unique capacity of hypergraphs to probe multiple dimensions of both strong and weak dynamic variability, show that hypergraph analysis is a valuable tool for understanding age-related changes and other individual differences in dynamic brain function.
Supporting Information {#supporting-information .unnumbered}
======================
The following information is included in this supplementary document to support the claims presented in the main work:
1. A discussion of observed effects of concatenating multiple time series.
2. Figure \[figS1\] and Figure \[figS2\]: Cumulative size distributions for several methods for minimizing the effect of concatenation.
3. A discussion of the effect of time series length on hyperedge size distributions for the age-memory data set.
4. Tables \[Table\_All\], \[Table\_MT\], and \[Table\_AM\]: Tables of individual difference measures grouped by category for the full analysis, multi-task data, and age-memory data.
5. Figure \[figS3\]: $R^2$ changes for the task-specific hypergraph cardinality regression analysis.
#### Methodological Considerations. {#S1_Appendix .unnumbered}
**Edge Effects in Task Concatenation:** In this paper, we investigate dynamic functional connectivity changes across multiple cognitive tasks and two separate imaging data sets. In order to capture changes across tasks in the multi-task data set, we concatenate the time series for all tasks, as in [@davison_brain_2015]. In our analysis of the age-memory data, we concatenate time series from three functional runs of the word memory task, and remove time windows from the ends of the time series of each task to reduce edge effects. Edge effects appear to be confined to the data points adjacent to the beginning and end of each run, but we remove the full $N \times N$ adjacency matrix to ensure we are not including any edge effects in the analysis. The resulting change in the cumulative size distribution is depicted in Figure \[figS1\]. With the edge blocks removed, there are fewer system-size hyperedges and more small hyperedges.
Figure \[figS1\] includes a comparison with another method for treating edge effects. In this case, the time series data for each of the three tasks is filtered separately before concatenation. This approach dramatically reduces the number of hyperedges. If filtering is responsible for introducing edge effects that drive hyperedges, the number of hyperedges are likely to increase when we employ this method. Instead, only 13 subjects had non-singleton hyperedges. We choose to not analyze these results further because there are too few subjects with hyperedge data.
Two further efforts to understand the effects of concatenating across functional runs on the cumulative size distribution are depicted in Figure \[figS2\]. In the trial-by-trial analysis, we performed the hypergraph method separately on each edge time series (10 data points each) for the three trials. Only 30 subjects have significant non-singleton hyperedges in at least one of the three trials and the number of large hyperedges is much lower than the original result. This decrease may be a result of our removal edge effects, but it is likely the shorter task length is driving the difference, as we discuss in the next section. To explicitly investigate the effect on the size distribution caused by each transition, we also split the time series data into three sets of 18 edge time series data points. The first includes the transition between the first and second trials, the last includes the transition between the second and third trials, and the middle includes both transitions. These distributions are also plotted in Figure \[figS2\]. We see that the overall number of hyperedges is greater than both the original age-memory hypergraph over all individuals, which is driven by a decrease in the number of system-size hypergraphs in the 18-split analysis. The distributions for all three follow similar patterns, indicating there is not a large discontinuity in the pattern of the distribution when we include both transitions.
![**Edge compensation comparison:** Cumulative size distributions for the original age-memory data set (with no changes to remove effects of the edges) and two methods for removing potential effects from the edges. The “edge blocks removed" method is used in all analyses in the main text.[]{data-label="figS1"}](S1_Fig.pdf){width="0.9\linewidth"}
![**Trial separation comparison:** Cumulative size distributions for two different methods for separating edge effects. In the trial-by-trial method, hypergraphs are constructed separately for each trial, while in the 18-split analysis, hypergraphs are constructed from the first, middle, or last 18 edge time series data points.[]{data-label="figS2"}](S2_Fig.pdf){width="0.9\linewidth"}
#### Methodological Considerations. {#S2_Appendix .unnumbered}
**Edge Time Series Length in Hypergraph Construction:** When we construct hypergraphs from the much shorter single task measurements within the multi-task data set, the number of large hyperedges is greatly reduced, with fewer hyperedges in the population near the system size (see Panel A of Figure \[fig6\]). We see a similar effect when we compare the distributions seen in Figure \[figS2\] for the split data sets. The trial-by-trial hypergraphs contain fewer hyperedges overall and far fewer system-size hyperedges than the 18-split hypergraphs. However, this increase is not driven by inclusion of the transitions alone, since the middle 18-split hypergraph contains approximately half the number of system-size hyperedges when compared to the full analysis. Since both hypergraphs are constructed across both transitions, this indicates that the edge time series length is more influential to population-level hypergraph properties than concatenation.
Further work is needed to elucidate the relationships between hyperedge size and the overall length and composition of the data set. Additionally, it remains to be determined whether there is an analogue to the scan length proposed for reliable FC estimates [@birn2013effect]; an edge time series length that ensures minimal fluctuations in the size distributions for longer scans. However, the very close correspondence between small-size hyperedges found during the word memory task in both data sets suggests that these hyperedges are capturing important characteristics of the dynamics within this task that are robust across imaging sessions and populations.
Performance (Word) Demographics Personality Cognitive Factors State of Mind
----------------------- ------------------- ------------- -------------------- ---------------------------
Criterion shift score Age PANAS (6) OSIQ-S/O Arrival time
Liberal Dprime Sex Big 5 (5) VVQ-W/P Meal (hours since)
Conservative Dprime Education (years) BIS/BAS (4) Need for cognition Hours of sleep
Overall Dprime Dominant hand SBCSQ visual Physical/mental comfort
SBCSQ verbal Beck Depression Inventory
Paper folding Alcohol (Y/N)
Card rotation Exercise (Y/N)
Smoking (Y/N)
Caffeine (Y/N)
: **Common behavioral measures in both data sets:** Categories containing measures of interest (42). For the state of mind measures, (Y/N) indicates measures where participants were asked whether they had performed the activity in the past 24 hours.[]{data-label="Table_All"}
Performance Demographics Personality Cognitive Factors State of Mind
-------------------- --------------- ------------- ------------------- ------------------------
Attention CS Military rank EPQ-R (4) Working memory MSW/MSF
Face memory CS Vocabulary test PTSD Score
Attention Dprime PTSD (Y/N)
Face memory Dprime Concussion score
Concussion 5 inventory
: **Additional behavioral measures in multi-task data:** Categories containing measures of interest. For the state of mind measures, (Y/N) indicates measures where participants were asked whether they had performed the activity in the past 24 hours.[]{data-label="Table_MT"}
Performance Demographics Personality Cognitive Factors State of Mind
--------------- -------------------- ------------- ------------------- ----------------------
Hit rates Height Distracted Stressed (Y/N)
Failure rates Weight Motivated Days since period
Reaction time Contraceptive use Usual hours of sleep
Children (Y/N) Drugs past 48h (Y/N)
Number of children MMSE (dementia)
: **Additional behavioral and brain measures in age-memory data:** Categories containing measures of interest. For the state of mind activity measures, yes indicates measures where participants were asked whether they had performed the activity in the past 24 hours. Questions about daily, weekly, and monthly amounts of activity, including whether activity in the past 24 hours were more or less than usual were also recorded for all (Y/N) state of mind activities in the age-memory study.[]{data-label="Table_AM"}
![**Task-specific multi-task $R^2$ changes:** Normalized $R^2$ changes with respect to task-specific hypergraph cardinality for each of the four task-specific hypergraphs. Rest-specific hypergraph cardinality is included as an independent variable for the other three tasks and is the only significant predictor, which is denoted with a bold outline.[]{data-label="figS3"}](S3_Fig.pdf){width="0.9\linewidth"}
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank John Bushnell for technical support.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Marketplaces specializing in malicious hacking products - including malware and exploits - have recently become more prominent on the darkweb and deepweb. We scrape 17 such sites and collect information about such products in a unified database schema. Using a combination of manual labeling and unsupervised clustering, we examine a corpus of products in order to understand their various categories and how they become specialized with respect to vendor and marketplace. This initial study presents how we effectively employed unsupervised techniques to this data as well as the types of insights we gained on various categories of malicious hacking products.'
author:
-
bibliography:
- 'IEEEabrv.bib'
- 'bib.bib'
title: Product Offerings in Malicious Hacker Markets
---
**Acknowledgments.** Some of the authors were supported by the Office of Naval Research (ONR) Neptune program, the Arizona State University Global Security Initiative (ASU GSI), and CNPq-Brazil.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Usual formulations of the clustering coefficient can be shown to be insufficient in the task of describing the local topology of very simple networks. Motivated by this, we review some alternatives in order to present an extension, the [*clustering profile*]{}. We show, both conceptually and through applications to well studied networks, that this measure is a more complete and robust measure of clustering. It imposes stringent constraints on theoretical growth models, specially on aspects of the network structure that play a central role in dynamics on networks. In addition, we study how it provides a richer perspective of phenomena such as hierarchy, small-worlds and clusterization.'
author:
- 'Alexandre H. Abdo'
- 'A. P. S. de Moura'
bibliography:
- 'abdo.bib'
title: Clustering as a measure of the local topology of networks
---
Introduction
============
Physicists have been greatly interested in studying the interplay between the topology and growth dynamics of complex networks, for they are a universal framework to understand various processes in previously distant areas ranging from cell metabolism to linguistics[@graphshandbook-editors; @evolutionnetworks-dorogovtsev]. The unusual nature of their structure, represented by graphs[@graphs-diestel], required the development of novel methods, and one of the core difficulties has been to understand what measurements are robust and indeed represent universal quantities by which we can compare and classify networks, or assert how effectively a network growth model reflects the network it tries to mimic.
Average distance, degree distribution and degree correlations[@networksreview-newman], loopiness[@loopiness-bianconi], motifs[@motifs-milo] and the clustering coefficient[@smallworld-watts] are some quantities which have established themselves as useful. But, among them, only the latter gives any description of the local topology — how the network is organized close to some vertex — measuring how many of its neighbors are also neighbors among themselves. It has also been established that local structure is not only an important topological characteristic, but also a main concern when studying searchability[@identitysearch-watts; @freenet-zhang] and dynamics[@epidemicstructured-eguiluz; @clusteringdisease-cross] on networks.
Still, important as it stands, the usual clustering coefficient has serious limitations, and is commonly replaced by *ad hoc* definitions. We first focus on these shortcomings, as they’ll let us understand why so many variations emerged, and by summarizing this variations propose a consistent and robust improvement, which should provide for a more complete and less specialized characterization of the local topology of networks.
We remark that our focus is not on clustering as transitivity[@serrano-generalclustering], but as a general measure of small-scale structural organization.
Background
==========
When first introduced[@smallworld-watts] to formalize the small-world phenomena, the clustering coefficient was defined as the fraction of connected pairs among the neighbors of a vertex, averaged over all vertices.
Following that, it was also defined by taking the fraction of connected pairs over all pairs of neighbors of all vertices[@generatingfunctions-newman], since that frequently is more tractable. In this form, it was used to derive analytical expressions that distinguish social networks from other networks[@socialnetworks-newman].
A first problem with the coefficient is that these two seemingly close definitions may yield very different results[@handbook-bollobas], as can be seen from the fact that the individual ratio for vertices of high degree has greater impact on the second case.
Later, a better description of clustering was introduced by considering the coefficient as a function of vertex degree, and doing so revealed important structural features such as hierarchical organization[@hierarchical-barabasi].
Another interesting consequence is that this also removes the ambiguity of the first two definitions: if averages are restricted to a set of vertices of same degree, averaging the coefficient over vertices or calculating the fraction of connections over all pairs of neighbors is the same thing.
Still, other definitions have become necessary, as in the case of bipartite networks used in the study of sexually transmitted diseases[@bipartite-herrmann], where no odd cycles exist and thus, because connected neighbors of a vertex would form cycles of length $3$, the clustering coefficient is always zero, even though a bipartite network may have a very complex local structure.
The underlying issue becomes more evident when noting that both a square lattice and a single large cycle, which are locally very different networks, have their clustering coefficient equal to zero.
We take this as evidence that a general treatment of clustering cannot expect to rest on a simple scalar quantity; however, it should remain a local property that can be calculated for each vertex, and comparable among them. Thus the source of the problem seems to be the question asked, of “how many of my neighbors are connected?”, instead of “how closely related are my neighbors?”, which we think better comprehends the concept of clustering as a measure of local topology.
In order to implement this idea we turn back to the original definition of clustering for a vertex, the fraction of connected pairs among its neighbors, and notice that “connected” can stand for “whose distance is $1$”. Then the usual clustering coefficient can be understood as the first term of a sequence, the term accounting for the fraction of pairs of neighbors whose distance is $1$. The second term would be the fraction of pairs of neighbors whose distance is $2$, and so on. Now, since those are all neighbors of the same vertex, they are always connected by a path of length $2$ going through that vertex, and so we must discard paths going through the vertex in question when calculating this distance between its neighbors.
These higher order clustering coefficients were first explored by A. Fronczak *et al.*[@higherorder-fronczak] in order to show, in a stronger sense, that the model of preferential attachment[@scalefree-barabasi] is blind to clustering mechanisms.
The meaning of this extended clustering can, perhaps, be more clearly understood in terms of cycles: the $n$th term of the sequence would correspond to the fraction of pairs of neighbors of a vertex whose smallest cycle shared with it has length $n+2$; as can be seen on FIG. \[fig-cp\].
![\[fig-cp\]Higher orders of clustering, $C^{d}(v)$: the fraction of pairs of neighbors whose smallest cycle shared with the vertex has length $d$. In the example above, $v$ has $10$ pairs of neighbors and a non-zero clustering for $d=1,2$. $C^{1}(v)$ is, by definition, the usual clustering coefficient.](fig-cp.eps){width="\columnwidth"}
So, by combining these dependencies on degree and distance, we obtain a measure that is neither ambiguous nor specialized; which not only applies to all networks, but is a richer description of their local topology. For practical purposes, instead of referring to higher order clustering coefficients averaged as a function of degree, we name this quantity the clustering profile, and proceed to a formal definition.
Clustering Profile
==================
In a network $G$ composed of vertices $V$ and edges $E$, so that $G=(V,E)$, we denote the clustering for a vertex $u\in V$ as $C^{d}(u)$, defined as the number of pairs of neighbors of $u$ whose distance in the induced network $G(V\setminus \{u\})$ is $d$, divided by the total number of pairs of neighbors of $u$. Thus
$$C^{d}(u) = \frac{ |\{ \{ v,w \}; v,w \in N(u) | d_{ G( V \setminus \{ u\} ) }(v,w) = d \} | }{ { |N(u)| \choose 2} } ,$$
where $N(\cdot )$ is the set of neighbors of a vertex, the modulus $|\cdot |$ represents the cardinality (number of elements) of a set, and so $|N(u)|$ is the degree of $u$, also denoted $deg(u)$.
This leads to a generalized description of how the network is organized around that vertex, reflecting the contribution of more distant neighbors in higher terms, while still preserving the good property that, when summed over all terms, it ranges between $0$ and $1$.
We can then define the clustering profile for a network, being the average of $C^{d}(u)$ over all vertices of same degree $k$, and denote it
$$C^{d}_{k} = \frac{ \sum_{ \{u|deg(u)=k\} } C^{d}(u) }{ |\{u|deg(u)=k\}| } .$$
It should be noticed that the usual clustering coefficient as a function of degree is simply $C^{1}_{k}$.
And although numerically calculating the clustering profile is a more expensive computer operation than calculating the usual coefficient, each step of the calculation is parallelizable, so even large networks can be treated with relatively small computer resources.
Applications
============
In order to illustrate the consequences of the clustering profile, we choose a well known network: the set of metabolic networks of bacteria first studied in references [@largescalemetabolic-barabasi; @hierarchical-barabasi], where a growth model[@hierarchical-barabasi] is provided to illustrate their hierarchical organization. Comparing the results for the real network and this hierarchical model will help us better understand the phenomena associated with this system. We first focus on the profile’s absolute value, then its variation with distance, and finally its variation with degree, as each of these will have distinct implications.
Small-worlds
------------
\[Bc\]\[Bc\][$k$]{} \[Bc\]\[Bc\][$C^{d}_{k}$]{} \[Br\]\[Br\]\[0.7\][$d\ 1$]{} \[Br\]\[Br\]\[0.7\][$2$]{} \[Br\]\[Br\]\[0.7\][$3$]{} \[Br\]\[Br\]\[0.7\][$4$]{} \[Br\]\[Br\]\[0.7\][$5$]{} \[Br\]\[Br\]\[0.7\][$sum$]{} ![\[fig-met-vdegree-noscale\]The well studied metabolic network, above, has a richer local topology than one can see with the usual clustering coefficient, since $C^{1}_{k}$ decreases quickly ($d=1$) with degree, while deeper measures ($d=2,3$) reveal the structure present even in high degrees. \[fig-hier-vdegree-noscale\]The hierarchical model, shown below, has a much simpler structure and does not reflect the importance of distance $2$ relationships present in the real network.](fig-met-vdegree-noscale.eps "fig:"){width="\columnwidth"} ![\[fig-met-vdegree-noscale\]The well studied metabolic network, above, has a richer local topology than one can see with the usual clustering coefficient, since $C^{1}_{k}$ decreases quickly ($d=1$) with degree, while deeper measures ($d=2,3$) reveal the structure present even in high degrees. \[fig-hier-vdegree-noscale\]The hierarchical model, shown below, has a much simpler structure and does not reflect the importance of distance $2$ relationships present in the real network.](fig-hier-vdegree-noscale.eps "fig:"){width="\columnwidth"}
FIG. \[fig-met-vdegree-noscale\](metabolic) shows the clustering profile for the metabolic network. We note that based on the usual coefficient alone ($d=1$), one could suggest that more connected metabolites live in an almost unclustered world.
In contrast, the profile not only shows that more pairs of neighbors are related by the second order of clustering than the first, but makes it clear that, as degree raises, clustering only migrates from distance one and two to distance three and four.
In fact, we verify that the sum of $C^{d}_{k}$ up to $d=5$ is close to $1.0$ for all $k$, which implies that almost every two neighbors of a vertex share with it a cycle of length in the order of the network’s average distance.
Now, as can be seen from FIG. \[fig-hier-vdegree-noscale\](model), the growth model shares neither of these properties.
Contrary to the metabolic network, distance $2$ does not play a major role on its local topology. Also, the model’s sum of all orders of the clustering profile quickly converges to $\frac{1}{2}$ for vertices of high degree, meaning that close to them it has a very different organization from that of the real network, being more open or tree-like.
So, although these networks are both considered small-world networks, in the sense that they have small average distance while maintaining a significant usual clustering coefficient, our observations with the clustering profile sets them clearly apart: there are networks where almost every vertex experiences, in a generalized sense, the small world phenomena, meaning their neighbors are all closely related, while on other networks this effect is restricted to a subset, out of which nodes have groups of neighbors who, if not for them, would live in distant clusters.
This is not only evidence of very different network-growth processes, but we note that this distinction reflects a strikingly different local topology around the high degree vertices, also called “hubs” in the literature, which play a central role in many dynamical processes, notably in disease spreading[@epidemicstructured-eguiluz] and information retrieval[@searchpowerlaw-adamic]. Therefore, establishing this distinction is an important step towards a more structured understanding of both the growth of, and the dynamics on, small-world networks.
From the arguments above, it is useful to define “complete-small-world networks” as those networks with small average distance and, for all degrees, the sum of their clustering profile up to an order of that average distance close to 1. As we noted for the metabolic network, this is equivalent to requiring that vertices share short cycles with most of their pairs of neighbors, regardless of degree.
Hierarchy
---------
\[Bc\]\[Bc\][$k$]{} \[Bc\]\[Bc\][$C^{d}_{k}$]{} \[Br\]\[Br\]\[0.7\][$d\ 1$]{} \[Br\]\[Br\]\[0.7\][$2$]{} \[Br\]\[Br\]\[0.7\][$3$]{} \[Br\]\[Br\]\[0.7\][$4$]{} \[Br\]\[Br\]\[0.7\][$5$]{} ![\[fig-met-vdegree-dilog\]In a log-log scale, we can see the metabolic network has a wide range with power-law behavior even for higher distances. \[fig-hier-vdegree-dilog\]The model, however, lacks this deep hierarchical structure, and has a zero profile for orders above $3$.](fig-met-vdegree-dilog.eps "fig:"){width="\columnwidth"} ![\[fig-met-vdegree-dilog\]In a log-log scale, we can see the metabolic network has a wide range with power-law behavior even for higher distances. \[fig-hier-vdegree-dilog\]The model, however, lacks this deep hierarchical structure, and has a zero profile for orders above $3$.](fig-hier-vdegree-dilog.eps "fig:"){width="\columnwidth"}
Fig.\[fig-met-vdegree-dilog\] shows the same data scaled to a log-log scale, in order to visualize the power-law behavior of clustering which characterizes hierarchy. We note that, on the metabolic network, $C^{d}_{k}$ for all $d$ varies as a power law in $k$ over a wide range of degrees, indicating a deeper hierarchical structure than previously known.
For the model, however, other than the usual clustering coefficient $C^{1}_{k}$, all additional orders differ from the behavior of the real network: $C^{2}_{k}$ and $C^{3}_{k}$ can hardly be considered hierarchical, and all orders greater than $3$ are constantly equal to zero.
While clearly this model was crafted only to illustrate the idea of hierarchy, we have been able to spot another important feature of the metabolic network’s topology, its deep hierarchical structure, that is missing from the model and would be relevant when studying, for example, flow-dynamics[@jamming-zoltan; @gradientnetworks-zoltan] in bacteria metabolism.
Clusterization dynamics
-----------------------
At last, we consider the behavior of the profile as a function of distance, for specific ranges of degrees, in order to examine the change in influence of growth dynamics over varying orders of clustering. We choose not to use the hierarchical model this time because, its profile being zero beyond order $3$, it is of questionable significance to consider its variation.
Instead, we compare the metabolic network with another small-world network of similar global characteristics, the World Wide Web[@diameterweb-barabasi]. Both are scale-free dissortative[@mixing-newman] hierarchical[@hierarchical-ravasz] networks. Our purpose here is to ask if the evolution from one order to another might give more clues about their underlying clusterization dynamics, since these networks are similar in every other aspect.
However, given that the absolute value of clustering depends on degree, there is little reason to suppose its variation with order would remain the same. Therefore, we split each network into three fractions: small degrees, medium degrees and large degrees, based on the behavior of the degree variation for the clustering profile on these networks. Medium degrees are those where it behaves well as a power law for various orders, small and high degrees are those under and above that range, respectively.
We verify that, for low and medium degrees, these networks show the same behavior, namely an exponential decrease of the coefficients with increasing order. But as we see on FIG. \[fig-vdistance\], for high degrees the metabolic network remains exponential, while the WWW behaves as a power law, revealing a structural change around its most connected vertices.
\[Bc\]\[Bc\][$d$]{} \[Bc\]\[Bc\][$<C^{d}(u)>$]{} \[Bc\]\[Bc\][$d^2$]{} \[Bc\]\[Bc\][$<C^{d}(u)>$]{} ![\[fig-vdistance\]Variation with distance, in log-log scale, of the average clustering over vertices with high degree. The metabolic network shows exponential behavior (see inset in monolog scale), while the WWW appears as a power law.](fig-vdistance.eps "fig:"){width="\columnwidth"}
This adds a new approach for evaluating and improving various network growth models that use community structure[@highlyclustered-newman], edge addition bias[@triadclustering-holme; @growingsocial-jin] and other strategies to explain the diverse mechanisms of clustering formation found in natural, social and technological networks. In the present case, it suggests a mechanism of clusterization dynamics more sensible to degree variation for the WWW.
We also observe that the first one or two orders of clustering on these networks do not uniformly follow such scales. In the metabolic network the first order is always lower than extrapolation would suggest and, for high degrees, so is the second order. In the WWW, however, that only happens on high degrees, and only to the first order of clustering, as can be seen in the picture.
It is not clear why such a deviation occurs, but we suppose it is a consequence of superimposed dynamic rules affecting clusterization.
On the metabolic network, we suggest this effect might be related to selective pressure against congestion[@jamming-zoltan; @jamming-park] of metabolic pathways, since it is always present and affecting deeper levels of clustering, with the effect getting stronger towards highly connected metabolites.
As for the WWW, this poses one interesting question: whether the competing rule is only suppressing the lower orders of the profile on high degrees, or whether its interaction with the dynamics is also the cause of the profile’s deviation from exponential behavior.
We intend to pursue these questions on a future study of network growth models.
Conclusions
===========
In this paper we reviewed some issues, limitations, and alternatives to the clustering coefficient, a measure which is a pivot concept in the field of complex networks. Through a more comprehensive formulation of this concept, new insights were given for networks well studied in the literature, specially when considering network growth processes. Not limited to that, the new measure presents us a broader view of usual phenomena related to network structure dynamics, such as the emergence of clustering, hierarchy, and the small-world effect. It also seems to be most consequential for well connected vertices, which are central in many networked processes. As a final statement, we believe that less specialized, richer descriptions will allow local topology to play a more significant role in understanding the interaction between structure and dynamics on networks.
Material and methods
====================
The data used in this paper for the metabolic network and the WWW was that available on the website of the CCNR at Univ. of Notre Dame, <http://www.nd.edu/~networks/> . The metabolic network was reduced according to the procedure in the supplemental online material of reference [@hierarchical-barabasi].
Non scaled graphics for the profile (Fig. \[fig-met-vdegree-noscale\]) were rebinned, log-log graphics for the profile (Fig. \[fig-met-vdegree-dilog\]) were log-rebinned, and the error bars are those from the rebinning process.
Higher order clustering coefficients and the clustering profile were calculated with use of the *graph-tool*, which is publicly available as *Free Software*[^1] at <http://graph-tool.forked.de/> .
The authors thank Tiago Peixoto and Al Scandar Solstag for useful conversations and advice. This work was funded in part by Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP, process 2005/00933-0) and CNPq.
[^1]: Under the terms of the GNU General Public License
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'It is known by a formula of Hasse-Sondow that the Riemann zeta function is given, for any $s=\sigma+it \in \mathbb{C}$, by $\sum_{n=0}^{\infty} \widetilde{A}(n,s)$ where $$\widetilde{A}(n,s):=\frac{1}{2^{n+1}(1-2^{1-s})} \sum_{k=0}^n \binom{n}{k}\frac{(-1)^k}{(k+1)^s} .$$ We prove the following approximate functional equation for the Hasse-Sondow presentation: For $\vert t \vert = \pi xy $ and $2y \neq (2N-1)\pi $ then $$\zeta(s)= \sum_{n \leq x } \widetilde{A}(n,s)+\frac{\chi(s)}{2^{s-1}-1} \left (\sum_{k \leq y} (2k-1)^{s-1} \right ) +O \left (e^{-\omega(x,y) t} \right ),$$ where $0 <\omega(x,y)$ is a certain transcendental number determined by $x$ and $y$. A central feature of our new approximate functional equation is that its error term is of exponential rate of decay. The proof is based on a study, via saddle point techniques, of the asymptotic properties of the function $$\widetilde{A}(u,s):= \frac{1}{2^{u+1} (1-2^{1-s}) \Gamma(s)} \int_{0}^{\infty} \left ( e^{-w} \left ( 1- e^{-w} \right)^u \right ) w^{s-1} dw,$$ and integrals related to it.'
address: 'Yochay Jerby, Faculty of Sciences, Holon Institute of Technology, Holon, 5810201, Israel'
author:
- Yochay Jerby
title: An approximate functional equation for the Riemann zeta function with exponentially decaying error
---
Introduction
============
The classical approximate functional equation for the Riemann zeta function was proved by Hardy and Littlewood in the series of works [@HL; @HL2; @HL3]. Let $A(n,s):=n^{-s}$ and $s=\sigma +it$ with $0 \leq \sigma \leq 1$. The theorem states that the following holds $$\label{eq:HL-AFE}
\zeta(s)= \sum_{n \leq x } A(n,s) + \chi(s) \left ( \sum_{k \leq y } k^{s-1} \right ) + O \left ( x^{-\sigma} + x^{\frac{1}{2} - \sigma} y^{-\frac{1}{2}} \right ),$$ when $x,y \geq 1$ are such that $\vert t \vert = 2 \pi xy$ and $$\chi(s):=2^s \pi^{s-1}sin \left ( \frac{\pi s}{2}\right ) \Gamma(1-s)$$ is the function appearing in the functional equation $\zeta(s) = \chi(s) \zeta(1-s)$. An asymptotic expansion for the error term in (\[eq:HL-AFE\]), is further described by the celebrated Riemann-Siegel formula, published in 1932 by C. L. Siegel, based on original manuscripts of Riemann, see [@Si] and also [@E; @I; @T].
In 1994, Sondow proved, via Euler’s transformation of series, that the Riemann zeta function can be expressed as $$\label{eq:HS1}
\zeta(s) := \sum_{n=0}^{\infty} \widetilde{A}(n,s)$$ where $$\label{eq:HS2}
\widetilde{A}(n,s):=\frac{1}{2^{n+1}(1-2^{1-s})} \sum_{k=0}^n \binom{n}{k}\frac{(-1)^k}{(k+1)^{s}},
.$$ for any $s \in \mathbb{C}$, see [@SE]. Formula (\[eq:HS1\]) also appears in the appendix to a work by Hasse from 1930, where he proved results of similar nature, see [@H]. In the appendix, Hasse attributes the conjecture of formula (\[eq:HS1\]) to Knopp. In this work we prove the following analog of the classical approximate functional equation, for the Hasse-Sondow representation of zeta:
\[thm:main-theorem\] For $x,y \geq 1$ such that $\vert t \vert = \pi xy $ and $2y \neq (2N-1)\pi $ the following holds $$\label{eq:NEW-AFE}
\zeta(s)= \sum_{n \leq x } \widetilde{A}(n,s)+\frac{\chi(s)}{2^{s-1}-1} \left (\sum_{k \leq y} (2k-1)^{s-1} \right ) +O \left (e^{-\omega(x,y) t} \right ),$$ where $\omega(x,y)>0$ is a given transcendental number, uniquely determined by $x$ and $y$.
A central feature of our new approximate functional equation (\[eq:NEW-AFE\]) is that its error term, $O \left (e^{-\omega(x,y) t} \right )$, is of exponential rate of decay, in contrast to the fractional power rate of decay in the classical case. In particular, the main term of our new equation (\[eq:NEW-AFE\]) gives an approximation of $\zeta(s)$ which surpasses the level of accuracy of the asymptotic Riemann-Siegel formula, expanded to any order. The exact definition of $\omega(x,y)$ is given below.
The proof of Theorem \[thm:main-theorem\], as the proof of the classical approximate functional equation, (\[eq:HL-AFE\]), in [@HL], utilizes the Euler-Maclaurin summation formula of order one which states that $$\label{eq:EM}
\sum_{n=M}^N A(n,s) = \int_M^N A(u,s) du + \int_M^N \psi(u) \left ( \frac{\partial}{\partial u} A(u,s) \right ) du+\frac{1}{2} \left [ A(M,s)+A(N,s) \right ],$$ where $A(u,s)$ is a continuous function in the $u$-variable and $$\psi(u):=u - [u]-\frac{1}{2}.$$ In the classical Hardy-Littlewood case[^1] of (\[eq:HL-AFE\]), the function $A(u,s)=u^{-s}$ is readily taken. However, in our case of the Hasse-Sondow presentation, the required interpolating function turns to be $$\label{eq:GL-integral}
\widetilde{A}(u,s):= \frac{1}{2^{u+1} (1-2^{1-s}) \Gamma(s)} \int_{0}^{\infty} \left ( e^{-w} \left ( 1- e^{-w} \right)^u \right ) w^{s-1} dw.$$ Let us note that $\widetilde{A}(u,s)$ is related to, $\Phi^{\ast}_u(z,s,a)$, the generalized Hurwitz-Lerch zeta function of Goyal and Laddha, via $$\widetilde{A}(u,s)=\frac{\Phi^{\ast}_{-u}(1,s,1)}{2^{u+1}(1-2^{1-s})},$$ see [@GL]. In Section 2, we include a self-contained derivation of (\[eq:GL-integral\]) from the corresponding N[ø]{}rlund-Rice integral, see [@FS].
Applying the Euler-Maclaurin formula (\[eq:EM\]) to $\widetilde{A}(u,s)$, gives $$\label{eq:EM2}
\zeta(s)-\sum_{n \leq x} \widetilde{A}(u,s) = \int_x^{\infty} \widetilde{A}(u,s) du + \int_x^{\infty} \psi(u) \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s) \right ) du+\frac{1}{2} \widetilde{A}(x,s).$$
The proof of Theorem \[thm:main-theorem\], thus, relies on the asymptotic saddle-point approximation of the three integrals appearing in the right hand side of (\[eq:EM2\]). That is:
1. The function $\widetilde{A}(x,s)$ itself. It is shown that $\widetilde{A}(u,s)=O \left ( e^{- \omega(x,y) t} \right)$ in Corollary \[cor:2.3\].
2. The integral $\widetilde{I}_1(x,s):=\int_x^{\infty} \widetilde{A}(u,s) du $. It is shown that $$\label{eq:11}
\widetilde{I}_1(x,s):= \frac{\chi(s)}{2^{s-1}-1} \left (\sum_{k \leq y} (2k-1)^{s-1} \right ) +O \left (e^{-\omega(x,y) t} \right )$$ in Theorem \[thm:first-integral\]. The main sum in (\[eq:11\]) arises from pole contributions.
3. The integral $\widetilde{I}_2(x,s):= \int_x^{\infty} \psi(u) \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s) \right ) du$. In Theorem \[thm:int2\] it is shown that also $\widetilde{A}(u,s)=O \left ( e^{- \omega(x,y) t} \right)$.
A key feature is that the saddle points of all three integrals (1)-(3) are shown to be given as the solutions of the following transcendental equation $$e^{w}-\alpha \cdot i w -1=0.$$ The solutions $$w_k(\alpha)=x_k(\alpha)+y_k(\alpha) \cdot i$$ can be ordered by $k \in \mathbb{Z}$ according to the value of $y_k(\alpha)$. Moreover, the main contribution for the asymptotic behavior of the integrals is shown to come from the $k(\alpha)$-th saddle point, which is the saddle point whose $\vert x_{k(\alpha)}(\alpha) \vert$ value is minimal. Finally, the exponential decay rate of Theorem \[thm:main-theorem\] is given, in Corollary \[cor:2.3\], as follows $$\omega(x,y):=- \alpha \cdot log \abs*{\frac{1-e^{-w_{k(\alpha)}(\alpha)}}{2}} + \theta_{k(\alpha)}(\alpha)-\frac{\pi}{2},$$ where $$\alpha=\alpha(x,y):=\frac{t}{x}=\frac{1}{\pi y},$$ and $w_k(\alpha):=r_k(\alpha)e^{i \theta_k(\alpha)}$ is the polar presentation of the $k(\alpha)$-th saddle-point. In particular, the reason for excluding the case $2y \neq (2N-1)\pi $ in Theorem \[thm:main-theorem\] is that these are the degenerate values for which $\omega(x,y)=0$.
The rest of the work is organized as follows: In Section 2 we review the definition of $\widetilde{A}(u,s)$ and study its asymptotic properties. The integral $\widetilde{I}_1(x,s)$ is investigated in Section 3. The integral $\widetilde{I}_2(x,s)$ is considered in section 4 together with the conclusion of the proof of the main theorem. In Section 5 we present concluding remarks.
The Function $A(u,s)$ and its Asymptotic Properties {#sec:2}
===================================================
Recall that the $n$-th difference of a continuous function $f(u)$ is given by the sum $$D_n [f]:= \sum_{k=0}^n \binom{n}{k} (-1)^k f(k).$$ For instance, the $n$-th coefficient of the Hasse-Sondow presentation (\[eq:HS2\]), could be written as $$\label{eq:def}
\widetilde{A}(n,s)=\frac{D_n[A(v+1,s)]}{2^{n+1}(1-2^{1-s})},$$ where $A(v,s)=v^{-s}$.
In general, one observes that $n$-th differences satisfy the following trivial “naive” bound $$\vert D_n [f] \vert \leq 2^{n} \max_{0 \leq k \leq n} \vert f(k) \vert,$$ which implies, assuming that $f(u)$ is bounded, that $D_n[f]=O(2^n)$. In the mid 1960’s, the study of asymptotic properties of $n$-th differences, $D_n[f]$, was found to be of central importance for various questions in algorithm analysis related to the works of De Bruijn, Knuth and Rice, see for instance Section 5.2.2 and 6.3 of [@K]. As a result, various saddle-point techniques and countour integration methods were developed and applied for the study of such asymptotics. In particular, one of the central phenomena, observed by the algorithm analysis community, is that $n$-differences actually tend to exhibit asymptotic behavior that is essentially slower than $O(2^n)$, a phenomena known as exponential cancellation, see [@FS].
As mentioned in the introduction, in order to apply analytic methods for the investigation of the asymptotic behavior of $D_n[f]$, a continuous function, interpolating the values of $D_n[f]$ is required. In general, such a function is given via the N[ø]{}rlund-Rice integral as follows: If $f(v)$ is an analytic function in a domain containing $0<Re(s)$, one has $$Res_{v=k} f(v) \frac{n!}{v(v-1)...(v-n)}=(-1)^{n-k} \binom{n}{k} f(k),$$ for $0 \leq k \leq n$. Note that $$\frac{n!}{v(v-1)...(v-n)}= (-1)^{k-1} B(n+1,-v),$$ where $B(z,w)$ is the beta function. Hence, the $n$-th difference could be expressed in terms of the following N[ø]{}rlund-Rice integral formula $$D_n[f]=\frac{1}{2 \pi i} \int_C B(n+1,-v)f(v) dv$$ where $C$ is a contour of integration encircling $[0,n]$, avoiding integral values between zero and $n$. Moreover, under mild conditions, the N[ø]{}rlund-Rice integral further simplifies, and takes the following form $$\label{eq:NR} D_n[f] :=\frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} B(n+1,-v)f(v) dv,$$ for $c<0$. In our case we have:
\[prop:integral\] For $s=\sigma+it$ with $\sigma>1-u$ the required interpolating function of (\[eq:def\]) is given by $$\label{eq:GL}
\widetilde{A}(u,s):= \frac{1}{2^{u+1} (1-2^{1-s}) \Gamma(s)} \int_{0}^{\infty} \left ( e^{-w} \left ( 1- e^{-w} \right)^u \right ) w^{s-1} dw.$$
In view of the above, we can express $$D_n[A(v+1,s)]:= \frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} B(n+1,-v)(v+1)^{-s} dv.$$ From the definition of the beta function $$B(z,w)= \int_0^1 t^{z-1} (1-t)^{w-1}dt,$$ for $Re(z),Re(w)>0$, it follows that $$\begin{gathered}
\label{eq:Mellin1}
D_n[A(v+1,s)] =\frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} B(n+1,-v)(v+1)^{-s} dv =\\= \frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} \left ( \int_0^1 t^n (1-t)^{-(v+1)} dt \right ) (v+1)^{-s} dv =\\
= \int_0^1 t^n \left ( \frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} (1-t)^{-(v+1)} (v+1)^{-s} dv \right ) dt.\end{gathered}$$ Consider the function $$G(t,s):= \frac{1}{2 \pi i} \int_{c-i \infty}^{c+i \infty} t^{-v} (v+1)^{-s} dv,$$ which is, by definition, the inverse Mellin transform of $(v+1)^{-s}$. On the other hand, note that $$\begin{gathered}
\int_0^1 \left (ln(t)^{s-1} \right ) t^v dt =\left [\begin{array}{cc} z=ln(t) \\ dz =\frac{dt}{t} \end{array} \right ] =- e^{i \pi s} \int_{0}^{\infty} z^{s-1} e^{-(z+1)u} du= \\ =\left [\begin{array}{cc} w=(v+1)z \\ dw=(v+1)dz \end{array} \right ]=-e^{i \pi s} (v+1)^{-s} \int_0^{\infty} w^{s-1} e^{-w} dw = -e^{i \pi s}\Gamma(s) (v+1)^{-s}.
\end{gathered}$$ Hence, for $Re(s)>0$, we have $$G(t,s) := \left \{
\begin{array}{cc} -\frac{ln(t)^{s - 1} \cdot t }{e^{ \pi i s}\Gamma(s)} & 0 \leq t \leq 1 \\ 0 & 1<t \end{array} \right.$$ In particular, taking $c>-1$, equation (\[eq:Mellin1\]) could be written as $$\begin{gathered}
D_n[A(v+1,s)]
= \int_0^1 \left (\frac{G(1-t,s)}{1-t} \right ) t^n dt= \\ =-\frac{1}{e^{i \pi s} \Gamma(s)} \int_0^1 \left ( ln(1-t)^{s-1} \right ) t^n dt=-\frac{1}{e^{i \pi s}\Gamma(s)} \int_0^1 (1-t)^n \left ( ln(t)^{s-1} \right ) dt=\\=\left [\begin{array}{cc} w=ln(t) \\ dw =\frac{dt}{t} \end{array} \right ] =-\frac{1}{e^{i \pi s} \Gamma(s)} \int_{-\infty}^0 \left (e^w(1-e^w)^n \right ) w^{s-1} dw=\\=
\frac{1}{\Gamma(s)} \int_{0}^{\infty} \left (e^{-w}\left (1-e^{-w} \right )^n \right ) w^{s-1} dw,\end{gathered}$$ as required.
Proposition \[prop:integral\] enables us to apply saddle-point techniques[^2] in order to describe the asymptotic behvior of $\widetilde{A}(u,s)$. We refer the reader to [@BH; @DB] for standard references on the subject. First, we are interested in the asymptotic behavior of the $1$-parametric family of integrals $$\label{eq:saddle}
I(\sigma,t ; \alpha):= \int_{0}^{\infty}
e^{-w} \left ( 1- e^{-w} \right)^{\alpha t} w^{s-1} dw,$$ as $t \rightarrow \infty$. Note that for $s=\sigma+it$ and $u=\alpha \cdot t$ equation (\[eq:GL\]) can be written as $$\label{eq:A}
\widetilde{A}(u,s):=\frac{I(\sigma,t; \alpha)}{2^{u+1}(1-2^{1-s}) \Gamma(s)}$$ Let us express $$I(\sigma,t ;\alpha) := \int_0^{\infty} g(\sigma,w) e^{t f(w ; \alpha)} dw$$ where the two functions are given by $$\label{eq:29}
\begin{array}{ccc} g(\sigma,w):=e^{-w} w^{\sigma -1}
& ; & f(w ;\alpha):=\alpha \cdot ln \left (1-e^{-w} \right ) + i \cdot ln(w).
\end{array}$$ When expressed in such a form, the saddle point technique implies that the asymptotic properties of the integrals $I(\sigma,t ;\alpha)$ are determined by the location of saddle points of $f(w ;\alpha)$ in the complex plane, that is, the by the location of the solutions of the transcendental equation $$\label{eq:equ30} \frac{\partial f}{\partial w} (w ; \alpha) = \frac{ \alpha \cdot e^{-w}}{1-e^{-w}}+\frac{i}{w}=0.$$ Note that the solutions of (\[eq:equ30\]) are equivalent to the solutions of $$\label{eq:trans2}
e^w- \alpha \cdot i w -1=0,$$ excluding the trivial solution at $w=0$. Set $w=x+yi$ and express (\[eq:trans2\]) as the system of two equations $$\label{eq:sys2}
\begin{array}{ccc}
H^{\alpha}_1(x,y):=e^x cos(y)+ \alpha y-1=0 & ; & H^{\alpha}_2(x,y):=e^x sin(y)- \alpha x=0. \end{array}$$ Figure 1 illustrates the solutions of (\[eq:sys2\]) in the region $-2 \leq x \leq 2 $ and $-20 \leq y \leq 20$ for $\alpha=\frac{1}{\pi}$. In particular, the saddle points appear at the intersection of the two curves:
We order the solutions $w_k(\alpha)=x_k(\alpha)+y_k(\alpha) i$, with $k \in \mathbb{Z}$, according to the value of $y_k(\alpha)$. As a convention we set $w_0(\alpha)=0$ even though it is a solution only of (\[eq:trans2\]) and not of (\[eq:equ30\]). As a direct application of the saddle point technique, we have:
\[prop:2.2\] The following holds $$\vert
I(\sigma,t ;\alpha) \vert \sim \abs*{ \sqrt{\frac{2 \pi}{t \vert f''(w_{k(\alpha)}(\alpha) ;\alpha ) \vert }} \cdot e^{-w_{k(\alpha)}(\alpha)} \left ( 1-e^{-w_{k(\alpha)}(\alpha)} \right )^{\alpha t} w_{k(\alpha)}(\alpha)^{\sigma +it -1} },$$ where for $N>1$ and $\alpha \in \left (\frac{2}{(2N+1)\pi},\frac{2}{(2N-1) \pi} \right )$ $$k(\alpha): = \left \{ \begin{array}{cc} N & \abs{x_N}<\abs{x_{N+1}} \\ N+1 & otherwise \end{array} \right.$$ and $k(\alpha)=1$ for $\alpha \in \left ( \frac{2}{ \pi} , \infty \right )$.
Proposition \[prop:2.2\] expresses the fact that the main contribution to the asymptotic behavior of $\abs{I(\sigma,t ;\alpha}$ is determined by the saddle point $w_k(\alpha)$ with $\vert x_k(\alpha) \vert$ minimal and $y_k(\alpha)>0$, which is the $k(\alpha)$-th saddle point. For instance, $k(\alpha)=2$ for $\alpha=\frac{1}{\pi}$, as seen in Fig. 1.
In order to further describe of the general position of the required saddle points, let us consider the graphs of $H_i^{\alpha}(x,y)$, for $i=1,2$, as the value of $x$ varies. Starting with $x<<0$ the two graphs are essentially linear. As the value of $x$ grows, the exponent and trigonometric factors begin to contribute, as illustrated in the following Fig. 2:
Assume $\alpha \in \left (\frac{2}{(2N+1)\pi},\frac{2}{(2N-1) \pi} \right )$, as in Proposition \[prop:2.2\]. From the above observation, we see that saddle points below the $\left [\frac{N}{2} \right ]$-th point are to occur around intersections of local maxima of $H^{\alpha}_1(x,y)$ with the axis which, in turn, correspond to local maxima of $cos(y)$. That is, we have $
y_k(\alpha) \approx 2k \pi$ for $k<\frac{N}{2}$.
For the higher saddle points, the parity of $N$ needs to be considered. For $N=2M$ even, we have $$\begin{array}{ccc}
y_M(\alpha) \approx \left (2M-\frac{1}{4} \right ) \pi & ; & y_{M+1}(\alpha) \approx \left (2M+\frac{5}{4} \right ) \pi.
\end{array}$$ While for $N=2M+1$ odd, we have $ y_{M+1}(\alpha) = \left ( 2M+\frac{3}{2} \right ) \pi$. In both cases, $y_k(\alpha) \approx (2k-1) \pi$ for $k > M+1$, as the saddle points above the $(M+1)$-th point are to occur around intersection of the local minima of $H^{\alpha}_1(x,y)$ with the axis, corresponding to local minima of $cos(y)$ (as seen in Fig. 2). Note also that by direct substitution, for $\alpha = \frac{2}{(2N-1) \pi}$, we have $$\label{eq:zero}
w_N \left (\alpha \right )= (2N-1) \pi i.$$
It should be remarked that approximations of $x_k(\alpha)$ can be obtained via the Lagrange inversion formula (see, for instance, 2.2 of [@DB]), by evaluating the solution of the equation $
H^{\alpha}_1(x,\widetilde{y}_k(\alpha))=0$, where $\widetilde{y}_k(\alpha)$ is the approximated value of $y_k(\alpha)$ discussed above. In particular, the growth of $x_k$ is $O(log(k))$ as $k \rightarrow \infty$. The following result could be viewed as a manifestation of the “exponential cancellation” phenomena for $\widetilde{A}(u,s)$.:
\[cor:2.3\] For $s=\sigma+it$ and $u= \alpha \cdot t$ the following holds $$\widetilde{A}(u,s)= O \left ( e^{ -\omega(\alpha) t} \right ),$$ where $$\omega(\alpha):=- \alpha \cdot log \abs*{\frac{1-e^{-w_{k(\alpha)}(\alpha)}}{2}} + \theta_{k(\alpha)}(\alpha)-\frac{\pi}{2},$$ and $w_k(\alpha):=r_k(\alpha)e^{i \theta_k(\alpha)}$ is the polar presentation of the $k(\alpha)$-th saddle-point.
By (\[eq:A\]) and Proposition \[prop:2.2\] $$\begin{gathered}
\label{eq:42}
\vert
\widetilde{A}(u,s)\vert = O \left ( \sqrt{\frac{2 \pi}{t \vert f''(w_{k(\alpha)}(\alpha) ;\alpha ) \vert }} \cdot \frac{e^{-w_{k(\alpha)}(\alpha)}}{2 (1-2^{1-s}) \Gamma(s)} \left (\frac{ 1-e^{-w_{k(\alpha)}(\alpha)}}{2} \right )^{\alpha t} w_{k(\alpha)}(\alpha)^{\sigma +it -1} \right )= \\ =O \left ( \sqrt{\frac{1}{t }} \cdot \left (\frac{ 1-e^{-w_{k(\alpha)}(\alpha)}}{2} \right )^{\alpha t} \frac{w_{k(\alpha)}(\alpha)^{\sigma +it -1}}{\Gamma( \sigma+it)} \right )=\\= O \left ( \left (\frac{ 1-e^{-w_{k(\alpha)}(\alpha)}}{2} \right )^{\alpha t} \left (\frac{w_{k(\alpha)}(\alpha)}{\sigma+it} \right)^{\sigma +it }\right ),\end{gathered}$$ where in the second step we use $\Gamma(\sigma+it)=O \left ( (\sigma+it)^{\sigma+it-\frac{1}{2}} \right )$, which follows from Stirling’s formula.
In particular, in the setting of Theorem \[thm:main-theorem\] of the introduction, where $\vert t \vert = 2 \pi x y$, we set $$\omega(x,y):=\omega(\alpha),$$ with $\alpha = \frac{1}{2\pi y}$.
Let us conclude this section by noting that, for a given $\alpha>0$, the solutions $w_k(\alpha)$ can be evaluated numerically to any level of percision. For instance, by the Luck and Stevens formula, see [@LS; @LZC]. For instance, for $\alpha=\frac{1}{\pi}$, we have for the leading saddle point $$\label{eq:Spoint}
w_2 \left ( \alpha \right ) \approx 0.68154 + 9.31481 i,$$ and hence $\omega(\alpha) \approx 0.017728$. The following Fig.3 illustrates Corollary \[cor:2.3\] in this case:
The Asymptotics for the Integral of $I_1(x,s)$
==============================================
In this section we consider the asymptotic behavior of the integral $$\widetilde{I}_1(x,s):=\int_x^{\infty} \widetilde{A}(u,s) du.$$ We have:
\[thm:first-integral\] Let $s=\sigma+it,x = \alpha \cdot t$ and $N \in \mathbb{N}$. In $ \alpha \in \left [ \frac{2 }{ (N+1) \pi}, \frac{2 }{N \pi } \right ]$ the following holds $$\widetilde{I}_1(x,s):= \frac{\chi(s)}{2^{s-1}-1} \left (\sum_{k=1}^{N} (2k-1)^{s-1} \right ) +O \left (e^{-\omega(\alpha) t} \right ).$$
Integrating (\[eq:GL\]) we obtain $$\begin{gathered}
\label{eq:Int1}
\widetilde{I}_1(x,s)= \frac{1}{2(1-2^{1-s})\Gamma(s)} \int_x^{\infty} \left ( \int_{0}^{\infty}
e^{-w} \left ( \frac{ 1- e^{-w}}{2} \right)^u w^{s-1} dw \right ) du
= \\ =
\frac{1}{2(1-2^{1-s})\Gamma(s)} \int_0^{\infty}
e^{-w} \left ( \int_x^{\infty} \left ( \frac{ 1- e^{-w}}{2} \right)^u du \right ) w^{s-1} dw = \\ =\frac{1}{2(1-2^{1-s})\Gamma(s)} \int_{0}^{\infty} \left [ \frac{ e^{-w}\left (\frac{1-e^{-w}}{2} \right )^x }{ln \left ( \frac{1-e^{-w}}{2} \right ) } \right ] w^{s-1} dw. \end{gathered}$$ As in (\[eq:29\]), let us set $$\begin{array}{ccc} g(\sigma,w):=e^{-w} w^{\sigma -1}
& ; & f(w; \alpha):=\alpha \cdot ln \left (1-e^{-w} \right ) + i \cdot ln(w),
\end{array}$$ and $$h(w):= ln \left (\frac{1-e^{-w}}{2} \right ).$$ For $\alpha \in \mathbb{R}$ let us set consider the $1$-parametric family of integrals: $$\widetilde{I}_1(t,\sigma ; \alpha):= \frac{1}{2(1-2^{1-s})\Gamma(s)} \int_{0}^{\infty} \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} dw.$$ The integrand of $\widetilde{I}_1(t,\sigma; \alpha)$ admits poles[^3] at $(2k-1) \pi i$, for $k \in \mathbb{Z}$, coming from the zeros of $h(w)$. In particular, the saddle point technique requires us to deform the original contour of integration $[0,\infty)$, to a new contour, $C_{\alpha} \subset \mathbb{C}$, passing through the main contributing saddle point. Considering $C_{\alpha}$ as a continuous $1$-parametric family, the contour $C_{\alpha}$ is required to cross the first $N$ poles. Figure 3 shows a schematic illustration:
By Cauchy’s theorem we have $$\int_{0}^{\infty} \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} dw =\int_{C_{\alpha}} \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} dw +\sum_{k=1}^N Res \left ( \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} ; (2k-1) \pi i \right )$$ In particular, each pole contributes the following residue $$\begin{gathered}
\label{eq:Residue}
Res \left ( \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} ; (2k-1) \pi i \right ) = \\ = - 2 \pi i \cdot Res \left ( \frac{1}{h(w)}; (2k-1)\pi i \right ) \cdot g(\sigma, (2k-1)\pi i) e^{t f((2k-1)
\pi i; \alpha) }= \\ =- 2 \pi i \cdot (-2) \cdot
e^{-\pi i} \left ( \frac{ 1- e^{-\pi i}}{2} \right)^u ((2k-1) \pi i)^{s-1} =4 (2k-1)^{s-1} \pi^s e^{ \frac{-\pi i s}{2}}.\end{gathered}$$ The functional equation implies $\chi(s) = (\chi(1-s))^{-1}$. Hence, one can express $$\chi(z)=\frac{2^{s-1} \pi^{s}}{ cos \left ( \frac{s \pi}{2} \right ) \Gamma(s) }.$$ Clearly, $$\frac{1}{ cos \left ( \frac{s \pi}{2} \right ) }=2 e^{-\frac{\pi i s}{2}}+O \left ( e^{-\frac{3 \pi s}{2} } \right ).$$ Hence, $$\begin{gathered}
\label{eq:Residue}
\frac{1}{2(1-2^{1-s}) \Gamma(s)} Res \left ( \frac{g(\sigma,w)}{h(w)} e^{t f(w; \alpha)} ; (2k-1) \pi i \right ) = \\ =\frac{4 (2k-1)^{s-1} \pi^s e^{ \frac{-\pi i s}{2}}}{2(1-2^{1-s}) \Gamma(s)}= \frac{\chi(s)}{2^{s-1}-1} \left ( (2k-1)^{s-1} +O \left ( e^{-\frac{3 \pi s}{2}} \right ) \right ).\end{gathered}$$ Finally, the main saddle point contributing to the integral over $C_{\alpha}$ is the $k(\alpha)$-th, as required.
The Asymptotics for the Integral of $I_2(x,s)$
==============================================
In this section we consider the asymptotic behavior of the integral $$\label{eq:der} \widetilde{I}_2(x,s):= \int_x^{\infty} \psi(u) \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s) \right ) du,$$ where $
\psi(u):=u-[u]-\frac{1}{2}$. We have:
\[thm:int2\] Let $s=\sigma+it$ and $x=\alpha \cdot t$. Then the following holds $$\widetilde{I}_2(x,s):= O \left (e^{-\omega(\alpha) t} \right ).$$
As in the classical proof of Hardy and Littlewood in [@HL], consider the Fourier series expansion of $\psi(u)$, which is given by $$\psi(u)= - \sum_{n=1}^{\infty} \frac{1}{n \pi} sin(2n \pi u).$$ Hence, we have $$\label{eq:60} \widetilde{I}_2(x,s):=- \sum_{n=1}^{\infty} \frac{1}{n \pi} \widetilde{I}_2^n(x,s),$$ where $$\widetilde{I}_2^n(x,s):= \int_x^{\infty} sin(2n \pi u) \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s) \right ) du.$$ Let us express $$2i \widetilde{I}_2^n(x,s)= \widetilde{J}^+_n(x,s)+\widetilde{J}^-_n(x,s).$$ where $$\widetilde{J}^{\pm}_n(x,s):= \int_x^{\infty} e^{\pm 2n \pi i u} \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s) \right ) du.$$ Expanding, we have $$\begin{gathered}
\widetilde{J}_n^{\pm}(x,s)= \int_x^{\infty} e^{\pm 2n \pi i u} \left ( \frac{\partial}{\partial u} \widetilde{A}(u,s)\right ) du=\\= \frac{1}{2(1-2^{1-s})} \int_x^{\infty} e^{\pm 2n \pi i u} \left (\frac{\partial}{\partial u} \left [ \frac{1}{\Gamma(s)} \int_{0}^{\infty}
e^{-w} \left ( \frac{ 1- e^{-w}}{2} \right)^u w^{s-1} dw \right ] \right ) du= \\ =
\frac{1}{2(1-2^{1-s}) \Gamma(s) } \int_0^{\infty} \left ( \int_x^{\infty} \frac{\partial}{\partial u} \left [ \left ( \frac{ 1- e^{-w}}{2} \right)^u \right ] e^{\pm 2 n \pi i u} du \right ) e^{-w} w^{s-1} dw = \\ =
\frac{1}{2(1-2^{1-s}) \Gamma(s) } \int_0^{\infty} \left ( \int_x^{\infty} \left ( \frac{ 1- e^{-w}}{2} \right)^u e^{\pm 2 n \pi i u} du \right ) ln \left ( \frac{1-e^{-w}}{2} \right ) e^{-w} w^{s-1} dw =\\=
\frac{1}{2(1-2^{1-s}) \Gamma(s) } \int_0^{\infty} \left ( \frac{ ln \left ( \frac{1-e^{-w}}{2} \right ) .\left ( \frac{ 1- e^{-w}}{2} \right)^x e^{-w \pm 2 n \pi i x}}{ ln \left ( \frac{1-e^{-w}}{2} \right ) \pm 2n\pi i } \right ) w^{s-1} dw \end{gathered}$$ For $s=\sigma+it$ and $u = \alpha \cdot t$ with $t \rightarrow \infty$ set $$\widetilde{J}_n^{\pm}(\sigma,t ; \alpha)=\frac{\widetilde{k}^{\pm}_n(\sigma,t ; \alpha)}{2^{x+1}(1-2^{1-s}) \Gamma(s) },$$ where $$\label{eq:II}
\widetilde{k}^{\pm}_n(\sigma,t ;\alpha ) := \int_0^{\infty} \widetilde{g}_n^{\pm}(\sigma,w) e^{t \widetilde{f}_n^{\pm}(w ; \alpha )} dw,$$ with $$\begin{array}{ccc} \widetilde{g}^{\pm}_n(\sigma,w):=\frac{ln \left ( \frac{1-e^{-w}}{2} \right ) e^{-w} w^{\sigma -1}}{ln \left ( \frac{1-e^{-w}}{2} \right ) \pm 2n \pi i}
& ; & \widetilde{f}^{\pm}_n(w ; \alpha):= \alpha \cdot ln \left (1-e^{-w} \right ) + i \cdot \left ( ln(w) \pm 2n \right ).
\end{array}$$ On the one hand, the integrand, in this case, has no poles since, as for $n \neq 0$, there are no solutions of $$ln \left ( \frac{1-e^{-w}}{2} \right ) = \pm 2n \pi i,$$ for the principle branch of the complex logarithm (in contrast to the case of $\widetilde{I}_1(x,s)$ of the previous Section 3). On the other hand, note that for $k \in \mathbb{Z}$, the integrand does admit branch cuts, starting at $2k \pi i$ and extending to infinity from the left. However, these branch cuts do not contribute to the asymptotic behavior of $\widetilde{J}^{\pm}_n(\sigma,t ;\alpha)$, since it is not necessary to cross the cuts when deforming $[0,\infty)$ to the saddle-point contour. Figure 4 shows a schematic illustration of the saddle point contour, avoiding crossing of the branch cuts.
Hence, only saddle points contribute to the asymptotic properties of the integral $\widetilde{k}^{\pm}_n(\sigma,t,\alpha)$. The saddle points are given by the solutions of $$\frac{\partial \widetilde{f}^{\pm}_n}{\partial w} (w ; \alpha) = \frac{ \alpha \cdot e^{-w}}{1-e^{-w}}+\frac{i}{w}=0.$$ which is similar to the saddle-point equation (\[eq:equ30\]) of Section 2. We, hence, have $$\begin{gathered}
\widetilde{J}^{\pm}_n(x,s)= O \left ( \sqrt{\frac{2 \pi}{t \vert f''(w_k(\alpha); \alpha) \vert}} \frac{ ln \left ( \frac{1-e^{-w_{k(\alpha)}(\alpha)}}{2} \right ) \left ( \frac{ 1- e^{-w_{k(\alpha)}(\alpha)}}{2} \right)^x e^{-w_{k(\alpha)}(\alpha) \pm 2 n \pi i x}}{ 2(1-2^{1-s}) (ln \left ( \frac{1-e^{-w_{k(\alpha)}(\alpha)}}{2} \right ) \pm 2n\pi i )\Gamma(s) } w_{k(\alpha)}(\alpha)^{s-1} \right ) =\\ = O \left ( \frac{1}{ n \cdot \Gamma(s) } \sqrt{\frac{1}{t }} \left ( \frac{ 1- e^{-w_{k(\alpha)}(\alpha)}}{2} \right)^x e^{-w_{k(\alpha)}(\alpha)} w_{k(\alpha)}(\alpha)^{s-1} \right )=O \left ( \frac{ e^{-\omega(\alpha) t}}{n} \right ).\end{gathered}$$ By definition, also $\widetilde{I}_2^n(x,s)=O \left ( \frac{ e^{-\omega(\alpha) t}}{n} \right )$. Hence, by (\[eq:60\]) we have $$\widetilde{I}_2(x,s)=O \left (\sum_{n=1}^{\infty} \frac{ e^{-\omega(\alpha) t}}{n^2} \right )=O \left (e^{-\omega(\alpha) t} \right ),$$ as required.
Summary and Concluding Remarks
==============================
In this work we have proved a new approximate functional equation, whose error term is of exponentional rate of decay (\[eq:NEW-AFE\]). Our proof, as the proof of Hardy and Littelwood in the classical case [@HL], relied on the Euler-Maclaurain formula (\[eq:EM\]). However, it is interesting to draw a few distinctions and similarities between the two settings.
First, let us note that, in the classical case, a direct application of the Euler-Maclaurin formula, as in the proof of Theorem \[thm:main-theorem\], does not lead to a proof of the full exceptional functional equation (\[eq:HL-AFE\]). Instead, this approach, in the classical case, leads only to the so-called “simplest approximate functional equation”[^4] which states $$\label{AFE-simple}
\zeta(s)= \sum_{n \leq x} n^{-s} +\frac{x^{1-s}}{s-1}+O(x^{-\sigma}),$$ for $\vert t \vert \leq \pi x$, see Lemma 2 of [@HL]. In fact, the main bulk of [@HL], as well as [@HL3], involves the development of ingenious analytic methods to extend the proof of the “simple” equation to a proof of the full classical approximate functional equation (\[eq:HL-AFE\]).
Concretely, as mentioned in the introduction, in the classical case, one applies the Euler-Maclaurain formula (\[eq:EM\]) to the function $A(u,s)=u^{-s}$. As a result, the corresponding first integral, in this case, is given by $$I_1(x,s)= \int_x^{\infty}A(u,s) du = -\frac{x^{1-s}}{1-s}$$ in the domain $\sigma>1$. In particular, the summation formula can be written in the following form $$\label{eq:simple2}
\zeta(s)-\sum_{n \leq x} A(u,s) =-\frac{x^{1-s}}{1-s}+ I_2(x,s) +O(x^{-s}),$$ where $$I_2(x,s):=\int_x^{\infty} \psi(u) \left ( \frac{\partial}{\partial u} A(u,s) \right ) du.$$ Note that (\[eq:simple2\]) still holds, by analytic continuation, for $\sigma >-2$. The proof of the “simple approximate equation” is thus based on showing that $$I_2(x,s)=O(x^{-s}),$$ when $\vert t \vert \leq \pi x$. However, complications arise when considering the values in the range $\pi x \leq \vert t \vert$. In this case, the difference between the value of $\zeta(s)$ and the $\left [x \right]$-th partial sum of the series $$D(x,s):=\zeta(s)-\sum_{n \leq x } A(u,s)$$ starts to largely deviate from the first integral $I_1(x,s)=-\frac{x^{1-s}}{1-s}$. Hence, the power of the full (classical) approximate functional equation is required.
In contrast, in the Hasse-Sondow setting our Theorem \[thm:int2\] shows that the analog second integral is of exponential rate of decay $$\widetilde{I}_2(x,s)=O(e^{-\omega(x,y)}),$$ with $\vert t \vert = \pi x y$ for any $y \geq 1$. Correspondingly, Theorem \[thm:first-integral\] shows that the first integral $$\widetilde{I}_1(x,s):= \frac{\chi(s)}{2^{s-1}-1} \left (\sum_{k=1}^{N} (2k-1)^{s-1} \right ) +O \left (e^{-\omega(x,y) t} \right )$$ serves as an approximation of $$\widetilde{D}(x,s):=\zeta(s)-\sum_{n \leq x} \widetilde{A}(u,s),$$ for any value of $x$. In particular, this sensitivity to the values of $x$, comes from the contribution of the added poles of the integrand, as seen in the proof of Theorem \[thm:first-integral\]. These unique features of the Hasse-Sondow setting, enable us to actually obtain the proof of Theorem \[thm:main-theorem\], which is a full approximate functional equation, along the lines of the proof of the “simple approximate functional equation”, in the classical case. However, with the trade-off that more involved saddle-point techniques are required to be applied.
[10]{}
N. Bleistein and R. A. Handelsman. Asymptotic Expansions of Integrals. Holt, Rinehart & Winston, New York, 1977.
N. G. de Bruijn. Asymptotic methods in analysis. Bibliotheca Mathematica. Vol. 4, North-Holland Publishing Co., Amsterdam; P. Noordhoff Ltd., Groningen; Interscience Publishers Inc., New York, 1958.
H. M. Edwards. Riemann’s zeta function. Pure and Applied Mathematics, 1974, 58, New York-London: Academic Press.
L. B. Felsen, N. Marcuvitz. Radiation and Scattering of Waves. Prentice-Hall, Microwaves and Fields Series, 1973.
P. Flajolet and R. Sedgewick. Mellin transforms and asymptotics: Finite differences and Rice’s integrals. Theoretical Computer Science 144 (1995) pp 101–124.
S. Goyal and R. K. Laddha. On the generalized Riemann zeta funcion and the generalized Lambert transform. Ganita Sandesh, 11, 99-108, 1997.
G. H. Hardy and J. E. Littlewood. The zeros of Riemann’s zeta function on the critical line. Math. Z. 10, 283–317 (1921).
G. H. Hardy and J. E. Littlewood. The approximate functional equation in the theory of the zeta function, with an application to the divisor-problems of Dirichlet and Piltz. Proceedings of the London Mathematical Society, Volume s2-21, Issue 1, 1923, Pages 39–74.
G. H. Hardy and J. E. Littlewood. The approximate functional equations for $\zeta(s)$ and $\zeta^2(s)$. Proceedings of the London Mathematical Society, S2-29 (1), 81-97, 1929.
H. Hasse. Ein Summierungsverfahren fur die Riemannsche $\zeta$-Reihe. Math. Z. 32: 458–464 (1930).
A. Ivic. The Riemann Zeta Function. John Wiley & Sons, 1985.
D. E. Knuth. The art of computer programming, Vol. 3: Sorting and Searching. Addison-Wesley Reading, MA, 1973.
R. Luck and J.W. Stevens. Explicit solutions for transcendental equations. SIAM Review, vol. 44, no. 2, pp. 227–233, 2002.
R. Luck, G. J. Zdaniuk, and H. Cho. An Efficient Method to Find Solutions for Transcendental Equations with Several Roots. International Journal of Engineering Mathematics, vol. 2015, Article ID 523043, 4 pages, 2015.
N. E. N[ø]{}rlund. Vorlesungen uber Differenzenrechnung. (1954) Chelsea Publishing Company, New York.
B. Riemann. Sullo svolgimento del quoziente di due serie ipergeometriche in frazione continua infinita. Gesammelte mathematische Werke, ed. H. Weber, Berlin/New York: Springer- Verlag, 1990, pp. 456–462.
J. Sondow. Analytic continuation of Riemann’s zeta function and values at negative integers via Euler’s transformation of series. Proceedings of the American mathematical society, Volume 120, Issue 2, 1994, 421–425.
C. L. Siegel. Über Riemanns Nachlaß zur analytischen Zahlentheorie. Quellen Studien zur Geschichte der Math. Astron. und Phys. Abt. B: Studien 2: 45–80, 1932. (Also in Gesammelte Abhandlungen, Vol. 1. Berlin: Springer-Verlag, 1966).
A. Sommerfeld. ber die Ausbreitung der Wellen in der drahtlosen Telegraphie. Ann. Physik .8 (1909) 665–736.
E. C. Titchmarsh and D. R. Heath-Brown (ed.). The Theory of the Riemann Zeta Function, 1984, (2nd rev. ed.). Oxford University Press.
B. L. van der Waerden. On the method of saddle points. Applied Scientific Research, Section B, December 1952, Volume 2, Issue 1, 33–45.
H. Weyl. Ausbreitung elektromagnetischer Wellen [ü]{}ber einem ebenen Leiter. Ann. Physik 60 (1919) 481
[^1]: Hardy and Littlewood gave two proofs of the approximate functional equation (\[eq:HL-AFE\]). The “first proof” is based on real analysis and the Eular-Maclaurin summation formula and is presented in [@HL; @HL3] while the “second proof” is based on complex analysis and contour integration and appears in [@HL2]. The Riemann-Siegel formula is obtained by further applying saddle point techniques to the integrals of [@HL2]. In our case, of formula (\[eq:NEW-AFE\]), the interpolating function $\widetilde{A}(u,s)$, of the Euler-Mclaurain summation itself, is represented by an integral whose asymptotic analysis requires application of the saddle-point technique. Therefore, although our starting point is the Euler-Mclaurain formula (as in the “first proof” of [@HL]), the proof of Theorem \[thm:main-theorem\] eventually involves a combination the various methods.
[^2]: The asymptotics of $\sum_{k=1}^n\binom{n}{k} \frac{(-1)^k}{k^s}$ (a variant of our $D_n[(A(v+1,s)]$) was studied by Flajolet and Sedgewick via the Rice method for N[ö]{}rlund-Rice integrals, in [@FS]. It was shown in [@FS] that for fixed $s =\sigma+it \in \mathbb{C}$ non-integer, and $n$ arbitrarly large, one has $$\sum_{k=1}^n\binom{n}{k} \frac{(-1)^k}{k^s}= \frac{1}{2 \pi i} \int_{\frac{1}{2}-i \infty}^{\frac{1}{2}+i \infty} B(u+1,-v)v^{-s} dv = O\left ((log(u)^{\sigma} \right ).$$ In our case, we are concerned with the asymptotics of $D_n[A(v+1,s)]$ when $n$ is proportional to $t=Im(s)$, for which classical saddle-point methods applied to (\[eq:GL\]) are adequate.
[^3]: Historically, the asymptotic evaluation of integrals with poles was first studied in relation to the investigation of propagation of radio waves over a plane earth, see [@So]. The question attracted the attention of mathematicians such as Van der Werden [@vanderw] and H. Weyl [@We]. We refer the reader also to the book by Felsen and Marcuvitz [@FM] which contains a detailed discussion on saddle-point techniques in the case of poles and branches.
[^4]: See Theorem 4.11 of [@T] and Theorem 1.8 of [@I]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A ${\ensuremath{\mathbb{T}}}$-variety is an algebraic variety $X$ with an effective regular action of an algebraic torus ${\ensuremath{\mathbb{T}}}$. Altmann and Hausen [@AlHa06] gave a combinatorial description of an affine ${\ensuremath{\mathbb{T}}}$-variety $X$ by means of polyhedral divisors. In this paper we compute the higher direct images of the structure sheaf of a desingularization of $X$ in terms of this combinatorial data. As a consequence, we give a criterion as to when a ${\ensuremath{\mathbb{T}}}$-variety has rational singularities. We also provide a partial criterion for a ${\ensuremath{\mathbb{T}}}$-variety to be Cohen-Macaulay. As an application we characterize in this terms quasihomogeneous elliptic singularities of surfaces.'
address: 'Université Grenoble I, Institut Fourier, UMR 5582 CNRS-UJF, BP 74, 38402 St. Martin d’Hères cédex, France'
author:
- Alvaro Liendo
bibliography:
- 'math.bib'
title: 'Rational singularities of normal ${\ensuremath{\mathbb{T}}}$-varieties'
---
[^1]
Introduction {#introduction .unnumbered}
============
Let ${\ensuremath{k}}$ be an algebraically closed field of characteristic 0. In this paper we study ${\ensuremath{\mathbb{T}}}$-varieties i.e., varieties endowed with an action of an algebraic torus ${\ensuremath{\mathbb{T}}}=({\ensuremath{k}}^*)^n$. A ${\ensuremath{\mathbb{T}}}$-action on $X={\ensuremath{\operatorname{Spec}}}\,A$ gives rise to an $M$-grading on $A$, where $M$ is a lattice of rank $n$ and vice versa, any effective $M$-grading appears in this way (see e.g. [@KaRu82]). The complexity of a ${\ensuremath{\mathbb{T}}}$-action is the codimension of a general orbit. For an effective ${\ensuremath{\mathbb{T}}}$-action, the complexity equals $\dim X-\dim {\ensuremath{\mathbb{T}}}$.
There are well known combinatorial descriptions of normal ${\ensuremath{\mathbb{T}}}$-varieties. For toric varieties see e.g., [@Dem70], [@KKMS73 Ch. 1], and [@Oda88]. For affine ${\ensuremath{k}}^*$-surfaces see [@FlZa03c]. For complexity 1 case see [@KKMS73 Ch. 2 and 4], and more generally [@Tim08]. Finally, for arbitrary complexity see [@AlHa06; @AHS08].
We let $N={\ensuremath{\operatorname{Hom}}}(M,{\ensuremath{\mathbb{Z}}})$, $M_{{\ensuremath{\mathbb{Q}}}}=M\otimes{\ensuremath{\mathbb{Q}}}$, and $N_{{\ensuremath{\mathbb{Q}}}}=N\otimes{\ensuremath{\mathbb{Q}}}$. Any affine toric variety can be described via the weight cone $\sigma^\vee\subseteq M_{{\ensuremath{\mathbb{Q}}}}$ spanned over ${\ensuremath{\mathbb{Q}}}_{\geq0}$ by all $m\in M$ such that $A_m\neq\{0\}$ or, alternatively, via the dual cone $\sigma\subseteq N_{{\ensuremath{\mathbb{Q}}}}$. Similarly, the description of normal affine ${\ensuremath{\mathbb{T}}}$-varieties due to Altmann and Hausen deals with a polyhedral cone $\sigma\subseteq N_{{\ensuremath{\mathbb{Q}}}}$ (dual to the weight cone $\sigma^\vee\subseteq M_{{\ensuremath{\mathbb{Q}}}}$), a normal semiprojective variety $Y$, and a divisor ${\ensuremath{\mathfrak{D}}}$ on $Y$ whose coefficients are polyhedra in $N_{{\ensuremath{\mathbb{Q}}}}$ invariant by translations in $\sigma$.
Let $X$ be a normal variety and let $\psi:Z\rightarrow X$ be a desingularization. Usually, the classification of singularities involves the higher direct images of the structure sheaf $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z$. In particular, $X$ has rational singularities if $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=0$ for all $i\geq 1$, see e.g., [@Art66; @KKMS73; @Elk78].
Rational singularities are Cohen-Macaulay. Recall that a local ring is Cohen-Macaulay if its Krull dimension equals to the depth. A variety $X$ is called Cohen-Macaulay if all the local rings ${\ensuremath{\mathcal{O}}}_{X,x}$ are Cohen-Macaulay. By a well known theorem of Kempf [@KKMS73 p. 50], a variety $X$ has rational singularities if and only if $X$ is Cohen-Macaulay and the induced map $\psi_*:\omega_Z\hookrightarrow \omega_X$ is an isomorphism.
The content of the paper is as follows. In Section 1 we recall the combinatorial description of ${\ensuremath{\mathbb{T}}}$-varieties due to Altmann and Hausen. In Section 2 we obtain our main results concerning the higher direct images of the structure sheaf of a desingularization of a ${\ensuremath{\mathbb{T}}}$-variety. More precisely, in Theorem \[Tdirect\] we compute the higher direct image sheaves $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z$ for a ${\ensuremath{\mathbb{T}}}$-variety $X$ in terms of the Altmann-Hausen combinatorial data $(Y,{\ensuremath{\mathfrak{D}}},\sigma)$. In Theorem \[Trat\] we give a criterion in this terms as to when a ${\ensuremath{\mathbb{T}}}$-variety has rational singularities. In Corollary \[cmgen\] we apply Kempf’s result to give a condition for a ${\ensuremath{\mathbb{T}}}$-variety to be Cohen-Macaulay.
Finally in Section 4 we apply previous results to characterize elliptic singularities of ${\ensuremath{k}}^*$-surfaces. A normal surface singularity $(X,x)$ is called elliptic if $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=0$ for all $i\geq 2$ and $R^1\psi_*{\ensuremath{\mathcal{O}}}_Z={\ensuremath{k}}$, see e.g., [@Lau77; @Wat80; @Yau80]. An elliptic singularity is called minimal if it is Gorenstein i.e., is Cohen-Macaulay and the canonical sheaf $\omega_X$ is invertible. In Proposition \[gor\] we give a criterion for a surface with a 1-torus action to be Gorenstein. In Theorem \[ellip\] we characterize (minimal) elliptic singularities in the combinatorial terms.
The author is grateful to Mikhail Zaidenberg for posing the problem and permanent encouragement, and to Hubert Flenner for useful discussions.
Preliminaries
=============
Combinatorial description of affine ${\ensuremath{\mathbb{T}}}$-varieties {#AHdes}
-------------------------------------------------------------------------
In this section we recall briefly the combinatorial description of affine ${\ensuremath{\mathbb{T}}}$-varieties given by Altmann and Hausen in [@AlHa06].
Let $N$ be a lattice of rank $n$ and $M={\ensuremath{\operatorname{Hom}}}(N,{\ensuremath{\mathbb{Z}}})$ be its dual lattice. We fix dual bases $\{\nu_1,\cdots,\nu_n\}$ and $\{\mu_1,\cdots,\mu_n\}$ for $N$ and $M$, respectively. We also let $N_{{\ensuremath{\mathbb{Q}}}}=N\otimes{\ensuremath{\mathbb{Q}}}$, $M_{{\ensuremath{\mathbb{Q}}}}=M\otimes{\ensuremath{\mathbb{Q}}}$, and we consider the natural duality $M_{{\ensuremath{\mathbb{Q}}}}\times N_{{\ensuremath{\mathbb{Q}}}}\rightarrow {\ensuremath{\mathbb{Q}}}$, $(m,p)\mapsto \langle m,p\rangle$.
Let ${\ensuremath{\mathbb{T}}}={\ensuremath{\operatorname{Spec}}}{\ensuremath{k}}[M]$ be the $n$-dimensional algebraic torus associated to $M$ and let $X={\ensuremath{\operatorname{Spec}}}\,A$ be an affine ${\ensuremath{\mathbb{T}}}$-variety. It is well known that the morphism $A\rightarrow A\otimes {\ensuremath{k}}[M]$ induces an $M$-grading on $A$ and, conversely, every $M$-grading on $A$ arises in this way. Furthermore, a ${\ensuremath{\mathbb{T}}}$-action is effective if an only if the corresponding $M$-grading is effective.
Let $\sigma$ be a pointed polyhedral cone in $N_{{\ensuremath{\mathbb{Q}}}}$. We denote the dual cone by $\sigma^\vee$ and we let $\sigma_M^\vee=\sigma^\vee\cap M$. We define ${\ensuremath{\operatorname{Pol}}}_{\sigma}(N_{{\ensuremath{\mathbb{Q}}}})$ to be the set of all polyhedra in $N_{{\ensuremath{\mathbb{Q}}}}$ which can be decomposed as the Minkowski sum of a compact polyhedron and $\sigma$. The Minkowsi sum endows ${\ensuremath{\operatorname{Pol}}}_{\sigma}(N_{{\ensuremath{\mathbb{Q}}}})$ with a structure of semigroup with neutral element $\sigma$.
To any polyhedron $\Delta\in{\ensuremath{\operatorname{Pol}}}_{\sigma}(N_{{\ensuremath{\mathbb{Q}}}})$ we associate its support function $h_{\Delta}:\sigma^\vee\rightarrow {\ensuremath{\mathbb{Q}}}$ defined by $h_{\Delta}(m)=\min\langle m,\Delta\rangle$. Clearly the function $h_{\Delta}$ is piecewise linear. Furthermore, $h_{\Delta}$ is upper convex and positively homogeneous i.e., $$h_{\Delta}(m+m')\geq h_{\Delta}(m)+h_{\Delta}(m'),\ \mbox{and}\ h_{\Delta}(\lambda m)=\lambda h_{\Delta}(m),\forall m,m'\in \sigma^{\vee},\ \forall\lambda\in {\ensuremath{\mathbb{Q}}}_{\geq 0}\,.$$
\[ppd\] A variety $Y$ is called semiprojective if it is projective over an affine variety. A *$\sigma$-polyhedral divisor* on $Y$ is a formal finite sum ${\ensuremath{\mathfrak{D}}}=\sum_{D}\Delta_D\cdot D$, where $D$ runs over prime divisors on $Y$, $\Delta_D\in{\ensuremath{\operatorname{Pol}}}_{\sigma}(N_{{\ensuremath{\mathbb{Q}}}})$ and $\Delta_D=\sigma$ for all but finitely many $D$. For $m\in\sigma^{\vee}$ we can evaluate ${\ensuremath{\mathfrak{D}}}$ in $m$ by letting ${\ensuremath{\mathfrak{D}}}(m)$ be the ${\ensuremath{\mathbb{Q}}}$-divisor $${\ensuremath{\mathfrak{D}}}(m)=\sum_{D} h_D(m) D\,,$$ where $h_D=h_{\Delta_D}$. A $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}$ is called *proper* if the following hold.
(i) The evaluation ${\ensuremath{\mathfrak{D}}}(m)$ is a semiample[^2] ${\ensuremath{\mathbb{Q}}}$-Cartier divisor for all $m\in\sigma_M^\vee$, and
(ii) ${\ensuremath{\mathfrak{D}}}(m)$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}(\sigma^\vee)$.
The following theorem gives a combinatorial description of ${\ensuremath{\mathbb{T}}}$-varieties analogous to the classical combinatorial description of toric varieties.
\[AH\] Let ${\ensuremath{\mathfrak{D}}}$ be a proper $\sigma$-polyhedral divisor on a semiprojective variety $Y$. Consider the schemes[^3] $${\ensuremath{\widetilde{X}}}={\ensuremath{\operatorname{Spec}}}_Y {\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}], \quad \mbox{where} \quad {\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}]=\bigoplus_{m\in\sigma_M}{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m))\,,$$ $$\mbox{and}\quad X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}], \quad \mbox{where} \quad A[Y,{\ensuremath{\mathfrak{D}}}]=H^0({\ensuremath{\widetilde{X}}},{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}})=H^0(Y,{\ensuremath{\widetilde{A}}})\,.$$ Then the following hold.
(i) The schemes $X$ and ${\ensuremath{\widetilde{X}}}$ are normal varieties of dimension $\dim Y+{\ensuremath{\operatorname{rank}}}M$.
(ii) The $M$-gradings on $A[C,{\ensuremath{\mathfrak{D}}}]$ and ${\ensuremath{\widetilde{A}}}[C,{\ensuremath{\mathfrak{D}}}]$ induce effective ${\ensuremath{\mathbb{T}}}$-actions on $X$ and ${\ensuremath{\widetilde{X}}}$, respectively. Moreover, the canonical morphism $\pi:{\ensuremath{\widetilde{X}}}\rightarrow Y$ is a good quotient for the ${\ensuremath{\mathbb{T}}}$-action on ${\ensuremath{\widetilde{X}}}$.
(iii) There is a ${\ensuremath{\mathbb{T}}}$-equivariant proper contraction $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$.
Conversely, any normal affine variety with an effective ${\ensuremath{\mathbb{T}}}$-action is equivariantly isomorphic to some $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ for a quasiprojective variety $Y$ and a proper $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}$ on $Y$.
\[aff\] In the case where $Y$ is affine the contraction $\varphi$ is an isomorphism since any quasi-coherent sheaf on an affine variety is generated by global sections.
Divisors on ${\ensuremath{\mathbb{T}}}$-varieties {#T-div}
-------------------------------------------------
In [@FlZa03c] a characterization of ${\ensuremath{\mathbb{T}}}$-invariant divisors on an affine ${\ensuremath{k}}^*$-surface is given. In [@PeSu08] some of these results are generalized to the case of a ${\ensuremath{\mathbb{T}}}$-variety of complexity 1. In this section we recall results from [@PeSu08] and add some minor generalizations that we need.
Letting ${\ensuremath{\mathfrak{D}}}$ be a proper polyhedral divisor on a semiprojective variety $Y$ we let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ and ${\ensuremath{\widetilde{X}}}={\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}]$. Since the contraction $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ is equivariant, the ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on $X$ are in bijection with the ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on ${\ensuremath{\widetilde{X}}}$ not contracted by $\varphi$.
We first apply the orbit decomposition of the variety ${\ensuremath{\widetilde{X}}}$ in Proposition 7.10 and Corollary 7.11 of [@AlHa06] to obtain a description of the ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors in ${\ensuremath{\widetilde{X}}}$. There are 2 types of ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on ${\ensuremath{\widetilde{X}}}$: the horizontal type corresponding to families of ${\ensuremath{\mathbb{T}}}$-orbits closures of dimension ${\ensuremath{\operatorname{rank}}}M-1$ over $Y$; and the vertical type corresponding to families of ${\ensuremath{\mathbb{T}}}$-orbits closures of dimension ${\ensuremath{\operatorname{rank}}}M$ over a prime divisor on $Y$.
Let ${\ensuremath{\mathfrak{D}}}=\sum_D\Delta_D\cdot D$ be a proper $\sigma$-polyhedral divisor on a normal semiprojective variety $Y$. Letting ${\ensuremath{\widetilde{X}}}={\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}]$, the following hold.
(i) The ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on ${\ensuremath{\widetilde{X}}}$ of horizontal type are in bijection with the extremal rays of $\rho\subseteq\sigma$.
(ii) The ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on ${\ensuremath{\widetilde{X}}}$ of vertical type are in bijection with pairs $(D,p)$ where $D$ is a prime Weil divisor on $Y$ and $p$ is a vertex of $\Delta_D$.
The lemma follows from Proposition 7.10 and Corollary 7.11 of [@AlHa06]. See also the proof of Proposition 3.13 in [@PeSu08].
The following lemma is a reformulation of Proposition 3.13 in [@PeSu08].
\[Tcont\] Let ${\ensuremath{\mathfrak{D}}}=\sum_D\Delta_D\cdot D$ be a proper $\sigma$-polyhedral divisor on a normal semiprojective variety $Y$. The following hold.
(i) Let $\rho\subseteq\sigma$ be an extremal ray and let $\tau$ be is dual codimension 1 face. The ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors of horizontal type on ${\ensuremath{\widetilde{X}}}$ corresponding to $\rho$ is not contracted by $\varphi$ if and only if ${\ensuremath{\mathfrak{D}}}(m)$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}(\tau)$.
(ii) Let $D$ be a prime Weil divisor on $Y$ and let $p$ be a vertex of $\Delta_D$. The ${\ensuremath{\mathbb{T}}}$-invariant prime Weil divisors on ${\ensuremath{\widetilde{X}}}$ of vertical type $(D,p)$ is not contracted by $\varphi$ if and only if ${\ensuremath{\mathfrak{D}}}(m)|_D$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}((\Delta_D-p)^\vee)$.
\[Tiso1\] The morphism $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ is an isomorphism in codimension 1 if and only if the following conditions hold.
1. For every codimension 1 face $\tau\subseteq\sigma^\vee$, the divisor ${\ensuremath{\mathfrak{D}}}(m)$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}(\tau)$.
2. For every prime Weil divisor $D$ on $Y$ and every vertex on $\Delta_D$, the divisor ${\ensuremath{\mathfrak{D}}}(m)|_Z$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}((\Delta_D-p)^\vee)$.
To prove that the equivariant contraction $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ is an isomorphism in codimension 1 we only need to prove that no ${\ensuremath{\mathbb{T}}}$-invariant Weil divisor is contracted by $\varphi$. The first condition ensures that no divisor of horizontal type is contracted and the second condition ensures that no divisor of vertical type is contracted.
In the case of a complexity 1 ${\ensuremath{\mathbb{T}}}$-action i.e., when $Y$ is a smooth curve, the condition $(ii)$ in Lemma \[Tcont\] and Corollary \[Tiso1\] is trivially verified.
For one of our applications we need the following lemma concerning the Picard group of ${\ensuremath{\mathbb{T}}}$-varieties, see Proposition 3.1 in [@PeSu08] for a particular case.
\[spic\] Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a proper $\sigma$-polyhedral divisor on a normal semiprojective variety $Y$. If $Y$ is projective then ${\ensuremath{\operatorname{Pic}}}(X)$ is trivial.
Let $D$ be a Cartier divisor on $X$, and let $f$ be a local equation of $D$ in an open set $U\subseteq X$ containing $\bar{0}$. By [@Bou65 §1, Exercise 16] we may assume that $D$ and $U$ are ${\ensuremath{\mathbb{T}}}$-invariant. Since $\bar{0}$ is an attractive fixed point, every ${\ensuremath{\mathbb{T}}}$-orbit closure contains $\bar{0}$ and so $U=X$, proving the lemma.
Smooth polyhedral divisors
--------------------------
The combinatorial description in Theorem \[AH\] is not unique. The following Lemma is a specialization of Corollary 8.12 in [@AlHa06]. For the convenience of the reader, we provide a short argument.
\[proj\] Let ${\ensuremath{\mathfrak{D}}}$ be a proper $\sigma$-polyhedral divisor on a normal semiprojective variety $Y$. Then for any projective birational morphism $\psi:Z\rightarrow Y$ the variety ${\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ is equivariantly isomorphic to ${\ensuremath{\operatorname{Spec}}}A[Z,\psi^*{\ensuremath{\mathfrak{D}}}]$.
We only need to show that $$H^0(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))\simeq H^0(Z,{\ensuremath{\mathcal{O}}}_Z(\psi^*{\ensuremath{\mathfrak{D}}}(m))),\mbox{ for all }m\in\sigma^\vee_M\,.$$ Letting $r$ be such that $r{\ensuremath{\mathfrak{D}}}(m)$ is Cartier $\forall m\in\sigma_M^\vee$, we have $$H^0(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))=\{f\in k(Y): f^r\in H^0(Y,r{\ensuremath{\mathfrak{D}}}(m))\},\quad \forall m\in\sigma_M^\vee\,.$$ Since $Y$ is normal and $\psi$ is projective, by Zariski main theorem $\psi_*{\ensuremath{\mathcal{O}}}_Z={\ensuremath{\mathcal{O}}}_Y$ and by the projection formula, for all $m\in\sigma_M^\vee$ we have $$\begin{aligned}
H^0(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))&\simeq\big\{f\in k(Z): f^r\in H^0(Z,{\ensuremath{\mathcal{O}}}_Z(\psi^*r{\ensuremath{\mathfrak{D}}}(m)))\big\} \\
&=H^0(Z,{\ensuremath{\mathcal{O}}}_Z(\psi^*{\ensuremath{\mathfrak{D}}}(m)))\,.\end{aligned}$$ This completes the proof.
In the previous Lemma, ${\ensuremath{\widetilde{X}}}={\ensuremath{\operatorname{Spec}}}{\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}]$ is not equivariantly isomorphic to ${\ensuremath{\operatorname{Spec}}}{\ensuremath{\widetilde{A}}}[Z,\psi^*{\ensuremath{\mathfrak{D}}}]$, unless $\psi$ is an isomorphism.
To restrict further the class of $\sigma$-polyhedral divisor we introduce the following notation.
We say that a $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}$ on a normal semiprojective variety $Y$ is smooth if $Y$ is smooth, ${\ensuremath{\mathfrak{D}}}$ is proper, and ${\ensuremath{\operatorname{Supp}}}{\ensuremath{\mathfrak{D}}}$ is SNC.
\[smooth1\] In the case of complexity one i.e., when $Y$ is a curve, any proper $\sigma$-polyhedral divisor is smooth. Indeed, any normal curve is smooth and any divisor on a smooth curve is SNC.
For any ${\ensuremath{\mathbb{T}}}$-variety $X$ there exists a smooth $\sigma$-polyhedral divisor on a smooth semiprojective variety $Y$ such that $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$.
By Theorem \[AH\], there exists a proper $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}'$ on a normal semiprojective variety $Y'$ such that $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$. Let $\psi:Y\rightarrow Y'$ be a resolution of singularities of $Y$ such that ${\ensuremath{\operatorname{Supp}}}{\ensuremath{\mathfrak{D}}}'$ is SNC. By Chow Lemma we can assume that $Y$ is semiprojective. By Lemma \[proj\], ${\ensuremath{\mathfrak{D}}}=\psi^*{\ensuremath{\mathfrak{D}}}'$ is a smooth $\sigma$-polyhedral divisor such that $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$.
In the sequel, unless the converse is explicitly stated, we restrict to smooth $\sigma$-polyhedral divisors.
Singularities of ${\ensuremath{\mathbb{T}}}$-varieties
======================================================
In this section we elaborate a method to effectively compute an equivariant partial resolution of singularities of an affine ${\ensuremath{\mathbb{T}}}$-variety in terms of the combinatorial data $(Y,{\ensuremath{\mathfrak{D}}})$. As a consequence, we compute the higher direct image of the structure sheaf for any resolution of singularities, and we give a criterion as to when a ${\ensuremath{\mathbb{T}}}$-variety has rational singularities.
A key ingredient for our results is the following example (cf. Example 3.19 in [@Lie08]).
\[extor\] Let $H_i$, $i\in\{1,\ldots,n\}$ be the coordinate hyperplanes in $Y={\ensuremath{\mathbb{A}}}^n$, and let a smooth divisor ${\ensuremath{\mathfrak{D}}}$ on $Y$ given by $${\ensuremath{\mathfrak{D}}}=\sum_{i=0}^{n}\Delta_i\cdot H_i,\quad \mbox{where } \Delta_i\in{\ensuremath{\operatorname{Pol}}}_\sigma(N_{{\ensuremath{\mathbb{Q}}}})\,.$$ Letting $h_i=h_{\Delta_i}$ be the support function of $\Delta_i$ and $k(Y)=k(t_1,\ldots, t_n)$, we have $$\begin{aligned}
H^0(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))&=\big\{f\in k(Y):{\ensuremath{\operatorname{div}}}(f)+{\ensuremath{\mathfrak{D}}}(m)\geq 0\big\} \\
&=\left\{f\in k(Y):{\ensuremath{\operatorname{div}}}(f)+\sum_{i=1}^n h_i(m)\cdot H_i\geq 0\right\} \\
&=\bigoplus_{r_i\geq -h_i(m)}k\cdot t_1^{r_1}\cdots t_n^{r_n}\,.\end{aligned}$$ Let $\widehat{N}=N \times {\ensuremath{\mathbb{Z}}}^n$, $\widehat{M}=M\times {\ensuremath{\mathbb{Z}}}^n$ and $\widehat{\sigma}$ be the cone in $\widehat{N}_{{\ensuremath{\mathbb{Q}}}}$ spanned by $(\sigma,\overline{0})$ and $(\Delta_i,e_i)$, $\forall i\in\{1,\ldots,n\}$, where $e_i$ is the $i$-th vector in the standard base of ${\ensuremath{\mathbb{Q}}}^n$. A vector $(m,r)\in M'$ belongs to the dual cone $(\sigma')^\vee$ if and only if $m\in\sigma^\vee$ and $r_i\geq -h_i(m)$.
With this definitions we have $$A[Y,{\ensuremath{\mathfrak{D}}}]=\bigoplus_{m\in\sigma_M^\vee} H^0(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))=\bigoplus_{(m,r) \in\widehat{\sigma}_{\widehat{M}}^\vee}k\cdot t_1^{r_1}\cdots t_n^{r_n}\simeq k[\widehat{\sigma}^\vee_{\widehat{M}}]\,.$$
Hence $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ is isomorphic as an abstract variety to the toric variety with cone $\widehat{\sigma}\subseteq \widehat{N}_{{\ensuremath{\mathbb{Q}}}}$. Since $Y$ is affine ${\ensuremath{\widetilde{X}}}\simeq X$ is also a toric variety.
Recall [@KKMS73] that a variety $X$ is toroidal if for every $x\in X$ there is a formal neighborhood isomorphic to a formal neighborhood of a point $y\in U_{\omega}$ in a toric variety.
\[toroidal\] Let ${\ensuremath{\mathfrak{D}}}=\sum_D \Delta_D\cdot D$ be a proper $\sigma$-polyhedral divisor on a semiprojective normal variety $Y$. If ${\ensuremath{\mathfrak{D}}}$ is smooth then ${\ensuremath{\widetilde{X}}}={\ensuremath{\operatorname{Spec}}}_Y {\ensuremath{\widetilde{A}}}[Y,{\ensuremath{\mathfrak{D}}}]$ is a toroidal variety.
For $y\in Y$ we consider the fiber $X_y$ over $y$ for the morphism $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow Y$. We let also $\mathfrak{U}_y$ be a formal neighborhood of $X_y$.
We let $S_y=\{D\in\operatorname{Wdiv}(Y), y\in D\mbox{ and } \Delta_D\neq\sigma\}$ and $n=\dim Y$. Since ${\ensuremath{\operatorname{Supp}}}{\ensuremath{\mathfrak{D}}}$ is SNC, we have that $\operatorname{card}(S_y)\leq n$. We let $j:S_y\rightarrow \{1,\ldots,n\}$ be any injective function.
We consider the smooth $\sigma$-polyhedral divisor $${\ensuremath{\mathfrak{D}}}_y'= \sum_{D\in S_y} \Delta_D\cdot H_{j(D)}, \quad \mbox{on} \quad {\ensuremath{\mathbb{A}}}^n\,.$$
Since $Y$ is smooth, $\mathfrak{U}_y$ is isomorphic to a formal neighborhood of the fiber over zero for the morphism $\pi':{\ensuremath{\operatorname{Spec}}}{\ensuremath{\widetilde{A}}}[{\ensuremath{\mathbb{A}}}^n,{\ensuremath{\mathfrak{D}}}_y']\rightarrow {\ensuremath{\mathbb{A}}}^n$ (see Theorem \[AH\] $(ii)$).
Finally, Example \[extor\] shows that ${\ensuremath{\operatorname{Spec}}}{\ensuremath{\widetilde{A}}}[{\ensuremath{\mathbb{A}}}^n,{\ensuremath{\mathfrak{D}}}_y']$ is toric for all $y$ and so $X$ is toroidal. This completes the proof.
Since the contraction $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ in Theorem \[AH\] is proper and birational, to obtain a desingularization of $X$ it is enough to have a desingularization of ${\ensuremath{\widetilde{X}}}$. If further ${\ensuremath{\mathfrak{D}}}$ is smooth, then ${\ensuremath{\widetilde{X}}}$ is toroidal and there exists a toric desingularization.
Rational Singularities
----------------------
In the following we use the partial desingularization $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ where ${\ensuremath{\widetilde{X}}}$ is toroidal as in Theorem \[AH\]. This allows us to provide information about the singularities of $X$ in terms of the combinatorial data $(Y,{\ensuremath{\mathfrak{D}}})$. We recall the following notion.
A variety $X$ has rational singularities if there exists a desingularization $\psi:Z\rightarrow X$, such that $$\psi_*{\ensuremath{\mathcal{O}}}_Z={\ensuremath{\mathcal{O}}}_X,\quad\mbox{and}\quad R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=0, \quad \forall i>0\,.$$
\[anydes\] This definition is correct since the higher direct image sheaves $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z$ are independent of the particular choice of the desingularization of $X$. The first condition $\psi_*{\ensuremath{\mathcal{O}}}_Z={\ensuremath{\mathcal{O}}}_X$ is equivalent to $X$ is normal.
The following well known Lemma, similarly as in [@Vieh77], follows by applying the Leray spectral sequence. For the convenience of the reader we provide a short argument.
\[rati\] Let $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ be a proper surjective, birational morphism, and let $\psi:Z\rightarrow X$ be a desingularization of $X$. If ${\ensuremath{\widetilde{X}}}$ has only rational singularities, then $$R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=R^i\varphi_*{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}}, \quad \forall i\geq0\,.$$
By Remark \[anydes\], we may assume that the desingularization $\psi$ is such that $\psi=\varphi\circ\widetilde{\psi}$, where $\widetilde{\psi}:Z\rightarrow{\ensuremath{\widetilde{X}}}$ is a desingularization of ${\ensuremath{\widetilde{X}}}$. The question is local on $X$, so we may assume that $X$ is affine. Then, by [@Har77 Ch. III, Prop. 8.5] we have[^4] $$R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=H^i(Z,{\ensuremath{\mathcal{O}}}_Z)^{\sim}\quad\mbox{and}\quad R^i\varphi_*{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}}=H^i({\ensuremath{\widetilde{X}}},{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}})^{\sim},\quad \forall i\geq 0\,.$$ Since ${\ensuremath{\widetilde{X}}}$ has rational singularities $$\widetilde{\psi}_*{\ensuremath{\mathcal{O}}}_Z={\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}},\quad\mbox{and}\quad R^i\widetilde{\psi}_*{\ensuremath{\mathcal{O}}}_Z=0, \quad \forall i>0\,.$$ By Leray spectral sequence for $(p,q)=(i,0)$ we have $$H^i(Z,{\ensuremath{\mathcal{O}}}_Z)=H^i({\ensuremath{\widetilde{X}}},\widetilde{\psi}_*{\ensuremath{\mathcal{O}}}_Z)=H^i({\ensuremath{\widetilde{X}}},{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}}),\quad \forall i\geq 0\,,$$ proving the Lemma.
In the following Theorem, which is our main result, for a ${\ensuremath{\mathbb{T}}}$-variety $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ and a desingularization $\psi:Z\rightarrow X$ we provide an expression for $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z$ in terms of the combinatorial data $(Y,{\ensuremath{\mathfrak{D}}})$.
\[Tdirect\] Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a smooth polyhedral divisor on $Y$. If $\psi:Z\rightarrow X$ is a desingularization, then $R^i\psi_*{\ensuremath{\mathcal{O}}}_Z$ is the sheaf associated to $$\bigoplus_{m\in\sigma_M^\vee} H^i(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))$$
Let $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ be as in Theorem \[AH\]. By Lemma \[toroidal\] ${\ensuremath{\widetilde{X}}}$ is toroidal, thus it has only rational singularities. By Lemma \[rati\] we have $$R^i\psi_*{\ensuremath{\mathcal{O}}}_Z=R^i\varphi_*{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}}, \quad \forall i\geq0\,.$$
Since $X$ is affine, we have $$R^i\varphi_*{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}}=H^i({\ensuremath{\widetilde{X}}},{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}})^{\sim}, \quad \forall i\geq0\,,$$ see [@Har77 Ch. III, Prop. 8.5]. Let $\pi:{\ensuremath{\widetilde{X}}}\rightarrow Y$ be the good quotient in Theorem \[AH\] and let ${\ensuremath{\widetilde{A}}}=\bigoplus_{m\in\sigma_M^\vee} {\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m))$ so that ${\ensuremath{\widetilde{X}}}={\ensuremath{\operatorname{Spec}}}_Y{\ensuremath{\widetilde{A}}}$. Since the morphism $\pi$ is affine, we have $$H^i({\ensuremath{\widetilde{X}}},{\ensuremath{\mathcal{O}}}_{{\ensuremath{\widetilde{X}}}})= H^i(Y,{\ensuremath{\widetilde{A}}})= \bigoplus_{m\in\sigma_M^\vee} H^i(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m))),\quad \forall i\geq0$$ by [@Har77 Ch III, Ex. 4.1], proving the Theorem.
As an immediate consequence of Theorem \[Tdirect\], in the following theorem, we characterize ${\ensuremath{\mathbb{T}}}$-varieties having rational singularities.
\[Trat\] Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a smooth $\sigma$-polyhedral divisor on $Y$. Then $X$ has rational singularities if and only if for every $m\in\sigma_M^\vee$ $$H^i(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))=0, \quad \forall i\in \{1,\ldots,\dim Y\}\,.$$
Since $X$ is normal, by Theorem \[Tdirect\] we only have to prove that $$\bigoplus_{m\in\sigma_M^\vee} H^i(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))=0,\quad \forall i>0$$ This direct sum is trivial if and only if each summand is. Hence $X$ has rational singularities if and only if $H^i(Y,{\ensuremath{\mathcal{O}}}_Y({\ensuremath{\mathfrak{D}}}(m)))=0$, for all $i>0$ and all $m\in\sigma_M^\vee$.
Finally, $H^i(Y,{\ensuremath{\mathscr{F}}})=0$, for all $i>\dim Y$ and for any sheaf ${\ensuremath{\mathscr{F}}}$, see [@Har77 Ch III, Th. 2.7]. Now the Lemma follows.
In particular, we have the following Corollary.
\[acyc\] Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$ for some smooth $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}$ on $Y$. If $X$ has only rational singularities, then the structure sheaf ${\ensuremath{\mathcal{O}}}_Y$ is acyclic i.e., $H^i(Y,{\ensuremath{\mathcal{O}}}_Y)=0$ for all $i>0$.
This is the “only if” part of Theorem \[Trat\] for $m=0$.
In the case of complexity 1 i.e., when $Y$ is a curve, there is a more explicit criterion.
\[rcom1\] Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a smooth $\sigma$-polyhedral divisor on a smooth curve $Y$. Then $X$ has only rational singularities if and only if
1. $Y$ is affine, or
2. $Y={\ensuremath{\mathbb{P}}}^1$ and $\deg\lfloor{\ensuremath{\mathfrak{D}}}(m)\rfloor \geq -1$ for all $m\in\sigma_M^\vee$.
If $Y$ is affine then by Remark \[aff\], the morphism $\varphi$ in Theorem \[AH\] is an isomorphism. By Lemma \[toroidal\] $X$ is toroidal and thus $X$ has only rational singularities.
If $Y$ is projective of genus $g$, we have $\dim H^1(Y,{\ensuremath{\mathcal{O}}}_Y)=g$. So by Corollary \[acyc\] if $X$ has rational singularities then $C={\ensuremath{\mathbb{P}}}^1$. Furthermore, for the projective line we have $H^1({\ensuremath{\mathbb{P}}}^1,{\ensuremath{\mathcal{O}}}_{{\ensuremath{\mathbb{P}}}^1}(D))\neq 0$ if and only if $\deg D\leq-2$ [@Har77 Ch. III, Th 5.1]. Now the corollary follows from Theorem \[Trat\].
Let ${\ensuremath{\mathfrak{D}}}$ be a smooth divisor on $Y$. Assume that $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ has log-terminal (or canonical or terminal) singularities. Then $X$ has rational singularities and so $(Y,{\ensuremath{\mathfrak{D}}})$ satisfies the assumptions of Theorem \[Trat\]. In particular, if $Y$ is a projective curve then $Y$ must be rational. This complements Proposition 3.9 in [@Sus08] by showing that the case with $Y$ an elliptic curve cannot happen.
Cohen-Macaulay singularities
----------------------------
In the following, we apply Theorem \[Trat\] to give a partial classification of Cohen-Macaulay ${\ensuremath{\mathbb{T}}}$-varieties in terms of the combinatorial description $(Y,{\ensuremath{\mathfrak{D}}})$.
Recall that a local ring is Cohen-Macaulay if its Krull dimension is equal to its depth. A variety is Cohen-Macaulay if all its local rings are. The following lemma is well known, see for instance [@KKMS73 page 50].
\[kempf\] Let $\psi:Z\rightarrow X$ be a desingularization of $X$. Then $X$ has rational singularities if and only if $X$ is Cohen-Macaulay and $\psi_*\omega_Z\simeq\omega_X$ [^5].
\[rm-cm\]
1. As in Lemma \[rati\], applying the Leray spectral sequence the previous lemma is still valid if we allow $Z$ to have rational singularities.
2. If $\psi$ is an isomorphism in codimension 1 then $\psi_*\omega_Z\simeq\omega_X$.
In the next corollary, we give a partial criterion as to when a ${\ensuremath{\mathbb{T}}}$-variety has only Cohen-Macaulay singularities.
\[cmgen\] Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}=\sum_D\Delta_D\cdot D$ is a smooth $\sigma$-polyhedral divisor on $Y$. Assume that the following conditions hold.
1. For every codimension 1 face $\tau\subseteq\sigma^\vee$, the divisor ${\ensuremath{\mathfrak{D}}}(m)$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}(\tau)$.
2. For every prime Weil divisor $D$ on $Y$ and every vertex on $\Delta_D$, the divisor ${\ensuremath{\mathfrak{D}}}(m)|_Z$ is big for all $m\in{\ensuremath{\operatorname{rel.int}}}((\Delta_D-p)^\vee)$.
Then $X$ is Cohen-Macaulay if and only if $X$ has rational singularities.
By Corollary \[Tiso1\], the contraction $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ is an isomorphism in codimension 1. The result now follows from Lemma \[kempf\] and Remark \[rm-cm\] (ii).
In the next corollary, we provide more explicit criterion for the case of complexity 1.
\[cm1\] Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$, where $Y$ is a smooth curve and ${\ensuremath{\mathfrak{D}}}$ is a smooth $\sigma$-polyhedral divisor on $Y$. If one of the following conditions hold,
1. $Y$ is affine, or
2. ${\ensuremath{\operatorname{rank}}}M=1$, or
3. $Y$ is projective and $\deg{\ensuremath{\mathfrak{D}}}\cap\rho=\emptyset$, for every extremal ray $\rho\subseteq \sigma$.
Then $X$ is Cohen-Macaulay if and only if $X$ has rational singularities.
If $Y$ is affine then $X$ has rational singularities by Corollary \[rcom1\], and so $X$ is Cohen-Macaulay. If ${\ensuremath{\operatorname{rank}}}M=1$ then $X$ is a normal surface. By Serre S$_2$ normality criterion any normal surface is Cohen-Macaulay, see Theorem 11.5 in [@Eis95]. Finally, if $(iii)$ holds then by Lemma \[Tiso1\], $\varphi:{\ensuremath{\widetilde{X}}}\rightarrow X$ is an isomorphism in codimension 1. Now the result follows by Lemma \[kempf\].
For isolated singularities we can give a full classification whenever ${\ensuremath{\operatorname{rank}}}M\geq 2$.
Let $X=A[Y,{\ensuremath{\mathfrak{D}}}]$, where $D$ is a smooth $\sigma$-polyhedral divisor on $Y$. If ${\ensuremath{\operatorname{rank}}}M\geq 2$ and $X$ has only isolated singularities, then $X$ is Cohen-Macaulay if and only if $X$ has rational singularities.
The only thing we have to prove is the “only if” part. Assume that $X$ is Cohen-Macaulay and let $\psi:Z\rightarrow X$ be a resolution of singularities. Since $X$ has only isolated singularities we have that $R^i\psi_*O_Z$ vanishes except possibly for $i=\dim X-1$, see [@Kov99 Lemma 3.3]. Now Theorem \[Tdirect\] shows that $R^i\psi_*O_Z$ vanishes also for $i=\dim X-1$ since $\dim Y=\dim X -{\ensuremath{\operatorname{rank}}}M$ and ${\ensuremath{\operatorname{rank}}}M\geq 2$.
The last two corollaries give a full classification of isolated Cohen-Macaulay singularities on ${\ensuremath{\mathbb{T}}}$-varieties of complexity 1.
Quasihomogeneous surfaces singularities
=======================================
In this section we study in more detail the particular case of a one dimensional torus action of complexity one i.e., the case of ${\ensuremath{k}}^*$-surfaces. We characterize Gorenstein and elliptic singularities in terms of the combinatorial data as in Theorem \[AH\].
Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ be a ${\ensuremath{k}}^*$-surface, so that $Y$ is a smooth curve and $M\simeq {\ensuremath{\mathbb{Z}}}$. There are only two non-equivalent pointed polyhedral cones in $N_{\ensuremath{\mathbb{Q}}}\simeq{\ensuremath{\mathbb{Q}}}$ corresponding to $\sigma=\{0\}$ and $\sigma={\ensuremath{\mathbb{Q}}}_{\geq 0}$, and any $\sigma$-polyhedral divisor ${\ensuremath{\mathfrak{D}}}$ on $Y$ is smooth.
With the notation of Theorem \[AH\] suppose that $Y$ is affine. Then $X\simeq{\ensuremath{\widetilde{X}}}$ by Remark \[aff\] and so $X$ is toroidal by Lemma \[toroidal\]. In this case the singularities of $X$ can be classified by toric methods. In particular they are all rational, see for instance [@Oda88].
If $Y$ is projective, then $\sigma\neq \{0\}$ and so we can assume that $\sigma={\ensuremath{\mathbb{Q}}}_{\geq 0}$. In this case ${\ensuremath{\mathfrak{D}}}(m)=m{\ensuremath{\mathfrak{D}}}(1)$. Hence ${\ensuremath{\mathfrak{D}}}$ is completely determined by ${\ensuremath{\mathfrak{D}}}_1:={\ensuremath{\mathfrak{D}}}(1)$.
Furthermore, $$A[Y,{\ensuremath{\mathfrak{D}}}]=\bigoplus_{m\geq0}A_m\chi^m,\quad \mbox{where} \quad A_m=H^0(Y,m{\ensuremath{\mathfrak{D}}}_1)\,.$$ and there is an unique atractive fixed point $\bar{0}$ corresponding to the augmentation ideal $\mathfrak{m}_0=\bigoplus_{m>0}A_m\chi^m$.
This is exactly the setting studied in [@FlZa03c], where all ${\ensuremath{k}}^*$-surfaces are divided in three types: elliptic, parabolic and hyperbolic. In combinatorial language these correspond, respectively, to the cases where $Y$ is projective and $\sigma={\ensuremath{\mathbb{Q}}}_{\geq 0}$, $Y$ is affine and $\sigma={\ensuremath{\mathbb{Q}}}_{\geq 0}$, and finally $Y$ is affine and $\sigma=\{0\}$.
In particular, in [@FlZa03c] invariant divisors on ${\ensuremath{k}}^*$-surfaces are studied. The results in *loc.cit.* are stated only for the hyperbolic case. However, similar statements for the remaining cases can be obtained with essentially the same proofs. In the recent preprint [@Sus08] some of the results in *loc.cit.* have been generalized to the case of ${\ensuremath{\operatorname{rank}}}M>1$. Let us recall the necessary results from [@FlZa03c §4], see also [@Sus08].
Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a smooth $\sigma$-polyhedral divisor on be a projective smooth curve $Y$ (the elliptic case), and let as before ${\ensuremath{\mathfrak{D}}}_1={\ensuremath{\mathfrak{D}}}(1)$. We can write $${\ensuremath{\mathfrak{D}}}_1=\sum_{i=1}^{\ell} \frac{p_i}{q_i}z_i,\quad\mbox{where}\quad \gcd(p_i,q_i)=1,\mbox{ and }q_i>0\,.$$
In this case, with the notation of Theorem \[AH\] the birational morphism $\rho:=\pi\circ\varphi^{-1}:X\rightarrow Y$ is surjective and its indeterminacy locus consists of the unique fixed point corresponding to the maximal ideal $\bigoplus_{m>0}A_i$. The ${\ensuremath{k}}^*$-invariant prime divisors are $D_{z}:=\rho^{-1}(z)$, $\forall z\in Y$. The total transforms are: $\rho^*(z)=D_z$ for all $z\notin{\ensuremath{\operatorname{Supp}}}{\ensuremath{\mathfrak{D}}}_1$, and $\rho^*(z_i)=q_iD_{z_i}$, for $i=1,\ldots,\ell$. We let $D_i=D_{z_i}$ for $i=1,\ldots,\ell$.
The canonical divisor of $X$ is given by $$K_X=\rho^*(K_Y)+\sum_{i=1}^\ell (q_i-1)D_i\,.$$ For a rational semi-invariant function $f\cdot\chi^m$, where $f\in K(Y)$ and $m\in{\ensuremath{\mathbb{Z}}}$, we have $${\ensuremath{\operatorname{div}}}(f\cdot\chi^m)=\rho^*({\ensuremath{\operatorname{div}}}f)+m\sum_{i=1}^{\ell} p_iD_i\,.$$
For our next result we need the following notation.
\[nog\] We let $$\begin{aligned}
\label{nog-1}
m_G=\frac{1}{\deg{\ensuremath{\mathfrak{D}}}_1}\left(\deg K_Y+ \sum_{i=1}^{\ell}\frac{q_i-1}{q_i} \right)\,,\end{aligned}$$ and $$\begin{aligned}
\label{nog-2}
D_G=\sum_{i=1}^{\ell}d_iz_i,\quad\mbox{where}\quad d_i=\frac{p_im_G+1}{q_i}-1,\quad \forall i\in\{1,\ldots,\ell\}\,.\end{aligned}$$
Recall that a variety $X$ is *Gorenstein* if it is Cohen-Macaulay and the canonical divisor $K_X$ is Cartier. By Serre S$_2$ normality criterion, all normal surface singularities are Cohen-Macaulay. In the following proposition we give a criterion for a ${\ensuremath{k}}^*$-surface to have Gorenstein singularities.
\[gor\] Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$, where ${\ensuremath{\mathfrak{D}}}$ is a smooth $\sigma$-polyhedral divisor on a smooth projective curve $Y$. With the notation as in \[nog\], the surface $X$ has Gorenstein singularities if and only if $m_G$ is integral and $D_G-K_Y$ is a principal divisor on $Y$.
By Lemma \[spic\], $X$ is Gorenstein if and only if $K_X$ is a principal divisor i.e., there exist $m_G\in{\ensuremath{\mathbb{Z}}}$ and a principal divisor $D={\ensuremath{\operatorname{div}}}(f)$ on $Y$ such that $$K_X=\rho^*(K_Y)+\sum_{i=1}^\ell (q_i-1)D_i=\rho^*D+m_G\sum_{i=1}^{\ell}p_iD_i={\ensuremath{\operatorname{div}}}(f\cdot\chi^{m_G}),.$$ Clearly ${\ensuremath{\operatorname{Supp}}}(K_Y-D)\subseteq\{z_1\ldots,z_\ell\}$. Letting $$K_Y-D=\sum_{i=1}^{\ell}d_iz_i$$ we obtain $$\sum_{i=1}^\ell q_id_iD_i=\sum_{i=1}^\ell (mp_i-q_i+1)D_i\,.$$ Hence the $d_i$ satisfy in \[nog\]. Furthermore, since $$\deg K_Y=\deg(K_Y-D)=\sum_{i=1}^\ell d_i\,,$$ $m_G$ satisfies in \[nog\]. So $X$ is Gorenstein if and only if $m_G$ is integral and $D=K_Y-D_G$ is principal, proving the proposition.
Let $(X,x)$ be a normal surface singularity, and let $\psi:Z\rightarrow X$ be a resolution of the singularity $(X,x)$. One says that $(X,x)$ is an *elliptic singularity*[^6] if $R^1\psi_*{\ensuremath{\mathcal{O}}}_Z\simeq {\ensuremath{k}}$. An elliptic singularity is *minimal* if it is Gorenstein. e.g., [@Lau77], [@Wat80], and [@Yau80].
In the following theorem we characterize quasihomogeneous (minimal) elliptic singularities of surfaces.
\[ellip\] Let $X={\ensuremath{\operatorname{Spec}}}A[Y,{\ensuremath{\mathfrak{D}}}]$ be a normal affine surface with an effective elliptic 1-torus action, and let $\bar{0}\in X$ be the unique fixed point. Then $(X,\bar{0})$ is an elliptic singularity if and only if one of the following two conditions holds:
1. $Y={\ensuremath{\mathbb{P}}}^1$, $\deg\lfloor m{\ensuremath{\mathfrak{D}}}_1\rfloor\geq-2$ and $\deg\lfloor m{\ensuremath{\mathfrak{D}}}_1\rfloor=-2$ for one and only one $m\in{\ensuremath{\mathbb{Z}}}_{>0}$.
2. $Y$ is an elliptic curve, and for every $m\in{\ensuremath{\mathbb{Z}}}_{>0}$, the divisor $\lfloor m{\ensuremath{\mathfrak{D}}}_1\rfloor$ is not principal and $\deg\lfloor m{\ensuremath{\mathfrak{D}}}_1\rfloor\geq 0$.
Moreover, $(X,\bar{0})$ is a minimal elliptic singularity if and only if $(i)$ or $(ii)$ holds, $m_G$ is integral and $D_G-K_Y$ is a principal divisor on $Y$, where $m_G$ and $D_G$ are as in \[nog\].
Assume that $Y$ is a projective curve of genus $g$, and let $\psi:Z\rightarrow X$ be a resolution of singularities. By Theorem \[Tdirect\], $$R^1\psi_*{\ensuremath{\mathcal{O}}}_Z=\bigoplus_{m\geq0}H^1(Y,{\ensuremath{\mathcal{O}}}_Y(m{\ensuremath{\mathfrak{D}}}_1))\,.$$ Since $\dim R^1\psi_*{\ensuremath{\mathcal{O}}}_Z\geq g=\dim H^1(Y,{\ensuremath{\mathcal{O}}}_Y)$, if $X$ has an elliptic singularity then $g\in\{0,1\}$.
If $Y={\ensuremath{\mathbb{P}}}^1$ then $(X,\bar{0})$ is an elliptic singularity if and only if $H^1(Y,{\ensuremath{\mathcal{O}}}_Y(m{\ensuremath{\mathfrak{D}}}_1)={\ensuremath{k}}$ for one and only one value of $m$. This is the case if and only if $(i)$ holds.
If $Y$ is an elliptic curve, then $H^1(Y,{\ensuremath{\mathcal{O}}}_Y)={\ensuremath{k}}$. So the singularity $(X,\bar{0})$ is elliptic if and only if $H^1(Y,m{\ensuremath{\mathfrak{D}}}_1)=0$ for all $m>0$. This is the case if and only if $(ii)$ holds.
Finally, the last assertion concerning maximal elliptic singularities follows immediately form Proposition \[gor\].
By applying the criterion of Theorem \[ellip\], the following combinatorial data gives rational ${\ensuremath{k}}^*$-surfaces with an elliptic singularity at the only fixed point.
1. $Y={\ensuremath{\mathbb{P}}}^1$ and ${\ensuremath{\mathfrak{D}}}_1=-\tfrac{1}{4}[0]-\tfrac{1}{4}[1]+\tfrac{3}{4}[\infty]$. In this case $X={\ensuremath{\operatorname{Spec}}}A[Y,m{\ensuremath{\mathfrak{D}}}_1]$ is isomorphic to the surface in ${\ensuremath{\mathbb{A}}}^3$ with equation $$x_1^4x_3+x_2^3+x_3^2=0\,.$$
2. $Y={\ensuremath{\mathbb{P}}}^1$ and ${\ensuremath{\mathfrak{D}}}_1=-\tfrac{1}{3}[0]-\tfrac{1}{3}[1]+\tfrac{3}{4}[\infty]$. In this case $X={\ensuremath{\operatorname{Spec}}}A[Y,m{\ensuremath{\mathfrak{D}}}_1]$ is isomorphic to the surface in ${\ensuremath{\mathbb{A}}}^3$ with equation $$x_1^4+x_2^3+x_3^3=0\,.$$
3. $Y={\ensuremath{\mathbb{P}}}^1$ and ${\ensuremath{\mathfrak{D}}}_1=-\tfrac{2}{3}[0]-\tfrac{2}{3}[1]+\tfrac{17}{12}[\infty]$. In this case $X={\ensuremath{\operatorname{Spec}}}A[Y,m{\ensuremath{\mathfrak{D}}}_1]$ is isomorphic to the surface $$V(x_1^4x_2x_3-x_2x_3^2+x_4^2\ ;\, x_1^5x_3-x_1x_3^2+x_2x_4\ ;\,x_2^2-x_1x_4)\subseteq {\ensuremath{\mathbb{A}}}^4\,.$$
This last example is not a complete intersection since otherwise $(X,\bar{0})$ would be Gorenstein i.e., minimal elliptic which is not the case by virtue of Theorem \[ellip\]. In the first two examples the elliptic singlarities are minimal, since every normal hypersurface is Gorenstein.
[^1]: : 14J17, 14E15.\
: torus actions, ${\ensuremath{\mathbb{T}}}$-varieties, rational singularities, Cohen-Macaulay singularities, elliptic surface singularities.
[^2]: A ${\ensuremath{\mathbb{Q}}}$-divisor $D$ is called semiample if there exists $r>1$ such that the linear system $|rD|$ is base point free.
[^3]: For a ${\ensuremath{\mathbb{Q}}}$-divisor $D$, ${\ensuremath{\mathcal{O}}}(D)$ stands for ${\ensuremath{\mathcal{O}}}(\lfloor D\rfloor)$, where $\lfloor D\rfloor$ is the integral part of $D$.
[^4]: As usual for a $A$-module $M$, $M^{\sim}$ denontes the associated sheaf on $X={\ensuremath{\operatorname{Spec}}}A$.
[^5]: As usual $\omega_Z$ and $\omega_X$ denote the canonical sheaf of $Z$ and $X$ respectively.
[^6]: Some authors call such $(X,x)$ a strongly elliptic singularity.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- |
Riuji Mochizuki[^1] and Kenji Ikegami[^2]\
[Laboratory of Physics, Tokyo Dental College, Chiba 261-8502, Japan]{}
title: A necessary condition for brane solutions to have their origins in string field theory
---
Introduction
============
In many brane-world models[@ngreview], our 4-dimensional spacetime is thought to be described as a low-energy approximation of a solution of string field theory. Unfortunately, however, since we do not understand string field theory well, we cannot conclude that a given supergravity solution has its origin in string field theory. If it is a low-energy solution of string theory, dimensional reduction (or oxidation) and T-duality operations should be well-defined. Dimensional reduction and T-duality operation rules are given for time-independent brane solutions [@Ortin]. Nevertheless, the situation is rather different when we consider time-dependent solutions. Although the dimensional reduction of spacelike brane solutions has been discussed [@Roy][@ohtan], it is not well-defined at all times. Furthermore, their T-duality is hardly discussed in any papers. It is the time-dependence of the dilaton expectation values that makes this problem so complicated. If an S-brane solution does not have T-duality, we should not regard it as a low-energy solution of string theory. In our previous papers[@mochi][@mochi2][@mi1], we constructed singular (i.e. neither general nor particular) S-brane solutions of supergravity. Furthermore, most of them possess static, flat dimensions. This is of merit, as it enables us to define the dimensional reduction and T-duality of time-dependent solutions. In this paper, we examine these solutions and find that compactification of dimensions whose metric tensors depend on time is ill-defined and that the time-dependence of dilaton expectation values spoils T-duality. These observations lead us to conclude that singular solutions of 10-dimensional supergravity solutions should be regarded as low-energy solutions of string field theory only if the dilaton expectation values are independent of time. More generally, we present a necessary condition for general brane solutions which should be satisfied if they are to be regarded as low-energy solutions of string field theory.
The organization of this paper is as follows: our starting point is Einstein gravity coupled to a dilaton and $n$-form fields in 11 dimensions (M-theory) and 10 dimensions (IIA and IIB superstring theories). In section 2 we write a singular, spacelike brane solution by following [@mochi2][@mi1]. In section 3, the T-duality of these solutions is discussed and our conclusion derived.
Singular spacelike solutions
============================
Firstly, we present singular S-brane solutions according to our previous papers[@mochi][@mochi2][@mi1]. We consider Einstein gravity coupled to a dilaton field $\phi$ and $m$ kinds of [*n*]{}-form field $F_{n}$, whose action $I$ is
$$I={1\over 16\pi G}\int d^Dx\surd \overline{- g}\Big[ R-{1\over 2}g^{\mu\nu}\partial_{\mu} \phi\partial_{\nu}\phi
-\sum_{A=1}^{m}{1\over 2\cdot n_A!}e^{\alpha_A\phi}F_{n_A}^{\ 2}\Big],\label{eq:action}$$
where $\alpha_A$ is the dilaton coupling constant given by $$\alpha_A = \left\{
\begin{array}{cl}
0 & ({\rm M-theory}) \\
-1 & ({\rm NS-NS\ sector}) \\
{5-n_A\over 2} & ({\rm R-R\ sector})
\end{array}
\right.$$ and [*D*]{}=11 for M-theory and [*D*]{}=10 for superstring theories.
We can write a set of solutions of the field equations and the Bianchi identity. These solutions may include S[*w*]{} and S[*m*]{}, which are spacelike counterparts of wave and monopole solutions, respectively. The metric form is $$ds^2 = H_1du^2 + H_2dv^2 + \sum_{i=3}^{p+1}H_idx^idx^i+\sum_{a,b=p+2}^D H_0\eta_{ab}dy^ady^b,\label{eq:setting}$$ where $$\eta_{ab} = \{diag.(+ , \cdots , + , - )\},\label{eq:singletime}$$ $$du = dx^1 + 2i\delta_m\tilde B_ady^a,$$ $$dv = dx^2 + i\delta_w(H-1)dx^3,$$ with $$\delta_{m(w)} = \left\{
\begin{array}{cl}
1 & ({\rm S}m(w){\rm \ is\ included.}) \\
0 & ({\rm S}m(w){\rm \ is\ not\ included.})
\end{array},
\right.$$ $$\partial_a\tilde{B}_{b}-\partial_b\tilde{B}_{a}=\eta_{ac}\eta_{bd}\epsilon^{cde}H^{-2}\partial_eH,$$ $$\partial^2H^{-1}(y)\equiv\eta^{ab}\partial_a\partial_bH^{-1}(y)=0.\label{eq:hh}$$ We use $\{ x^i;\ i=1,\cdots,p+1\}$ as the coordinates of the space where the branes exist. General orthogonally intersecting S[*p*]{}-brane solutions have been given [@gen], where the metric functions and fields depend only on $y^D$. A D-brane solution which depends on all the extra space coordinates has been suggested[@multi].
The metric functions are $$H_1(y)=H^{\delta_m + \sum_{A=1}^m{\delta_{A,i}\over D-2}}(y),\label{eq:solu1}$$ $$H_2(y)=H^{-\delta_w + \sum_{A=1}^m{\delta_{A,i}\over D-2}}(y),\label{eq:solu2}$$ $$H_3(y)=H^{\delta_w + \sum_{A=1}^m{\delta_{A,i}\over D-2}}(y),\label{eq:solu3}$$ $$H_i(y)=H^{\sum_{A=1}^m{\delta_{A,i}\over D-2}}(y)\ \ \ \ (i=4, \cdots ,p+1),\label{eq:solu4}$$ $$H_0(y)=H^{-\delta_m -\sum_{A=1}^m{q_A+1\over D-2}}(y),\label{eq:solv}$$ where $$\delta_{A,i} = \left\{
\begin{array}{cl}
D-q_A-3 & (i\in q_A) \\
-(q_A+1) & (i\notin q_A)
\end{array}
\right..\label{eq:deltai}$$
The field strength for an electrically charged S[*q*]{}-brane is given by $$(F_n)_{i_{1}\cdots i_{n-1}a}(y)=\epsilon_{i_1\cdots i_{n-1}}\partial_a E(y),$$ $$(n = q+2),$$ while the magnetically charged case is given by $$(F_n)^{a_1\cdots a_{n}}(y)=\frac{1}{\surd \overline{-g}}e^{-\alpha\phi}\epsilon^{a_1\cdots a_{n}b}\partial_b E(y),$$ $$(n = D-q-2),$$ where $$E(y)=iH(y).\label{eq:solE}$$ The dilaton field is $$\phi(y) =H^{-\sum_{A=1}^m{\varepsilon_A\alpha_A\over 2}}(y),\label{eq:solphi}$$ where $$\varepsilon_A = \left\{
\begin{array}{cl}
+1 & (F_{n_A}\ {\rm is\ an\ electric\ field\ strength}) \\
-1 & (F_{n_A}\ {\rm is\ a\ magnetic\ field\ strength})
\end{array}
\right..$$ For S[*w*]{} and S[*m*]{}, the dilaton coupling constant $\alpha_A=0$.
The intersection rule, which has been suggested for general solutions in other papers [@gen][@int1][@int2], should be satisfied: $$-\varepsilon_A\varepsilon_B\alpha_A\alpha_B-2(\bar q+1)+{2(q_A+1)(q_B+1)\over D-2}=0,\label{eq:intersection}$$ where $\bar q+1$ is the number of dimensions which $q_A$-brane and $q_B$-brane are crossing on. An S[*w*]{} can be put in any two isometric directions, while an S[*m*]{} needs a direction with no other branes.
To discuss dimensional reduction and T-duality, we construct a singular solution with static dimensions. To this end, we consider solutions which depend only on the scale parameter $r$ of the entire, or a part of, spacetime perpendicular to the brane: $$r \equiv \sqrt{-\eta_{ab}y^{a}y^{b}},\ \ \ \ -\eta_{ab}y^{a}y^{b}>0,$$ Note that $r$ is a timelike coordinate. To satisfy (\[eq:hh\]), $$H = r^{D-p-3}.\label{eq:HH}$$ Then, the metric of this spacetime becomes $$ds^2=\sum_{i=1}^{p+1}H_idx^idx^i-H_0\Big(dr^2 - r^2d\Sigma^2_{D-p-2} \Big), \label{eq:metric}$$ after some coordinate transformation if S[*w*]{} and/or S[*m*]{} are included. $d\Sigma_{D-p-2}$ is the line element of a $(D-p-2)$-dimensional hyperbolic space whose scale factor is unity.
We define cosmic time ([*our time*]{}) $t$ as $$\begin{aligned}
dt&\equiv&H_0^{\ 1/2}dr\nonumber\\
&=&r^{(D-p-3)\big[-{\delta_m\over 2}-\sum{q_A+1\over 2(D-2)}\big]}dr,\end{aligned}$$ and impose a condition: $$(D-p-3)\big[{\delta_m\over 2}+\sum_{A=1}^m{q_A+1\over 2(D-2)}\big]=1.\label{eq:joukenn}$$
In this case, since $$t= \ln r,$$ $$H_0=r^{-2}.$$ $$H = {\rm e}^{(D-p-3)t},\label{eq:ht}$$ the metric (\[eq:metric\]) becomes $$ds^2=-dt^2+\sum_{i=1}^{p+1}H_i(t)dx^idx^i+d\Sigma^2_{D-p-2}\ , \label{eq:metric2}$$ where $H_i(t)$s are defined by substituting (\[eq:ht\]) for $H(y)$ in (\[eq:solu1\])$\sim$(\[eq:solv\]). Note that the scale factor of the extra space is independent of $t$ and some $H_i$s may be independent of time in this metric.
All the solutions satisfying the above condition (\[eq:joukenn\]) and possessing static flat dimensions are given in [@mi1].
A necessary condition for brane solutions
==========================================
In this section we discuss the T-duality of S-brane solutions and derive a necessary condition for them to have their origins in string field theory. We expect that the time-independent rule of dimensional reduction can be applied to static flat dimensions of S-brane solutions. In fact, we can easily confirm this. Dimensional reduction is well-defined and other dimensions behave in exactly the same way as they did before dimensional reduction for all the static flat dimensions. For example, let us consider a $D=11$, $p=6$ solution in which 3 SM2 branes are located on the 1st, 2nd and 4th, on the 1st, 3rd and 5th and on the 2nd, 3rd and 6th dimensions. Its metric is $$ds^2 = -dt^2 + \sum_{i=1}^3 {\rm e}^{2t}dx^idx^i + \sum_{j=4}^6dx^jdx^j + {\rm e}^{-2t}dx^7dx^7 + d\Sigma^2_3.\label{eq:example}$$ We can compactify the 4th, 5th and 6th dimensions. When the number of the spacetime dimensions is reduced from 11 to 8, the metric of the 8 remaining dimensions stays as it was in 11 dimensional spacetime: $$ds^2 = -dt^2 + \sum_{i=1}^3 {\rm e}^{2t}dx^idx^i + {\rm e}^{-2t}dx^7dx^7 + d\Sigma^2_3.\label{eq:remain}$$ If only the 4th dimension is compactified, we obtain $$ds^2 = -dt^2 + \sum_{i=1}^3 {\rm e}^{2t}dx^idx^i + \sum_{j=5}^6dx^jdx^j + {\rm e}^{-2t}dx^7dx^7 + d\Sigma^2_3.$$ This is a IIA supergravity solution which consists of 1 SF1 brane and 2 SD2 branes. Similarly, all the $D=11$ M-theory solutions which satisfiy (\[eq:joukenn\]) and have static dimensions become $D=10$ IIA supergravity solutions[@mi1] by dimensional reduction. Oxidation is naturally defined as the inverse operation of dimensional reduction. On the other hand, if we compactify a time-dependent dimension (one of the 1st, 2nd, 3rd and 7th dimensions), metric functions of the other dimensions change in accordance with reduction of the dimensions. We, therefore, conclude that reduction of dimensions which depends on time cannot be defined.
Using dimensional reduction, we can define T-duality between IIA and IIB solutions. If IIA and IIB solutions are reduced to the same 9-dimensional solution, the IIA solution is the T-dual of the IIB solution and vice versa. Obeying these rules, we confirm that each IIA solution that satisfies (\[eq:joukenn\]) and has static dimensions can find a IIB solution as its T-dual partner. Nevertheless, some of such IIB solutions have no T-dual partners. An example is a $p=5$ IIB solution in which 2 SD1 branes are located on the 1st and 3rd and on the 2nd and 4th dimensions, and 1 SD3 brane on the 1st, 2nd, 5th and 6th dimensions. Its metric is $$ds^2 = -dt^2 + \sum_{i=1}^2 {\rm e}^{2t}dx^idx^i + \sum_{j=3}^6dx^jdx^j + d\Sigma^2_3.\label{eq:dame}$$ If the 3rd dimension is compactified, its metric becomes $$ds^2 = -dt^2 + \sum_{i=1}^2 {\rm e}^{2t}dx^idx^i + \sum_{j=4}^6dx^jdx^j + d\Sigma^2_3.$$ No IIA solutions are, however, obtained from this $D=9$ solution by oxidation. That is, (\[eq:dame\]) has no T-dual partner.
It is the dilaton expectation values that cause this difference. The T-duality of a solution is well-defined if its dilaton expectation value is independent of time. If not, T-duality does not exist for the solution. We are convinced of this reason if we examine T-duality operations in the Einstein frame. A formula for converting a IIB solution to IIA by compactifying the $z$ direction is $$\begin{aligned}
g^{A}_{\mu\nu}&=&(g^B_{zz})^{\frac{2}{D-2}} \exp\big\{\frac{8}{(D-2)^2}\phi^B\big\}\nonumber \\
&&\times\Big[
g^B_{\mu\nu}
-\frac{g^B_{z\mu}g^B_{z\nu}-4B^{(1)}_{z\mu}B^{(1)}_{z\nu}\exp\{-\frac{8}{D-2}\phi^B\}}
{g^B_{zz}}\Big],\label{eq:convert}\end{aligned}$$ where $g^A$ and $g^B$ are IIA and IIB metric tensors, respectively. If a IIB metric tensor in (\[eq:convert\]) is static and $\phi^B$ depends on time, a IIA metric tensor must depend on time. That is, there exists no IIA solution that can be reduced to the same $D=9$ solution that this IIB solution is reduced to. Thus, we argue that such a IIB solution has no T-duality. A constant dilaton expectation value in 10-dimensional spacetime corresponds to a constant coupling in our 4-dimensional spacetime. This condition is agreeable and sometimes imposed by hand on phenomenological models.
Moreover, taking account of (\[eq:solphi\]) and assuming that string field theory would determine brane distribution, we conclude that supergravity solutions should satisfy a condition $$\sum^{m}_{A=1}\varepsilon_A\alpha_A = 0,\label{eq:necessary}$$ if they have their origins in string field theory. We hypothesise that this necessary condition also applies to general S-brane solutions, even if no dimensions can be reduced. We would like to thank Associate Professor Jeremy Williams, Tokyo Dental College, for his assistance with the English of this manuscript.
[99]{} for a review, N. Ohta, [*Accelerating cosmologies and inflation from M/superstring theories, Int. J. Mod. Phys.*]{} [**A20**]{} (2005) 1, \[hep-th/0411230\] E. Bergshoeff, C. Hull and T. Ortin [*Duality in the type-II superstring effective action, Nucl. Phys.*]{} [**B451**]{} (1995) 547, \[hep-th/9504081\] S. Roy [*Dimensional reductions of M-theory S-branes to string theory S-branes, Phys. Lett.*]{} [**B576**]{} (2003) 199, \[hep-th/0305175\] N. Ohta, [*A study of accelerating cosmologies from superstring/M theories, Prog. Theor. Phys.*]{} [**110**]{} (2003) 269, \[hep-th/0304172\] R. Mochizuki, [*Eternally accelerating spacelike braneworld cosmologies, Mod. Phys. Lett.*]{} [**A21**]{} (2006) 2989, \[hep-th/0606009\] R. Mochizuki, [*Eternally inflating cosmologies from intersecting spacelike branes, Mod. Phys. Lett.*]{} [**A23**]{} (2008) 487, \[hep-th/0710.3619\] R. Mochizuki and K. Ikegami, [T-duality and dimensional reduction of S-brane solutions]{} \[hep-th/1008.5220\] N. Ohta, [*Intersection rules for S-brane, Phys. Lett.*]{} [**B558**]{} (2003) 213, \[hep-th/0301095\] K. Maeda and M. Tanabe, [*Stationary spacetime from intersecting M-branes, Nucl. Phys.*]{} [**B738**]{} (2006) 184, \[hep-th/0510082\] R. Argurio, F. Englert and L. Houart, [*Intersection rules for p-branes, Phys. Lett.*]{} [**B 398**]{} (1997) 61, \[hep-th/9701042\] N. Ohta, [*Intersection rules for non-extreme $p$-branes, Phys. Lett.*]{} [**B403**]{} (1997) 218, \[hep-th/9702164\]
[^1]: e-mail address:rjmochi@tdc.ac.jp
[^2]: e-mail address:kikegami@tdc.ac.jp
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper investigates the application of non-orthogonal multiple access in millimeter-Wave communications (mmWave-NOMA). Particularly, we consider downlink transmission with a hybrid beamforming structure. A user grouping algorithm is first proposed according to the channel correlations of the users. Whereafter, a joint hybrid beamforming and power allocation problem is formulated to maximize the achievable sum rate, subject to a minimum rate constraint for each user. To solve this non-convex problem with high-dimensional variables, we first obtain the solution of power allocation under arbitrary fixed hybrid beamforming, which is divided into intra-group power allocation and inter-group power allocation. Then, given arbitrary fixed analog beamforming, we utilize the approximate zero-forcing method to design the digital beamforming to minimize the inter-group interference. Finally, the analog beamforming problem with the constant-modulus constraint is solved with a proposed boundary-compressed particle swarm optimization algorithm. Simulation results show that the proposed joint approach, including user grouping, hybrid beamforming and power allocation, outperforms the state-of-the-art schemes and the conventional mmWave orthogonal multiple access system in terms of achievable sum rate and energy efficiency.'
author:
- 'Lipeng Zhu, Jun Zhang, Zhenyu Xiao, Xianbin Cao, Dapeng Oliver Wu, and Xiang-Gen Xia, [^1] [^2] [^3] [^4]'
bibliography:
- 'IEEEabrv.bib'
- 'Xiao60GHz.bib'
- 'Xiao5GnNOMA.bib'
title: 'Millimeter-Wave NOMA with User Grouping, Power Allocation and Hybrid Beamforming'
---
mmWave communications, NOMA, user grouping, power allocation, hybrid beamforming.
Introduction
============
-wave (mmWave) communication has been proposed as one of the candidate key techniques for the fifth-generation (5G) wireless communications and beyond [@andrews2014will; @niu2015survey; @rapp2013mmIEEEAccess; @XiaoM2017survmmWave]. The abundant spectrum (30-300GHz) in mmWave-band can provide great potentials to meet the requirements of high data rates and low transmission latency. Due to the high path loss, large antenna array is usually utilized in mmWave communications, where beamforming techniques are required to increase the spectrum efficiency [@XiaoM2017survmmWave; @xiao2017mmWaveFD; @andrews2016modeling]. Fully digital beamforming (DBF) is one of the signal processing approaches in baseband [@Rusek2013MIMO; @Gao2017MIMO], where each antenna is driven by an independent radio frequency (RF) chain, and multiple data streams can be transmitted simultaneously. However, the DBF architecture results in unaffordable hardware cost and energy consumption in the mmWave-band with large antenna array [@Daill2017]. In contrast, analog beamforming (ABF), where the antennas share only one RF chain, is an energy-efficient alternative [@Ding2017random; @xiao2016codebook]. However, one RF chain can support only one data stream in general, which limits the spectrum efficiency. In consideration of the compromission between energy efficiency and spectrum efficiency, hybrid analog and digital beamforming (HBF) was proposed and preferred [@Gao2016hyb; @Daill2017; @Dai2018MIMONOMA]. With a small number of RF chains connected to a large number of antennas, beam gain and interference management can be achieved simultaneously.
One of the typical application scenarios for the 5G wireless communications is the massive connectivity. However, for mmWave communications with the conventional orthogonal multiple access (OMA) schemes, such as time division multiple access (TDMA), code division multiple access (CDMA), orthogonal frequency division multiple access (OFDMA), and space division multiple address (SDMA), the number of the users for each data stream in the same time-frequency-code-space resource block (RB) is one [@Daill2017; @Dai2018MIMONOMA; @xiao2018mmWaveNOMA; @Zhu2018UplinkNOMA]. Thus, the total number of served users is limited, which is no more than the number of RF chains in each RB [@Daill2017; @Dai2018MIMONOMA; @xiao2018mmWaveNOMA; @Zhu2018UplinkNOMA]. To address this problem, non-orthogonal multiple access (NOMA) was proposed to combine with mmWave communications [@Ding2017random; @Daill2017; @Ding2017survNOMA; @Dai2018NOMAsurvey]. In contrast to the conventional OMA schemes, NOMA can transmit the signals for different users in the same RB, while distinguishing them in the power domain. By employing superposition coding at the transmitter and successive interference cancellation (SIC) at the receiver, the users with different channel conditions can be served simultaneously. The number of served users in the same time-frequency-code-space RB can be improved manyfold [@Benjebbour2013ConceptNOMA; @Dai2015NOMA5G; @Ding2017survNOMA; @Dai2018NOMAsurvey; @Choi2014NOMA; @Zhu2018NOMAPSO]. Note that the implementation of NOMA does not result in extra delay caused by channel estimation and feedback compared with OMA [@sun2018delay]. Although the SIC at the receiver brings in supererogatory computation for demodulation and decoding at the NOMA user, the corresponding latency in the physical layer is negligible compared with the delay in the network layer. The performance analysis of NOMA for Ultra-Reliable and Low-Latency Communications (URLLC) has been investigated in [@Amjad2018delay], where grant-free NOMA with short-packet communications has significantly reduced the latency and improved the reliability for URLLC to support the time-critical applications. Moreover, it has been verified that NOMA with short-packet communications can significantly outperform OMA by achieving a higher effective throughput with the same latency requirement [@sun2018delay]. Besides, several schemes have been proposed to realize the tradeoff between the capacity (or energy efficiency) and the delay [@Choi2017delay; @Ding2018delay; @Ning2019delay]. The analysis and optimization of the delay for NOMA are beyond the scope of this paper.
It has been verified that applying NOMA in mmWave communications (mmWave-NOMA) can significantly improve the throughput capacity compared with mmWave-OMA [@Ding2017random; @Daill2017; @Dai2018MIMONOMA; @xiao2018mmWaveNOMA; @Zhu2018UplinkNOMA]. Due to the directional feature of mmWave transmission, it is ideal for the users whose channels are highly correlated to perform NOMA. There are several prior works on mmWave-NOMA with ABF. Using random ABF, mmWave-NOMA could outperform mmWave-OMA in terms of outage sum rates, respecting to a targeted data rate of the strong user [@Ding2017random]. In [@xiao2018mmWaveNOMA], a 2-user downlink mmWave-NOMA scenario with ABF was considered. A joint Tx beamforming and power allocation problem was formulated and solved to maximize the achievable sum rate (ASR), subject to a minimum rate constraint for each user. In [@Zhu2018UplinkNOMA], a joint Rx beamforming and power control problem was solved in a 2-user uplink mmWave-NOMA system. Furthermore, a joint Tx-Rx beamforming and power allocation problem was solved for $K$-user downlink mmWave-NOMA in [@Zhu2018NOMAPSO]. The closed-form optimal power allocation and Rx beamforming were obtained under arbitrary fixed Tx beamforming, and a boundary-compressed particle swarm optimization (BC-PSO) algorithm was proposed to solve the ABF problem with the constant modulus (CM) constraint.
In addition, mmWave-NOMA with HBF was also investigated in several literatures. In [@Daill2017], a new transmission scheme of beamspace multiple-input multiple-output NOMA (MIMO-NOMA) was proposed, where the number of users can be larger than the number of RF chains. Based on the equivalent-channel hybrid precoding scheme, an iterative algorithm was developed to obtain the optimal power allocation for the users. In [@Dai2018MIMONOMA], a user grouping algorithm and an HBF algorithm were proposed for mmWave-MIMO-NOMA system with simultaneous wireless information and power transfer. Then, the optimization for power allocation and power splitting factors was operated to maximize the ASR. The optimal power allocation and user scheduling were obtained with the branch and bound approach in [@Cui2018mmWaveNOMA], where HBF is random and fixed. In [@Wu2017hybridBF], the authors considered the problems of user pairing, hybrid beamforming and power allocation separately in an mmWave-NOMA system. In [@Zhang2017mmWaveMIMONOMA], a capacity analysis for the integrated NOMA-mmWave-massive-MIMO systems was provided based on a simplified mmWave channel model. In [@Wei2018NOMA], a multi-beam NOMA framework for hybrid mmWave systems was proposed, where a beam splitting technique was introduced to generate multiple analog beams to facilitate the NOMA transmission.
In this paper, we investigate mmWave-NOMA with HBF structures. Different from the works above, we consider user grouping and jointly optimize HBF and power allocation. Particularly, we consider a single-cell downlink system, where the base station (BS) is equipped with a large antenna array, and serves multiple single-antenna users. The contributions of this paper are summarized as follows[^5].
1. To implement NOMA in mmWave communications with HBF, we propose a user grouping algorithm first, where K-means algorithm is utilized and the normalized channel correlation is defined as the measure. The users with high channel correlation are assigned to the same group, while the users with low channel correlation are assigned to different groups, which can significantly mitigate the interference between different groups of users. Then, a problem jointly optimizing power allocation and HBF is formulated to maximize the ASR of the users, subject to a minimum rate constraint for each user.
2. We obtain a sub-optimal solution of the power allocation problem under arbitrary and fixed HBF. Since the power allocation problem is non-convex, we divide it into two sub-problems, i.e., intra-group power allocation (intra-GPA) and inter-group power allocation (inter-GPA). Significantly, we prove the proposed solution of power allocation is globally optimal under ideal beam pattern (i.e., no inter-group interference).
3. We design the HBF matrix to suppress the inter-group interference as well as maximize the ASR. In the proposed solution, DBF is designed by using the approximate zero-forcing (AZF) method under arbitrary and fixed ABF. Then, substituting the obtained power allocation and DBF as the function of the ABF matrix, we utilize the boundary-compressed particle swarm optimization (BC-PSO) algorithm to solve the ABF problem, which realizes the joint optimization of power allocation and HBF.
4. We evaluate the performance of the proposed user grouping, power allocation and HBF algorithm for mmWave-NOMA through simulations. The simulation results show that the proposed solution is significantly better than those of state-of-the-art schemes and the conventional mmWave-OMA system in terms of ASR. The energy-efficiency (EE) performance of the proposed mmWave-NOMA scheme with an HBF structure outperforms the fully digital MIMO structure[^6]. The ASR of the proposed solution is close to the ideal case with no inter-group interference, which demonstrates that the designed HBF can significantly achieve low inter-group interference.
The rest of the paper is organized as follows. In Section II, we present the system model. In Section III, we first propose the user grouping algorithm and formulate the problem. Then, we provide a solution of power allocation with an arbitrary fixed HBF in Section IV. In Section V, we design DBF and ABF. In Section VII, we summarize the complete solution and provide the computational complexity. Simulation results are given to demonstrate the performance of the proposed solution in Section VII, and the paper is concluded finally in Section VIII.
Symbol Notation: $a$, $\mathbf{a}$, $\mathbf{A}$ and $\mathcal{A}$ denote a scalar, a vector, a matrix and a set, respectively. $(\cdot)^{\rm{T}}$, $(\cdot)^{\rm{H}}$ and $(\cdot)^{\dag}$ denote transpose, conjugate transpose and pseudo inverse, respectively. $|a|$ and $\|\mathbf{a}\|$ denote the absolute value of $a$ and Frobenius norm of $\mathbf{a}$, respectively, while $|\mathcal{A}|$ denotes the number of elements in set $\mathcal{A}$. $\mathbb{E}(\cdot)$ denotes the expectation operation. $[\mathbf{a}]_i$, $[\mathbf{A}]_{i,:}$, $[\mathbf{A}]_{:,j}$ and $[\mathbf{A}]_{i,j}$ denote the $i$th entry of $\mathbf{a}$, the $i$th row, the $j$th column, and the entry in the $i$th row and the $j$th column of $\mathbf{A}$, respectively. $\mathbf{I}_K$ is the $K \times K$ identity matrix and $\Phi$ denotes the empty set.
System Model
============
System model
------------
![Illustration of the architecture of the BS, which is equipped with $M$ RF chains and $N$ antennas.[]{data-label="fig:system"}](system.eps){width="8.8"}
In this paper, we consider a single-cell downlink mmWave-NOMA system. The BS is equipped with HBF structure, where $N$ antennas share $M$ RF chains. $K$ single-antenna users are served simultaneously, where $K>M$. The architecture of the BS is shown in Fig. \[fig:system\], which is a fully connected HBF structure[^7]. $N_{S}$ data streams in the baseband are precoded by the DBF matrix $\mathbf{D}^{M\times N_{S}}$. After passing through the corresponding RF chain, the digital-domain signal from each RF chain is delivered to $N$ phase shifters (PSs) to perform ABF. Thus, the ABF matrix is $\mathbf{A}^{N\times M}$.
In order to achieve a higher multiplexing gain, the number of data streams is assumed to be equal to the number of RF chains in this paper, i.e., $N_{S}=M$. Thus, the $K$ users should be first scheduled into $M$ groups, and each group is corresponding to an independent data stream. The users in the same group can perform NOMA and implement SIC, while the signals from different groups of users are treated as interference. The details of user grouping will be shown later. Denote the user set of the $m$th group as $\mathcal{G}_{m}$. As a result, we have $\mathcal{G}_{i}\cap \mathcal{G}_{j}=\Phi$ for $i\neq j$ and $\sum \limits_{m=1}^{M} |\mathcal{G}_{m}|=K$, where $|\mathcal{G}_{m}|$ denotes the number of users for $\mathcal{G}_{m}$. Since $M$ RF chains can support $M$ data streams at most, there should be at least one user in each group to avoid the idleness of the RF resource, and thus we have $|\mathcal{G}_{m}|\geq 1$. Then, the received signal for the $n$th user in the $m$th group is $$y_{m,n}=\mathbf{h}_{m,n}^{\rm{H}}\mathbf{ADPs}+u_{m,n},$$ where $\mathbf{h}_{m,n}$ with $N \times 1$ dimension is the channel response vector between the BS and the $n$th user in the $m$th group. $u_{m,n}$ is the Gaussian white noise at the user with average power $\sigma^2$. $\mathbf{s}^{K \times 1}$ is the vector of the transmission signals, where $\mathbf{s}=[s_{1,1}, \cdots,s_{1,|\mathcal{G}_{1}|},\cdots,s_{M,1}, \cdots,s_{M,|\mathcal{G}_{M}|}]^{\mathrm{T}}$ and $E(\mathbf{ss}^{\rm{T}})=\mathbf{I}_{K}$, and $\mathbf{P}$ is the $M \times K$ power allocation matrix: $\mathbf{P}=\mathrm{diag}\{\mathbf{p}_1,\mathbf{p}_2,\cdots,\mathbf{p}_M\}$ and $\mathbf{p}_m=[\sqrt{p_{m,1}},\sqrt{p_{m,2}},\cdots,\sqrt{p_{m,|\mathcal{G}_{m}|}}]$. $\mathbf{D}$ is the DBF matrix. $\mathbf{A}$ is the ABF matrix with the CM constraint of [@xiao2018mmWaveNOMA; @Zhu2018UplinkNOMA; @Zhu2018NOMAPSO] $$|[\mathbf{A}]_{i,j}|=\frac{1}{\sqrt{N}},~1\leq i \leq N,~1\leq j \leq M.$$
We define the HBF matrix as $$\mathbf{W}=\mathbf{AD}=[\mathbf{w}_{1},\mathbf{w}_{2},\cdots,\mathbf{w}_{M}].$$ Since we separate the transmission power from HBF, it is without loss of generality to assume that each column of the HBF matrix has a unit norm, i.e., $$\|\mathbf{w}_{m}\|=1,~1\leq m \leq M.$$
Subject to limited scattering in the mmWave band, multipath is mainly caused by reflection. As the number of the multipath components (MPCs) is small in general, the mmWave channel has directionality and appears spatially sparse in the angle domain [@peng2015enhanced; @wang2015multi; @Lee2014exploiting; @Gao2016ChannelEst; @xiao2016codebook; @alkhateeb2014channel]. Different MPCs have different angles of departure (AoDs) and angles of arrival (AoAs). Without loss of generality, we adopt the directional mmWave channel model assuming a uniform linear array (ULA) with a half-wavelength antenna spacing. For the $N\times 1$ channel response vector $\mathbf{h}_{m,n}$, we adopt the widely used Saleh-Valenzuela channel for mmWave communications [@Ding2017random; @Daill2017; @Dai2018MIMONOMA], which is[^8] $$\label{eq_oriChannel}
\mathbf{h}_{m,n}=\sum_{\ell=1}^{L_{m,n}}\lambda_{m,n}^{(\ell)}\mathbf{a}(N,\theta_{m,n}^{(\ell)}).$$ Note that for convenience, we denote the channel coefficients in terms of both the indexes $m$ and $n$ in , where $m~(1\leq m \leq M)$ represents the $m$th group, and the index $n~(1\leq n \leq |\mathcal{G}_{m}|)$ represents the $n$th user in each group. $\lambda_{m,n}^{(\ell)}$ is the complex coefficient of the $\ell$-th MPC of the channel response vector for the $n$th user in the $m$th group. $\theta_{m,n}^{(\ell)}$, within the range $(-1,1]$, is the cosine of the AoD [@balanis2016antenna]. $L_{m,n}$ is the total number of the MPCs. $\mathbf{a}(\cdot)$ is the steering vector functions defined as $$\begin{aligned}
\label{eq_steeringVCT}
&\mathbf{a}(\theta)=[e^{j2\pi0(d/\lambda)\theta},e^{j2\pi(d/\lambda)\theta},\cdots,e^{j2\pi(N-1)(d/\lambda)\theta}]^{\mathrm{T}},\end{aligned}$$ which depends on the array geometry. $d$ is the antenna spacing, and $\lambda$ is the signal wavelength. For a half-wavelength antenna spacing array, we have $d=\lambda/2$.
Achievable Rate
---------------
In general, the optimal decoding order for NOMA is the increasing order of the users’ channel gains[@saito2013non; @Dai2018NOMAsurvey]. However, for the mmWave-NOMA with HBF structure in this paper, the effective channel gains of the users are determined by both the channel gains and the beamforming gains. Thus, we need to sort the effective channel gains first, and then determine the decoding order. For notational simplicity and without loss of generality, we assume that the order of the effective channel gains in the $m$th group is $|\mathbf{h}^{\mathrm{H}}_{m,1}\mathbf{w}_{m}|^{2} \geq |\mathbf{h}^{\mathrm{H}}_{m,2}\mathbf{w}_{m}|^{2} \geq \cdots \geq |\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|}\mathbf{w}_{m}|^{2}$ [^9], and thus the optimal decoding order is the increasing order of the effective channel gains [@xiao2018mmWaveNOMA; @Ding2017random; @Daill2017]. Therefore, the $n$th user in the $m$th group can decode $s_{m,j}~(n+1 \leq j \leq |\mathcal{G}_{m}|)$ and then remove them from the received signal in a successive manner. The other signals are treated as interference. Thus, the signal to interference plus noise power ratio (SINR) of the $n$th user in the $m$th group can be written as $$\label{eq_SINR}
\gamma_{m,n}=\frac{|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{m}|^{2}p_{m,n}}{|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{m}|^{2}\sum \limits_{j=1}^{n-1}p_{m,j}+\sum \limits_{i\neq m}\sum \limits_{k=1}^{|\mathcal{G}_{i}|}|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{i}|^{2}p_{i,k}+\sigma^{2}}.$$
Note that Gaussian signalling is assumed for transmitting data here. As a result, the achievable rate of the $n$th user in the $m$th group is $$\label{eq_Rate}
R_{m,n}=\log_{2}(1+ \gamma_{m,n}).$$
Finally, the ASR of the proposed mmWave-NOMA system is $$\label{eq_ASR}
R_{\rm{sum}}=\sum\limits_{m=1}^{M} \sum\limits_{n=1}^{|\mathcal{G}_m|} R_{m,n}.$$
Note that in the proposed downlink mmWave-NOMA system, we assume that the CSI between the BS and the users is known by the BS, and thus user grouping, power allocation and beamforming can be accomplished at the BS. The channel-gain information and beamforming-gain information of the other users are not required at the user side. However, compared with the conventional OMA system, information about the decoding order and codebook of the prior users in the same group should be transmitted to each user to accomplish SIC, which results in extra overhead. The amount of overhead depends on the number of users with in the same NOMA group. In the proposed solution of this paper, a great number of users are divided into many NOMA groups, and the number of users within the same NOMA group is usually not large so as to maintain the performance. Hence, the extra overhead is in fact not high, especially in slow varying channel, where the decoding order and codebook are also slow varying, the overhead can be further reduced.
User Grouping and Problem Formulation
=====================================
As the number of the users is larger than that of the RF chains, i.e., $K>M$, we need to schedule the user into $M$ groups. To this end, we propose an intuitive algorithm for user grouping first, and then formulate a problem to jointly optimize HBF and power allocation.
User Grouping
-------------
Due to the spacial directivity of the mmWave channel, the users whose channels are highly correlated should be assigned to the same group to make full use of the multiplexing gain, while the users whose channels are uncorrelated should be assigned to different groups to decrease the interference. The normalized channel correlation between User $i$ and User $j$ is defined as $$C_{i,j}=\frac{\mathbf{h}^{\mathrm{H}}_{i}\mathbf{h}_{j}}{\|\mathbf{h}_{i}\|\|\mathbf{h}_{j}\|}.$$
We use the K-means clustering algorithm to implement the user grouping, where the normalized channel correlation is defined as the measure [@kanungo2002efficient]. First, we select $M$ users randomly, denoted by $\{\Omega_{1},\Omega_{2},\cdots,\Omega_{M}\}$, as the representatives of the $M$ clusters. Then, the other users can be assigned to the cluster according to the normalized channel correlation. For instance, User $k$ should be assigned to the $m^{\star}$th cluster, where $$\label{cluster}
m^{\star}=\mathop{\mathrm{arg~max}}\limits_{1\leq m \leq M} C_{k,\Omega_{m}}.$$
After that, the representative of each cluster should be updated. To further decrease the correlation of the channels between different clusters, the representative of each cluster is updated as the one with the lowest correlation with the other clusters. The correlation between a user to the other clusters is defined as the summation of the normalized channel correlation between this user to the users of the other clusters, i.e., $$\bar{C}_{k}=\sum \limits^{j \notin \mathcal{G}^{(k)} }_{1 \leq j \leq K}C_{k,j},$$ where $\mathcal{G}^{(k)}$ denotes the cluster which includes User $k$, and the representative of the $m$th cluster is updated as $$\label{representative}
\Omega_{m}=\mathop{\mathrm{arg~min}}\limits_{1\leq n \leq |\mathcal{G}_{m}|} \bar{C}_{n},$$ where $\mathcal{G}_{m}$ denotes the $m$th cluster. After updating the representative of each cluster, the other users are reassigned to the clusters according to . The iteration is stopped if the representatives of the clusters are unchanged. The details of the proposed user grouping algorithm are summarized in Algorithm \[alg\_grouping\].
\
$K$, $M$, $\{\mathbf{h}_{k}\}$, and $\{C_{i,j}\}$. The user grouping scheme: $\{\mathcal{G}_1,\mathcal{G}_2,\cdots,\mathcal{G}_M\}$.\
$\mathcal{K}=\{1,2,\cdots,K\}$. Initialize $\Omega_{m}^{(1)}=k_{m} \in \mathcal{K}$ randomly for $m=1,2,\cdots,M$. $t=1$. Initialize $\mathcal{G}_{m}=\Omega_{m}^{(t)}$ for $m=1,2,\cdots,M$. $m^{\star}=\mathop{\mathrm{arg~max}}\limits_{1\leq m \leq M} C_{k,\Omega_{m}^{(t)}}$. $\mathcal{G}_{m^{\star}}=\mathcal{G}_{m^{\star}}\bigcup k$. $t=t+1$. Update $\Omega_{m}^{(t)}$ for $m=1,2,\cdots,M$ according to . $\{\mathcal{G}_1,\mathcal{G}_2,\cdots,\mathcal{G}_M\}$.
Problem Formulation
-------------------
Generally, there are mainly two categories of optimizing the overall rate performance in a communication system. One is to maximize the ASR. However, when maximizing the sum rate, the BS tends to allocate most power and beam gains to the users with the strong channels. Then, the users with the low channel gains can not be served by the BS. The other category is to ensure the user fairness, where the max-min fairness or proportion fairness are considered to improve the performance of the users with worse channel conditions. However, the fairness among the users may result in a performance loss of the sum rate. To realize the tradeoff between the sum-rate performance and the user fairness, we maximize the achievable sum rate while ensuring the minimum achievable rate of each user in this paper, which is also adopted in the related mmWave-NOMA systems [@Daill2017; @Dai2018MIMONOMA; @Cui2018mmWaveNOMA]. Then, the problem is formulated as $$\label{eq_problem}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{p_{m,n}\},\mathbf{A},\mathbf{D}}~~~~ &R_{\rm{sum}}\\
\mathrm{s.t.}~~~~~~~~ &C_1~:~R_{m,n} \geq r_{m,n},~~\forall m,n, \\
&C_2~:~p_{m,n} \geq 0, ~~\forall m,n,\\
&C_3~:~\sum\limits_{m=1}^{M} \sum\limits_{n=1}^{|\mathcal{G}_m|} p_{m,n} \leq P, \\
&C_4~:~|[\mathbf{A}]_{i,j}| = \frac{1}{\sqrt{N}},~~\forall i,j,\\
&C_5~:~\|[\mathbf{AD}]_{:,m}\| = 1,~~\forall m,
\end{aligned}$$ where the constraint $C_1$ is the minimum rate constraint for each user. The constraint $C_2$ indicates that the power allocated to each user should be non-negative. The constraint $C_3$ is the total transmission power constraint, where the total power at the BS is no more than $P$. $C_4$ is the CM constraint for the ABF matrix, and $C_5$ is the unit power constraint for the HBF matrix.
The total dimension of the variables in Problem is $K+MN+M^2$, which is large in general. Exhaustive search for the optimal solution results in heavy computational load, which is hard to accomplish in practice. To solve Problem , there are two main challenges. One is that the optimized variables are entangled with each other, which makes the formulation non-convex. The other is that the expression of $R_{\rm{sum}}$ depends on the decoding order. In general, the optimal decoding order is the increasing order of the users’ effective channel gains. However, the order of effective channel gains varies with different beamforming matrixes. In other words, given different HBF matrixes, the objective function in Problem , i.e., the ASR of the users, has different expressions. The two challenges make it infeasible to solve Problem by using the existing optimization tools. Next, we will propose a sub-optimal solution with promising performance but low computational complexity.
The proposed solution of Problem can be obtained with two stages. In the first stage, we provide a low-complexity algorithm to obtain the sub-optimal power allocation with an arbitrary fixed HBF. In the second stage, we design the HBF, where the DBF matrix and the ABF matrix are obtained using the AZF method and the proposed BC-PSO algorithm, respectively.
Solution of Power Allocation
============================
As we have analyzed before, an essential challenge to solve Problem is the variation of the decoding order. However, given an arbitrary fixed ABF matrix ${\bf{A}}$ and an arbitrary fixed DBF matrix ${\bf{D}}$, the order of the effective channel gains is fixed. For notational simplicity and without loss of generality, we assume $|\mathbf{h}^{\mathrm{H}}_{m,1}\mathbf{w}_{m}|^{2} \geq |\mathbf{h}^{\mathrm{H}}_{m,2}\mathbf{w}_{m}|^{2} \geq \cdots \geq |\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|}\mathbf{w}_{m}|^{2}$ for any $1\leq m\leq M$, where $\mathbf{w}_{m}=[\mathbf{AD}]_{:,m}$. The original problem can be simplified as $$\label{eq_problem2}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{p_{m,n}\}}~~~~ &R_{\rm{sum}}\\
\mathrm{s.t.}~~~~~~ &C_1~:~R_{m,n} \geq r_{m,n},~~\forall m,n, \\
&C_2~:~p_{m,n} \geq 0, ~~\forall m,n,\\
&C_3~:~\sum\limits_{m=1}^{M} \sum\limits_{n=1}^{|\mathcal{G}_m|} p_{m,n} \leq P,
\end{aligned}$$ where ${\bf{A}}$ and ${\bf{D}}$ are arbitrary and fixed.
According to the expression of the achievable rate in , a user may suffer the interference from both the intra-group users and the inter-group users. Although the HBF matrix is fixed, the objective function and the constraint $C_1$ of Problem are still non-convex. To address this problem, we divide it into two sub-problems, i.e., intra-GPA and inter-GPA. Define $\sum\limits_{n=1}^{|\mathcal{G}_m|} p_{m,n}=P_{m}$ for $1\leq m\leq M$, which means the allocated power for the $m$th group, and then Problem is equivalent to $$\label{eq_problem3}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{P_{m}\}}\mathop{\mathrm{Max}}\limits_{\{p_{m,n}\}}~~~~ &R_{\rm{sum}}\\
\mathrm{s.t.~~~~~~~}~~~ &C_1~:~R_{m,n} \geq r_{m,n},~~\forall m,n, \\
&C_2~:~p_{m,n} \geq 0, ~~\forall m,n,\\
&C_3~:~\sum\limits_{n=1}^{|\mathcal{G}_m|} p_{m,n} = P_{m}, ~~\forall m,\\
&C_4~:~\sum\limits_{m=1}^{M} P_{m} \leq P,
\end{aligned}$$
Note that the introduced inter-GPA variables, i.e., $\{P_{m}\}$, have no influence on the optimality of the power allocation problem, because there is no loss of the degree of freedom in Problem compared with Problem , and Problem is more tractable. First, given arbitrary and fixed inter-GPA, a closed-form sub-optimal intra-GPA can be obtained. Then, substituting the intra-GPA into Problem , we can obtain a sub-optimal inter-GPA solution. Although the proposed solution of power allocation is not globally optimal, we will prove that it is near-to-optimal when the inter-group interference is small through the theoretical analysis and simulation verification.
The Intra-GPA Problem
---------------------
As shown in and , one user may suffer the interference from the users in the same group and the users in other groups, which are called intra-group interference and inter-group interference, respectively. Considering that HBF can be well designed in general, such that the inter-group interference is small and can be neglected. Thus, we have the following proposition to solve the intra-GPA problem.
Given an arbitrary fixed inter-GPA of $\{P_1,P_2,\cdots,P_M\}$, if the inter-group interference can be neglected, the optimal intra-GPA in Problem should always satisfy $$R_{m,n}=r_{m,n}~(1\leq m \leq M,~2\leq n \leq |\mathcal{G}_m|).$$
If the inter-group interference is small and can be neglected, Problem can be divided into $M$ independent intra-GPA problems. For the $m$th group, the intra-GPA problem is simplified as $$\label{eq_problem4}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{p_{m,n}\}}~~~~ &\sum \limits _{n=1}^{|\mathcal{G}_{m}|}R_{m,n}\\
\mathrm{s.t.~~~}~~~ &C_1~:~R_{m,n} \geq r_{m,n},~~\forall n, \\
&C_2~:~p_{m,n} \geq 0, ~~\forall n,\\
&C_3~:~\sum\limits_{n=1}^{|\mathcal{G}_{m}|} p_{m,n} = P_{m},
\end{aligned}$$ which is a power allocation problem without inter-group interference. This problem has been solved in [@Zhu2018NOMAPSO], where the optimal power allocation always satisfies $R_{m,n}=r_{m,n} ~(2\leq n \leq |\mathcal{G}_m|)$.
By solving the equation sets of $R_{m,n}=r_{m,n}~(1\leq m \leq M,~2\leq n \leq |\mathcal{G}_m|)$ and $\sum\limits_{n=1}^{|\mathcal{G}_{m}|} p_{m,n} = P_{m}~(1\leq m \leq M)$, we can obtain a sub-optimal intra-GPA for each group of users, which is shown in on the top of the next page, where $\eta_{m,n}=2^{r_{m,n}}-1$. Note that although the inter-group interference is neglected in Proposition 1, it is included when solving the equation sets. Thus, the minimal rate constraints for the users (from the 2nd one to the last one in each group) are always satisfied. The impact of the approximation on the inter-group interference will be evaluated in the simulation.
$$\label{opt_power}
\left \{
\begin{aligned}
&p_{m,|\mathcal{G}_{m}|}^{\circ}=\frac{\eta_{m,|\mathcal{G}_{m}|}}{\eta_{m,|\mathcal{G}_{m}|}+1}(P_{m}+\frac{\sum \limits_{i\neq m}|\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|}\mathbf{w}_{i}|^{2}P_{i}+\sigma^2}{|\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|}\mathbf{w}_{m}|^{2}}),\\
&p_{m,|\mathcal{G}_{m}|-1}^{\circ}=\frac{\eta_{m,|\mathcal{G}_{m}|-1}}{\eta_{m,|\mathcal{G}_{m}|-1}+1}(P_{m}-p_{m,|\mathcal{G}_{m}|}^{\circ}+\frac{\sum \limits_{i\neq m}|\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|-1}\mathbf{w}_{i}|^{2}P_{i}+\sigma^2}{|\mathbf{h}^{\mathrm{H}}_{m,|\mathcal{G}_{m}|-1}\mathbf{w}_{m}|^{2}}),\\
&~~~~\vdots\\
&p_{m,2}^{\circ}=\frac{\eta_{m,2}}{\eta_{m,2}+1}(P_{m}-\sum\limits_{k=3}^{|\mathcal{G}_{m}|}p_{m,k}^{\circ}+\frac{\sum \limits_{i\neq m}|\mathbf{h}^{\mathrm{H}}_{m,2}\mathbf{w}_{i}|^{2}P_{i}+\sigma^2}{|\mathbf{h}^{\mathrm{H}}_{m,2}\mathbf{w}_{m}|^{2}}),\\
&p_{m,1}^{\circ}=P_{m}-\sum\limits_{k=2}^{|\mathcal{G}_{m}|}p_{m,k}^{\circ},
\end{aligned}
\right.$$
Under Proposition 1, the ASR in Problem can be simplified as $$\label{eq_ASR2}
R_{\rm{sum}}=\sum\limits_{m=1}^{M}R_{m,1}+\sum\limits_{m=1}^{M} \sum\limits_{n=2}^{|\mathcal{G}_m|}r_{m,n}.$$
Substituting into Problem , Problem can be transformed to $$\label{eq_problem5}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{P_{m}\}}~~~~ &\sum\limits_{m=1}^{M}R_{m,1}\\
\mathrm{s.t.~~~}~~~ &C_1~:~R_{m,1} \geq r_{m,1},~~\forall m, \\
&C_2~:~\sum\limits_{m=1}^{M} P_{m} \leq P,
\end{aligned}$$ which is an inter-GPA problem.
The Inter-GPA Problem
---------------------
Due to the inter-group interference in the expression of the objective function, it is still challenging to solve Problem . We propose an iterative algorithm here. First, we initialize the group power $P_{m}$ equally. Then, we start iteration. In each iteration, the inter-group interference is assumed to be invariable, and we update the inter-GPA by maximizing the ASR in Problem , where the inter-group interference is defined as $$\begin{aligned}
I^{\mathrm{(inter)}}_{m,n}\triangleq \sum \limits_{i\neq m}\sum \limits_{k=1}^{|\mathcal{G}_{i}|}|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{i}|^{2}p_{i,k}=\sum \limits_{i\neq m}|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{i}|^{2}P_{i}.
\end{aligned}$$
Thus, the SINR for the first user in each group is linear to its signal power, i.e., $$\gamma_{m,1}=\frac{|\mathbf{h}^{\mathrm{H}}_{m,1}\mathbf{w}_{m}|^{2}p_{m,1}^{\circ}}{I^{\mathrm{(inter)}}_{m,1}+\sigma^2}$$ where $p_{m,1}^{\circ}$ is defined in . Furthermore, according to the expression in , if the inter-group interference is invariable, $p_{m,1}^{\circ}$ is also linear to $P_{m}$. Thus, we can obtain the relationship between $\gamma_{m,1}$ and $P_{m}$ as $$\label{SINR_linear2}
\gamma_{m,1}=k_{m}P_{m}+b_{m},$$ where $k_{m}$ and $b_{m}$ are given by $$\begin{aligned}
&k_{m}=\frac{|\mathbf{h}^{\mathrm{H}}_{m,1}\mathbf{w}_{m}|^{2}}{I^{\mathrm{(inter)}}_{m,1}+\sigma^2}\Bigg{(}1-\sum \limits_{n=2}^{|\mathcal{G}_{m}|} \Big{[} \eta_{m,n}\prod \limits_{j=2}^{n}\frac{1}{(\eta_{m,j}+1)}\Big{]} \Bigg{)},\\ \nonumber
&b_{m}=-\frac{|\mathbf{h}^{\mathrm{H}}_{m,1}\mathbf{w}_{m}|^{2}}{I^{\mathrm{(inter)}}_{m,1}+\sigma^2} \times \sum \limits_{n=2}^{|\mathcal{G}_{m}|} \Big{[} \eta_{m,n}\frac{I^{\mathrm{(inter)}}_{m,n}+\sigma^2}{|\mathbf{h}^{\mathrm{H}}_{m,n}\mathbf{w}_{n}|^{2}}\prod \limits_{j=2}^{n}\frac{1}{(\eta_{m,j}+1)}\Big{]}.\end{aligned}$$
It is easy to verify that $k_{m}>0$ and $b_{m}<0$. Then, the objective function in Problem is equal to $$\begin{aligned}
\sum\limits_{m=1}^{M}R_{m,1}&=\sum\limits_{m=1}^{M}\log_{2}(1+\gamma_{m,1}) =\sum\limits_{m=1}^{M}\log_{2}(k_{m}P_{m}+b_{m}+1)\triangleq f(\{P_{m}\}).
\end{aligned}$$
Constraint $C_1$ in Problem is equivalent to $$\begin{aligned}
&R_{m,1} \geq r_{m,1}
\Leftrightarrow \gamma_{m,1} \geq \eta_{m,1}
\Leftrightarrow P_{m} \geq \frac{\eta_{m,1}-b_{m}}{k_{m}}.
\end{aligned}$$
As the objective function becomes concave now and the constraints are linear, Problem can be directly solved by using the convex optimization tools [@boyd2004convex]. In order to explore the essential principle of the inter-GPA for mmWave-NOMA, we propose a method with low computation complexity here. We begin from the case without constraint $C_1$ in Problem and give the following Lemma.
If the inter-group interference is assumed to be invariant in Problem , without the constraint $C_1$, the globally optimal solution is $$\label{inter_power}
P_{m}^{\star}=\frac{P+\sum \limits_{i=1}^{M}\frac{b_{i}+1}{k_{i}}}{M}-\frac{b_{m}+1}{k_{m}}, ~1\leq m \leq M.$$
See Appendix A.
According to Lemma 1, if $P_{m}^{\star}$ in is located in the feasible domain of the constraint $C_1$ in Problem , i.e., $P_{m}^{\star}\geq \frac{\eta_{m,1}-b_{m}}{k_{m}}$ for all $1\leq m\leq M$, $P_{m}^{\star}$ is the optimal solution of Problem . However, if $P_{m}^{\star}$ in is not located in the feasible domain of the constraint $C_1$ in Problem , i.e. $P_{m}^{\star}>\frac{\eta_{m,1}-b_{m}}{k_{m}}$ for any one of $1\leq m\leq M$, $P_{m}^{\star}$ is not the optimal solution of Problem . We may find the optimal solution by using the following Lemma.
If the inter-group interference is assumed to be invariant in Problem , with the constraint $C_1$, the globally optimal solution should always satisfy $$\label{inter_power2}
P_{m}^{\circ}=\frac{\eta_{m,1}-b_{m}}{k_{m}}, ~\forall m \in \mathcal{U},$$ where $\mathcal{U}=\{i|1\leq i\leq M,~P_{i}^{\star}<\frac{\eta_{i,1}-b_{i}}{k_{i}}\}$ and $P_{i}^{\star}$ is defined in .
See Appendix B.
Lemma 2 provides the globally optimal power allocation for $m \in \mathcal{U}$. For $m \notin \mathcal{U}$, the optimal power allocation can be obtained by solving the following problem. $$\label{eq_problem5.2}
\begin{aligned}
\mathop{\mathrm{Max}}\limits_{\{P_{m}\}}~~~~ &\sum\limits_{m \notin \mathcal{U}} R_{m,1}\\
\mathrm{s.t.~~~}~~~ &C_1~:~R_{m,1} \geq r_{m,1},~~m \notin \mathcal{U}, \\
&C_2~:~\sum\limits_{m \notin \mathcal{U}} P_{m} \leq P-\sum\limits_{j \in \mathcal{U}} P_{j}^{\circ},
\end{aligned}$$ which has a similar formulation with Problem . Thus, Lemma 1 and Lemma 2 can also be used to solve Problem , which forms a closed loop. In summary, we give Algorithm \[alg\_powerallo\] to accomplish the inter-GPA.
$K$, $M$, $\{\mathcal{G}_m\}$, $P$, $\{\mathbf{h}_{k}\}$, $\{r_{k}\}$, $\mathbf{W}$, and $F_{\mathrm{max}}$.\
Inter-GPA: $\{P_{m}^{\circ}\}$.\
$P_{m}^{\circ(0)}=\frac{P}{M} ~(1\leq m \leq M)$. $\mathcal{M}=\{1,2,\cdots,M\}$. $\mathcal{U}=\mathcal{M}$. Obtain $k_{m},~b_{m} ~(\forall m \in \mathcal{M})$ in . Obtain $P_{m}^{\star} ~(\forall m \in \mathcal{M})$ according to . $\mathcal{U}=\{i|i \in \mathcal{M},~P_{i}^{\star}<\frac{\eta_{i,1}-b_{i}}{k_{i}}\}$. $P_{m}^{\circ(t)}=\frac{\eta_{m,1}-b_{m}}{k_{m}} ~(\forall m \in \mathcal{U})$. $\mathcal{M}=\mathcal{M}/\mathcal{U}$. $P_{m}^{\circ(t)}=P_{m}^{\star} ~(\forall m \in \mathcal{M})$. $P_{m}^{\circ}=P_{m}^{\circ(T_{\mathrm{max}})} ~(1 \leq m \leq M)$. $\{P_{m}^{\circ}\}$.
Hereto, the power allocation is solved. Given an arbitrary fixed HBF, we can obtain the inter-GPA using Algorithm \[alg\_powerallo\] and obtain the intra-GPA according to . Since the proposed intra-GPA and inter-GPA solutions are both sub-optimal, we provide the following theorem to evaluate the optimality of the proposed power allocation solution.
If the inter-group interference in Problem is small and approaches to zero, the proposed solution of power allocation in Algorithm \[alg\_powerallo\] and is globally optimal.
If the inter-group interference is zero, the intra-GPA problems are independent for different groups. According to the conclusion in \[21, Theorem 1\], is the optimal intra-GPA solution with the given fixed inter-GPA. Substituting into Problem , the inter-GPA problem is concave and can be solved by using Algorithm \[alg\_powerallo\] with only one iteration. Due to the concavity, the inter-GPA solution is also optimal. Thus, the globally optimal power allocation can be obtained by using the proposed scheme if the inter-group interference is zero.
Based on Theorem 1, we can find that the optimality of the power allocation solution depends on the inter-group interference, which can be restrained through the elaborate beamforming design. Thus, the design of HBF should take both decreasing the interference and increasing the ASR into account. The details will be shown in the next section.
Solution of Hybrid Beamforming
==============================
In this Section, we provide the solution of HBF in Problem . As we have analyzed previously, the design of HBF should guarantee the suppression of the inter-group interference, as well as the improvement of the ASR. For mmWave-NOMA, there may exist more than one users in each group. The traditional unidirectional beamforming cannot support all the users. Thus, a multi-directional beamforming scheme is required in the analog domain. However, the non-convex modulus constraint for ABF makes the beamforming problem challenging. Besides, as shown in , due to the superposition of the inter-group interference and the intra-group interference, it is difficult to obtain the optimal HBF solution. To this end, we propose a sub-optimal approach. First, the DBF is designed using the AZF method to reduce the inter-group interference, where the ABF matrix is arbitrary and fixed. Then, we use the BC-PSO algorithm in [@Zhu2018NOMAPSO] to solve the ABF problem, where the power allocation and DBF matrix are substituted as the function of the ABF matrix.
DBF with Arbitrary Fixed ABF
----------------------------
As each group of users have a unique DBF vector, we may design the DBF with the AZF method to reduce the inter-group interference, where the ABF is arbitrary and fixed. Since the rank of the DBF matrix is no more than the number of the users, i.e., $M\leq K$, the inter-group interference cannot be completely suppressed through DBF. Recalling that when optimizing the power allocation, the rate gains are acquired at the first user in each group. Thus, we select the channel response vector of the user with the highest channel gain in each group as the equivalent channel vector. Note that the channel gain utilized here corresponds to the power of the channel response vector before beamforming, which differs from the effective channel gain after beamforming. Then, the $N\times M$ equivalent channel matrix is $$\mathbf{\tilde{H}}=[\mathbf{h}_{1,1},\mathbf{h}_{2,1},\cdots,\mathbf{h}_{M,1}].$$
Consequently, the DBF matrix can be generated by the AZF method as [^10] $$\label{DBF}
\mathbf{\tilde{D}}=(\mathbf{\tilde{H}}^{\mathrm{H}}\mathbf{A})^{\dag}.$$
Due to the unit power constraint for the HBF matrix, each column of the DBF matrix should be normalized as $$\label{DBF_norm}
[\mathbf{D}^{\circ}]_{:,m}=\frac{[\mathbf{\tilde{D}}]_{:,m}}{\|\mathbf{A}[\mathbf{\tilde{D}}]_{:,m}\|}.$$
Although the inter-group interference cannot be completely eliminated with DBF, it can be further suppressed with ABF, which has a higher degree of freedom.
ABF Using BC-PSO Alogrithm
--------------------------
Given an arbitrary fixed ABF matrix, we can obtain the DBF matrix according to and . Then, the inter-GPA can be obtained by Algorithm \[alg\_powerallo\], and meanwhile the intra-GPA is given by . It is hard to optimize ABF with the conventional approaches, since the closed-form expression of $R_{\mathrm{sum}}$ over $\mathbf{A}$ is complicated. In addition, the ABF matrix $\mathbf{A}$ with CM constraint is high-dimensional, i.e., $N \times M$, which makes the ABF design difficult.
To solve this difficult problem, particle swarm optimization (PSO) is a good approach [@fukuyama2008fundamentals]. In the $N\times M$-dimensional search space $\mathcal{S}$, the $I$ particles in the swarm are randomly initialized with position $\mathbf{A}$ and velocity $\mathbf{V}$. Each particle has a memory for its best found position $\mathbf{P}_{\text{best}}$ and the globally best position $\mathbf{G}_{\text{best}}$, where the goodness of a position is evaluated by the fitness function. For each iteration, the velocity and position of each particle are updated based on $$\label{eq_PSOregular}
\begin{aligned}
&[\mathbf{V}]_{i,j}=\omega[\mathbf{V}]_{i,j}+c_{1}\text{rand()}*([\mathbf{P}_{\text{best}}]_{i,j}-[\mathbf{A}]_{i,j})+c_{2}\text{rand()}*([\mathbf{G}_{\text{best}}]_{i,j}-[\mathbf{A}]_{i,j})\\
&[\mathbf{A}]_{i,j}=[\mathbf{A}]_{i,j}+[\mathbf{V}]_{i,j}
\end{aligned}$$ for $i=1,2,\cdots,N;~j=1,2,\cdots,M$. The parameter $\omega$ is the inertia weight of velocity. In general, $\omega$ is decreasing linearly from the maxima to the minima for each time of iteration to improve the convergence speed. The parameters $c_{1}$ and $c_{2}$ are the cognitive ratio and social ratio, respectively. The random number function rand() returns a number between 0.0 and 1.0 with uniform distribution.
Due to the CM constraint, the search space for $\mathbf{A}$, i.e., $\{\mathbf{A}\big{|}|[\mathbf{A}]_{i,j}| = \frac{1}{\sqrt{N}}\}$, is highly non-convex. It has been shown that the BC-PSO algorithm outperforms the classic PSO algorithm in the ABF problem [@Zhu2018NOMAPSO]. The key idea of the BC-PSO algorithm is to relax the search space as a convex set, i.e., $\mathcal{S}=\{\mathbf{A}\big{|}|[\mathbf{A}]_{i,j}| \leq \frac{1}{\sqrt{N}}\}$, and adjust the particles onto the boundaries for each iteration to satisfy the CM constraint. The outer boundary is defined as $\{\mathbf{A}\big{|}|[\mathbf{A}]_{i,j}| = d_{\mathrm{out}}\}$, where $d_{\mathrm{out}}=\frac{1}{\sqrt{N}}$ is fixed. The inter boundary is defined as $\{\mathbf{A}\big{|}|[\mathbf{A}]_{i,j}| = d_{\mathrm{in}}\}$, where $d_{\mathrm{in}}=\frac{t}{T_{\mathrm{max}}}\frac{1}{\sqrt{N}}$ is dynamic. $T_{\mathrm{max}}$ is the maximum number of iterations and $t=1,2,\cdots,T_{\mathrm{max}}$. For each iteration, the particles out of the boundaries are adjusted onto the boundaries. Then, after calculating the fitness function for each particle, the locally and globally best positions, i.e., $\mathbf{P}_{\text{best}}$ and $\mathbf{G}_{\text{best}}$, are updated. With this implementation, the particles can move throughout the relaxed search space and converge to satisfy the CM constraint eventually. Compared with the classic PSO algorithm, the BC-PSO algorithm has enhanced search capabilities.
Summary of the Complete Solution and Computational Complexity
=============================================================
Summary of The Complete Solution
--------------------------------
In the above sections, we have presented the algorithms and formulas, respectively, for user grouping, power allocation, digital beamforming and analog beamforming. Based on these algorithms and formulas, we give the complete solution to realize an arbitrary mmWave-NOMA system. As shown in in Algorithm \[alg\_PSO\], we firstly use Algorithm \[alg\_grouping\] to divide the users into $M$ groups, and obtain $\{\mathcal{G}_m\}$. Then, we use the BC-PSO algorithm to iteratively optimize the position of the particle, i.e., the ABF matrix, where the fitness function is defined as the ASR in . Note that in the part of power allocation and DBF, we assume that the ABF matrix is arbitrary and fixed. Thus, the power allocation and DBF can be substituted as the function of the analog beamforming matrix in Algorithm \[alg\_PSO\]. Given different ABF matrixes, we should calculate the power allocation and DBF matrixes first, and then obtain the ASR. In each iteration, the computations of the DBF matrix $\mathbf{D}^{\circ}$ using and , the inter-GPA $\{P_{m}^{\circ}\}$ using Algorithm \[alg\_powerallo\], and the intra-GPA $\{p_{m,n}^{\circ}\}$ using are performed sequentially after determining the ABF matrix. Hence, after $T_{\mathrm{max}}$ iterations, the sub-optimal overall solution $\mathbf{A}^{\circ}$, $\mathbf{D}^{\circ}$ and $\{p_{m,n}^{\circ}\}$ are jointly obtained.
$K$, $M$, $N$, $P$, $\{\mathbf{h}_{k}\}$, $\{r_{k}\}$, and parameters\
for BC-PSO $\{I$, $T_{\mathrm{max}}$, $c_1$, $c_2$, $\omega_{\text{max}}$, $\omega_{\text{min}}\}$. $\{\mathcal{G}_m\}$, $\mathbf{A}^{\circ}$, $\mathbf{D}^{\circ}$ and $\{p_{m,n}^{\circ}\}$.\
Obtain the user grouping $\{\mathcal{G}_m\}$ using Algorithm \[alg\_grouping\]. Initialize the position $\mathbf{A}_{i}$ and velocity $\mathbf{V}_{i}$. Find the globally best position $\mathbf{G}_{\text{best}}$. $\omega=\omega_{\text{max}}-\frac{t}{T}(\omega_{\text{max}}-\omega_{\text{min}})$. $d_{\mathrm{out}}=\frac{1}{\sqrt{N}},~d_{\mathrm{in}}=\frac{t}{T_{\mathrm{max}}}\frac{1}{\sqrt{N}}$. Update $[\mathbf{V}_{l}]_{i,j}$ and $[\mathbf{A}_{l}]_{i,j}$ based on . $[\mathbf{A}_{l}]_{i,j}=d_{\mathrm{out}}\frac{[\mathbf{A}_{l}]_{i,j}}{|[\mathbf{A}_{l}]_{i,j}|}$. $[\mathbf{A}_{l}]_{i,j}=d_{\mathrm{in}}\frac{[\mathbf{A}_{l}]_{i,j}}{|[\mathbf{A}_{l}]_{i,j}|}$. $[\mathbf{P}_{\text{best},l}]_{i,j}=d_{\mathrm{in}}\frac{[\mathbf{P}_{\text{best},l}]_{i,j}}{|[\mathbf{P}_{\text{best},l}]_{i,j}|}$. Obtain the DBF matrix $\mathbf{D}^{\circ}$ according to and . Reorder the effective channel gains of the users in each group. Obtain the inter-GPA $\{P_{m}^{\circ}\}$ using Algorithm \[alg\_powerallo\]. Obtain the intra-GPA $\{p_{m,n}^{\circ}\}$ according to . Obtain the fitness function $R_{\mathrm{sum}}$ according to . Update $\mathbf{P}_{\text{best},l}$. Update $\mathbf{G}_{\text{best}}$. $\mathbf{A}^{\circ}=\mathbf{G}_{\text{best}}$. $\{\mathcal{G}_m\}$, $\mathbf{A}^{\circ}$, $\mathbf{D}^{\circ}$ and $\{p_{m,n}^{\circ}\}$.
Computational Complexity
------------------------
When operating the user grouping in Algorithm \[alg\_grouping\], the complexities of calculating the channel correlation and the norm channel vector are $\mathcal{O}(K^{2}N)$ and $\mathcal{O}(KN)$, respectively. In each iteration, the complexities of updating the cluster representative and the user grouping are $\mathcal{O}(K^{2})$ and $\mathcal{O}(KM)$, respectively. Since the number of antennas is much larger than that of the RF chains, i.e., $N\gg M$, the maximal complexity of Algorithm \[alg\_grouping\] is $\mathcal{O}(K^{2}N)$. In Algorithm \[alg\_powerallo\], the complexity of calculating the effective channel gains of the users is $\mathcal{O}(MKN)$. For each time of updating the inter-GPA, the maximal number of iterations to update the inter-GPA from Step 5 to 11 is $M$, and the complexity of computing the inter-GPA in each subcycle is no more than $\mathcal{O}(K^{2})$. Thus, the complexity of Algorithm \[alg\_powerallo\] is $\mathcal{O}(MKN+F_{\mathrm{max}}MK^{2})$. In Algorithm \[alg\_PSO\], the numbers of invoking Algorithm \[alg\_grouping\] and Algorithm \[alg\_powerallo\] are 1 and $T_{\mathrm{max}}IMN$, respectively. Consequently, the total computational complexity of the proposed user pairing algorithm, HBF and power allocation algorithm is $\mathcal{O}(T_{\mathrm{max}}IM^{2}KN^{2}+T_{\mathrm{max}}F_{\mathrm{max}}IM^{2}K^{2}N)$, which is a polynomial complexity. In comparison, The total computational complexity of the algorithm in [@Dai2018MIMONOMA] is $\mathcal{O}(MK^2+MN+TK^{4.5}\log_2(1/\varepsilon))$, where $T$ is the maximum iteration times and $\varepsilon$ is the solution accuracy. Since the number of the antennas is much larger than those of the users and the RF chains, i.e., $N\gg K, N\gg M$, the computational complexity in [@Dai2018MIMONOMA] is lower compared with our algorithm, because the HBF is not jointly optimized with the power allocation.
Simulation Results
==================
In this section, we provide some simulation results to verify the performance of the proposed mmWave-NOMA scheme. We adopt the channel model shown in , where the users are uniformly distributed from 10m to 100m away from the BS, and the channel gain of the node 30m away from the BS has an average power of 0 dB to noise power. The number of MPCs for all the users are $L=4$. Both LOS and NLOS channel models are considered. For the LOS channel, the average power of the NLOS paths is 15 dB weaker than that of the LOS path. For the NLOS channel, the coefficient of each path has an average power of $1/\sqrt{L}$. The cosine of the AoD for each path of the users is generated by a uniformly distributed random variable ranging from -1 to 1. Each point of the figures are the average performance of 100 channel realizations. The corresponding parameter settings are $I=800, F_{\mathrm{max}}=6, T_{\mathrm{max}}=200, c_{1}=c_{2}=1.4, \omega_{\text{max}}=0.9, \omega_{\text{min}}=0.4$.
In the simulations, we consider the following six typical mmWave communication schemes: “mmWave-NOMA Proposed” is corresponding to the proposed joint approach, including user grouping, power allocation, and HBF. “mmWave-NOMA Ideal” is based on the proposed joint approach and with assumption of none inter-group interference, i.e., $I^{\mathrm{(inter)}}_{m,n}=0$. Besides, “mmWave-NOMA \[13\]” and “fully digital MIMO” are corresponding to the approach for mmWave-NOMA with fully connected HBF structure in [@Dai2018MIMONOMA] and the mmWave-fully-digital-MIMO structure with zero-forcing precoding, respectively. For fair comparison, the power splitting part in [@Dai2018MIMONOMA] is neglected in the simulations, which means that all the power is used for wireless information transmission. “TDMA-ZF” corresponds to the performance of mmWave time division multiple access (TDMA) system, where $M$ out of $K$ users are served in each time slot. Each user is served by an independent analog beamformer with steering vector, and ZF and water-filling method is adopted for digital beamforming. While for “mmWave-FDMA”, the users are assigned into $M$ groups, and the users in the same group perform frequency division multiple access (FDMA) [@Dai2018MIMONOMA]. Then, the achievable rate of the mmWave-FDMA scheme for the $k$th user is $$\label{eq_OMA}
R_{k}^{\mathrm{FDMA}}=\frac{1}{\left|\mathcal{G}^{k}\right|}\log_{2}\left(1+\frac{|\mathbf{h}^{\mathrm{H}}_{k}\mathbf{w}_{k}|^{2}p_{k}}{\sum \limits_{j\notin \mathcal{G}^{k}}|\mathbf{h}^{\mathrm{H}}_{k}\mathbf{w}_{j}|^{2}p_{j}+\frac{\sigma^{2}}{\left|\mathcal{G}^{k}\right|}}\right),$$ where $\mathcal{G}^{k}$ represents the group which the $k$th user belongs to. The beamforming vector $\mathbf{w}_{k}$ and the power allocation $\{p_{k}\}$ are generated by using the approach in [@Dai2018MIMONOMA].
In addition, we also evaluate the performance of the EE, which is defined as the ratio between the ASR and total power consumption, i.e., $$\label{EE}
EE=\frac{R_{\mathrm{sum}}}{P+N_{\mathrm{RF}}P_{\mathrm{RF}}+N_{\mathrm{PS}}P_{\mathrm{PS}}},$$ where $R_{\mathrm{sum}}$ is the ASR. $P$ is the transmission power. $P_{\mathrm{RF}}$ is the power consumption of each RF chain, and $N_{\mathrm{RF}}$ is the number of the RF chains, where $N_{\mathrm{RF}}=N$ for the fully digital structure and $N_{\mathrm{RF}}=M$ for the hybrid structure. $P_{\mathrm{PS}}$ is the power consumption of each PS, and $N_{\mathrm{RF}}$ is the number of the PSs, where $N_{\mathrm{PS}}=0$ for the fully digital structure and $N_{\mathrm{PS}}=MN$ for the hybrid structure. In the simulations, we select the typical parameter settings of $P=$1 W, $P_{\mathrm{RF}}=$250 mW, and $P_{\mathrm{PS}}=$1 mW[@Gao2016hyb].
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying minimum rate constraint under the NLOS channel model, where $N=64$, $M=2$, $K=6$, and $P/\sigma^2=30$ dB.[]{data-label="Comp_r_NLOS"}](Comp_r_LOS_R2.eps){width="\figwidth"}
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying minimum rate constraint under the NLOS channel model, where $N=64$, $M=2$, $K=6$, and $P/\sigma^2=30$ dB.[]{data-label="Comp_r_NLOS"}](Comp_r_NLOS_R2.eps){width="\figwidth"}
Figs. \[Comp\_r\_LOS\] and \[Comp\_r\_NLOS\] show the ASR and EE comparisons between the proposed mmWave-NOMA approach, the mmWave-NOMA scheme in [@Dai2018MIMONOMA], mmWave-OMA and fully digital MIMO with varying minimum rate constraint under the LOS channel and the NLOS channel models, respectively. The minimum rate constraints for all the users are equal to $r$. Clearly, the performance of the proposed mmWave-NOMA system is distinctly better than that of the mmWave-OMA system, TDMA, and the solution of mmWave-NOMA in [@Dai2018MIMONOMA]. Particularly, when the minimum rate constraint $r$ ranges from 1 to 2 bps/Hz, the ASR of the proposed approach is nearly 10 bps/Hz larger than that of the scheme in [@Dai2018MIMONOMA]. The reason is as follows. When $r$ is small, according to the NOMA principle, more beam gains and power can be allocated to the user with the highest channel gain in each group [@Zhu2018NOMAPSO]. Thus, the beamforming scheme in [@Dai2018MIMONOMA] is effective, where the beam in analog domain is steering to the first user in each group. When $r$ becomes larger, the users with worse channel conditions can only be served by the sidelobe of the beam in [@Dai2018MIMONOMA]. In contrast, the proposed solution in this paper can allocate more beam gains in analog domain to the users with worse channel conditions in each group. Thus, the proposed approach outperforms the scheme in [@Dai2018MIMONOMA]. However, when $r$ is large, there may exist some channel realizations in which the minimum rate constraint cannot be satisfied. In such a case, the ASR is set to be zero. This operation is also adopted in the scheme of [@Dai2018MIMONOMA], which ensures the fairness of the comparison between the two methods. Therefore, the ASR tends to be zero for both of the two schemes, when $r$ is sufficiently large. Since the average ASR of the proposed scheme is larger than that of the scheme in [@Dai2018MIMONOMA], it can be concluded that our method can find a better solution and achieve a higher feasibility. Besides, we can also find that the ASR of the proposed approach is close to the ideal case, which indicates that the inter-group interference is small by using the proposed user grouping and HBF schemes and has little influence on the ASR. This result also verifies that the approximation of neglecting the inter-group interference when optimizing the intra-GPA is reasonable. We have also provided an enlarged view of the ASR curve in Fig. \[Comp\_r\_NLOS\], it can be seen that there is a small gap between the ideal curve and the designed curve, which is caused by the inter-group interference. The performance gap is no more than 0.5 bps/Hz, which is very small compared with the total ASR. In the two figures, we can also find that, although the ASR of the fully digital MIMO structure is higher than that of both the mmWave-NOMA and mmWave-OMA, the EE of the fully digital MIMO structure is low compared with the HBF structure. Particularly, the EE of the proposed mmWave-NOMA scheme can achieve nearly fourfold EE compared with the fully digital MIMO structure when the minimal rate constraint is no more than 1.5 bps/Hz.
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying total power to noise ratio under the NLOS channel model, where $N=64$, $M=2$, $K=6$, and $r_{k}=1$ bps/Hz.[]{data-label="Comp_P_NLOS"}](Comp_P_LOS_R2.eps){width="\figwidth"}
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying total power to noise ratio under the NLOS channel model, where $N=64$, $M=2$, $K=6$, and $r_{k}=1$ bps/Hz.[]{data-label="Comp_P_NLOS"}](Comp_P_NLOS_R2.eps){width="\figwidth"}
Figs. \[Comp\_P\_LOS\] and \[Comp\_P\_NLOS\] compare the ASRs/EEs between the proposed mmWave-NOMA approach, the mmWave-NOMA scheme in [@Dai2018MIMONOMA], mmWave-OMA and fully digital MIMO with varying total power to noise ratio under the LOS channel and the NLOS channel models, respectively. From the two figures, we can find again that the proposed mmWave-NOMA approach can achieve a higher ASR than that of mmWave-NOMA in [@Dai2018MIMONOMA], as well as the mmWave-OMA system. Particularly, when $P/\sigma^2$ is low, i.e., the mmWave-NOMA system is power limited, the superiority of the proposed algorithm is more conspicuous compared with the approach in [@Dai2018MIMONOMA]. When $P/\sigma^2$ is larger than 35 dB, the performance gap between the proposed solution and the solution in [@Dai2018MIMONOMA] stabilises around 5 bps/Hz in Fig. \[Comp\_P\_LOS\], while the performance gap stabilises around 7.5 bps/Hz in Fig. \[Comp\_P\_NLOS\]. From the two figures, we can find again that the EE of the proposed mmWave-NOMA scheme with a HBF structure is larger than that of the fully digital MIMO structure, as well as larger than the EE of mmWave-OMA. When $P/\sigma^2$ becomes large, the curves of the EE for different schemes all tend to be linear, and the increasing velocity, i.e., the slope of the EE curve, for mmWave-NOMA is larger than that for both fully digital MIMO and mmWave-OMA.
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying number of RF chains under the NLOS channel model, where $N=16$, $K=6$, $r_{k}=1$ bps/Hz and $P/\sigma^2=30$ dB.[]{data-label="Comp_RF_NLOS"}](Comp_RF_LOS_R2.eps){width="\figwidth"}
![ASR/EE comparison between the mmWave-NOMA and mmWave-OMA systems with varying number of RF chains under the NLOS channel model, where $N=16$, $K=6$, $r_{k}=1$ bps/Hz and $P/\sigma^2=30$ dB.[]{data-label="Comp_RF_NLOS"}](Comp_RF_NLOS_R2.eps){width="\figwidth"}
Figs. \[Comp\_RF\_LOS\] and \[Comp\_RF\_NLOS\] compare the ASRs/EEs between mmWave-NOMA and mmWave-OMA systems with varying number of RF chains under the LOS channel and the NLOS channel models, respectively. It can be observed that the proposed mmWave-NOMA approach outperforms the mmWave-OMA. In Fig. \[Comp\_RF\_LOS\], when the number of RF chains is no larger than 4, the ASR of the proposed approach is larger than that of mmWave-NOMA in [@Dai2018MIMONOMA]. When the number of RF chains is 5, the scheme in [@Dai2018MIMONOMA] behaves slightly better than the proposed scheme, and both of them are close to the performance of the fully digital structure. The reason is that the total number of users is 6 in Fig. \[Comp\_RF\_LOS\]. When the number of RF chains becomes larger, i.e., approximately equal to the number of users, the number of users in each group is usually one. Thus, the beamforming scheme in [@Dai2018MIMONOMA] is more effective, where the analog beams steer to the first user in each group. Moreover, the proposed solution always outperforms the mmWave-NOMA scheme in [@Dai2018MIMONOMA] in Fig. \[Comp\_RF\_NLOS\]. Comparing the two figures, we can find that the ASRs of the proposed approach are almost not influenced by the channel models. In contrast, the ASRs of the mmWave-NOMA scheme in [@Dai2018MIMONOMA] under the NLOS channel model is lower than that under the LOS channel model. The results indicate that the proposed approach is more robust against the channel model. We can also find that the EE of the proposed mmWave-NOMA scheme increases for the number of the RF chains, and it is significantly larger than the EE of the fully digital MIMO structure.
Conclusion
==========
In this paper, we investigated the application of NOMA in mmWave communications. Particularly, we considered downlink transmission with HBF structure. First, we proposed the K-means based user grouping algorithm according to the channel correlations of the multiple users. Whereafter, a joint hybrid beamforming and power allocation problem was formulated to maximize the ASR, subject to a minimum rate constraint for each user. To solve this non-convex problem with high-dimensional variables, we first obtained a sub-optimal solution of power allocation under arbitrary fixed HBF, where the intra-GPA and inter-GPA sub-problems are solved, respectively. Then, given an arbitrary fixed ABF, we utilized the approximately zero-forcing method to design the DBF matrix to minimize the inter-group interference. Finally, the ABF problem with the CM constraint was solved by using the proposed BC-PSO algorithm. Simulation results showed that the proposed mmWave-NOMA scheme outperforms mmWave-OMA in terms of ASR, and the proposed mmWave-NOMA scheme with the HBF structure is more energy efficient compared with the fully digital MIMO structure. The proposed solution for mmWave-NOMA, including user grouping, joint power allocation and HBF, can achieve a better performance in terms of ASR and EE compared with the benchmark scheme, with the expending of a higher computational complexity.
Proof of Lemma 1
================
It is obvious that $f(\{P_{m}\})$ is increasing for $P_{m} ~(1\leq m \leq M)$. Thus, the optimal solution always satisfies $\sum\limits_{m=1}^{M} P_{m} = P$. Then, Problem without the constraint $C_1$ can be solved by Lagrange Multiplier Method, where the KKT equation set is $$\left\{\begin{aligned}
&\frac{\partial f}{\partial P_{m}} = \lambda,~(1\leq m \leq M)\\
&\sum\limits_{m=1}^{M} P_{m} = P.
\end{aligned}
\right.$$
Solve the equation sets above and we can obtain the optimal solution of Problem as shown in .
Proof of Lemma 2
================
We prove Lemma 2 by using contradiction. Denote the optimal solution of Problem is $\{P_{m}^{\circ}\}$. Assume there exists an index $m_{1}~(m_{1} \in \mathcal{U})$ which satisfies $P_{m_1}^{\circ}>\frac{\eta_{m_1,1}-b_{m_1}}{k_{m_1}}$. Since $m_{1} \in \mathcal{U}$, we have $P_{m_1}^{\circ}>\frac{\eta_{m_1,1}-b_{m_1}}{k_{m_1}}>P_{m_1}^{\star}$. Then, there always exists another index $m_{2}~(m_{2}\neq m_{1})$ which satisfies $P_{m_2}^{\circ}<P_{m}^{\star}$, because $\sum\limits_{m=1}^{M} P_{m}^{\circ}\leq P=\sum\limits_{m=1}^{M}P_{m}^{\star}$. Consider the power allocation of $$\left\{
\begin{aligned}
&P_{m_1}^{\prime}=P_{m_1}^{\circ}-\epsilon \\
&P_{m_2}^{\prime}=P_{m_2}^{\circ}+\epsilon \\
&P_{m}^{\prime}=P_{m}^{\circ}, ~ m \neq m_1,m_2,
\end{aligned}
\right.$$ where $\epsilon$ is a nonnegative and small number.
The partial derivative of the objective function is $$\label{pdiff}
\frac{\partial f}{\partial P_{i}}=\frac{1}{\ln 2}\frac{k_{i}}{ (k_{i}P_{i}+b_{i}+1)}, ~1\leq i\leq M,$$ which is a monotone decreasing function of $P_{i}$. Since $P_{m_1}^{\circ}>P_{m_1}^{\star}$ and $P_{m_2}^{\circ}<P_{m_2}^{\star}$ , we have $$\begin{aligned}
&\frac{\partial f}{\partial P_{m_1}}|\{P_{m}=P_{m}^{\circ}\}<\frac{\partial f}{\partial P_{m_1}}|\{P_{m}=P_{m}^{\star}\} \\
&\frac{\partial f}{\partial P_{m_2}}|\{P_{m}=P_{m}^{\circ}\}>\frac{\partial f}{\partial P_{m_2}}|\{P_{m}=P_{m}^{\star}\}
\end{aligned}$$
Define $g(\epsilon)=f (\{P_{m}^{\prime}\})-f (\{P_{m}^{\circ}\})$. It is easy to verify that $g(0)=0$. The derivative of the function $g(\epsilon)$ is $$\begin{aligned}
\frac{d~g}{d~\epsilon}&= \frac{d~f (\{P_{m}^{\prime}\})}{d~\epsilon}= \frac{\partial f}{\partial P_{m_2}}|\{P_{m}=P_{m}^{\circ}\}- \frac{\partial f}{\partial P_{m_1}}|\{P_{m}=P_{m}^{\circ}\}\\
&> \frac{\partial f}{\partial P_{m_2}}|\{P_{m}=P_{m}^{\star}\}- \frac{\partial f}{\partial P_{m_1}}|\{P_{m}=P_{m}^{\star}\}=0,
\end{aligned}$$ which means that $g(\epsilon)$ is a monotone increasing function of $\epsilon$. We can select a sufficiently small $\epsilon$ which satisfies $P_{m_1}^{\prime}>P_{m_1}^{\star}$, $P_{m_2}^{\prime}<P_{m_2}^{\star}$ and $g(\epsilon)=f (\{P_{m}^{\prime}\})-f (\{P_{m}^{\circ}\})>0$. In other words, $\{P_{m}^{\prime}\}$ is better than $\{P_{m}^{\circ}\}$. It contradicts to the assumption that $\{P_{m}^{\circ}\}$ is the optimal solution of Problem . Thus, we can conclude that for $\forall m \in \mathcal{U}$, the optimal solution of Problem should always satisfy $P_{m}^{\circ}=\frac{\eta_{m,1}-b_{m}}{k_{m}}$.
[^1]: L. Zhu, Z. Xiao and X. Cao are with the School of Electronic and Information Engineering, Beihang University, Beijing 100191, China. {zhulipeng@buaa.edu.cn, xiaozy@buaa.edu.cn, xbcao@buaa.edu.cn}.
[^2]: J. Zhang is with the Advanced Research Institute of Multidisciplinary Science, Beijing Institute of Technology, Beijing, 100081, China. {buaazhangjun@vip.sina.com}.
[^3]: D. O. Wu is with the Department of Electrical and Computer Engineering, University of Florida, Gainesville, FL 32611, USA. {dpwu@ufl.edu}.
[^4]: X.-G. Xia is with the Department of Electrical and Computer Engineering, University of Delaware, Newark, DE 19716, USA. {xianggen@udel.edu}.
[^5]: In our previous work [@Zhu2018NOMAPSO], a mmWave-NOMA system with the pure analog beamforming structure was considered, where several key problems for HBF structure were not included, e.g., the user grouping, digital beamforming, inter-group interference suppression and power allocation among different groups, which bring new challenges for the multi-group mmWave-NOMA system.
[^6]: In the simulation, we compare the ASR/EE performance of mmWave-NOMA with the scheme proposed in [@Dai2018MIMONOMA], which is regarded as the benchmark.
[^7]: It is worthy of noting that the proposed approach in this paper can also be directly used for the partially connected HBF structure [@Dai2018MIMONOMA].
[^8]: Since we concentrate on the user grouping and resource allocation for mmWave-NOMA, the channel estimation problem is beyond the scope of this paper. We assume that the channel state information (CSI) between the BS and the users is known by the BS. A number of approaches on mmWave channel estimation have been proposed and could be referred, such as, [@xiao2016codebook; @Kokshoorn2017ChannelEstim; @xiao2018codebook; @Hu2018channelEstim].
[^9]: We can always define the user with the $n$th highest effective channel gain in the $m$th group as the $n$th user in this group. Thus, this simplified subscript has no influence on the solution in this paper.
[^10]: Since the DBF design implements an approximate zero-forcing method, i.e., only to the first user in each group, the inclusion of inter-group interference in the previous section is relevant.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The topological susceptibility, $\chi^4$, following the work of Witten and Veneziano, plays a key role in identifying the relative magnitude of the $\eta^{\prime}$ mass, the so-called $U(1)_{A}$ problem. A nonzero $\chi^4$ is caused by the Veneziano ghost, the occurrence of an unphysical massless pole in the correlation function of the topological current $K_{\mu}$. In a recent paper, [@Kharzeev:2015xsa], an explicit relationship between this Veneziano ghost and color confinement was proposed, by connecting the dynamics of the Veneziano ghost, and thus the topological susceptibility, with Gribov copies. However, the analysis of [@Kharzeev:2015xsa] is incompatible with BRST symmetry [@Dudal:2015khv]. In this paper, we investigate the topological susceptibility, $\chi^4$, in $SU(3)$ and $SU(2)$ Euclidean Yang-Mills theory using an appropriate Padé approximation tool and a non-perturbative gluon propagator, within a BRST invariant framework and by taking into account Gribov copies in a general linear covariant gauge.'
author:
- 'D. Dudal$^{\dag,\ddag}$, C. P. Felix$^{\dag}$, M. S. Guimaraes$^{\S}$, S. P. Sorella$^{\S}$'
title: '[**Accessing the topological susceptibility via the Gribov horizon**]{}'
---
Introductory remarks
====================
There are two important properties of QCD that are decisive in determining its particle spectrum: [*confinement*]{}, the fact that quarks and gluons are not observable as free particles, and [*chiral symmetry breaking*]{}, answering the question why hadrons composed of $u$, $d$ or $s$ quarks or antiquarks are so massive while these (light) quark flavours themselves are almost massless.
Already 40 years ago, Gribov [@Gribov:1977wm] showed that the Faddeev-Popov construction is not valid at the non-perturbative level, i.e. in the low energy limit where the coupling constant, $g$, is large. In this regime, we have Gribov copies, caused by multiple intersections of gauge orbits with the hypersurface corresponding to a given gauge condition $f(A)=0$. This means that we have to deal in some way with equivalent field configurations obeying the same gauge fixing condition. Thus, in a non-perturbative non-Abelian gauge theory setting, the Faddeev-Popov procedure is incomplete as it stands. For reviews, see [@Vandersickel:2012tz; @Sobreiro:2005ec]. The Gribov problem for covariant gauges was also put on a mathematical footing in [@Singer:1978dk], at the same time showing that it cannot be avoided. Loosely speaking, Gribov copies imply that:
- we are overcounting equivalent gauge configurations, since we have more than one gauge fixed configuration for each gauge orbit. This implies the Faddeev-Popov $\delta$-function implementing the gauge condition will have multiple zeros of its argument, complicating its interpretation as a unity being inserted into the a priori gauge invariant partition function.
- the Faddeev-Popov measure is ill-defined, since there are zero-modes of the Faddeev-Popov operator when considering the infinitesimal copies. This implies a vanishing Faddeev-Popov (Jacobian) determinant. Indeed, considering 2 infinitesimally connected gauge configurations,$$\Tilde{A}^{a}_{\mu}=A^{a}_{\mu}+D^{ab}_{\mu}\theta^{b},
\label{gauge_equiv}$$ $\Tilde{A}^{a}_{\mu}$ will obey the same (linear) gauge condition, ${\partial}_\mu A_\mu^a=f^a$, as $A_\mu^a$ if $$-\partial_{\mu}D^{ab}_{\mu}\theta^{b}=0,
\label{gange_A}$$ i.e. whenever the Faddeev-Popov operator exhibits normalizable zero modes, see [@Sobreiro:2005ec] for some explicit example.
To solve this problem, Gribov proposed to restrict the domain of integration in the path integral to a certain region $\Omega$ in field space, called the Gribov region, which is free from infinitesimal Gribov copies: $$\Omega=\{A_{\mu}^a;\ \partial_\mu A_{\mu}^a=0,\qquad \mathcal{M}^{ab}(A)=-\partial_\mu D_\mu^{ab}(A)>0 \}.
\label{GR}$$ The Gribov region is the set of all field configurations which obey the Landau gauge, $\partial_\mu A_{\mu}^a=0$, and where the Hermitian Faddeev-Popov operator, $\mathcal{M}^{ab}(A)$, is positive. Inside the Gribov region, there are no infinitesimal copies, since $\mathcal{M}^{ab}(A)>0$. The region $\Omega$ is also known to be convex, bounded and intersected at least once by every gauge orbit [@Dell'Antonio:1989jn]. Its boundary, $\partial\Omega$, where the first vanishing eigenvalue of $\mathcal{M}^{ab}(A)$ (i.e. the first zero-mode of Faddeev-Popov operator) appears, is called the first Gribov horizon. Unfortunately, even if the restriction of the domain of integration in the functional integral to the region $\Omega$ allows to get rid of a large number of copies, there still remain additional copies inside $\Omega$ [@vanBaal:1991zw; @semenov]. A subregion of $\Omega$, known as the fundamental modular region $\Lambda$, has been proven to be fully free from Gribov copies. Though, unlike the Gribov region $\Omega$, a concrete setup to implement the restriction of the domain of integration in the functional integral to $\Lambda$ within a local and renormalizable framework is still far beyond our present capabilities. Therefore, we shall focus on the Gribov region $\Omega$, which enables us to already capture quite useful non-perturbative aspects, see [@applications] for recent applications. Let us proceed by giving a sketchy overview of the construction of the local action which emerges from the restriction to $\Omega$, referring to [@Gribov:1977wm; @Zwanziger:1989mf; @Zwanziger:1992qr; @Maggiore:1993wq; @Dudal:2010fq; @Capri:2015mna; @Dudal:2007cw; @Dudal:2008sp; @Dudal:2011gd] for the specific details.
In the Landau gauge, the effective implementation of the restriction to $\Omega$ is accomplished by means of the Gribov-Zwanziger action [@Gribov:1977wm; @Zwanziger:1989mf; @Zwanziger:1992qr], $$\begin{aligned}
S&=&S_{YM}+S_{GF}+S_{GZ}.
\label{S_gamma}\end{aligned}$$ $S_{YM}$ is the Yang-Mills action, $$S_{YM} = \frac{1}{4} \int d^4x F^a_{\mu\nu} F^a_{\mu\nu} , \label{ym}$$ $S_{GF}$ is the Landau gauge Faddeev-Popov action, $$S_{GF} = \int d^{4}x \left(
ib^{a}\,\partial_{\mu}A^{a}_{\mu}
+\bar{c}^{a}\partial_{\mu}D^{ab}_{\mu}(A)c^{b} \right) , \label{sfp}$$ and finally, $$\begin{aligned}
S_{GZ}&=&\int d^{4}x\left[\bar{\varphi}^{ac}_{\mu}\partial_{\nu}D^{ab}_{\nu}\varphi^{bc}_{\mu}-\bar{\omega}^{ac}_{\mu}\partial_{\nu}(D^{ab}_{\nu}\omega^{bc}_{\mu})-g(\partial_{\nu}\bar{\omega}^{an}_{\mu})f^{abc}D^{bm}_{\nu}c^{m}\varphi^{cn}_{\mu}\right]
\nonumber\\&&-\gamma^{2}g\int d^4 x\left[f^{abc}A^{a}_{\mu}\varphi^{bc}_{\mu}+f^{abc}A_{\mu}^{a}\bar{\varphi}_{\mu}^{bc}+\frac{d}{g}(N_{c}^{2}-1)\gamma^{2}\vphantom{\frac{1}{2}}\right],
\label{S_gamma2}\end{aligned}$$ is the part of the action which enables us to restrict the functional integral to $\Omega$. In expression , $(\bar{\varphi}^{ac}_{\mu},~\varphi^{ac}_{\mu})$ is a pair of complex-conjugate bosonic fields, $(\bar{\omega}^{ac}_{\mu},~\omega^{ac}_{\mu})$ a pair of anti-commuting complex-conjugate fields, while $\gamma$ is the Gribov parameter, dynamically fixed by means of its gap equation [@Zwanziger:1989mf; @Zwanziger:1992qr], $$\label{gap}
\braket{f^{abc}A_\mu^a({\varphi}_{\mu}^{bc}+\bar{\varphi}_{\mu}^{bc})}=2d(N^2-1)\frac{\gamma^2}{g^2}.$$ Remarkably, the action $S$ is multiplicative renormalizable to all orders, as discussed in [@Zwanziger:1992qr; @Maggiore:1993wq; @Dudal:2010fq; @Capri:2015mna].
After the formulation of the Gribov-Zwanziger action, it was realized that it is plagued by non-perturbative dynamical instabilities, caused by the formation of the dimension two condensates, $\langle A_{\mu}^{a}A_{\mu}^{a}\rangle$ and $\langle \bar{\varphi}^{ab}_{\mu}\varphi^{ab}_{\mu}-\bar{\omega}^{ab}_{\mu}\omega^{ab}_{\mu} \rangle$, which are energetically favoured [@Dudal:2007cw; @Dudal:2008sp; @Dudal:2011gd]. These condensates can be taken into account from the beginning, leading to the so-called Refined-Gribov-Zwanziger action [@Dudal:2007cw; @Dudal:2008sp; @Dudal:2011gd] $$S_{R} = S_{YM}+S_{GF}+S_{RGZ} \;, \label{sref}$$ where $$S_{RGZ}=S_{GZ}+\frac{m^2}{2}\int d^4 x A_{\mu}^{a}A_{\mu}^{a}+M^{2}\int d^4 x(\bar{\varphi}^{ab}_{\mu}\varphi^{ab}_{\mu}-\bar{\omega}^{ab}_{\mu}\omega^{ab}_{\mu}).
\label{RGZ_action}$$ As much as the Gribov mass $\gamma^2$, the new parameters $(m^2, M^2)$ are dynamically determined by their own gap equations [@Dudal:2007cw; @Dudal:2008sp; @Dudal:2011gd]. As the action $S$, the refined action $S_{R}$ is multiplicative renormalizable to all orders [@Dudal:2007cw; @Dudal:2008sp; @Dudal:2011gd]. Moreover, the introduction of the aforementioned condensates allows for a nice agreement with the lattice data, see e.g. [@Dudal:2010tf; @Cucchieri:2011ig; @Dudal:2012zx].
Recently, the Gribov-Zwanziger formalism was generalized to the linear covariant gauges [@Capri:2015ixa]. Simultaneously, an exact BRST invariance of the Landau gauge actions , was found, with immediate extension to the linear covariant gauges [@Capri:2015ixa]. Following [@Capri:2015ixa], the action is replaced by $$S_{\rm lc} = S_{YM} + S_{GF} + S_{RGZ} + S_{\tau} , \label{stact}$$ where $S_{GF}$ now denotes the Faddeev-Popov gauge-fixing in linear covariant gauges, i.e. $$S_{GF} = \int d^{4}x \left( \frac{\alpha}{2}\,b^{a}b^{a}
+ib^{a}\,\partial_{\mu}A^{a}_{\mu}
+\bar{c}^{a}\partial_{\mu}D^{ab}_{\mu}(A)c^{b} \right) , \label{sfpb}$$ with $\alpha$ denoting the gauge parameter. The value $\alpha=0$ corresponds to the Landau gauge. For the Refined-Gribov-Zwanziger action in linear covariant gauges we have [@Capri:2015ixa] $$\begin{aligned}
\label{lact}
S_{RGZ} &=& \int d^{4}x \, \left(
- \bar\varphi^{ac}_{\nu}{\cal M}^{ab}(A^h)\varphi^{bc}_{\nu}
+\bar\omega^{ac}_{\nu}{\cal M}^{ab}(A^h)\omega^{bc}_{\nu}
+\gamma^{2}g\,f^{abc}(A^h)^{a}_{\mu}(\varphi^{bc}_{\mu}+\bar\varphi^{bc}_{\mu})+\frac{m^2}{2}\int d^4 x (A^h)_{\mu}^{a}(A^h)_{\mu}^{a}\right.\nonumber\\&&+\left.M^{2}\int d^4 x(\bar{\varphi}^{ab}_{\mu}\varphi^{ab}_{\mu}-\bar{\omega}^{ab}_{\mu}\omega^{ab}_{\mu}) -d(N^2-1)\gamma^4 \right),\end{aligned}$$ and ${\cal M}^{ab}(A^h)$ is the Hermitian, gauge invariant operator $${\cal M}^{ab}(A^h) = - \delta^{ab} \partial^2 + g f^{abc} (A^h)^c_\mu \partial_\mu . \label{fpop}$$ The configuration $A_\mu^h$ is a non-local power series in the gauge field, obtained by minimizing the functional $f_A[u]$ along the gauge orbit of $A_{\mu }$ [@Dell'Antonio:1989jn; @vanBaal:1991zw; @Lavelle:1995ty], with $$\begin{aligned}
f_A[u] &\equiv &\min_{\{u\}}\mathrm{Tr}\int d^{4}x\,A_{\mu
}^{u}A_{\mu }^{u},
\nonumber \\
A_{\mu }^{u} &=&u^{\dagger }A_{\mu }u+\frac{i}{g}u^{\dagger }\partial _{\mu
}u. \label{Aminn0}\end{aligned}$$ One finds that a local minimum is given by $$\begin{aligned}
A_{\mu }^{h} &=&\left( \delta _{\mu \nu }-\frac{\partial _{\mu }\partial
_{\nu }}{\partial ^{2}}\right) \phi _{\nu }\;, \qquad \partial_\mu A^h_\mu= 0 \;, \nonumber \\
\phi _{\nu } &=&A_{\nu }-ig\left[ \frac{1}{\partial ^{2}}\partial A,A_{\nu
}\right] +\frac{ig}{2}\left[ \frac{1}{\partial ^{2}}\partial A,\partial
_{\nu }\frac{1}{\partial ^{2}}\partial A\right] +O(A^{3}). \label{min0}\end{aligned}$$ It can be checked that the quantity $A_\mu^h$ is gauge invariant order by order [@Capri:2015ixa]. To get control on the renormalization properties of this seemingly highly non-local quantum field theory, an equivalent local formulation can be obtained. Following [@Capri:2015ixa; @Fiorentini:2016rwx], we set $$A^{h}_{\mu}=(A^{h})^{a}_{\mu}T^{a}=h^{\dagger}A^{a}_{\mu}T^{a}h+\frac{i}{g}\,h^{\dagger}\partial_{\mu}h, \label{st}$$ while $$h=e^{ig\,\xi^{a}T^{a}},
\label{hxi}$$ with the role of the fields $\xi^a$ akin to that of the Stueckelberg formulation. The local gauge invariance of $A^{h}_{\mu}$ under a gauge transformation $u\in SU(N)$ is now immediately clear from $$h\to u^\dagger h\,,\qquad\ h^\dagger\to h^\dagger u\,,\qquad A_\mu \to u^\dagger A_\mu u + \frac{i}{g}u^\dagger {\partial}_\mu u.$$ The term $$S_{\tau} = \int d^4x\; \tau^{a}\,\partial_{\mu}(A^h)^{a}_{\mu} \label{stau}$$ implements, through the Lagrange multiplier $\tau$, the transversality of the composite operator $(A^h)_\mu^a$, namely $\partial_{\mu}(A^h)^{a}_{\mu}=0$. Solving the latter constraint gives back the non-local expression for the field $A^h_\mu$ expressed in . This constraint also plays a crucial role to maintain the ultraviolet renormalizability of the theory [@Fiorentini:2016rwx], in sharp contrast with the standard Stueckelberg formulation. It can also be shown that at the practical level, and give rise to identical dynamics when working in the Landau gauge, due to the gauge condition ${\partial}_\mu A_\mu^a=0$ [@Capri:2015ixa].
The action $S_{\rm lc}$ enjoys an exact nilpotent BRST invariance, $s S_{lc} = 0$, expressed by [@Capri:2015ixa] $$\begin{aligned}
s A^{a}_{\mu}&=&-D^{ab}_{\mu}c^{b}\,,\;\;\;\; s c^{a}=\frac{g}{2}f^{abc}c^{b}c^{c}\,,\nonumber\\
s \bar{c}^{a}&=&ib^{a}\,,\;\;\;\;
s b^{a}= 0\,, \nonumber\\
s h^{ij} &=& -ig c^a (T^a)^{ik} h^{kj} \;, \nonumber\\
s \varphi^{ab}_{\mu}&=& 0 \,,\;\;\;\; s \omega^{ab}_{\mu}=0\,,\nonumber\\
s\bar\omega^{ab}_{\mu}&=&0 \,,\;\;\;\; s\bar\varphi^{ab}_{\mu}=0\,,\nonumber\\
s\tau^{a}&=&0.
\label{brstgamma}\end{aligned}$$ A detailed analysis of the consequences of this BRST invariance at the quantum level can be found in [@Fiorentini:2016rwx], where e.g. the Nielsen identities were discussed. Notice that this BRST transformation is different from the one adopted before in e.g. [@Zwanziger:1992qr]. In particular, the auxiliary fields $(\varphi^{ab}_{\mu},\bar\varphi^{ab}_{\mu}, \omega^{ab}_{\mu}, \bar\omega^{ab}_{\mu})$ are now BRST singlets, i.e. they are BRST invariant fields. Though, as shown in details in [@Capri:2017bfd] they possess their own Ward identities which ensure the all orders renormalizability of the action $S_{\rm lc}$, eq.. The gap equation gets replaced by its gauge invariant counterpart $\braket{f^{abc}(A^h)_\mu^a({\varphi}_{\mu}^{bc}+\bar{\varphi}_{\mu}^{bc})}=2d(N^2-1)\frac{\gamma^2}{g^2}$. For the current work, we are mostly interested in the general form of the gluon propagator [@Fiorentini:2016rwx], which can be proven to be $$D_{\mu\nu}(p)=D(p)P_{\mu\nu}(p)+L(p)\frac{p_{\mu}p_{\nu}}{p^{2}},
\label{gluon_prop_faynman}$$ with the transverse form factor, $$\begin{aligned}
\label{rgz}
D(p)&=&\frac{p^2+M^2}{p^4+(M^2+m^2)p^2+M^2m^2+\lambda^4},\end{aligned}$$ containing all non-trivial information. At tree level, this factor stems from the quadratic part of the action , where we set $\lambda^4=2g^2N\gamma^4$. Next to $$L(p)=\frac{\alpha}{p^2},
\label{gluon_prop_faynmanbis}$$ with $$P_{\mu\nu}(p)=\delta_{\mu\nu}-\frac{p_{\mu}p_{\nu}}{p^{2}}\,,\qquad L_{\mu\nu}(p)=\frac{p_{\mu}p_{\nu}}{p^{2}}
\label{tranverse_prop}$$ the transversal and longitudinal projectors. Analogously as in perturbation theory, the longitudinal gluon propagator is exactly known, being equal to its tree level expression, as it follows from the BRST symmetry . The same behavior for the longitudinal component of the propagator is found in other non-perturbative approaches to the linear covariant gauges [@Aguilar:2015nqa].
It is worth mentioning here that expression fits very well the lattice data, see [@Dudal:2010tf; @Cucchieri:2011ig; @Dudal:2012zx]. Moreover, when the various mass parameters are estimated by means of the direct lattice comparison [@Dudal:2010tf; @Cucchieri:2011ig; @Dudal:2012zx], it turns out that expression exhibits 2 complex-conjugate poles, indicating that the corresponding degrees of freedom cannot be associated to excitations of the physical spectrum, similar to the rationale of e.g. [@Kharzeev:2015xsa] or [@Hawes:1993ef]. Said otherwise, expression can be seen as a manifestation of gluon confinement. Propagators with complex poles such as have also been used to fit finite temperature lattice data [@Aouane:2011fv], based on which the confinement-deconfinement transition was discussed using the Polyakov loop criterion, providing a reasonable estimate for the critical temperature, see e.g. [@Fukushima:2012qa; @Canfora:2015yia]. How to connect the (infrared) behaviour of the gluon, quark and ghost propagators and their mutual interaction vertices to the linear confining potential between static charges etc. is an open question, for whatever analytical scheme is considered to study the elementary $n$-point functions of QCD.
The topological susceptibility
===============================
Returning to QCD with 3 light flavours $(u, d, s)$, we recall it enjoys an (almost) $U_L(3)\times U_R(3)$ (left $\times$ right) symmetry. There is a dynamical chiral (axial $= L-R$) symmetry breaking, reducing the invariance to the vector part ($L+R$), $U_V(3)$. So one experimentally expects a nonet of (almost) massless Goldstone modes, viz the 3 pions, 4 kaons, $\eta$ and $\eta'$. Though, it turns out that the $\eta'$ particle is way too massive to be called the ninth “almost Goldstone” boson as $m_{\eta'}\approx 958~\text{MeV}$.
One solution to this apparent $U_A(1)$ problem was offered by ’t Hooft by means of instanton calculus, see [@tHooft:1976rip]. The validity of the instanton calculus in the large $N$ limit is a delicate issue (see [@marino] for a recent account), but another way to understand the anomalous $\eta'$ mass was, independently, worked out by Veneziano and Witten in [@Witten:1979vv; @Veneziano:1979ec] [^1]. Summarizing, the celebrated Veneziano-Witten formula reads $$\label{wz}
m_{\eta'}^2=\frac{4N_f}{f_\pi^2}\chi^4_{\theta=0,N_f=0}=\mathcal{O}(1/N),$$ where $\theta$ is the vacuum angle and $f_\pi$ the pion decay constant. Although simple at first sight, this is a very intricate formula, since the l.h.s. refers to QCD (with $N_f$ flavours), while the r.h.s. to the pure gauge theory. The relation thus explains the relatively large $\eta'$ mass, given that $\chi^4_{\theta=0,N_f=0}\equiv \chi^4$ is sufficiently large. Filling in the numbers requires $\chi^4\sim (200~\text{MeV})^4$, not far from the lattice $SU(3)$ estimates as reported in [@Alles:1996nm].
Returning to the work [@Kharzeev:2015xsa], it was suggested that a resolution of the Gribov copies problem, being intimately linked with the topology of the gauge group, might also be of direct importance to capture the non-trivial topological structure of the QCD vacuum. In Euclidean space-time, we have the classical instanton solutions, describing in Minkowski space-time the tunneling between the degenerate vacuum states with different Chern-Simons charge [@Kharzeev:2015ifa], $$X=\int d^3 xK_0,
\label{CS_number}$$ with $K_0$ the temporal component of topological Chern-Simons current, $$K_\mu =\frac{g^2}{16\pi^2}\epsilon_{\mu\nu\rho\sigma}A_{\nu,a}\left(\partial^\rho A^{\sigma,a}+\frac{g}{3}f^{abc}A^\rho _b A^\sigma _c\right).
\label{CS_current}$$ This current is related to the topological charge density, $$Q(x)=\partial_\mu K_\mu=\frac{g^2}{32\pi^2}F_{\mu\nu}\Tilde{F}_{\mu\nu}.
\label{top_charge_density}$$ Witten and Veneziano suggested that the vacuum topology fluctuations can be captured by the occurrence of an unphysical mass pole [@Witten:1979vv; @Veneziano:1979ec], the Veneziano ghost, in the topological current correlator, $$p_{\mu}p_{\nu}\Braket{K_{\mu}K_{\nu}}_{p=0}\neq 0.
\label{KK}$$ Thus, the Veneziano solution was to assume that $$K_{\mu\nu}(p)=i\int d^{4}x\ e^{ipx}\left<K_{\mu}(x)K_{\nu}(0)\right>\stackrel{p^{2}\thicksim 0}\thicksim -\frac{\chi^{4}}{p^{2}}g_{\mu\nu},
\label{vene_prop.}$$ where $\chi^4\geq0$ is the topological susceptibility of pure Yang-Mills theory. The negative sign in means that we are dealing with an unphysical ghost particle, so it cannot be directly measured in a physical process, though the couplings of the ghosts can influence physical amplitudes [@Kharzeev:2015xsa].
Kharzeev and Levin [@Kharzeev:2015xsa; @Kharzeev:2015ifa] interpreted the current correlator as resulting from an effective interaction between the gluon (in Feynman gauge $\alpha=1$) and the Veneziano ghost [@Kharzeev:2015xsa; @Kharzeev:2015ifa; @Dudal:2015khv]. An effective ghost-gluon vertex $\Gamma_{\mu}(q,p)$ was postulated, and then it was found that a dynamically corrected gluon propagator (the “glost”), $$\mathcal{G}_{\mu\nu}(p^{2})=\frac{p^{2}}{p^{4}+\chi^{4}}\delta_{\mu\nu},
\label{glost}$$ solves the Dyson-Schwinger equation, when using only this coupling [@Kharzeev:2015xsa; @Dudal:2015khv] in the deep infrared. Immediately, we notice that there is an inconsistency between and , indicating that the propagator is incompatible with BRST symmetry as well as with all independently obtained functional forms for the non-perturbative Feynman gauge gluon propagator, see also [@Dudal:2015khv] for more concerns concerning the construction proposed in [@Kharzeev:2015xsa].
Setup of a rational (Padé) approximation via the spectral Källén-Lehmann representation of the topological current correlator
=============================================================================================================================
From now on, we wish to find out if, by using Gribov type propagators, we can obtain a reasonable “semi-non-perturbative” estimate for the topological susceptibility $\chi^4$, without the need to introduce new effective vertices. We notice that transversal component $D(p)$ of the propagator in can be written as the linear combination of 2 standard massive propagators with complex-conjugate masses, which allows for standard Feynman diagram-computational manipulations.
We will follow the Euclidean conventions of [@Meggiolaro:1998bh]. More precisely, we have $$\label{v2b}
\chi^4 = -\lim_{p^2\to0} p_\mu p_\nu \braket{K_\mu K_\nu}\geq0.$$ That $\chi^4\geq 0$ follows from translational invariance applied to $$\chi^4 = \int d^4x \braket{F\tilde F(x),F\tilde F(0)}=V^{-1}\int d^4x \int d^4y \braket{F\tilde F(x),F\tilde F(y)}\geq 0.$$
As pointed out originally in [@Seiler:1987ig; @Seiler:2001je], see also [@Vicari:2008jw], it holds that $$\label{rf1}\braket{Q(x)Q(0)}<0\,,\qquad\text{for~} |x|>0,$$ due to the (Osterwälder-Schrader) reflection positivity and to the $t$-odd character of $Q$. Combined with $\int d^4x \braket{Q(x)Q(0)}>0$, this entails that $\braket{Q(x)Q(0)}$ must contain a positive contact term to compensate the negative rest of the integral. For example, setting $$\mathcal{Q}(x)=\braket{Q(x)Q(0)}\,,\qquad\text{for~} |x|>0,$$ then we need $$\braket{Q(x)Q(0)}=\mathcal{Q}(x) +C\delta(x)\,,\qquad\text{for~} |x|\geq0,$$ with $C>0$.
The contact term thus plays a pretty important role in the *definition* of the topological susceptibility, as recognized in [@Seiler:1987ig]. Contact terms are evidently also important to get a finite (cf. additive renormalization) value for $\braket{Q(x)Q(0)}$. As in principle these contact terms can be chosen, one might have the impression that this would reflect in a randomness in the definition of $\chi^4$. This is however not the case, see later.
Closely related to the above comment is that the $t$-odd character of $Q$ also means that the Källén-Lehmann spectral density of $\braket{QQ}$ is negative. The original remarks can be found in [@Seiler:1987ig; @Seiler:2001je]. Following the derivation of the usual Minkowski 2-point function of $Q$, we will find a positive spectral integral. Though, upon passing to Euclidean space-time via a Wick rotation, the $t$-odd nature of $Q$ will introduce an extra factor $i$ [^2], meaning that $\braket{QQ}$ in Euclidean spacetime will pick up an extra overall minus. Thus, actually, we have $$\label{KLQ}\hat{\mathcal{Q}}(p^2)\equiv\braket{Q(p)Q(-p)}=-\int_0^\infty d\mu\frac{\rho(\mu)}{\mu+p^2}\,,\qquad \rho(\mu)\geq 0.$$
It is interesting to remark that is consistent with when we think in terms of the usual temporal Schwinger function that is used to check for positivity violations [@Alkofer:2003jj]. Indeed, usually we introduce, for $t>0$, $$\label{temps}C(t)=\frac{1}{2\pi} \int_{-\infty}^{\infty} dp e^{-ipt} \hat{\mathcal{Q}}(p^2).$$ Substituting into , we can rewrite this as $$\label{temps2}C(t)=-\frac{1}{2} \int_{0}^{\infty}d\mu\frac{\rho(\mu)}{\sqrt{\mu}}e^{-t\sqrt{\mu}}= - \int_{0}^{\infty}d\nu\rho(\nu^2)e^{-t\nu}<0.$$ The definition can however be rewritten as, with $|x|=|(t,\vec{x})|>0$, $$\label{temps3}C(t)=\frac{1}{2\pi}\frac{1}{V_3} \int d^3x\int d^4p e^{-ipx} \hat{\mathcal{Q}}(p^2).$$ We recognize the Fourier transform of $\mathcal{Q}(x)$, so we recover $$\mathcal{Q}(t)\propto C(t)<0,$$ i.e. , the negative spectral density in the Källén-Lehmann representation indeed ensures .
Let us now carry out a power counting analysis. As $\dim\hat{\mathcal{Q}}(p^2)=4$, this correlation function needs 3 subtractions to make it well defined (finite), as $\rho(\tau)\sim \tau^2$ for $\tau\to\infty$. Without loss of generality, we can subtract at zero momentum, finding as proper version of $$\label{KLQsub}\hat{\mathcal{Q}}(p^2)=a_0+a_1 p^2+ a_2p^4-p^6\int_0^\infty d\tau\frac{\rho(\tau)}{(\tau+p^2)\tau^3}$$
Doing so, $\chi^4$ would be determined by the subtraction constant $a_0$. This is a somewhat undesirable feature, as we wish to derive an estimate for $\chi^4$ from the correlator. Though, usually one employs the a priori knowledge of the correlator, and possibly of its slope, at zero momentum via low energy theorems, to exchange constants like $a_0$, $a_1$ in terms of these a priori known numbers (like $\chi^4$) to completely fix the correlation at any momentum. This is rather the inverse order of the current analysis, where $\chi^4$ is unknown. For a few examples we refer to [@Narison:2002pw].
Next, including quarks in the analysis, it was shown in [@Crewther:1978kq] that the susceptibility, as defined via , is the one entering the anomalous chiral Ward identities. As discussed in [@Meggiolaro:1998bh], this definition also concurs with the one given in [@Witten:1979vv] via the double derivative of the $\theta$ term in the action.
Let us show, employing expression , also removes any ambiguity imposed by the subtraction procedure. We may in general set $$\begin{aligned}
\label{k1}
\braket{K_\mu(p) K_\nu(-p)}&=&\left(\delta_{\mu\nu}-\frac{p_\mu p_\nu}{p^2}\right)\mathcal{K}_\perp(p^2)+\frac{p_\mu p_\nu}{p^2}\mathcal{K}_\|(p^2)\nonumber\\&\equiv&\left(\delta_{\mu\nu}-\frac{p_\mu p_\nu}{p^2}\right)\int_0^\infty d\tau\frac{\rho_\perp(\tau)}{\tau+p^2}+\frac{p_\mu p_\nu}{p^2}\int_0^\infty d\tau \frac{\rho_\|(\tau)}{\tau+p^2},\end{aligned}$$ based on Euclidean invariance. Then, we already find that[^3] $$\label{k2}
\hat{\mathcal{Q}}(p^2)=-p^2\mathcal{K}_\|(p^2)=-p^2\int_0^\infty d\tau \frac{\rho_\|(\tau)}{\tau+p^2}$$ and thus $$\label{k3}
-\chi^4=\lim_{p^2\to0}p^2\mathcal{K}_\|(p^2)=\lim_{p^2\to0}p^2\int_0^\infty d\tau \frac{\rho_\|(\tau)}{\tau+p^2}.$$ As the l.h.s. of is gauge invariant, so should the r.h.s. be, meaning that the longitudinal form factor $\mathcal{K}_\|(p^2)$ and its associated spectral function $\rho_\|(\mu)$ ought to be gauge invariant. Likewise, the transversal piece may contain gauge variant contributions depending on the gauge parameter $\alpha$. This is the reason why we opt to work with a general linear covariant gauge, as we can then explicitly check how the gauge (in)variance manifests itself in the full $\braket{KK}$-correlation function.
Let us now be a bit more careful, and include the subtraction terms. We focus on the relevant longitudinal sector. From dimensional analysis, it is clear that this time we only need 2 subtractions ($\rho_\|(\tau)\sim \tau$ for $\tau\to\infty$), so a finite result[^4] is guaranteed from $$\label{k7}
\mathcal{K}_\|(p^2)=b_0+b_1p^2+p^4\int_0^\infty d\tau \frac{\rho_\|(\tau)}{(\tau+p^2)\tau^2}$$ and thus $$\label{k4}
-\chi^4=\lim_{p^2\to0}p^2\left(b_0+b_1p^2+p^4\int_0^\infty d\tau \frac{\rho_\|(\tau)}{(\tau+p^2)\tau^2}\right),$$ with $b_{0,1}$ subtraction constants. Obviously, we can rewrite as $$\label{k5}
-\chi^4=\lim_{p^2\to0}p^6\int_0^\infty d\tau \frac{\rho_\|(\tau)}{(\tau+p^2)\tau^2}$$ and all reference to subtraction constants is indeed gone.
In a loop expansion approach using the RGZ action and associated RGZ propagator , will still automatically vanish. It is important to stress here how difficult it is to get non-zero value for the topological susceptibility from a continuum viewpoint. Obviously, in the absence of a dynamical singularity at $p^2\to0$ in the correlation function of two topological currents, $\chi^4$ will always automatically vanish. This is precisely why one needs a dynamically generated massless (ghost) bound state in the $K_\mu$ channel. Usually, to get an estimate for a bound state mass, one makes a series of approximations and/or assumptions and obtain an estimate for e.g. the mass of a pion. Here, an approximation of the mass of the Veneziano ghost would be insufficient, as we need it to be *exactly* zero. Usually, a massless bound state can be probed when a symmetry (and symmetry-consistent approximations) protects the mass to be zero, needless to say we are thinking about the Goldstone mechanism here. Clearly, the latter does not apply to the Veneziano ghost particle, which is not the Goldstone degree of freedom created by a current corresponding to a spontaneously broken global symmetry. On the contrary, it is a topologically conserved current. The inherent difficulty just sketched to get a continuum handle on the Veneziano ghost is what, to our understanding, explains the lack of references trying to tackle the problem directly. Notice that this does not a priori exclude a truly non-perturbative lattice study of the relevant correlator, apart from the potential difficulties in precisely defining the topological current on the lattice [@Luscher:1981zq; @Seiberg:1984id; @Karsch:1993fv; @Luscher:2004fu].
Therefore, in this section we work out an approximation to the quantity appearing in , more precisely to $$\label{nieuwevgl1}
p^6\int_0^\infty d\tau \frac{\rho_\|(\tau)}{(\tau+p^2)\tau^2}.$$ Evidently, e.g. the one loop approximation to $\mathcal{K}_\|(p^2)$ is not exact, since RGZ is still meant to be an expansion in the YM coupling $g^2$, though on top of a nontrivial vacuum, encoded in the condensates, e.g. the dynamical mass scales, present in . Due to asymptotic freedom, the one loop result will be trustworthy at sufficiently large $p^2$. As such, we will only retain the leading $\frac{1}{p^2}$ power correction (proportional to the mass scales present in the RGZ propagator) relative to the standard perturbative one loop estimate for $\mathcal{K}_\|(p^2)$.
Assuming we temporarily rewrite the RGZ gluon propagator as $$D(p^{2})=\frac{p^{2}+M_{1}^2}{p^4+M_2^2p^2+M_3^4},
\label{Prop_gluon}$$ we first need to extract the spectral density associated with the Källén-Lehmann representation of the physical part of the $K_\mu$ correlation function. This has been worked out in full detail in the Appendix \[appA\], leading to , $$\label{herhaling}
\rho_\|(\tau)=-2A_+A_-\frac{g^{4}(N^{2}-1)}{2^{2d+5}\pi^{7/2}\Gamma(\frac{d-1}{2})}\frac{\left(\tau^{2}-4b^{2}-4a\tau\right)^{(d-1)/2}}{\tau^{d/2}}\qquad\text{for}~\tau\geq \tau_c=2(a+\sqrt{a^2+b^2}),$$ where $$\label{herhaling2}
a=\frac{M_2^2}{2}\,,\qquad b=\frac{\sqrt{4M_3^4-M_2^4}}{2}.$$ A formal way to derive the spectral density based on properties of the Stieltjes transform can be found in [@Dudal:2010wn]. The threshold $\tau_c$ is the natural generalization to the case of 2 complex conjugate masses of the usual threshold appearing at $(m+m')^2$ in the case of 2 standard particles with masses $m$ and $m'$.
It is perhaps interesting to notice here that other non-perturbative approaches do also give reasonably good descriptions of the lattice data, see f.i. [@Aguilar:2008xm; @Fischer:2008uz; @Cyrol:2016tym; @Serreau:2012cg], but in particular is the output of functional equations as used in [@Aguilar:2008xm; @Fischer:2008uz; @Cyrol:2016tym] purely numerical, which would make the determination of the here required spectral density much more cumbersome, perhaps possibly via the (rather involved) numerical routine of [@Windisch:2012zd]. Therefore, we will solely rely here on the discussed RGZ propagator, which allows for closed analytical expressions.
As we use lattice data to get an estimate of the parameters $M_i$, we need to work in a lattice compatible renormalization scheme, such as the momentum subtraction MOM scheme, defined by $$D(p^{2}=\mu^2)=\frac{1}{\mu^2}.
\label{prop_gluon_MOM}$$ The proper renormalization factor $Z$, at scale $\mu$, is thus given by $$D(p^{2})=Z\frac{p^{2}+M_{1}^2}{p^4+M_2^2p^2+M_3^4},
\label{Prop_gluon_MOM_scale}$$ with $$Z=\frac{1}{\mu^2}\frac{\mu^4+M^2_2 \mu^2+M_3^4}{\mu^2+M_1^2}.
\label{factor_Z}$$ Since the gluon propagator we will use is the one renormalized in MOM scheme at scale $\mu$, the coupling constant $g^2$ present in becomes $$g^2(\mu)=\frac{1}{\beta_{0}\log\left(\frac{\mu^2}{\Lambda_{\mbox{\tiny{MOM}}}^2}\right)} \,,\qquad \beta_0=\frac{11}{3}\frac{N}{16\pi^2}.
\label{g_mu}$$ We have checked the conversion formulae of [@Boucaud:2008gn], leading to the following relation, valid for generic $N$ and $N_f=0$ between the MOM and ${\overline{\mbox{MS}}}$ scales. $$\Lambda_{\mbox{\tiny{MOM}}}={\Lambda_{\overline{\mbox{\tiny{MS}}}}}e^{169/264}.$$ This relation follows from the general theory of [@Celmaster:1979km], using the following relation between the ${\overline{\mbox{MS}}}$ ($\overline g^2$) and MOM coupling ($g^2$), $$g^2=\overline g^2(1+c_1\overline g^2 + \ldots)\,,\qquad c_1=\frac{169}{36}N.$$ For ${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=2}$, we can use the estimate of [@Lucini:2008vi], $${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=2}\approx 0.752~\sqrt{\sigma}$$ with $\sigma$ the string tension to set the physical scale on a lattice. Using the standard value $\sqrt{\sigma}\approx 0.44~\text{GeV}$, we find $${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=2}\approx 331~\text{MeV}$$ and thus $$\Lambda_{\mbox{\tiny{MOM}}}^{N=2}\approx 628~\text{MeV}.
\label{Lambda_MOM_2}$$ Moreover, [@Lucini:2008vi] also reports ${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=3}\approx 0.538\sqrt{\sigma}\approx 237~\text{MeV}$, a value which compares favourably well with the estimate of [@Boucaud:2008gn], stating ${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=3}\approx 224~\text{MeV}$. The (somewhat older) work [@Boucaud:2001st] predicted ${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=3}\approx 233~\text{MeV}$. The more recent work [@Sternbeck:2010xu] gave a preliminary value of ${\Lambda_{\overline{\mbox{\tiny{MS}}}}}^{N=3}\approx 0.62r_0\approx 262~\text{MeV}$ by using $r_0\approx2.367/\text{GeV}$. Using the estimate of [@Boucaud:2008gn], we get $$\label{MOM3}
\Lambda_{\mbox{\tiny{MOM}}}^{N=3}\approx 425~\text{MeV}.$$ Let us first work out the $SU(3)$ case. Including the renormalization factors $Z$, the eventual spectral density can be obtained from , $$\begin{aligned}
\rho_\|(\tau)=-2A_+A_-\frac{g^{4}(\mu)Z^2}{2^{9}\pi^{4}}\frac{\left(\tau^{2}-4b^{2}-4a\tau\right)^{3/2}}{\tau^{2}}.
\label{rho_tau_Z}\end{aligned}$$ Using the lattice obtained values [@Oliveira:2012eh] $$\label{fitje}
M_{1}^2=4.473\pm0.021~\text{GeV}^2\,,\qquad M_{2}^{2}=0.704\pm0.029~\text{GeV}^2\,,\qquad M_{3}^4=0.3959\pm0.0054~\text{GeV}^4\,,$$ we get as central values $$\begin{aligned}
a=0.352~\text{GeV}^2\,,\qquad b=0.522~\text{GeV}^2\,,\qquad 2A_+A_-=31.719.\end{aligned}$$ The fitted estimates of the mass scales entering were obtained by matching the tree level propagator on top of the non-perturbative gluon lattice data, as a way to determine the size of the dynamical RGZ mass scales. The latter capture the non-perturbative nature of the nontrivial RGZ vacuum, around which we expect perturbation theory to work, as the RGZ mass scales offer a dynamical screening of the Landau pole. At the to be considered scales $\mu$, relative to the MOM scale , the corresponding MOM strong coupling expansion parameter is effectively very small, an indication that a perturbative treatment certainly makes sense in the considered momentum region, after which we, using the described Padé analysis, “extrapolate” to the deep infrared, where we can no longer trust our perturbative result, albeit corrected with power corrections, for the correlation function. This is also the region where we eventually have to consider the zero momentum limit to make contact with the topological susceptibility. The Padé extrapolation is thus used to estimate the hard to access small momentum behaviour of a correlation function from its controllable behaviour at higher momentum. This shares a certain resemblance with the (Laplace or other) sum rules approaches to estimate the topological susceptibility, where the ultimate source of non-perturbative effects is also tracing back to non-trivial QCD vacuum condensates that enter the Operator Product Expansion of a specific correlation function, from which is, after transforming, then also extracted the information of interest by scanning for an optimal parameter space [@Narison:2002pw; @Narison:1994hv].
More precisely, we approximated with an $[\mathcal{M}+2,\mathcal{M}]$ Padé rational function in variable $p^2$, which are the ones having the same large $p^2$ behavior, viz. $\mathcal{O}(p^4)$. We opted to do the Padé approximation around $p^2=\mu^2$. In general, given a function $f(x)$, its $[\mathcal{N},\mathcal{M}]$ Padé approximant $R_{\mathcal{N},\mathcal{M}}^{x_0}(x)$ is given by $$\begin{aligned}
R_{\mathcal{N},\mathcal{M}}^{x_0}(x) &=& \frac{a_0+\ldots + a_{\mathcal{N}}x^{\mathcal{N}}}{1+\ldots+ b_{\mathcal{M}}x^{\mathcal{M}}}\end{aligned}$$ such that the Taylor series around $x_0$ of $R_{\mathcal{N},\mathcal{M}}^{x_0}(x)$ and $f(x)$ coincide up to order $\mathcal{N}+\mathcal{M}$.
With this, we can study the function $\chi(\mu^2)$ using the previous ingredients and search for optimal values, in the sense of minimal dependence, on the scale $\mu^2$. We remark here that the scale at which we do the Padé approximation should be not too small, so that we can trust the (perturbatively) computed r.h.s. of , and not too large so that we are taking into account sizable non-perturbative effects from the presence of the RGZ mass scales in , and perform a sensible extrapolation of the approximant to zero momentum to get an estimate for $\chi^4$. A natural choice is to expand around the renormalization scale $\mu^2$, since the MOM renormalization scale is subject to the same assumptions when used to renormalize lattice data, see e.g. [@Boucaud:2000ey]. The results are shown in FIG. 1 for $\mathcal{M}=1,2,3$ for a reasonable interval for $\mu^2$, to be compared with the lattice ballpark value of $\chi\sim 200~$MeV [@Alles:1996nm]. It is not a surprise that the results are pushed down as $\cal M$ grows, since for ${\cal M}\to\infty$ the approximant will converge to the original propagator which we know to have a trivial $\chi$. So, although Padé approximation suggests a nonzero value for $\chi$, it is difficult to provide a definite estimate.
To get an error estimation from the uncertainty on $\vec{x}\equiv(M_1^2,M_2^2, M_3^4)$, we compute the corresponding standard deviation on $\chi(\mu^2)$ in the standard way, $$\begin{aligned}
\sigma_\chi(\mu^2)=\sqrt{\sum_i \left(\frac{{\partial}\chi}{{\partial}x_i}\right)^2 \sigma_{x_i}^2},\end{aligned}$$ as the errors on the $x_i$ are small; the $\sigma_i$ can be read off from . We have displayed $\sigma_\chi(\mu^2)$ in Fig. 2.
![The $SU(3)$ topological susceptibility $\chi$ for variable $\mu^2$ for $\mathcal{M}=1,2,3$ (full, dashed, dotted).[]{data-label="chi"}](figuur1.pdf){width="70.00000%"}
![Estimated error on $\chi$ for $SU(3)$ due to the uncertainty on the fitting parameters for variable $\mu^2$ for ${\cal M}=1,2,3$ (full, dashed, dotted).[]{data-label="chi"}](figuur2.pdf){width="70.00000%"}
For $N=2$, $\Lambda_{\mbox{\tiny{MOM}}}$ is given by and the spectral density thence reads $$\begin{aligned}
\rho_\|(\tau)=-2A_+A_-\frac{3g^{4}(\mu)Z^2(\mu)}{2^{12}\pi^{4}}\frac{\left(\tau^{2}-4b^{2}-4a\tau\right)^{3/2}}{\tau^{2}}.
\label{rho_tau_Z_2}\end{aligned}$$ Following the same procedure as for $N=3$, we get the graphs of FIG. 3 and FIG. 4 in the $N=2$ case. Here, we used $$M_{1}^2=2.508\pm0.078~\text{GeV}^2\,,\qquad M_{2}^{2}=0.590\pm0.026~\text{GeV}^2\,,\qquad M_{3}^4=0.518\pm0.013~\text{GeV}^4,$$ as can be inferred from the largest volume data of Table II in [@Cucchieri:2011ig], yielding as central values $$\begin{aligned}
a=0.295~\text{GeV}^2\,,\qquad b=0.657~\text{GeV}^2\,,\qquad 2A_+A_-=6.176.\end{aligned}$$
![The $SU(2)$ topological susceptibility $\chi$ for variable $\mu^2$ for ${\cal M}=1,2,3$ (full, dashed, dotted).[]{data-label="chi"}](figuur3.pdf){width="70.00000%"}
![Estimated error on $\chi$ for $SU(2)$ due to the uncertainty on the fitting parameters for variable $\mu^2$ for ${\cal M}=1,2,3$ (full, dashed, dotted).[]{data-label="chi"}](figuur4.pdf){width="70.00000%"}
For the record, let us mention that the $SU(2)$ lattice prediction for the topological susceptibility sets $\chi=200-230~\text{MeV}$, see [@Teper:1999wp]. A compatible value was also found in [@Campagnari:2008yg] using a non-perturbative continuum Hamiltonian approach in Coulomb gauge.
Conclusion
==========
We have analyzed the topological susceptibility, $\chi^4$, in $SU(2)$ and $SU(3)$ Euclidean Yang-Mills theory in a generic linear covariant gauge taking into account the Gribov ambiguity. We employed a recently constructed effective action that implements a restriction of the gauge field path integration to a suitable subregion so that at least the infinitesimal gauge copies are eliminated, this without violating the BRST symmetry. As a consequence, the topological susceptibility is, as required, gauge invariant in this non-perturbative framework, explicitly checked to leading order in the present work. In an attempt to get estimates for the topological susceptibility, we developed a particular Padé rational function approximation based on the Källén-Lehmann spectral integral representation of the topological current correlation function. To improve upon the presented crude estimates, we plan to include the next order correction in future work. Notice this will be computationally challenging, thanks to the significantly enlarged set of vertices in the now considered Refined Gribov-Zwanziger action for the linear covariant gauge.
Acknowledgments {#acknowledgments .unnumbered}
===============
The Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq-Brazil), the Faperj, Funda[ç]{}[ã]{}o de Amparo [à]{} Pesquisa do Estado do Rio de Janeiro, the SR2-UERJ and the Coordena[ç]{}[ã]{}o de Aperfei[ç]{}oamento de Pessoal de N[í]{}vel Superior (CAPES) are gratefully acknowledged for financial support. S. P. Sorella is a level PQ-1 researcher under the program Produtividade em Pesquisa-CNPq, 300698/2009-7; M. S. Guimaraes is supported by the Jovem Cientista do Nosso Estado program - FAPERJ E-26/202.844/2015, is a level PQ-2 researcher under the program Produtividade em Pesquisa-CNPq, 307905/2014-4 and is a Procientista under SR2-UERJ. C. P. Felix is a PhD student supported by the program Ci[ê]{}ncias sem Fronteiras - CNPq, 234112/2014-0.
Explicit evaluation of the spectral density {#appA}
===========================================
In this Appendix, we compute at leading order the spectral density of the topological current correlator. Using $$K_{\mu}=\frac{g^{2}}{16\pi^{2}}\varepsilon_{\mu\nu\lambda\sigma}A_{\nu}^{a}\left(\partial_{\lambda}A_{\sigma}^{a}+\frac{g}{2}f^{abc}A_{\lambda}^{b}A_{\sigma}^{c}\right),
\label{Chern_Simons_current}$$ we obtain for the current-current correlator $\left<KK\right>$ at one loop order $$\braket{K_{\mu}(x)K_{\nu}(y)}=\frac{g^{4}}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha} \braket{A_{\beta}^{a}(x)\partial_{\rho}A_{\sigma}^{a}(x)A_{\lambda}^{d}(y)\partial_{\theta}A_{\alpha}^{d}(y)}.
\label{KK_correlator}$$ Now, in Fourier space $$\begin{aligned}
\braket{K_\mu(x) K_\nu(y)}&=&\frac{g^{4}}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha} \braket{A_{\beta}^{a}(x)\partial_{\rho}A_{\sigma}^{a}(x)A_{\lambda}^{d}(y)\partial_{\theta}A_{\alpha}^{d}(y)}\nonumber\\
&=&\frac{g^{4}}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha}\int dkdpdudq \mathrm{e}^{iqy}\mathrm{e}^{ipx}\mathrm{e}^{ikx}\mathrm{e}^{iuy}p_{\rho}p_{\theta} \braket{A_{\beta}^{a}(k)A_{\sigma}^{a}(p)A_{\lambda}^{d}(u)A_{\alpha}^{d}(q)}.
\label{KK_cor_Fourier}\end{aligned}$$ We perform the contractions using Wick’s theorem and disregarding the disconnected contributions, we get $$\begin{aligned}
\braket{K_\mu(x) K_\nu(y)}=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha}\left[\int dkdq \mathrm{e}^{i(x-y)(k-q)}q_{\rho}q_{\theta}D_{\beta\lambda}(k)D_{\sigma\alpha}(-q)+\right.\nonumber\\
\left.+\int dkdp\mathrm{e}^{i(x-y)(p+k)}p_{\rho}k_{\theta}D_{\beta\alpha}(k)D_{\sigma\lambda}(p)\right].
\label{KK_cor_Fourier_solution}\end{aligned}$$ Via the substitutions $$\ell=k-q \quad \mathrm{and} \quad \ell'=p+k,
\label{vai_chan}$$ we can rewrite $$\braket{K_\mu(x) K_\nu(y)}=\int\frac{d^{d}\ell}{(2\pi)^{d}}\mathrm{e}^{i\ell(x-y)}\mathcal{F}_{\mu\nu}(\ell)+\int\frac{d^{d}\ell'}{(2\pi)^{d}}\mathrm{e}^{i\ell'(x-y)}\mathcal{G}_{\mu\nu}(\ell'),
\label{KK_cor_Fourier_transform}$$ by which $$\mathcal{F}_{\mu\nu}(\ell)=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha}\int\frac{d^{d}k}{(2\pi)^{d}}\left[(k-\ell)_{\rho}(k-\ell)_{\theta}D_{\beta\lambda}(k)D_{\sigma\alpha}(\ell-k)\right]
\label{fl}$$ and $$\mathcal{G}_{\mu\nu}(\ell')=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\varepsilon_{\mu\beta\rho\sigma}\varepsilon_{\nu\lambda\theta\alpha}\int\frac{d^{d}k}{(2\pi)^{d}}\left[(\ell'-k)_{\rho}k_{\theta}D^{\beta\alpha}(k)D^{\sigma\lambda}(\ell'-k)\right].
\label{glprime}$$ Using [FeynCalc]{} [@feyncalc], these expressions can be further simplified to $$\begin{aligned}
\mathcal{F}_{\mu\nu}(\ell)=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}\frac{D(\ell-k)}{k^{2}}&&\left\{\delta_{\mu\nu}\left[k^{2}\left(\ell^{2}(D(k)-L(k))-2D(k)k^{2}\right)+\right.\right.\nonumber\\
&&+\left.\left.k.\ell\left((L(k)-D(k))k.\ell+4D(k)k^{2}\right)-2D(k)k^{2}\ell^{2}\right]+\right.\nonumber\\
&&+\left.
\ell_{\mu}\left[\vphantom{\frac{1}{2}}
\ell_{\nu}\left(k^{2}(L(k)-D(k))+2D(k)k^{2}\right)+\right.\right.\nonumber\\
&&+\left.\left.k_{\nu}\left((D(k)-L(k))k.\ell-2D(k)k^{2}\right)\vphantom{\frac{1}{2}}\right]+\right.\nonumber\\
&&+\left.k_{\mu}\left[\vphantom{\frac{1}{2}}k_{\nu}\left(\ell^{2}(L(k)-D(k))+2D(k)k^{2}\right)+\right.\right.\nonumber\\
&&+\left.\left.\ell_{\nu}\left((D(k)-L(k))k.\ell-2D(k)k^{2}\right)
\vphantom{\frac{1}{2}}\right]
\right\}
\label{flM}\end{aligned}$$and $$\begin{aligned}
\mathcal{G}_{\mu\nu}(\ell')=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}&&2D(k)D(\ell'-k)\left[\delta_{\mu\nu}(k.\ell'-k^{2})+k_{\mu}(k_{\nu}-\ell'_{\nu})\right].
\label{glprimeM}\end{aligned}$$ We are only interested in the longitudinal component, which can be extracted by acting with the appropriate projector $L_{\mu\nu}(p)$. As it can be easily verified, the $\alpha$-dependent contributions do cancel, as expected from gauge invariance. As such, from here on, we can set $L(k)=0$. Doing so, reduces to $$\begin{aligned}
\mathcal{F}_{\mu\nu}(\ell)=\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}\frac{D(k)D(\ell-k)}{k^{2}}&&\left\{-\delta_{\mu\nu}\left[2k^{2}\ell^{2}+k^{2}(2k^{2}-\ell^{2})-4k^{2}k.\ell+(k.\ell)^{2}\right]+\right.\nonumber\\
&&+\left.\ell_{\mu}\left[k^{2}\ell_{\nu}+k_{\nu}(k.\ell-2k^{2})\right]+k_{\mu}\left[(2k^{2}-\ell^{2})k_{\nu}+\right.\right.\nonumber\\
&&+\left.\left.\ell_{\nu}(k.\ell-2k^{2})\right]
\right\}.
\label{flMLG}\end{aligned}$$ Eventually, in Fourier space we obtain $$\begin{aligned}
\braket{K_\mu(p) K_\nu(-p)}&=&\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}D(k)D(p-k) \Biggl\{6\delta_{\mu\nu}k.p+4k_{\mu}k_{\nu}-4\delta_{\mu\nu}k^{2}-\delta_{\mu\nu}p^{2}-\Biggr.\nonumber\\
&-&\Biggl.4k_{\mu}p_{\nu}+\delta_{\mu\nu}\frac{(k.p)^{2}}{k^{2}}+p_{\mu}p_{\nu}+p_{\mu}k_{\nu}\frac{k.p}{k^{2}}-2p_{\mu}k_{\nu}-p^{2}\frac{k_{\mu}k_{\nu}}{k^{2}}+k_{\mu}p_{\nu}\frac{k.p}{k^{2}}\Biggr\}
\label{KK_correlator_simple}\end{aligned}$$ The longitudinal part of is then given by $$\begin{aligned}
\braket{K_\mu(p) K_\nu(-p)}&=&\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\frac{p_\mu p_\nu}{p^2}\int\frac{d^{d}k}{(2\pi)^{d}}D(k)D(p-k) 4\frac{(k.p)^{2}-k^{2}p^{2}}{p^{2}}\nonumber\\
&=&\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}\frac{1}{k^{2}-m_{1}^{2}}\frac{1}{(p-k)^{2}-m_{2}^{2}}4\frac{(k.p)^{2}-k^{2}p^{2}}{p^{2}}.
\label{longitudinal_KK}\end{aligned}$$ The next step is to rewrite in spectral form, i.e. to extract the spectral density $\rho_\|(p^2)$, making use of $$\begin{aligned}
\label{decomp}
D(p^2)=\frac{p^2+M_1^2}{p^4+M_2^2p^2+M_3^4}=\frac{A_+}{p^2+m_+^2}+\frac{A_-}{p^2+m_-^2},\end{aligned}$$ where we assume $M_2^4<4M_3^4$ as motivated from all lattice fits, so that the poles and residues are complex-conjugate numbers. Using this decomposition into standard massive Feynman propagators, each propagator with mass $m_+^2$ needs to be combined with an accompanying propagator with complex-conjugate mass $m_-^2$ to assure a branch cut along the usual (real) half-axis, consistent with a standard Källén-Lehmann integral [@Baulieu:2009ha]. This means that we need to take into account the following contributions, $$\begin{aligned}
\braket{K_\mu(p) K_\nu(-p)}&=&\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}\left(\frac{A_+}{k^{2}+m_{+}^{2}}\frac{A_-}{(p-k)^{2}+m_{-}^{2}}+\frac{A_-}{k^{2}+m_{-}^{2}}\frac{A_+}{(p-k)^{2}+m_{+}^{2}}\right)4\frac{(k.p)^{2}-k^{2}p^{2}}{p^{2}}\nonumber\\
&=&\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}2A_+A_-\left(\frac{1}{k^{2}+m_{+}^{2}}\frac{1}{(p-k)^{2}+m_-^2}\right)4\frac{(k.p)^{2}-k^{2}p^{2}}{p^{2}}.
\label{longitudinal_KK10}\end{aligned}$$ We will for the time being forget about the prefactor $2A_+A_-$ and will restore it at the end.
To continue, we temporarily look at a general massive propagator in Minkowski space, i.e. at $$\begin{aligned}
\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d}}\frac{1}{k^{2}-m_{1}^{2}}\frac{1}{(p-k)^{2}-m_{2}^{2}}4\frac{(k.p)^{2}-k^{2}p^{2}}{p^{2}}.
\label{longitudinal_KK2}\end{aligned}$$ The reason is that, as discussed in [@Dudal:2010wn], to compute the spectral density entering the Källén-Lehmann representation, we can formally use the (Minkowski) Cutkosky cut rules [@Cutkosky:1960sp] pretending to work with particles with on-shell real masses, and at the end, move back to Euclidean space, simultaneously replacing the masses with their respective complex-conjugate values $$m_{1}^{2}\to m_+^2=a+ib \quad \mathrm{and} \quad m_{2}^{2}\to m_-^2=a-ib.
\label{cmp}$$ So, let us apply the Cutkosky rules via the usual replacement, $$\frac{1}{k^{2}-m_{1}^{2}}\rightarrow 2\pi\theta(k^{0})\delta({\vec{k}^{2}-m_{1}^{2}})
\label{p1_r}$$ and $$\frac{1}{(p-k)^{2}-m_{2}^{2}}\rightarrow 2\pi\theta((p-k)^{0})\delta({(\vec{p}-\vec{k})^{2}-m_{2}^{2}}).
\label{p2_r}$$ We know that $\rho(\tau)\propto Disc \mathcal{K}(\tau)=2\text{Im}~\mathcal{K}(\tau)$, thus $$\begin{aligned}
\text{Im}~\mathcal{K}(p,m_{1},m_{2})&=&\frac{1}{2}\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d-2}}\theta(k^{0})\delta({\vec{k}^{2}-m_{1}^{2}})\theta((p-k)^{0})\delta({(\vec{p}-\vec{k})^{2}-m_{2}^{2}})\left(\frac{4(k.p)^{2}}{p^{2}}-\right.\nonumber\\
&&\left.-4k^{2}\vphantom{\frac{1}{2}}\right).
\label{ImF}\end{aligned}$$ We will work in the center-of-mass frame, i.e. $p_{\mu}=(p^{0},0)=(E,0)$, then $$\begin{aligned}
\text{Im}~\mathcal{K}(E^{2})&=&\frac{1}{2}\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d-2}}\theta(k^{0})\delta({(\vec{k}^{0})^{2}-\omega_{k,1}^{2}})\theta((E-k)^{0})\delta({(E-k^{0})^{2}-\omega_{k,2}^{2}})\left(\frac{4(k^{0}E)^{2}}{E^{2}}-\right.\nonumber\\
&&\left.-4((k^{0})^{2}+\vec{k}^{2})\vphantom{\frac{1}{2}}\right)\nonumber\\
&=&\frac{1}{2}\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d-2}}\theta(k^{0})\delta({(\vec{k}^{0})^{2}-\omega_{k,1}^{2}})\theta((E-k)^{0})\delta({(E-k^{0})^{2}-\omega_{k,2}^{2}})\left(4(k^{0})^{2}-\right.\nonumber\\
&&\left.-4(k^{0})^{2}-4\vec{k}^{2}\right)\nonumber\\
&=&-2\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d^{d}k}{(2\pi)^{d-2}}\theta(k^{0})\delta({(\vec{k}^{0})^{2}-\omega_{k,1}^{2}})\theta((E-k)^{0})\delta({(E-k^{0})^{2}-\omega_{k,2}^{2}})\vec{k}^{2},
\label{ImF_CoM}\end{aligned}$$ with $\omega_{k,i}=\sqrt{\vec{k}^{2}+m_{i}^{2}}$. Integrating over $k^{0}$ and changing for spherical coordinates, we obtain $$\begin{aligned}
\text{Im}~\mathcal{K}(E^{2})=-\frac{\pi^{d-1/2}}{\Gamma(\frac{d-1}{2})}\frac{g^{4}(N^{2}-1)}{16^{2}\pi^{4}}\int\frac{d|\vec{k}|}{(2\pi)^{d-2}}|\vec{k}|^{d-2}\frac{1}{\omega_{k,1}\omega_{k,2}}\delta(E-\omega_{k,1}-\omega_{k,2})\vec{k}^{2}.
\label{ImF_k_0}\end{aligned}$$ Using the property $\delta(g(\vec{k}^{2}))=\frac{1}{|g\prime(\vec{k}_{0}^{2})|}\delta(\vec{k}^{2}-\vec{k}_{0}^{2})=\frac{1}{2|\vec{k}_{0}||g\prime(\vec{k}_{0}^{2})|}\delta(|\vec{k}|-|\vec{k}_{0}|)$, where $\vec{k}_{0}^{2}$ is such that $g(\vec{k}_{0}^{2})=0$, i.e. $$\begin{aligned}
g(\vec{k}_{0}^{2})&=&E-\sqrt{\vec{k}_{0}^{2}+m_{1}^{2}}-\sqrt{\vec{k}_{0}^{2}+m_{2}^{2}}=0\nonumber\\
&|\vec{k}_{0}|&=\frac{\sqrt{(E^{2}-m_{1}^{2}-m_{2}^{2})^{2}-4m_{1}^{2}m_{2}^{2}}}{2E},
\label{k_0}\end{aligned}$$ gives us $$\begin{aligned}
\text{Im}~\mathcal{K}(E^{2})=-\frac{g^{4}(N^{2}-1)}{2^{d+6}\pi^{5/2}\Gamma(\frac{d-1}{2})}\frac{|\vec{k}_{0}|^{d-3}\vec{k}_{0}^{2}}{E}.
\label{ImF_solved}\end{aligned}$$ With conjugate masses parameterized as in , can be rewritten as $$|\vec{k}_{0}|=\frac{\sqrt{E^{4}-4b^{2}-4aE^{2}}}{2E}
\label{k_0_cmp}$$ and with the two last equations, we can write as $$\begin{aligned}
\text{Im}~\mathcal{K}(E^{2})&=&-\frac{g^{4}(N^{2}-1)}{2^{d+6}\pi^{5/2}\Gamma(\frac{d-1}{2})}\frac{\left(\frac{\sqrt{E^{4}-4b^{2}-4aE^{2}}}{2E}\right)^{d-3}\left(\frac{\sqrt{E^{4}-4b^{2}-4aE^{2}}}{2E}\right)^{2}}{E}\nonumber\\
&=&-\frac{g^{4}(N^{2}-1)}{2^{2d+5}\pi^{5/2}\Gamma(\frac{d-1}{2})}\frac{\left(E^{4}-4b^{2}-4aE^{2}\right)^{(d-1)/2}}{E^{d}}.
\label{ImF_solved_cmp}\end{aligned}$$ Using $E^{2}\rightarrow \tau$ and the equivalence $\rho=\frac{1}{\pi}\text{Im}~ \mathcal{K}(\tau)$ we finally get $$\begin{aligned}
\rho_\|(\tau)=-2A_+A_-\frac{g^{4}(N^{2}-1)}{2^{2d+5}\pi^{7/2}\Gamma(\frac{d-1}{2})}\frac{\left(\tau^{2}-4b^{2}-4a\tau\right)^{(d-1)/2}}{\tau^{d/2}}\qquad\text{for}~\tau\geq \tau_c.
\label{rho_tau}\end{aligned}$$ The threshold is given by [@Dudal:2010wn] $$\begin{aligned}
\label{threshold}
\tau_c=(m_++m_-)^2=2(a+\sqrt{a^2+b^2}).\end{aligned}$$ We also restored the prefactor $2A_+A_-$.
[99]{}
D. E. Kharzeev and E. M. Levin, Phys. Rev. Lett. [**114**]{} (2015) 242001. D. Dudal and M. S. Guimaraes, Phys. Rev. D [**93**]{} (2016) 085010. V. N. Gribov, Nucl. Phys. B [**139**]{} (1978) 1. N. Vandersickel and D. Zwanziger, Phys. Rept. [**520**]{} (2012) 175. R. F. Sobreiro and S. P. Sorella, hep-th/0504095. I. M. Singer, Commun. Math. Phys. [**60**]{} (1978) 7. G. Dell’Antonio and D. Zwanziger, Nucl. Phys. B **326** (1989) 333. P. van Baal, Nucl. Phys. B **369** (1992) 259. M. Semenov-Tyan-Shanskii and V. Franke, “A variational principle for the Lorentz condition and restriction of the domain of path integration in non-abelian gauge theory,” Zap. Nauch. Sem. Leningrad. Otdeleniya Matematicheskogo Instituta im V. A. Steklov, AN SSSR 120 (1982) p159. (English translation: New York: Plenum Press 1986).
A. Maas, Phys. Rev. D [**79**]{} (2009) 014505; F. Canfora, A. Giacomini and J. Oliva, Phys. Rev. D [**84**]{} (2011) 105019; A. Sternbeck and M. Müller-Preussker, Phys. Lett. B [**726**]{} (2013) 396; J. Serreau and M. Tissier, Phys. Lett. B [**712**]{} (2012) 97; A. Cucchieri and T. Mendes, Phys. Rev. D [**88**]{} (2013) 114501; V. G. Bornyakov, V. K. Mitrjushkin and R. N. Rogalyov, Phys. Rev. D [**89**]{} (2014) 054504; J. Serreau, M. Tissier and A. Tresmontant, Phys. Rev. D [**89**]{} (2014) 125019; P. Y. Moshin and A. A. Reshetnyak, Nucl. Phys. B [**888**]{} (2014) 92; A. Cucchieri, D. Dudal, T. Mendes and N. Vandersickel, Phys. Rev. D [**90**]{} (2014) 051501. D. Dudal, M. S. Guimaraes, I. F. Justo and S. P. Sorella, Eur. Phys. J. C [**75**]{} (2015) 83; A. G. Duarte, O. Oliveira and P. J. Silva, Phys. Rev. D [**94**]{} (2016) 014502; A. Cucchieri and T. Mendes, Phys. Rev. D [**94**]{} (2016) 014505. F. Canfora, D. Hidalgo and P. Pais, Phys. Lett. B [**763**]{} (2016) 94; A. Maas, Phys. Rev. D [**93**]{} (2016) 054504; A. A. Reshetnyak and P. Y. Moshin, Russ. Phys, Journal 59 (2016) 165 (on Russian). D. Zwanziger, Nucl. Phys. B [**323**]{} (1989) 513. D. Zwanziger, Nucl. Phys. B [**399**]{} (1993) 477. N. Maggiore and M. Schaden, Phys. Rev. D [**50**]{} (1994) 6616. D. Dudal, S. P. Sorella and N. Vandersickel, Eur. Phys. J. C [**68**]{} (2010) 283.
M. A. L. Capri, M. S. Guimaraes, I. Justo, L. F. Palhares and S. P. Sorella, Eur. Phys. J. C [**76**]{} (2016) 141. D. Dudal, S. P. Sorella, N. Vandersickel and H. Verschelde, Phys. Rev. D [**77**]{} (2008) 071501.
D. Dudal, J. A. Gracey, S. P. Sorella, N. Vandersickel and H. Verschelde, Phys. Rev. D [**78**]{} (2008) 065047.
D. Dudal, S. P. Sorella and N. Vandersickel, Phys. Rev. D [**84**]{} (2011) 065039. D. Dudal, O. Oliveira and N. Vandersickel, Phys. Rev. D [**81**]{} (2010) 074505. A. Cucchieri, D. Dudal, T. Mendes and N. Vandersickel, Phys. Rev. D [**85**]{} (2012) 094513. D. Dudal, O. Oliveira and J. Rodriguez-Quintero, Phys. Rev. D [**86**]{} (2012) 105005.
M. A. L. Capri [*et al.*]{}, Phys. Rev. D [**92**]{} (2015) 045039; M. A. L. Capri [*et al.*]{}, Phys. Rev. D [**93**]{} (2016) 065019; M. A. L. Capri [*et al.*]{}, Phys. Rev. D [**94**]{} (2016) 025035; M. A. L. Capri, D. Fiorentini, A. D. Pereira, R. F. Sobreiro, S. P. Sorella and R. C. Terin, Annals Phys. [**376**]{} (2017) 40; M. A. L. Capri, D. Dudal, A. D. Pereira, D. Fiorentini, M. S. Guimaraes, B. W. Mintz, L. F. Palhares and S. P. Sorella, Phys. Rev. D [**95**]{} (2017) 045011. M. Lavelle and D. McMullan, Phys. Rept. **279** (1997) 1.
M. A. L. Capri, D. Fiorentini, M. S. Guimaraes, B. W. Mintz, L. F. Palhares and S. P. Sorella, Phys. Rev. D [**94**]{} (2016) 065009. M. A. L. Capri, D. Fiorentini, A. D. Pereira and S. P. Sorella, arXiv:1708.01543 \[hep-th\]. A. Cucchieri, T. Mendes and E. M. S. Santos, Phys. Rev. Lett. [**103**]{} (2009) 141602; F. Siringo, Phys. Rev. D [**90**]{} (2014) 094021; A. C. Aguilar, D. Binosi and J. Papavassiliou, Phys. Rev. D [**91**]{} (2015) 085014; M. Q. Huber, Phys. Rev. D [**91**]{} (2015) 085018; F. Siringo, Phys. Rev. D [**92**]{} (2015) 074034; P. Bicudo, D. Binosi, N. Cardoso, O. Oliveira and P. J. Silva, Phys. Rev. D [**92**]{} (2015) 114514; A. C. Aguilar, D. Binosi and J. Papavassiliou, Phys. Rev. D [**95**]{} (2017) 034017.
F. T. Hawes, C. D. Roberts and A. G. Williams, Phys. Rev. D [**49**]{} (1994) 4683. R. Aouane, V. G. Bornyakov, E. M. Ilgenfritz, V. K. Mitrjushkin, M. Muller-Preussker and A. Sternbeck, Phys. Rev. D [**85**]{} (2012) 034501. K. Fukushima and K. Kashiwa, Phys. Lett. B [**723**]{} (2013) 360. F. E. Canfora, D. Dudal, I. F. Justo, P. Pais, L. Rosa and D. Vercauteren, Eur. Phys. J. C [**75**]{} (2015) 326.
G. ’t Hooft, Phys. Rev. Lett. [**37**]{} (1976) 8; G. ’t Hooft, Phys. Rept. [**142**]{} (1986) 357.
M. Marino, *Instantons and large $N$: An introduction to non-perturbative methods in QFT*, see <http://laces.web.cern.ch/laces/LACES10/notes/instlargen.pdf>.
G. Veneziano, Nucl. Phys. B [**159**]{} (1979) 213.
E. Witten, Nucl. Phys. B [**156**]{} (1979) 269.
B. Alles, M. D’Elia and A. Di Giacomo, Nucl. Phys. B [**494**]{} (1997) 281 Erratum: \[Nucl. Phys. B [**679**]{} (2004) 397\];
D. E. Kharzeev, Int. J. Mod. Phys. A [**31**]{} (2016) 1645023.
E. Meggiolaro, Phys. Rev. D [**58**]{} (1998) 085002.
E. Seiler and I. O. Stamatescu, MPI-PAE-PTh-10-87. <http://ccdb5fs.kek.jp/cgi-bin/img_index?8705403> E. Seiler, Phys. Lett. B [**525**]{} (2002) 355.
E. Vicari and H. Panagopoulos, Phys. Rept. [**470**]{} (2009) 93.
R. Alkofer, W. Detmold, C. S. Fischer and P. Maris, Phys. Rev. D [**70**]{} (2004) 014014; A. Cucchieri, T. Mendes and A. R. Taurines, Phys. Rev. D [**71**]{} (2005) 051902; J. M. Cornwall, Mod. Phys. Lett. A [**28**]{} (2013) 1330035; P. J. Silva, D. Dudal and O. Oliveira, PoS LATTICE [**2013**]{} (2014) 366.
S. Narison, Camb. Monogr. Part. Phys. Nucl. Phys. Cosmol. [**17**]{} (2001) 1.
R. J. Crewther, Riv. Nuovo Cim. [**2N8**]{} (1979) 63.
D. Espriu and R. Tarrach, Z. Phys. C [**16**]{} (1982) 77.
M. Luscher, Commun. Math. Phys. [**85**]{} (1982) 39. N. Seiberg, Phys. Lett. [**148B**]{} (1984) 456. F. Karsch, M. L. Laursen, T. Neuhaus and B. Plache, Phys. Rev. D [**50**]{} (1994) 5912.
M. Luscher, Phys. Lett. B [**593**]{} (2004) 296.
D. Dudal and M. S. Guimaraes, Phys. Rev. D [**83**]{} (2011) 045013.
A. C. Aguilar, D. Binosi and J. Papavassiliou, Phys. Rev. D [**78**]{} (2008) 025010. C. S. Fischer, A. Maas and J. M. Pawlowski, Annals Phys. [**324**]{} (2009) 2408. A. K. Cyrol, L. Fister, M. Mitter, J. M. Pawlowski and N. Strodthoff, Phys. Rev. D [**94**]{} (2016) no.5, 054005. J. Serreau and M. Tissier, Phys. Lett. B [**712**]{} (2012) 97; M. Tissier and N. Wschebor, Phys. Rev. D [**82**]{} (2010) 1017. A. Windisch, R. Alkofer, G. Haase and M. Liebmann, Comput. Phys. Commun. [**184**]{} (2013) 109.
P. Boucaud, F. De Soto, J. P. Leroy, A. Le Yaouanc, J. Micheli, O. Pene, J. Rodriguez-Quintero, Phys. Rev. [**D79** ]{} (2009) 014508.
W. Celmaster, R. J. Gonsalves, Phys. Rev. [**D20** ]{} (1979) 1420.
B. Lucini, G. Moraitis, Phys. Lett. [**B668** ]{} (2008) 226-232. P. Boucaud, A. Le Yaouanc, J. P. Leroy, J. Micheli, O. Pene, J. Rodriguez-Quintero, Phys. Rev. [**D63** ]{} (2001) 114003. A. Sternbeck, E. -M. Ilgenfritz, K. Maltman, M. Muller-Preussker, L. von Smekal, A. G. Williams, PoS [**LAT2009** ]{} (2009) 210.
O. Oliveira and P. J. Silva, Phys. Rev. D [**86**]{} (2012) 114513. S. Narison, G. M. Shore and G. Veneziano, Nucl. Phys. B [**433**]{} (1995) 209.
P. Boucaud [*et al.*]{}, JHEP [**0004**]{} (2000) 006. M. Teper, Nucl. Phys. Proc. Suppl. [**83**]{} (2000) 146; P. de Forcrand, M. Garcia Perez and I. O. Stamatescu, Nucl. Phys. B [**499**]{} (1997) 409; T. A. DeGrand, A. Hasenfratz and T. G. Kovacs, Nucl. Phys. B [**505**]{} (1997) 417.
D. R. Campagnari and H. Reinhardt, Phys. Rev. D [**78**]{} (2008) 085001. V. Shtabovenko, R. Mertig and F. Orellana, Comput. Phys. Commun. [**207**]{} (2016) 432. See also <http://www.feyncalc.org/>. L. Baulieu, D. Dudal, M. S. Guimaraes, M. Q. Huber, S. P. Sorella, N. Vandersickel and D. Zwanziger, Phys. Rev. D [**82**]{} (2010) 025021.
R. E. Cutkosky, J. Math. Phys. [**1**]{} (1960) 429.
[^1]: Due to the intricate infrared problems accompanying instantons in an infinite volume, a large $N$ extrapolation is difficult when not working in a finite volume [@marino]
[^2]: Compare indeed to the complex $i\theta F\tilde F$ topological term in the Euclidean YM action.
[^3]: We temporarily ignored the necessary subtractions here, see later.
[^4]: It has been shown that the topological charge operator itself is a renormalization group invariant [@Vicari:2008jw; @Espriu:1982bw] in pure gauge theories (it is not when dynamical quarks are included, in which case even operator mixing occurs with the chiral current). Though we are interested in the topological susceptibility of the pure gauge theory. Notice that it is indeed this quantity which enters the Witten-Veneziano formula for the $\eta'$ mass [@Witten:1979vv; @Veneziano:1979ec]. As a consequence, the leading ultraviolet behaviour in the $p^2\to\infty$ limit of the correlator $\mathcal{K}_\|(p^2)$ will also be completely determined by its tree level behaviour, taking into account the asymptotic freedom and vanishing anomalous dimension of the topological charge operator.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The generalized multiple depot traveling salesmen problem (GMDTSP) is a variant of the multiple depot traveling salesmen problem (MDTSP), where each salesman starts at a distinct depot, the targets are partitioned into clusters and at least one target in each cluster is visited by some salesman. The GMDTSP is an NP-hard problem as it generalizes the MDTSP and has practical applications in design of ring networks, vehicle routing, flexible manufacturing scheduling and postal routing. We present an integer programming formulation for the GMDTSP and valid inequalities to strengthen the linear programming relaxation. Furthermore, we present a polyhedral analysis of the convex hull of feasible solutions to the GMDTSP and derive facet-defining inequalities that strengthen the linear programming relaxation of the GMDTSP. All these results are then used to develop a branch-and-cut algorithm to obtain optimal solutions to the problem. The performance of the algorithm is evaluated through extensive computational experiments on several benchmark instances.'
address:
- 'Graduate Student, Dept. of Mechanical Engineering, Texas A$\&$M University, College Station, TX 77843, USA'
- 'Assistant Professor, Dept. of Mechanical Engineering, Texas A$\&$M University, College Station, TX 77843, USA'
author:
- Kaarthik Sundar
- Sivakumar Rathinam
bibliography:
- 'GMDTSP.bib'
title: 'Generalized multiple depot traveling salesmen problem - polyhedral study and exact algorithm'
---
Generalized multiple depot traveling salesmen ,Routing ,Branch-and-cut ,Polyhedral study
Introduction\[sec:Introduction\]
================================
The generalized multiple depot travelling salesmen problem (GMDTSP) is an important combinatorial optimization problem that has several practical applications including but not limited to maritime transportation, health-care logistics, survivable telecommunication network design ([@Bektas2011]), material flow system design, postbox collection ([@Laporte1996]), and routing unmanned vehicles ([@Manyam2014; @Oberlin2010]). The GMDTSP is formally defined as follows: let $D:=\{d_1,\dots,d_k\}$ denote the set of depots and $T$, the set of targets. We are given a complete undirected graph $G=(V,E)$ with vertex set $V:=T\cup D$ and edge set $E:=\{(i,j):i\in V , j\in T\}$. In addition, a proper partition $C_1,\dots,C_m$ of $T$ is given; these partitions are called *clusters*. For each edge $(i,j)=e \in E$, we associate a non-negative cost $c_e=c_{ij}$. The GMDTSP consists of determining a set of at most $k$ simple cycles such that each cycle starts an ends at a distinct depot, at least one target from each cluster is visited by some cycle and the total cost of the set of cycles is a minimum. The GMDTSP reduces to a multiple depot traveling salesmen problem (MDTSP - [@Benavent2013]) when every cluster is a singleton set. The GMDTSP involves two related decisions:
1. choosing a subset of targets $S\subseteq T$, such that $|S\cap C_h|\geq 1$ for $h=1,\dots,m$;
2. solving a MDTSP on the subgraph of $G$ induced by $S\cup D$.
The GMDTSP can be considered either as a generalization of the MDTSP in [@Benavent2013] where the targets are partitioned into clusters and at least one target in each cluster has to be visited by some salesman or as a multiple salesmen variant of the symmetric generalized traveling salesman problem (GTSP) in [@Fischetti1995; @Fischetti1997]. [@Benavent2013] and [@Fischetti1995] present a polyhedral study of the MDTSP and GTSP polytope respectively, and develop a branch-and-cut algorithm to compute optimal solutions for the respective problem.
This is the first work in the literature that analyzes the facial structure and derives additional valid and facet-defining inequalities for the convex hull of feasible solutions to the GMDTSP. This paper presents a mixed-integer linear programming formulation and develops a branch-and-cut algorithm to solve the problem to optimality. This work generalizes the results of the two aforementioned problems namely the MDTSP ([@Benavent2013]) and the GTSP ([@Fischetti1995]).
Related work:\[subsec:litreview\]
---------------------------------
A special case of the GMDTSP with one salesman, the symmetric generalized traveling salesman problem (GTSP), was first introduced by [@Henry1969] and [@Srivastava1969] in relation to record balancing problems arising in computer design and to the routing of clients through agencies providing various services respectively. Since then, the GTSP has attracted considerable attention in the literature as several variants of the classical traveling salesman problem can be modeled as a GTSP ([@Laporte1996; @Feillet2005; @Oberlin2009; @Manyam2014]). [@Noon1989] developed a procedure to transform a GTSP to an asymmetric traveling salesman problem and the [@Laporte1987] investigated the asymmetric counterpart of the GTSP. Despite most of the aforementioned applications of the GTSP ([@Laporte1996]) extending naturally to their multiple depot variant, there are no exact algorithms in the literature to address the GMDTSP.
A related generalization of the GMDTSP can be found in the vehicle routing problem (VRP) literature. VRPs are capacitated counterparts for the TSPs where the vehicles have a limited capacity and each target is associated with a demand that has to be met by the vehicle visiting that target. The multiple VRPs can be classified based on whether the vehicles start from a single depot or from multiple depots. The generalized multiple vehicle routing problem (GVRP) is a capacitated version of the GMDTSP with all the vehicles starting from a single depot. [@Bektas2011] present four formulations for the GVRP, compare the linear relaxation solutions for them, and develop a branch-and-cut to optimally solve the problem. In [@Laporte1987a], @Laporte1987a models the GVRP as a location-routing problem. On the contrary, [@Ghiani2000] develop an algorithm to transform the GVRP into a capacitated arc routing problem, which therefore enables one to utilize the available algorithms for the latter to solve the former. In a more recent paper, [@Bautista2008] study a special case of the GVRP derived from a waste collection application where each cluster contains at most two vertices. The authors describe a number of heuristic solution procedures, including two constructive heuristics, a local search method and an ant colony heuristic to solve several practical instances. To our knowledge, there are no algorithms in the literature to compute optimal solutions to the generalized multiple depot vehicle routing problem or the GMDTSP.
The objective of this paper is to develop an integer programming formulation for the GMDTSP, study the facial structure of the GMDTSP polytope and develop a branch-and-cut algorithm to solve the problem to optimality. The rest of the paper is organized as follows: in Sec. \[sec:Formulation\] we introduce notation and present the integer programming formulation. In Sec. \[sec:polyhedral\], the facial structure of the GMDTSP polytope is studied and its relation to the MDTSP polytope ([@Benavent2013]) is established. We also introduce a general theorem that allows one to lift any facet of the MDTSP polytope into a facet of the GMDTSP polytope. We further use this result to develop several classes of facet-defining inequalities for the GMDTSP. In the subsequent sections, the formulation is used to develop a branch-and-cut algorithm to obtain optimal solutions. The performance of the algorithm is evaluated through extensive computational experiments on 116 benchmark instances from the GTSP library ([@Gutin2010]).
Problem Formulation\[sec:Formulation\]
======================================
We now present a mathematical formulation for the GMDTSP inspired by models in [@Benavent2013] and [@Fischetti1995]. We propose a two-index formulation for the GMDTSP. We associate to each feasible solution $\mathcal{F}$, a vector $\mathbf{x}\in\mathbb{R}^{|E|}$ (a real vector indexed by the elements of $E$) such that the value of the component $x_{e}$ associated with edge $e$ is the number of times $e$ appears in the feasible solution $\mathcal{F}$. Note that for some edges $e\in E$, $x_{e}\in\{0,1,2\}$ *i.e,* we allow the degenerate case where a cycle can only consist of a depot and a target. If $e$ connects two vertices $i$ and $j$, then $(i,j)$ and $e$ will be used interchangeably to denote the same edge. Similarly, associated to $\mathcal{F}$, is also a vector $\mathbf{y}\in\mathbb{R}^{|T|}$, *i.e.,* a real vector indexed by the elements of $T$. The value of the component $y_{i}$ associated with a target $i \in T$ is equal to one if the target $i$ is visited by a cycle and zero otherwise.
For any $S\subset V$, we define $\gamma(S)=\{(i,j)\in E:i,j\in S\}$ and $\delta(S)=\{(i,j)\in E:i\in S,\, j\notin S\}$. If $S=\{i\},$ we simply write $\delta(i)$ instead of $\delta(\{i\})$. We also denote by $C_{h(v)}$ the cluster containing the target $v$ and define $W:=\{v\in T:|C_{h(v)}|=1\}$. Finally, for any $\hat{{E}}\subseteq E$, we define $x(\bar{E})=\sum_{(i,j)\in\bar{E}}x_{ij}$, and for any disjoint subsets $A,B\subseteq V$, $(A:B) = \{(i,j)\in E: i\in A, j\in B\}$ and $x(A:B)=\sum_{e\in(A:B)} x_{ij}$. Using the above notations, the GMDTSP is formulated as a mixed integer linear program as follows:
& \_[eE]{}c\_[e]{}x\_[e]{}\[eq:obj\] &\
& &\
& x((i))=2y\_[i]{}iT,\[eq:degree\]\
& \_[iC\_h]{} y\_[i]{} 1h{1,…,m},\[eq:assignment\] &\
& x((S))2y\_[i]{}ST,iS,\[eq:sec\] &\
& x(D’:{j})+3x\_[jk]{}+x({k}:DD’)2(y\_[j]{}+y\_[k]{})j,kT;D’D,\[eq:4path\] &\
& x(D’:{j})+2x((S{j,k}))+x({k}:DD’)\_[vS]{}2 y\_[v]{}+2(y\_[j]{}+y\_[k]{})-y\_i&\
& iS;j,kT, jk;ST{j,k},S;D’D,\[eq:path\]\
& x\_[e]{}{0,1}e(T),\[eq:xinteger1\] &\
& x\_[e]{}{0,1,2}e(D:T),\[eq:xinteger2\] &\
& y\_[i]{}{0,1}iT.\[eq:yinteger\]
In the above formulation, the constraints in ensure the number edges incident on any vertex $i\in T$ is equal to $2$ if and only if target $i$ is visited by a cycle ($y_{i}=1$). The constraints in force at least one target in each cluster to be visited. The constraints in are the connectivity or sub-tour elimination constraints. They ensure a feasible solution has no sub-tours of any subset of customers in $T$. The constraints in and are the path elimination constraints. They do not allow for any cycle in a feasible solution to consist of more than one depot. The validity of these constraints is discussed in the subsection \[sub:Path-elimination\]. Finally, the constraints - are the integrality restrictions on the $\mathbf{x}$ and $\mathbf{y}$ vectors.
Path elimination constraints:\[sub:Path-elimination\]
------------------------------------------------------
The first version of the path elimination constraints was developed in the context of location routing problems [@Laporte1986]. @Laporte1986 named these constraints as chain-barring constraints. Authors in [@Belenguer2011] and [@Benavent2013] use similar path elimination constraints for the location routing and the multiple depot traveling salesmen problems. The version of path elimination constraints used in this article is adapted from [@Sundar2014]. Any path that originates from a depot and visits exactly two customers before terminating at another depot is removed by the constraint in . The validity of the constraint can be easily verified as in [@Laporte1986; @Sundar2014]. Any other path $d_{1},t_{1},\cdots,t_{p},d_{2}$, where $d_{1},d_{2}\in D$, $t_{1},\cdots,t_{p}\in T$ and $p\geq3$, violates inequality with $D'=\{d_{1}\},$ $S=\{t_{2},\cdots,t_{p-1}\}$, $j=t_{1}$, $k=t_{p}$ and $i=t_{r}$ where $2\leq r\leq p-1$. The proof of validity of the constraint in Eq. is discussed as a part of the polyhedral analysis of the polytope of feasible solutions to the GMDTSP in the next section (see proposition \[prop:path\]).
We note that our formulation allows for a feasible solution with paths connecting two depots and visiting exactly one customer. We refer to such paths as 2-paths. As the formulation allows for two copies of an edge between a depot and a target, 2-paths can be eliminated and therefore there always exists an optimal solution which does not contain any 2-path. In the following subsection, we prove polyhedral results and derive classes of facet-defining inequalities for the model in -.
Polyhedral analysis \[sec:polyhedral\]
======================================
In this section we analyse the facial structure of the GMDTSP polytope while leveraging the results already known for the multiple depot traveling salesmen problem (MDTSP).
If the number of targets $|T|=n$ and the number of depots $|D|=k$, then the number of $x_e$ variables is $|E| = \binom{n}{2}+nk$ ($\binom{n}{2}$ is the number of edges between the targets and $nk$ is the number of edges between targets and depots). Also the number of $y_i$ variables is $|T|=n$ and hence, the total number of variables used in the problem formulation is $|E|+|T| = \binom n2 + nk + n$. Let $P$ and $Q$ denote the GMDTSP and MDTSP as follows:
P &:= {()\^[|E|+|T|]{}: () } \[eq:P\]\
Q &:= {()P: y\_v=1 vT }. \[eq:Q\]
The dimension of the polytope $Q$ was shown to be $\binom n2 + n(k-1)$ in [@Benavent2013]. To relate the polytopes $P$ and $Q$, we define an intermediate polytope $P(F)$ as follows:
P(F):= {()P: y\_v = 1 v F}, \[eq:PF\]
where $\emptyset \subseteq F \subseteq T$. Observe that $P(\emptyset) = P$ and $P(T) = Q$. Now, we determine the dimension of the polytope $P(F)$. The number of variables in the equation system for $P(F)$ is $|E|+|T| = \binom n2 + nk + n$. The system also includes $|T|=n$ linear independent equations in and variable fixing equations given by $$y_v=1 \text{ for all } v\in F\cup W$$ where, $W$ is the set of targets that lie in clusters that are singletons (defined in Sec. \[sec:Formulation\]). The following lemma gives the dimension of $P(F)$.
\[lem:PFdim\] For all $F\subseteq T$, $\operatorname{dim}(P(F)) = \binom n2 + nk - |F\cup W|$.
Since the equation system for $P(F)$ has $\binom n2 + nk + n$ variables and $n+|F\cup W|$ linear independent equality constraints, the $\operatorname{dim}(P(F)) \leq \binom n2 + nk - |F\cup W|$. We claim that $P(F)$ contains $\binom n2 + nk - |F\cup W| + 1$ affine independent points. The claim proves $\operatorname{dim}(P(F)) \geq \binom n2 + nk - |F\cup W|$. Hence, the lemma follows. We prove the claim by induction on the cardinality of the set $F$.
For the base case, we have $F = T$ and $P(T)=Q$ where $Q$ is the the MDTSP polytope. Since $\operatorname{dim}(Q) = \binom n2 + nk - n$ ([@Benavent2013]), there are $\binom n2 + nk - n + 1$ affine independent points in $Q$. Assume that the claim holds for a set $F_i$ with $|F_i| = i$ and $i>0$, and consider a subset of targets $F_{i-1}$ such that $|F_{i-1}|=i-1$. Let $v$ be any target not in $F_{i-1}$, and define $F_i:= F_{i-1}\cup\{v\}$. The induction hypothesis provides $\binom n2 + nk - |F_i\cup W| + 1$ affine independent points belonging to $P(F_i)$ and hence, to $P(F_{i-1})$ (since $P(F_i)\subseteq P(F_{i-1})$). If $v\in W$, then $|F_{i-1}\cup W| = |F_i\cup W|$ and we are done. Otherwise, $|F_{i-1}\cup W| = |F_i\cup W| - 1$ and we need an additional point on the polytope $P(F_{i-1})$ that is affine independent with the rest of the $\mathcal L = \binom n2 + nk - |F_i\cup W| + 1$ points. All these $\mathcal L$ points satisfy the equation $y_v = 1$. An additional point that is affine independent with the $\mathcal L$ points always exists and is given by any feasible MDTSP solution in the subgraph induced by the set of vertices $(T\cup D)\setminus \{v\}$ because, any feasible MDTSP solution on the set of vertices $(T\cup D)\setminus \{v\}$ satisfies $y_v = 0$.
$\operatorname{dim}(P) = \binom n2 + nk - |W|$.
Lemma \[lem:PFdim\] indicates that for any given subset $F\subseteq T$ and $v\in F$, either $\operatorname{dim}(P(F\setminus \{v\})) = \operatorname{dim}(P(F))$ (if $v \in W$) or $\operatorname{dim}(P(F\setminus \{v\})) = \operatorname{dim}(P(F)) + 1$ (when $v \notin W$) *i.e.*, the dimension of the polytope $P(F)$ increases by at most one unit when a target is removed from $F$. Hence, we can lift any facet-defining valid inequality for $P(F)$ to be facet-defining for $P(F\setminus \{v\})$. In the ensuing proposition, we introduce a result based on the sequential lifting for zero-one programs ([@Padberg1975]) which we will use to lift facets of $Q$ into facets of $P$. The proposition generalizes a similar result in [@Fischetti1995] used to lift facets of the travelling salesman problem to facets of GTSP.
\[prop:lifting\] Suppose that for any $F\subseteq T$ and $u\in F$, $$\sum_{e\in E} \alpha_e x_e + \sum_{v\in T} \beta_v (1-y_v) \geq \eta$$ is any facet-defining inequality for $P(F)$. Then the lifted inequality $$\sum_{e\in E} \alpha_e x_e + \sum_{v\in T\setminus \{u\}} \beta_v (1-y_v) + \bar{\beta}_u(1-y_u) \geq \eta$$ is valid and facet-defining for $P(F\setminus \{u\})$, where $\bar{\beta}_u$ takes an arbitrary value when $u \in W$ and $$\bar{\beta}_u = \eta - \min \left\{\sum_{e\in E} \alpha_e x_e + \sum_{v\in T\setminus \{u\}} \beta_v (1-y_v): (\mathbf{x,y}) \in P(F\setminus \{u\}), y_u=0 \right\}$$ when $u \notin W$. Note that the statement can be trivially modified to deal with “$\leq$” inequalities.
The proof follows from the sequential lifting theorem in [@Padberg1975].
Proposition \[prop:lifting\] is used to derive facet-defining inequalities for the GMDTSP polytope $P$ by lifting the facet-defining inequalities for the MDTSP polytope $Q$ in [@Benavent2013]. For a given lifting sequence of the set of targets $T$, say $\{v_1,\dots,v_n\}$, the procedure is iteratively applied to derive a facet of $P(\{v_{t+1},\dots,v_n\})$ from a facet of $P(\{v_{t},\dots,v_n\})$ for $t=1,\dots,n$. Different lifting sequences produce different facets; hence the name, *sequence dependent* lifting. In the rest of the section, we use the lifting procedure to check if the constraints in - are facet-defining and derive additional facet-defining inequalities for the GMDTSP polytope.
\[prop:trivial\] The following results hold for the GMDTSP polytope $P$:
1. $x_e \geq 0$ defines a facet for every $e \in E$ if $|T|\geq 4$,
2. $x_e\leq 1$ defines a facet if and only if $e\in \gamma(W)$ and $|T|\geq 3$,
3. $x_e \leq 2$ does not define a facet for any $e\in (D:T)$,
4. $y_i\geq 0$ does not define a facet for any $i\in T$,
5. $y_i\leq 1$ defines a facet if and only if $i\notin W$, and
6. $\sum_{i\in C_h} y_{i} \geq 1$ does not define a facet for any $h\in \{1,\dots,m\}$.
We use the facet-defining results of the MDTSP polytope ([@Benavent2013]) in conjunction with Proposition \[prop:lifting\] to prove (1)–(3).
1. Observe that for every $e \in E$, $x_e\geq 0$ defines a facet of the MDTSP polytope $Q$ if $|T| \geq 4$. Now for any lifting sequence, Proposition \[prop:lifting\] produces $\bar{\beta}_v = 0$ for all $v\in T$ and the result follows.
2. Suppose that $e = (i,j)$. If $i,j\in W$ and $|T|\geq 3$, then the claim follows from the forthcoming Proposition \[prop:gsec\] by choosing $S = \{i,j\}$. Otherwise if $e=(i,j)\in \gamma(T)$, then $x_e \leq 1$ is dominated by $x_e \leq y_i$ if $i \notin W$ and $x_e \leq y_j$ if $j \notin W$.
3. Let $e = (d,i)$ where $d\in D, i\in T$. $x_e \leq 2$ defines a face of the MDTSP polytope $Q$. Hence neither of the lifted versions of the inequality *i.e.*, $x_e \leq 2$ (if $i \in W$) or $x_e \leq 2y_i$ (if $i\notin W$) defines a facet of $P$.
4. The inequality $y_i \geq \frac 12 x_e$ for $e \in \delta(i)$ dominates $y_i \geq 0$. Hence, $y_i\geq 0$ does not define a facet for any $i\in T$.
5. Observe that the valid inequality $y_i\leq 1$ induces a face, $P(\{i\}) = \{(\mathbf{x,y})\in P: y_i = 1\}$ of $P$. From the Lemma \[lem:PFdim\], $\operatorname{dim}(P(\{i\})) = \operatorname{dim}(P) - 1$ if and only if $i\notin W$. Hence, $y_i \leq 1$ is facet-defining for $P$ if and only if $i\notin W$. When $i\in W$, the inequality defines an improper face.
6. The constraint $\sum_{i\in C_h} y_{i} \geq 1$ can be reduced to $\sum_{e\in \delta(C_h)} x_e + 2\sum_{e\in \gamma(C_h)} x_e \geq 2$ using the degree constraints in . When $\gamma(C_h) \neq \emptyset$, the constraint $\sum_{e\in \delta(C_h)} x_e + 2\sum_{e\in \gamma(C_h)} x_e \geq 2$ is dominated by $\sum_{e\in \delta(C_h)} x_e \geq 2$. When $\gamma(C_h) = \emptyset$ (*i.e.,* $|C_h| = 1$), the constraint $\sum_{e\in \delta(C_h)} x_e = 2$ is satisfied by any feasible solution in $P$ and hence in this case, it is an improper face. Therefore, $\sum_{i\in C_h} y_{i} \geq 1$ does not define a facet for any $h\in \{1,\dots,m\}$.
In the next proposition, we prove that the sub-tour elimination constraints in Eq. define facets of $P$. To do so, we apply the lifting procedure in Proposition \[prop:lifting\] to the MDTSP sub-tour elimination constraints $$x(\delta(S)) \geq 2 \text{ for all } S\subseteq T.$$ In the process, we derive alternate versions of the sub-tour elimination constraints in Eq. which we will refer to as the generalized sub-tour elimination constraints (GSEC). To begin with, we observe that sub-tour elimination constraints given above define facets of the MDTSP poytope $Q$ when $|T|\geq 3$ (see [@Benavent2013]).
\[prop:gsec\] Let $S\subseteq T$ and $|T|\geq 3$. Then the following *Generalized Sub-tour Elimination Constraint* (GSEC) is valid and facet-defining for $P$: $$x(\delta(S)) + \bar{\beta}_i (1-y_i) \geq 2 \text{ for } i\in S,$$ where $$\bar{\beta}_i = \begin{cases} 2 & \text{ if } \mu(S)=0, \\ 0 & \text{ otherwise};\end{cases}$$ $\mu(S)$ is defined as $\mu(S) = |\{h:C_h \subseteq S\}|$.
We first observe that the inequality $x(\delta(S)) \geq 2$ with $S\subseteq T$ and $|T|\geq 3$ defines a facet for the MDTSP polytope. We lift this inequality using the lifting procedure in Proposition \[prop:lifting\]. Let $\{v_1,\dots,v_n\}$ be any lifting sequence of the set of targets such that $v_n = i$. The lifting coefficients $\bar{\beta}_{v_t}$ are computed iteratively for $t=1,\dots,n$. For $t=1,\dots,n-1$, it is trivial to see that $\bar{\beta}_{v_t} = 0$. Hence, $x(\delta(S)) \geq 2$ defines a facet of $P(\{v_n\})$. As to $\bar{\beta}_{v_n}$, we compute its value by performing the lifting procedure again and obtain a facet of $P$. We have $$\bar{\beta}_{v_n} = 2 - \min \left\{x(\delta(S)): (\mathbf{x,y}) \in P, \text{ and } y_{v_n} = 0\right\}.$$ Solving for $\bar{\beta}_{v_n}$ using the above equation, we obtain $\bar{\beta}_{v_n} = 2$ if a feasible GMDTSP solution visiting no target in $S$ exists (*i.e.*, no $C_h \subseteq S$ exists) and $\bar{\beta}_{v_n} = 0$ otherwise.
In summary, the Proposition \[prop:gsec\] results in the following facet-defining inequalities of $P$: suppose $S\subseteq T$ with $|T|\geq 3$. Then,
x((S)) &2 (S)0 \[eq:sec1\]\
x((S)) &2y\_i (S) = 0, iS. \[eq:sec2\]
Note that the inequality $x(\delta(S)) \geq 2y_i$ is valid for any $S \subseteq T$. It is facet-defining for $P$ only when $\mu(S) \neq 0$. When $\mu(S)=0$ it does not define a facet of $P$ as it is dominated by Eq. . Using the degree constraints in Eq. , the above GSECs can rewritten as
x((S)) &\_[vS]{} y\_v - 1 (S)0 \[eq:sec1a\]\
x((S)) &\_[vS{i}]{} y\_v (S) = 0, iS. \[eq:sec2a\]
In the forthcoming two propositions, we prove that the path elimination constraints in Eq. and are facet-defining of $P$ using Proposition \[prop:lifting\]. The corresponding path elimination constraints for the MDTSP polytope $Q$ are as follows: suppose that $j,k \in T$, $D'\subset D$ with $D'\neq \emptyset$, then
x(D’:{j})+3x\_[jk]{}+x({k}:DD’)&4 \[eq:4pathmdtsp\]\
x(D’:{j})+2x((S{j,k}))+x({k}:DD’)&2|S|+3 ST{j,k},S\[eq:pathmdtsp\]
We remark that Eq. and define facets for the MDTSP polytope $Q$ (see [@Benavent2013]).
\[prop:4path\] Suppose $j,k\in T$ and $D' \subset D$ with $D'\neq \emptyset$. Then the following path elimination constraint is valid and facet-defining for $P$: $$x(D':\{j\})+3x_{jk}+x(\{k\}:D\setminus D') + \bar{\beta}_j (1-y_j) + \bar{\beta}_k (1-y_k) \leq 4$$ where $\bar{\beta}_j = \bar{\beta}_k = 2$.
Let $\{v_1,\dots,v_n\}$ be any lifting sequence of the set of targets such that $v_{n-1} = j$ and $v_n = k$. The lifting coefficients are iteratively computed for $t=1,2,\dots,n$. Coefficients $\bar{\beta}_v$ for $v\in\{v_1,\dots,v_{n-2}\}$ are easily computed (tight GMDTSP solution is depicted in Fig. \[fig:4path\](a), showing that the value of $\bar{\beta}_v$ cannot be increased without producing a violated inequality).
![Tight feasible solutions for proof of Proposition \[prop:4path\]. The vertices $d_1$ and $d_2$ are depots and $j,k,$ and $v$ are targets.[]{data-label="fig:4path"}](fig1.pdf)
![Tight feasible solutions for proof of Proposition \[prop:4path\]. The vertices $d_1$ and $d_2$ are depots and $j,k,$ and $v$ are targets.[]{data-label="fig:4path"}](fig2.pdf)
Similarly for $t=n-1$ *i.e.*, $v_t = j$, the correctness of the coefficient $\bar{\beta}_j = 2$ can be checked with the help of Fig. \[fig:4path\](b). Analogously, we obtain $\bar{\beta}_{k} = 2$.
The inequality in Proposition \[prop:4path\] can be rewritten as $x(D':\{j\})+3x_{jk}+x(\{k\}:D\setminus D')\leq2(y_{j}+y_{k})$ which is the path elimination constraint in Eq. . We have proved that this inequality is valid and defines a facet of $P$.
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig3.pdf)
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig4.pdf)
\
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig5.pdf)
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig6.pdf)
\
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig7.pdf)
![Tight feasible solutions for proof of Proposition \[prop:path\].[]{data-label="fig:path"}](fig8.pdf)
\[prop:path\] Let $j,k\in T$, $D' \subset D$, $S\subseteq T\setminus \{j,k\}$ and $i\in S$ such that $D'\neq \emptyset$ and $S\neq \emptyset$. Also let $\bar S = S\cup \{j,k\}$. Then the following *Generalized Path Elimination Constraint* (PSEC) is valid and facet-defining for $P$: $$x(D':\{j\})+2x(\gamma(\bar S))+x(\{k\}:D\setminus D') + \sum_{v\in T} \bar{\beta}_v (1-y_v) \leq 2|S|+3$$ where $$\bar{\beta}_v = \begin{cases} 0 & \text{ if } v \in T\setminus{\bar S}, \\ 2 & \text{ if } v \in \bar S\setminus\{i\}, \\ 1 & \text{ if } v = i \text{ and } \mu(S) = 0, \\ 2 & \text{ if } v = i \text{ and } \mu(S) \neq 0;\end{cases}$$ $\mu(S)$ is defined as $\mu(S) = |\{h:C_h \subseteq S\}|$.
Consider any lifting sequence of the the set of targets $\{v_1,\dots,v_n\}$ such that each target in the set $S\setminus \{i\}$ follows all the targets in the set $|T\setminus \bar S|$ and $v_{n-2}=j$, $v_{n-1}=k$ and $v_n = i$. The coefficients $\bar\beta_v = 0$ for $v\in T\setminus \bar S$ and $\bar\beta_v = 2$ for $v\in S\setminus \{i\}$ are trivial to compute (tight GMDTSP solution is depicted in Fig. \[fig:path\](a) and \[fig:path\](b) respectively, showing that the value of $\bar{\beta}_v$ cannot be increased without producing a violated inequality). The correctness of coefficients $\bar \beta_j=2$ and $\bar \beta_k=2$ can be checked with the help of Fig. \[fig:path\](c) and \[fig:path\](d), respectively.
It remains to compute the value of coefficient $\bar\beta_i$. For computing $\bar\beta_i$, we have to take into account for the possibility of a GMDTSP solution not visiting any target in the set $S$. This can happen when $\mu(S) = 0$. In this case, we obtain $\bar \beta_i = 1$; see Fig. \[fig:path\](e). Likewise, when $\mu(S) \neq 0$, any GMDTSP solution has to have at least two edges in $\delta(S)$. This leads to $\bar \beta_i = 2$; tight GMDTSP solution is shown in Fig. \[fig:path\](f).
In summary, the Proposition \[prop:path\] results in the following facet-defining inequalities of $P$: suppose $j,k\in T$, $D' \subset D$, $S\subseteq T\setminus \{j,k\}$, $\bar S = S\cup \{j,k\}$ and $i\in S$ such that $D'\neq \emptyset$ and $S\neq \emptyset$, then
x(D’:{j})+2x((|S))+x({k}:DD’)&\_[v|S]{}2y\_v - y\_i (S)= 0 \[eq:pec1\]\
x(D’:{j})+2x((|S))+x({k}:DD’)&\_[v|S]{}2y\_v - 1 (S)0. \[eq:pec2\]
We note that the above PSECs can be rewritten in cut-set form as
x((|S)) &x(D’:{j})+x({k}:DD’)+ y\_i (S)= 0 \[eq:pec1a\]\
x((|S)) &x(D’:{j})+x({k}:DD’)+ 1 (S)0. \[eq:pec2a\]
As we will see in the forthcoming section, the GPECs in the above form are more amicable for developing separation algorithms. Next, we examine the comb inequalities that are valid and facet-defining for the MDTSP polytope. These inequalities were initially introduced for the TSP in [@Chvatal1973]. These inequalities were extended and proved to be facet-defining for the MDTSP polytope in [@Benavent2013]. We define a comb inequality using a comb, which is a family $C = (H,\mathcal T_1, \mathcal T_2,\dots, \mathcal T_t)$ of $t+1$ subsets of the targets; $t$ is an odd number and $t\geq 3$. The subset $H$ is called the handle and the subsets $\mathcal T_1,\dots,\mathcal T_t$ are called teeth. The handle and teeth satisfy the following conditions:
i. $H\cap \mathcal T_i \neq \emptyset \quad \forall i=1,\dots,t,$
ii. $\mathcal T_i\setminus H \neq \emptyset \quad \forall i=1,\dots,t,$
iii. $\mathcal T_i\cap \mathcal T_j = \emptyset \quad 1\leq i\leq j \leq t$.
The conditions i. and ii. indicate that every tooth $T_i$ intersects the handle $H$ and the condition iii. indicates that no two teeth intersect. We define the size of $C$ as $\sigma(C):= |H| + \sum_{i=1}^t|\mathcal T_i| - \frac{3t+1}{2}$. Then the comb inequality associated with $C$ is given by
x((H)) + \_[i=1]{}\^t x((T\_i)) &(C) \[eq:comba\]
The inequality in Eq. \[eq:comba\] is valid and facet-defining for the MDTSP (see [@Benavent2013]). A special case of the comb inequality, called *2-matching* inequality is obtained when $|\mathcal T_i| = 2$ for $i=1,\dots,t$. In the case of a 2-matching inequality, the size of the comb is $\sigma(C) = |H| + \frac{t+1}{2}$. We apply the lifting procedure in Proposition \[prop:lifting\] to the inequality in and obtain facet-defining inequality for the GMDTSP. The following proposition is adapted from [@Fischetti1995]; the proof of the proposition is omitted as it is similar to the proof of the corresponding theorem for GTSP in [@Fischetti1995].
\[prop:comb\] Suppose $\mu(S) = |\{h:C_h \subseteq S\}|$ for $S\subseteq T$ and let $C = (H,\mathcal T_1,\dots,\mathcal T_t)$ be a comb. For $i=1,\dots,t$, let $a_i$ be any target in $\mathcal T_i\cap H$ if $\mu(\mathcal T_i\cap H)=0$; $a_i=0$ (a dummy value) otherwise; and let $b_i$ be any target in $\mathcal T_i\setminus H$ if $\mu(\mathcal T_i\setminus H) = 0$; $b_i=0$ otherwise. Then the following comb inequality is valid and facet-defining for the GMDTSP polytope $P$:
x((H)) + \_[i=1]{}\^t x((T\_i)) + \_[vT]{} |\_v (1-y\_v) &(C), \[eq:combb\]
where $\bar \beta_v = 0$ for all $v\in T\setminus (H\cup \mathcal T_1 \cup \dots\cup \mathcal T_t)$, $\bar \beta_v = 1$ for all $v\in H\setminus (\mathcal T_1 \cup \dots\cup \mathcal T_t)$ and for $i=1,\dots,t$:
|\_v &= 2\
|\_[a\_i]{} &= 1\
|\_v &= 1\
|\_[b\_i]{} &= 0 .
See [@Fischetti1995].
Additional valid inequalities specific to multiple depot problems \[subsec:Tcomb\]
----------------------------------------------------------------------------------
In this section, we will examine a special type of comb inequality called the T-comb inequalities. The T-comb inequalities were introduced in [@Benavent2013] and proved to be valid and facet-defining for the MDTSP polytope. These inequalities are specific to problems involving multiple depots and hence, are important for the GMDTSP. A T-comb inequality $C$ is defined by an handle $H$ and teeth $\mathcal T_1, \dots, \mathcal T_t$ such that the following conditions are satisfied:
i. $H\cap \mathcal T_i \neq \emptyset \quad \forall i=1,\dots,t,$
ii. $\mathcal T_i\setminus H \neq \emptyset \quad \forall i=1,\dots,t,$
iii. $\mathcal T_i\cap \mathcal T_j = \emptyset \quad 1\leq i\leq j \leq t,$
iv. $\mathcal T_i \cap D \neq \emptyset \quad \forall i=1,\dots,t,$
v. $H \subset T,$
vi. $H\setminus \cup_{i=1}^t \mathcal T_i \neq \emptyset,$
vii. $D\setminus \cup_{i=1}^t \mathcal T_i \neq \emptyset$.
The difference between the T-comb inequalities and the comb inequalities defined in Eq. is that, the number of teeth are allowed to be even ($t\geq 1$) and each teeth must contain a depot. The comb size in this case is given by $\sigma(C)= |H| + \sum_{i=1}^t|\mathcal T_i| - (t+1)$. In this paper, we will only examine the T-comb inequalities with $|\mathcal T_i| = 2$ for every $i\in\{1,\dots,t\}$; the size of the comb in this case reduces to $\sigma(C) = |H| + t - 1$ and the corresponding T-comb inequality is given by
x((H)) + \_[i=1]{}\^t x((T\_i)) &|H| + t - 1, \[eq:Tcomba\]
The inequality in Eq. is valid and facet-defining for the MDTSP when $t\geq 2$. Again, we apply the lifting procedure in Proposition \[prop:lifting\] to the inequality in and obtain facet-defining inequality for the GMDTSP.
\[prop:Tcomb\] Let $C = (H,\mathcal T_1,\dots,\mathcal T_t)$ be a T-comb with $|\mathcal T_i| = 2$ for every $i\in\{1,\dots,t\}$ and $t\geq 2$. Also suppose $|H\setminus \cup_i \mathcal T_i| > 1$ $($the proposition can be trivially extended to the case where $|H\setminus \cup_i \mathcal T_i| = 1)$. Let $\bar a$ be any target in $H\setminus \cup_i \mathcal T_i$. Then the following T-comb inequality is valid and facet-defining for the GMDTSP polytope $P$:
x((H)) + \_[i=1]{}\^t x((T\_i)) + \_[vT]{} |\_v (1-y\_v) &|H| + t - 1, \[eq:Tcombb\]
where $\bar \beta_v = 0$ for all $v\in T\setminus (H\cup \mathcal T_1 \cup \dots\cup \mathcal T_t)$, $\bar \beta_v = 1$ for all $v\in H\setminus (\mathcal T_1 \cup \dots\cup \mathcal T_t \cup \{\bar a\})$, $\bar \beta_{\bar a} = 0$, and $\bar \beta_v = 2$ for all $v\in \mathcal T_i \cap H,i=1,\dots,t$.
![Tight feasible solutions for proof of Proposition \[prop:Tcomb\].[]{data-label="fig:Tcomb"}](fig9.pdf)
![Tight feasible solutions for proof of Proposition \[prop:Tcomb\].[]{data-label="fig:Tcomb"}](fig10.pdf)
\
![Tight feasible solutions for proof of Proposition \[prop:Tcomb\].[]{data-label="fig:Tcomb"}](fig11.pdf)
![Tight feasible solutions for proof of Proposition \[prop:Tcomb\].[]{data-label="fig:Tcomb"}](fig12.pdf)
Consider any lifting sequence for the set of targets $T$ in the following order: (i) targets in the set $T\setminus (H\cup \mathcal T_1 \cup \dots\cup \mathcal T_t)$, (ii) $v\in H\setminus (\mathcal T_1 \cup \dots\cup \mathcal T_t \cup \{\bar a\})$, (iii) $\bar a$, and (iv) $v\in \mathcal T_i \cap H, i=1,\dots,t$. The lifting coefficients $\bar\beta_v = 0$ and $\bar\beta_v = 1$ for the sets in (i) and (ii) respectively, are trivial to compute (tight feasible GMDTSP solutions are depicted in Fig. \[fig:Tcomb\](a) and \[fig:Tcomb\](b), respectively). Similarly, tight feasible GMDTSP solutions for the cases where $\bar\beta_{\bar a} = 0$ and $\bar\beta_v = 2$ (cases (iii) and (iv)) are shown in Fig. \[fig:Tcomb\](c) and \[fig:Tcomb\](d), respectively.
In the above proposition, for the case when $|H\setminus \cup_i \mathcal T_i| = 1$, the facet-defining inequality is given by
x((H)) + \_[i=1]{}\^t x((T\_i)) &\_[i=1]{}\^t \_[vHT\_i]{} 2y\_v. \[eq:Tcombc\]
Separation algorithms\[sec:separation\]
=======================================
In this section, we discuss the algorithms that are used to find violated families of all the valid inequalities introduced in Sec. \[sec:polyhedral\]. We denote by $G^* = (V^*,E^*)$ the *support graph* associated with a given fractional solution $(\mathbf{x}^*, \mathbf{y}^*) \in \mathbb{R}^{|E|\cup|T|}$ *i.e.,* $G^*$ is a capacitated undirected graph with vertex set $V^* := \{i\in T: y^*_i >0\}\cup D$ and $E^*:=\{e\in E:x_e^* > 0\}$ with edge capacities $x^*_e$ for each edge $e\in E^*$.
Separation of generalized sub-tour elimination constraints in Eq. and Eq. : {#subsec:sec}
---------------------------------------------------------------------------
We first develop a separation algorithm for constraints in Eq. : $x(\delta(S)) \geq 2y_i$ for $\mu(S) = 0, i\in S$ and $S\subseteq T$. Given a fractional solution $(\mathbf{x}^*, \mathbf{y}^*)$, the most violated constraint of the form can be obtained by computing a minimum capacity cut $(S,V^*\setminus S)$ with $i\in S$ and $D\subseteq V^*\setminus S$ on the graph $G^*$. The minimum capacity cut can be obtained by computing a maximum flow from $i$ to $t$, where $t$ is an additional vertex connected with each depot in the set $D$ through an edge having very large capacity. The algorithm is repeated for every target $i \in T \cap V^*$ and the target set $S$ obtained during each run of the algorithm defines a violated inequality if the capacity of the cut is strictly less than $2y^*_i$. This procedure can be implemented in $O(|T|^4)$ time.
Now we consider the constraint in Eq. : $x(\delta(S)) \geq 2$ for $\mu(S) \neq 0$ and $S\subseteq T$. Given a fractional solution $(\mathbf{x}^*, \mathbf{y}^*)$, the most violated inequality in this case is obtained by computing a minimum capacity cut $(S,V^*\setminus S)$ with a cluster $C_h \subseteq S$ and $D\subseteq V^*\setminus S$ on the graph $G^*$. This is in turn achieved by computing a maximum $s-t$ flow on $G^*$, where $s$ and $t$ are additional vertices connected with each $j\in C_h$ and each $d \in D$ respectively through an edge having very large capacity. The algorithm is repeated for every cluster $C_h$ and the set $S$ obtained on each run of the algorithm defines a violated inequality if the capacity of the cut is strictly less than $2$. The time complexity of this procedure is $O(m|T|^3)$, where $m$ is the number of clusters.
We remark that the violated inequality of the form using the above algorithm, is not necessarily facet-defining as the set $S$ computed using the algorithm might have $\mu(S) \neq 0$. When this happens, we reject the inequality in favour of its dominating and facet-defining inequality in Eq. .
Separation of path elimination constraints in Eq. , , and : {#subsec:pec}
-----------------------------------------------------------
We first discuss the procedure to separate violated constraints in Eq. . Consider every pair of targets $j,k \in V^*\cap T$. We rewrite the constraint in as $x(D':\{j\}) + x(\{k\}:D\setminus D') \leq 2(y_k + y_j)-3x_{jk}$. Given $j,k$ and a fractional solution $(\mathbf{x}^*, \mathbf{y}^*)$, the RHS of the above inequality is a constant and is equal to $2(y_k^* + y_j^*) - 3x_{jk}^*$. We observe that the LHS of the inequality is maximized when $D'=\{d\in D:x^*_{jd} \geq x^*_{kd}\}$. Furthermore, when $D'=\emptyset$ or $D' = D$, no path constraint in Eq. is violated for the given pair of vertices. With $D' = \{d\in D: x^*_{jd} \geq x^*_{kd}\}$, if $x^*(D':\{j\}) + x^*(\{k\}:D\setminus D')$ is strictly greater than $2(y_k^* + y_j^*) - 3x_{jk}^*$, the path constraint in Eq. is violated for the pair of vertices $j,k$ and the subset of depots $D'$. This procedure can be implemented in $O(|T|^2)$.
For constraints in Eq. and , we present two separation algorithms that are very similar to the algorithms presented in Sec. \[subsec:sec\]. We will use the equivalent constraints in Eq. and to develop the algorithms. We first consider the path elimination constraint in Eq. . Given $j,k$ and a fractional solution $(\mathbf{x}^*, \mathbf{y}^*)$, we first compute $D'$ to maximize $x^*(D':\{j\}) + x^*(\{k\}:D\setminus D') := \mathcal L$. Now, the most violated constraint of the form can be obtained by computing a minimum capacity cut $(\bar S, V^*\setminus \bar S)$ with $j,k\in \bar S$, a cluster $C_h \subseteq \bar S\setminus\{j,k\}$ and $D\subseteq V^*\setminus \bar S$. This algorithm is repeated for every target $j,k\in T$ and cluster $C_h$ such that $j,k \notin C_h$ and the target set $S = \bar S\setminus \{j,k\}$ obtained during each run of the algorithm defines a violated inequality if the capacity of the cut is strictly less than $\mathcal L + 1$. The time complexity of this algorithm is $O(m|T|^4)$. Similarly, the most violated constraint of the form can be obtained by computing a minimum capacity cut $(\bar S, V^*\setminus \bar S)$, with $i,j,k \in \bar S$ and $D\subseteq V^*\setminus \bar S$ on the graph $G^*$. This algorithm is repeated for very triplet of targets in $V^*$ and the set $S = \bar S\setminus \{j,k\}$ defines a violated inequality if the capacity of the cut is strictly less than $\mathcal L+y_i^*$. The time complexity of the algorithm is $O(|T|^5)$.
Similar to the separation of the sub-tour elimination constraints, we remark that the violated inequality of the form , computed using the above algorithm is not necessarily facet-defining as the set $S$ might have $\mu(S) \neq 0$. When this happens, we reject the inequality in favour of its dominating and facet-defining inequality in Eq. .
Separation of comb inequalities in Eq. {#subsec:comb}
---------------------------------------
For the comb-inequalities in Eq. , we use the separation procedures discussed in [@Fischetti1997]. We first consider the special case of the comb inequalities with $|\mathcal T_i|=2$ for $i=1,\dots,t$ *i.e.,* the 2-matching inequalities. Using a construction similar to the one proposed in [@Padberg1982] for the $b$-matching problem, the separation problem for the 2-matching inequalities can be transformed into a minimum capacity off cut problem; hence this separation problem is exactly solvable in polynomial time. But this procedure is computationally intensive, and so we use the following heuristic proposed by [@Grotschel1985]. Given a fractional solution $(\mathbf{x}^*, \mathbf{y}^*)$, the heuristic considers a graph $\bar G = (\bar V, \bar E)$ where $\bar V = V^* \cap T$ and $\bar E = \{e: 0<x_e^*<1\}$. Then, we consider each connected component $H$ of $\bar G$ as a handle of a possibly violated 2-matching inequality whose two-vertex teeth correspond to edges $e \in \delta(H)$ with $x_e^* = 1$. We reject the inequality if the number of teeth is even. The time complexity of this algorithm is $O(|\bar V| + |\bar E|)$. As for the comb inequalities, we apply the same procedure after shrinking each cluster into a single supernode.
Separation of T-comb inequalities in Eq. and {#subsec:Tcomb}
---------------------------------------------
We present a separation heuristic similar to the one used in [@Benavent2013] to identify violated T-comb inequalities of the form Eq. and . We first build a set of teeth, each containing a distinct depot according to the following procedure: a tooth $\mathcal T_i$ is built by starting with a set containing a depot $d\in D$; a target $v\in T$ is added to $\mathcal T_i$ such that $x(\delta(\mathcal T_i))$ is a minimum. Then, for every subset of this set of teeth such that: (i) they are pairwise disjoint, (ii) belong to the same connected component of the support graph $G^* = (V^*,E^*)$, and (iii) do not together contain all the targets of that connected component, an appropriate handle $H$ is built as follows: assume $H$ is the set of all the targets in the connected component and remove the targets in $H\setminus (\mathcal T_i\cup \dots \cup \mathcal T_t)$ sequentially. Every time a target is removed, the T-comb inequality of the appropriate form is checked for violation. The time complexity of this algorithm is $O(|T|)$.
Branch-and-cut algorithm \[sec:bandc\]
======================================
In this section, we describe important implementation details of the branch-and-cut algorithm for the GMDTSP. The algorithm is implemented within a CPLEX 12.4 framework using the CPLEX callback functions [@cplex124]. The callback functions in CPLEX enable the user to completely customize the branch-and-cut algorithm embedded into CPLEX, including the choice of node to explore in the enumeration tree, the choice of branching variable, the separation and the addition of user-defined cutting planes and the application of heuristic methods.
The lower bound at the root node of the enumeration tree is computed by solving the LP relaxation of the formulation in Sec. \[sec:Formulation\] that is further strengthened using the cutting planes described in Sec. \[sec:polyhedral\]. The initial linear program consisted of all constraints in -, except , and . For a given LP solution, we identify violated inequalities using the separation procedures detailed in Sec. \[sec:separation\] in the following order: (i) sub-tour elimination constraints in Eq. , (ii) sub-tour elimination constraints in Eq. (iii) path elimination constraints in Eq. , and, , (iv) generalized comb constraints in Eq. , and (v) T-comb constraints in Eq. and . This order of adding the constraints to the formulation was chosen after performing extensive computational experiments. Furthermore, we disabled the separation of all the cuts embedded into the CPLEX framework because enabling these cuts increased the average computation time for the instances. Once the new cuts generated using these separation procedures were added to the linear program, the tighter linear program was resolved. This procedure was iterated until either of the following conditions was satisfied: (i) no violated constraints could be generated by the separation procedures, (ii) the current lower bound of the enumeration tree was greater or equal to the current upper bound. If no constraints are generated in the separation phase, we create subproblems by branching on a fractional variable. First, we select a fractional $y_{i}$ variable, based on the *strong branching* rule ([@Achterberg2005]). If all these variables are integers, then we select a fractional $x_{e}$ variable using the same rule. As for the node-selection rule, we used the best-first policy for all our computations,*i.e.*, select the subproblem with the lowest objective value.
Preprocessing \[subsec:preprocessing\]
--------------------------------------
In this section, we detail a preprocessing algorithm that enables the reduction of size of the GMDTSP instances whose edge costs satisfy the triangle inequality *i.e.,* for distinct $i,j,k \in T$, $c_{ij} + c_{jk} \geq c_{ik}$. A similar algorithm is presented in [@Laporte1987; @Bektas2011] for the asymmetric generalized traveling salesman problem and generalized vehicle routing problem respectively. In a GMDTSP instance where the edge costs satisfy the triangle inequality, the optimal solution would visit exactly one target in each cluster. We utilize this structure of the optimal solution and reduce the size of a given GMDTSP instance, if possible. To that end, we define a target $i\in T$ to be *dominated* if there exits a target $j \in C_{h(i)}$, $j\neq i$ such that
1. $c_{pi} + c_{iq} \geq c_{pj} + c_{jq}$ for any $p,q \in T \setminus C_{h(i)}$,
2. $c_{di} \geq c_{dj}$ for all $d \in D$, and
3. $c_{di} + c_{ip} \geq c_{dj} + c_{jp}$ for any $d \in D, p\in T\setminus C_{h(i)}$.
\[prop:pp\] If a dominated target is removed from a GMDTSP instance satisfying triangle inequality, then the optimal cost to the instance does not change.
Let $i\in T$ be a dominated vertex. If the target $i$ is not visited in the optimal solution, then its removal does not change the optimal cost. So, assume that $i\in T$ is visited by the optimal solution. Since the edge costs of the instance satisfy the triangle inequality, exactly one target in each cluster is visited by the optimal solution. We now claim that it is possible to exchange the target $i$ with a target $j \in C_{h(i)}$ without increasing the cost of the optimal solution. This follows from the definition of a dominated target.
The preprocessing checks if a target is dominated and removes the target if it is found so. Then the other targets are checked for dominance relative to the reduced instance. The time complexity of the algorithm is $O(|T|^5)$.
LP rounding heuristic \[subsec:heuristic\]
------------------------------------------
We discuss an *LP-rounding* heuristic that aides to generate feasible solutions at the root node and to speed up the convergence of the branch-and-cut algorithm. The heuristic constructs a feasible GMDTSP solution from a given fractional LP solution. It is used only at the root node of the enumeration tree. The heuristic is based on a transformation method in [@Oberlin2009]. We are given $\mathbf{y}^*$, the vector of fractional $y_i$ values (denoted by $y_i^f$) for each target $i$. The algorithm proceeds as follows: for each cluster $C_k$ and every target $i\in C_k$, the heuristic sets the value of $y_i$ to $0$ or $1$ according to the condition $y_i^f \geq 0.5$ or $y_i^f < 0.5$ respectively. If every target $i \in C_k$ has $y_i^f < 0.5$, then we set the value of $y_j =1$ where $j = \operatorname{argmax}\{y_i^f:i\in C_k\}$. Once we have assigned the $y_i$ value for each target $i$, we define the set $\Pi := \{i\in T: y_i = 1\}$. We then solve a multiple depot traveling salesman problem (MDTSP) on the set of vertices $\Pi \cup D$. A heuristic based on the transformation method in [@Oberlin2009] and LKH heuristic (see [@Helsgaun2000]) is used to solve the MDTSP.
Computational results \[sec:results\]
=====================================
In this section, we discuss the computational results of the branch-and-cut algorithm. The algorithm was implemented in C++ (gcc version 4.6.3), using the elements of Standard Template Library (STL) in the CPLEX 12.4 framework. As mentioned in Sec. \[sec:bandc\], the internal CPLEX cut generation was disabled, and CPLEX was used only to mange the enumeration tree. All the simulations were performed on a Dell Precision T5500 workstation (Inter Xeon E5360 processor @2.53 GHz, 12 GB RAM). The computation times reported are expressed in seconds, and we imposed a time limit of 7200 seconds for each run of the algorithm. The performance of the algorithm was tested on a total of 116 instances, all of which were generated using the generalized traveling salesman problem library (see [@Fischetti1997; @Gutin2010]).
Problem instances \[subsec:instance\]
-------------------------------------
All the computational experiments were conducted on a class of 116 test instances generated from 29 GTSP instances. The GTSP instances are taken directly from the GTSP Instances Library (see [@Gutin2010]). The instances are available at <http://www.cs.nott.ac.uk/~dxk/gtsp.html>. For each of the 29 instances, GMDTSP instances with $|D| \in \{2,3,4,5\}$ were generated by assuming the first $|D|$ targets in a GTSP instance to be the set of depots; these depots were then removed from the target clusters. The number of targets in the instances varied from 14 to 105, and the maximum number of target clusters was 21. Hence we had 4 GMDTSP instances for each of the 29 GTSP instances totalling to 116 test instances. We also note that for 64/116 instances, the edge costs do not satisfy the triangle inequality and for the remaining 52 instances, the edge costs satisfy the triangle inequality. The name of the generated instances are the same but for a small modification to spell out the number of depots in the instances. The naming conforms to the format `GTSPinstancename-D`, where `GTSPinstancename` corresponds to the GTSP instance name from the library (the first and the last integer in the name corresponds to the number of clusters and the number of targets in the GTSP instance respectively) and `D` corresponds the number of depots in the instance.
The results are tabulated in tables \[tab:results\] and \[tab:times\]. The following nomenclature is used in the table \[tab:results\]
**name**: problem instance name (format: `GTSPinstancename-D`);\
**opt**: optimal objective value;\
**LB**: objective value of the LP relaxation computed at the root node of the enumeration tree;\
**%LB**: percentage LB/opt;\
**UB**: cost of the best feasible solution generated by the LP-rounding heuristic generated at the root node of the enumeration tree;\
**%UB**: percentage UB/opt;\
**sec1**: total number of constraints generated;\
**sec2**: total number of constraints generated;\
**4pec**: total number of constraints generated;\
**pec**: total number of constraints and generated;\
**comb**: total number of constraints , , and generated;\
**nodes**: total number of nodes examined in the enumeration tree.\
The table \[tab:times\] gives the computational time for each separation routine and the overall the branch-and-cut algorithm. The nomenclature used in table \[tab:times\] are as follows:
**name**: problem instance name (format: `GTSPinstancename-D`);\
**total-t**: CPU time, in seconds, for the overall execution of the branch-and-cut algorithm;\
**sep-t**: overall CPU time, in seconds, spent for separation;\
**sec-t**: CPU time, in seconds, spent for the separation of constraints and ;\
**4pec-t**: CPU time, in seconds, spent for the separation of constraints ;\
**pec-t**: CPU time, in seconds, spent for the separation of constraints and ;\
**comb-t**: CPU time, in seconds, spent for the separation of constraints , , and ;\
**%pec**: percentage of separation time spent for the separation of path elimination constraints and .\
[lrrrrrR[0.7cm]{}R[0.7cm]{}R[0.7cm]{}R[0.8cm]{}R[0.7cm]{}r]{}
\
name & opt & LB & %LB & UB & %UB & sec1 & sec2 & 4pec & pec & comb & nodes
\
name & opt & LB & %LB & UB & %UB & sec1 & sec2 & 4pec & pec & comb & nodes
3burma14-2 & 1939 & 1939.00 & 100.00 & 1939 & 100.00 & 51 & 8 & 0 & 2 & 0 & 03burma14-3 & 1664 & 1664.00 & 100.00 & 1664 & 100.00 & 11 & 15 & 0 & 2 & 0 & 03burma14-4 & 1296 & 1296.00 & 100.00 & 1296 & 100.00 & 8 & 14 & 0 & 0 & 0 & 03burma14-5 & 562 & 562.00 & 100.00 & 562 & 100.00 & 1 & 20 & 0 & 0 & 0 & 04br17-2 & 31 & 31.00 & 100.00 & 54 & 174.19 & 7 & 4 & 0 & 0 & 1 & 34br17-3 & 31 & 31.00 & 100.00 & 31 & 100.00 & 7 & 7 & 0 & 0 & 0 & 04br17-4 & 19 & 19.00 & 100.00 & 19 & 100.00 & 5 & 14 & 0 & 0 & 0 & 04br17-5 & 19 & 19.00 & 100.00 & 19 & 100.00 & 5 & 20 & 0 & 4 & 0 & 04gr17-2 & 958 & 846.33 & 88.34 & 965 & 100.73 & 22 & 187 & 8 & 335 & 0 & 974gr17-3 & 738 & 722.88 & 97.95 & 794 & 107.59 & 3 & 43 & 1 & 53 & 4 & 64gr17-4 & 611 & 611.00 & 100.00 & 611 & 100.00 & 2 & 14 & 0 & 3 & 0 & 04gr17-5 & 513 & 513.00 & 100.00 & 513 & 100.00 & 1 & 25 & 0 & 0 & 0 & 04ulysses16-2 & 4695 & 4695.00 & 100.00 & 4695 & 100.00 & 36 & 18 & 0 & 0 & 0 & 04ulysses16-3 & 4695 & 4695.00 & 100.00 & 4695 & 100.00 & 53 & 20 & 0 & 0 & 0 & 04ulysses16-4 & 4695 & 4695.00 & 100.00 & 4695 & 100.00 & 50 & 27 & 0 & 0 & 0 & 04ulysses16-5 & 3914 & 3884.00 & 99.23 & 4188 & 107.00 & 22 & 27 & 0 & 7 & 0 & 35gr21-2 & 1679 & 1531.67 & 91.22 & 1985 & 118.23 & 419 & 367 & 12 & 2158 & 0 & 4495gr21-3 & 1024 & 1024.00 & 100.00 & 1024 & 100.00 & 6 & 32 & 0 & 2 & 0 & 05gr21-4 & 953 & 953.00 & 100.00 & 953 & 100.00 & 9 & 20 & 0 & 1 & 0 & 05gr21-5 & 780 & 780.00 & 100.00 & 780 & 100.00 & 4 & 9 & 0 & 2 & 0 & 05gr24-2 & 377 & 340.53 & 90.33 & 828 & 219.63 & 25 & 169 & 0 & 366 & 0 & 135gr24-3 & 377 & 318.00 & 84.35 & 569 & 150.93 & 37 & 181 & 0 & 524 & 32 & 425gr24-4 & 371 & 325.17 & 87.65 & 753 & 202.96 & 39 & 157 & 8 & 303 & 6 & 265gr24-5 & 362 & 308.17 & 85.13 & 739 & 204.14 & 12 & 99 & 7 & 222 & 0 & 875ulysses22-2 & 5199 & 5199.00 & 100.00 & 5199 & 100.00 & 70 & 71 & 2 & 126 & 1 & 05ulysses22-3 & 5311 & 5310.50 & 99.99 & 5442 & 102.47 & 45 & 82 & 0 & 1 & 0 & 35ulysses22-4 & 5021 & 5021.00 & 100.00 & 5021 & 100.00 & 45 & 39 & 0 & 0 & 0 & 05ulysses22-5 & 3913 & 3913.00 & 100.00 & 3913 & 100.00 & 37 & 27 & 0 & 1 & 0 & 06bayg29-2 & 711 & 624.50 & 87.83 & 905 & 127.29 & 82 & 312 & 0 & 1526 & 0 & 1486bayg29-3 & 684 & 582.50 & 85.16 & 841 & 122.95 & 70 & 809 & 3 & 3489 & 28 & 3016bayg29-4 & 583 & 527.50 & 90.48 & 811 & 139.11 & 25 & 91 & 0 & 171 & 7 & 246bayg29-5 & 565 & 520.79 & 92.17 & 1888 & 334.16 & 40 & 103 & 0 & 360 & 6 & 216bays29-2 & 849 & 761.46 & 89.69 & 1194 & 140.64 & 123 & 178 & 0 & 1466 & 0 & 2966bays29-3 & 830 & 777.68 & 93.70 & 1092 & 131.57 & 80 & 145 & 1 & 959 & 17 & 486bays29-4 & 691 & 650.60 & 94.15 & 847 & 122.58 & 30 & 92 & 3 & 238 & 20 & 66bays29-5 & 622 & 591.55 & 95.10 & 1052 & 169.13 & 30 & 99 & 1 & 258 & 3 & 106fri26-2 & 480 & 471.50 & 98.23 & 541 & 112.71 & 54 & 184 & 1 & 519 & 0 & 156fri26-3 & 486 & 466.00 & 95.88 & 510 & 104.94 & 167 & 166 & 0 & 1923 & 3 & 3886fri26-4 & 440 & 414.57 & 94.22 & 446 & 101.36 & 92 & 128 & 0 & 355 & 9 & 386fri26-5 & 436 & 411.56 & 94.39 & 473 & 108.49 & 66 & 91 & 2 & 520 & 2 & 419dantzig42-2 & 413 & 413.00 & 100.00 & 413 & 100.00 & 114 & 300 & 0 & 0 & 0 & 09dantzig42-3 & 351 & 351.00 & 100.00 & 358 & 101.99 & 82 & 328 & 0 & 10 & 1 & 39dantzig42-4 & 350 & 345.75 & 98.79 & 396 & 113.14 & 81 & 272 & 1 & 442 & 33 & 69dantzig42-5 & 348 & 344.29 & 98.93 & 348 & 100.00 & 82 & 203 & 2 & 346 & 45 & 1210att48-2 & 4924 & 4284.05 & 87.00 & 5510 & 111.90 & 456 & 945 & 0 & 7563 & 0 & 26810att48-3 & 4913 & 4539.33 & 92.39 & 6054 & 123.22 & 177 & 880 & 8 & 10115 & 154 & 140610att48-4 & 4428 & 3980.11 & 89.89 & 5685 & 128.39 & 197 & 738 & 2 & 8555 & 138 & 87910att48-5 & 4204 & 3897.97 & 92.72 & 5515 & 131.18 & 87 & 690 & 9 & 12826 & 1077 & 59410gr48-2 & 1708 & 1707.00 & 99.94 & 1708 & 100.00 & 88 & 186 & 1 & 259 & 0 & 210gr48-3 & 1638 & 1628.14 & 99.40 & 2345 & 143.16 & 74 & 220 & 4 & 1011 & 0 & 1410gr48-4 & 1645 & 1629.23 & 99.04 & 2197 & 133.56 & 86 & 185 & 0 & 958 & 1 & 3310gr48-5 & 1638 & 1471.48 & 89.83 & 2243 & 136.94 & 108 & 405 & 5 & 2163 & 30 & 17910hk48-2 & 6401 & 6209.83 & 97.01 & 6753 & 105.50 & 357 & 418 & 7 & 3018 & 0 & 8210hk48-3 & 5872 & 5567.49 & 94.81 & 6211 & 105.77 & 234 & 364 & 1 & 2549 & 0 & 7510hk48-4 & 5642 & 5044.00 & 89.40 & 6359 & 112.71 & 269 & 474 & 1 & 2370 & 3 & 6910hk48-5 & 5641 & 5145.17 & 91.21 & 6702 & 118.81 & 282 & 399 & 0 & 3455 & 14 & 2711berlin52-2 & 3500 & 3425.00 & 97.86 & 4010 & 114.57 & 121 & 288 & 0 & 1 & 1 & 1711berlin52-3 & 3500 & 3376.17 & 96.46 & 3963 & 113.23 & 142 & 311 & 1 & 753 & 66 & 2011berlin52-4 & 3500 & 3280.00 & 93.71 & 3699 & 105.69 & 88 & 241 & 1 & 426 & 3 & 2511berlin52-5 & 3500 & 3273.92 & 93.54 & 4169 & 119.11 & 131 & 160 & 0 & 599 & 26 & 2611eil51-2 & 175 & 174.50 & 99.71 & 175 & 100.00 & 148 & 522 & 2 & 1071 & 0 & 311eil51-3 & 174 & 168.83 & 97.03 & 174 & 100.00 & 138 & 269 & 3 & 1160 & 54 & 1111eil51-4 & 175 & 165.24 & 94.42 & 183 & 104.57 & 175 & 273 & 11 & 1837 & 18 & 7411eil51-5 & 170 & 166.44 & 97.91 & 170 & 100.00 & 71 & 214 & 2 & 479 & 6 & 812brazil58-2 & 14939 & 14939.00 & 100.00 & 14939 & 100.00 & 141 & 278 & 3 & 834 & 0 & 012brazil58-3 & 14930 & 14840.50 & 99.40 & 15240 & 102.08 & 140 & 298 & 1 & 967 & 57 & 1812brazil58-4 & 13082 & 12680.46 & 96.93 & 16148 & 123.44 & 147 & 397 & 1 & 1447 & 126 & 4012brazil58-5 & 12613 & 11958.93 & 94.81 & 15546 & 123.25 & 153 & 1049 & 1 & 583 & 50 & 9814st70-2 & 304 & 288.01 & 94.74 & 307 & 100.99 & 392 & 576 & 2 & 3147 & 3 & 8114st70-3 & 301 & 292.57 & 97.20 & 312 & 103.65 & 313 & 600 & 6 & 2846 & 12 & 1714st70-4 & 298 & 287.25 & 96.39 & 298 & 100.00 & 182 & 372 & 4 & 1404 & 4 & 1914st70-5 & 298 & 282.28 & 94.73 & 325 & 109.06 & 313 & 670 & 9 & 3883 & 5 & 16316eil76-2 & 198 & 198.00 & 100.00 & 198 & 100.00 & 223 & 436 & 0 & 945 & 0 & 016eil76-3 & 197 & 197.00 & 100.00 & 197 & 100.00 & 174 & 258 & 3 & 727 & 6 & 016eil76-4 & 197 & 197.00 & 100.00 & 197 & 100.00 & 147 & 360 & 4 & 941 & 20 & 016eil76-5 & 188 & 180.42 & 95.97 & 196 & 104.26 & 233 & 386 & 5 & 1132 & 25 & 2720gr96-2$^\dagger$ & 29966 & 28357.03 & 94.63 & 30821 & 102.85 & 823 & 1220 & 1 & 3540 & 0 & 6220gr96-3$^\dagger$ & 29621 & 29263.93 & 98.79 & 30768 & 103.87 & 876 & 1326 & 2 & 3382 & 529 & 5020gr96-4 & 28705 & 27650.63 & 96.33 & 30121 & 104.93 & 866 & 1754 & 6 & 4268 & 7 & 14420gr96-5 & 28598 & 27768.50 & 97.10 & 29976 & 104.82 & 676 & 1269 & 1 & 2087 & 1 & 5220kroA100-2 & 9630 & 9265.75 & 96.22 & 9769 & 101.44 & 746 & 1080 & 5 & 3481 & 0 & 6620kroA100-3 & 9334 & 8935.25 & 95.73 & 9535 & 102.15 & 532 & 915 & 0 & 2801 & 0 & 9220kroA100-4 & 8897 & 8539.03 & 95.98 & 10243 & 115.13 & 935 & 1241 & 2 & 4490 & 0 & 12620kroA100-5 & 8827 & 8477.39 & 96.04 & 9020 & 102.19 & 520 & 1028 & 4 & 2480 & 0 & 4720kroB100-2 & 9800 & 9492.00 & 96.86 & 10382 & 105.94 & 510 & 955 & 4 & 3025 & 0 & 3020kroB100-3$^\dagger$ & 10218 & 9197.41 & 90.01 & 10300 & 100.80 & 903 & 1120 & 1 & 5373 & 0 & 13020kroB100-4 & 9564 & 9293.31 & 97.17 & 9637 & 100.76 & 361 & 714 & 0 & 2323 & 0 & 2020kroB100-5 & 9226 & 8525.71 & 92.41 & 11708 & 126.90 & 739 & 1058 & 10 & 7225 & 0 & 11920kroC100-2$^\dagger$ & 10089 & 9548.13 & 94.64 & 10089 & 100.00 & 420 & 974 & 0 & 1551 & 0 & 320kroC100-3 & 9244 & 9130.82 & 98.78 & 9346 & 101.10 & 494 & 1006 & 0 & 1940 & 1 & 820kroC100-4 & 9292 & 9061.20 & 97.52 & 9342 & 100.54 & 307 & 707 & 2 & 1132 & 3 & 1020kroC100-5 & 9252 & 8991.89 & 97.19 & 10437 & 112.81 & 380 & 956 & 3 & 2181 & 0 & 1920kroD100-2$^\dagger$ & 9353 & 8497.63 & 90.85 & 9381 & 100.30 & 886 & 1525 & 4 & 3221 & 6 & 6520kroD100-3 & 8813 & 8130.12 & 92.25 & 11404 & 129.40 & 1284 & 1664 & 5 & 11642 & 24 & 21220kroD100-4 & 8772 & 8283.74 & 94.43 & 8823 & 100.58 & 577 & 1067 & 11 & 3230 & 3 & 6720kroD100-5 & 8677 & 8233.85 & 94.89 & 9247 & 106.57 & 478 & 732 & 1 & 3277 & 0 & 4520kroE100-2 & 9526 & 9290.65 & 97.53 & 10207 & 107.15 & 599 & 1098 & 7 & 4461 & 0 & 4520kroE100-3 & 9262 & 9153.61 & 98.83 & 9854 & 106.39 & 612 & 1048 & 7 & 3974 & 19 & 2620kroE100-4 & 9262 & 9147.56 & 98.76 & 11046 & 119.26 & 513 & 1032 & 3 & 3410 & 4 & 2120kroE100-5 & 9081 & 8900.07 & 98.01 & 9707 & 106.89 & 391 & 925 & 3 & 2802 & 0 & 3220rat99-2 & 505 & 504.33 & 99.87 & 521 & 103.17 & 507 & 951 & 0 & 0 & 0 & 720rat99-3 & 504 & 498.23 & 98.85 & 543 & 107.74 & 528 & 977 & 4 & 1582 & 1 & 2020rat99-4 & 501 & 490.67 & 97.94 & 515 & 102.79 & 958 & 1259 & 5 & 10214 & 0 & 238320rat99-5 & 487 & 477.67 & 98.08 & 506 & 103.90 & 688 & 967 & 4 & 4320 & 0 & 37620rd100-2$^\dagger$ & 3459 & 3380.39 & 97.73 & 3714 & 107.37 & 742 & 1406 & 0 & 4119 & 0 & 4220rd100-3 & 3383 & 3218.89 & 95.15 & 3384 & 100.03 & 657 & 1456 & 2 & 4238 & 1 & 5520rd100-4 & 3298 & 3167.38 & 96.04 & 3398 & 103.03 & 530 & 889 & 2 & 2651 & 0 & 2920rd100-5 & 3234 & 3109.99 & 96.17 & 3327 & 102.88 & 559 & 1056 & 6 & 4114 & 1 & 6421eil101-2 & 248 & 245.41 & 98.96 & 255 & 102.82 & 387 & 812 & 0 & 1476 & 0 & 2021eil101-3 & 248 & 243.04 & 98.00 & 267 & 107.66 & 570 & 982 & 4 & 2371 & 6 & 3721eil101-4 & 233 & 230.2759 & 98.83 & 251 & 107.73 & 432 & 629 & 3 & 2586 & 0 & 1521eil101-5 & 232 & 226.33 & 97.56 & 257 & 110.78 & 275 & 527 & 0 & 1483 & 2 & 1621lin105-2 & 8358 & 8316.43 & 99.50 & 8726 & 104.40 & 652 & 1122 & 0 & 0 & 0 & 1621lin105-3$^\dagger$ & 8304 & 8164.21 & 98.32 & 8619 & 103.79 & 870 & 1298 & 3 & 25572 & 22 & 710321lin105-4 & 7827 & 7695.17 & 98.32 & 8365 & 106.87 & 619 & 941 & 2 & 888 & 12 & 8921lin105-5$^\dagger$ & 8052 & 7568.64 & 94.00 & 8110 & 100.72 & 745 & 1166 & 1 & 2419 & 6 & 145
[lR[1.1cm]{}R[1cm]{}R[1cm]{}R[1cm]{}R[1cm]{}R[1.2cm]{}r]{}
\
name & total-t & sep-t & sec-t & 4pec-t & pec-t & comb-t & %pec
\
name & total-t & sep-t & sec-t & 4pec-t & pec-t & comb-t & %pec
3burma14-2 & 0.07 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 3.133burma14-3 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 2.683burma14-4 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 1.973burma14-5 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 3.504br17-2 & 0.03 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 1.144br17-3 & 0.01 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.004br17-4 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.004br17-5 & 0.04 & 0.01 & 0.00 & 0.00 & 0.00 & 0.00 & 68.524gr17-2 & 1.16 & 0.33 & 0.10 & 0.00 & 0.22 & 0.01 & 65.714gr17-3 & 0.23 & 0.05 & 0.01 & 0.00 & 0.04 & 0.00 & 74.034gr17-4 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.004gr17-5 & 0.01 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.004ulysses16-2 & 0.05 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 1.714ulysses16-3 & 0.05 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 2.044ulysses16-4 & 0.08 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 1.934ulysses16-5 & 0.13 & 0.02 & 0.01 & 0.00 & 0.02 & 0.00 & 72.635gr21-2 & 12.89 & 3.63 & 1.00 & 0.00 & 2.54 & 0.09 & 69.985gr21-3 & 0.04 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 2.285gr21-4 & 0.02 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 2.865gr21-5 & 0.07 & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 2.815gr24-2 & 1.81 & 0.45 & 0.07 & 0.00 & 0.38 & 0.00 & 84.825gr24-3 & 3.51 & 0.92 & 0.18 & 0.00 & 0.73 & 0.01 & 79.175gr24-4 & 2.89 & 0.76 & 0.11 & 0.00 & 0.64 & 0.01 & 83.805gr24-5 & 1.63 & 0.38 & 0.12 & 0.00 & 0.25 & 0.01 & 65.265ulysses22-2 & 0.77 & 0.18 & 0.04 & 0.00 & 0.13 & 0.00 & 74.265ulysses22-3 & 0.43 & 0.03 & 0.03 & 0.00 & 0.00 & 0.00 & 0.645ulysses22-4 & 0.18 & 0.02 & 0.02 & 0.00 & 0.00 & 0.00 & 0.755ulysses22-5 & 0.06 & 0.01 & 0.01 & 0.00 & 0.00 & 0.00 & 1.826bayg29-2 & 18.69 & 4.97 & 0.73 & 0.00 & 4.17 & 0.08 & 83.796bayg29-3 & 20.50 & 5.66 & 1.31 & 0.00 & 4.19 & 0.15 & 74.106bayg29-4 & 1.26 & 0.31 & 0.06 & 0.00 & 0.24 & 0.01 & 77.326bayg29-5 & 1.19 & 0.27 & 0.08 & 0.00 & 0.18 & 0.01 & 68.116bays29-2 & 21.40 & 6.19 & 0.96 & 0.00 & 5.14 & 0.08 & 83.166bays29-3 & 10.60 & 2.78 & 0.33 & 0.00 & 2.43 & 0.02 & 87.506bays29-4 & 1.22 & 0.30 & 0.05 & 0.00 & 0.24 & 0.01 & 80.746bays29-5 & 0.97 & 0.22 & 0.04 & 0.00 & 0.18 & 0.00 & 79.986fri26-2 & 5.55 & 1.34 & 0.12 & 0.00 & 1.22 & 0.01 & 90.536fri26-3 & 18.32 & 5.55 & 1.11 & 0.00 & 4.31 & 0.13 & 77.686fri26-4 & 3.75 & 0.92 & 0.12 & 0.00 & 0.78 & 0.01 & 85.236fri26-5 & 3.26 & 0.83 & 0.12 & 0.00 & 0.70 & 0.01 & 84.679dantzig42-2 & 1.07 & 0.28 & 0.27 & 0.00 & 0.00 & 0.01 & 0.389dantzig42-3 & 1.26 & 0.34 & 0.16 & 0.00 & 0.18 & 0.00 & 51.779dantzig42-4 & 5.15 & 1.29 & 0.22 & 0.00 & 1.05 & 0.01 & 81.819dantzig42-5 & 7.97 & 1.93 & 0.20 & 0.00 & 1.71 & 0.01 & 88.7110att48-2 & 280.75 & 80.02 & 6.73 & 0.00 & 72.88 & 0.41 & 91.0810att48-3 & 243.27 & 71.62 & 9.29 & 0.00 & 60.66 & 1.67 & 84.7010att48-4 & 203.20 & 59.39 & 7.56 & 0.00 & 50.63 & 1.19 & 85.2610att48-5 & 130.36 & 38.93 & 5.95 & 0.00 & 31.74 & 1.23 & 81.5510gr48-2 & 9.25 & 2.26 & 0.21 & 0.00 & 2.04 & 0.01 & 90.5010gr48-3 & 31.81 & 7.87 & 0.54 & 0.00 & 7.30 & 0.03 & 92.7210gr48-4 & 39.36 & 9.62 & 0.60 & 0.00 & 8.96 & 0.06 & 93.1010gr48-5 & 43.79 & 11.76 & 1.39 & 0.00 & 10.17 & 0.20 & 86.4810hk48-2 & 273.81 & 69.58 & 3.29 & 0.00 & 66.15 & 0.14 & 95.0710hk48-3 & 170.99 & 43.05 & 1.76 & 0.00 & 41.19 & 0.10 & 95.6610hk48-4 & 35.98 & 9.64 & 1.04 & 0.00 & 8.51 & 0.09 & 88.2810hk48-5 & 92.75 & 24.49 & 1.57 & 0.00 & 22.84 & 0.08 & 93.2711berlin52-2 & 2.28 & 1.06 & 1.03 & 0.00 & 0.00 & 0.02 & 0.3711berlin52-3 & 67.95 & 16.48 & 0.95 & 0.00 & 15.48 & 0.05 & 93.9111berlin52-4 & 27.96 & 7.19 & 0.44 & 0.00 & 6.72 & 0.04 & 93.4111berlin52-5 & 19.57 & 5.17 & 0.46 & 0.00 & 4.66 & 0.05 & 90.1611eil51-2 & 200.63 & 48.72 & 1.39 & 0.00 & 47.29 & 0.03 & 97.0811eil51-3 & 100.95 & 24.48 & 0.98 & 0.00 & 23.47 & 0.03 & 95.8511eil51-4 & 142.50 & 37.00 & 1.94 & 0.00 & 34.95 & 0.11 & 94.4511eil51-5 & 33.19 & 8.25 & 0.36 & 0.00 & 7.87 & 0.02 & 95.4212brazil58-2 & 33.00 & 7.94 & 0.96 & 0.00 & 6.95 & 0.03 & 87.5112brazil58-3 & 56.51 & 13.29 & 0.93 & 0.00 & 12.31 & 0.06 & 92.6012brazil58-4 & 32.61 & 8.62 & 1.00 & 0.00 & 7.53 & 0.09 & 87.3512brazil58-5 & 3.48 & 1.06 & 0.52 & 0.00 & 0.44 & 0.10 & 41.5514st70-2 & 876.36 & 222.60 & 6.73 & 0.00 & 215.47 & 0.39 & 96.8014st70-3 & 1071.01 & 264.38 & 4.16 & 0.00 & 260.10 & 0.12 & 98.3814st70-4 & 354.16 & 87.56 & 1.86 & 0.00 & 85.61 & 0.08 & 97.7814st70-5 & 429.46 & 113.03 & 5.51 & 0.00 & 106.96 & 0.57 & 94.6316eil76-2 & 160.97 & 38.04 & 1.72 & 0.00 & 36.27 & 0.04 & 95.3616eil76-3 & 71.48 & 17.47 & 0.80 & 0.00 & 16.64 & 0.03 & 95.2416eil76-4 & 173.67 & 43.19 & 1.11 & 0.00 & 42.03 & 0.05 & 97.3116eil76-5 & 274.12 & 69.50 & 1.87 & 0.00 & 67.52 & 0.12 & 97.1520gr96-2$^\dagger$ & 7200.00 & 1901.87 & 44.02 & 0.00 & 1857.29 & 0.56 & 97.6620gr96-3$^\dagger$ & 7200.00 & 1862.37 & 38.38 & 0.00 & 1823.22 & 0.77 & 97.9020gr96-4 & 5467.42 & 1428.08 & 48.45 & 0.00 & 1378.35 & 1.28 & 96.5220gr96-5 & 6495.00 & 1643.50 & 35.00 & 0.00 & 1607.79 & 0.71 & 97.8320kroA100-2 & 4291.87 & 1091.52 & 22.62 & 0.00 & 1068.47 & 0.42 & 97.8920kroA100-3 & 4225.89 & 1060.29 & 14.82 & 0.00 & 1044.91 & 0.56 & 98.5520kroA100-4 & 5057.47 & 1300.82 & 28.19 & 0.00 & 1271.60 & 1.04 & 97.7520kroA100-5 & 6368.98 & 1606.81 & 20.13 & 0.00 & 1585.98 & 0.70 & 98.7020kroB100-2 & 3389.43 & 841.28 & 12.24 & 0.00 & 828.79 & 0.25 & 98.5220kroB100-3$^\dagger$ & 7200.04 & 1838.03 & 33.81 & 0.00 & 1803.14 & 1.08 & 98.1020kroB100-4 & 3120.43 & 778.88 & 9.44 & 0.00 & 769.15 & 0.29 & 98.7520kroB100-5 & 3397.49 & 883.26 & 24.75 & 0.00 & 857.50 & 1.01 & 97.0820kroC100-2$^\dagger$ & 7200.00 & 1821.34 & 15.18 & 0.00 & 1805.91 & 0.25 & 99.1520kroC100-3 & 3052.62 & 747.14 & 10.82 & 0.00 & 736.09 & 0.23 & 98.5220kroC100-4 & 1009.37 & 250.86 & 4.82 & 0.00 & 245.88 & 0.16 & 98.0120kroC100-5 & 2839.31 & 713.70 & 11.93 & 0.00 & 701.39 & 0.38 & 98.2820kroD100-2$^\dagger$ & 7200.00 & 1852.91 & 33.91 & 0.00 & 1818.46 & 0.54 & 98.1420kroD100-3 & 6287.9 & 1671.43 & 50.47 & 0.00 & 1619.66 & 1.30 & 96.9020kroD100-4 & 4716.98 & 1190.26 & 18.79 & 0.00 & 1170.92 & 0.55 & 98.3820kroD100-5 & 2669.25 & 671.32 & 13.10 & 0.00 & 657.78 & 0.44 & 97.9820kroE100-2 & 4718.14 & 1204.19 & 24.14 & 0.00 & 1179.63 & 0.41 & 97.9620kroE100-3 & 4737.91 & 1147.37 & 24.29 & 0.00 & 1122.59 & 0.49 & 97.8420kroE100-4 & 2624.53 & 641.08 & 17.04 & 0.00 & 623.69 & 0.35 & 97.2920kroE100-5 & 1892.52 & 476.91 & 10.32 & 0.00 & 466.24 & 0.35 & 97.7620rat99-2 & 65.57 & 12.65 & 12.55 & 0.00 & 0.02 & 0.09 & 0.1520rat99-3 & 2416.98 & 583.46 & 14.15 & 0.00 & 569.01 & 0.30 & 97.5220rat99-4 & 6091.56 & 1414.13 & 140.03 & 0.00 & 1245.85 & 28.26 & 88.1020rat99-5 & 3165.79 & 747.76 & 46.84 & 0.00 & 693.47 & 7.45 & 92.7420rd100-2$^\dagger$ & 7200.00 & 1846.05 & 37.12 & 0.00 & 1808.40 & 0.52 & 97.9620rd100-3 & 3815.24 & 969.42 & 23.26 & 0.00 & 945.69 & 0.47 & 97.5520rd100-4 & 3273.97 & 826.82 & 16.76 & 0.00 & 809.60 & 0.46 & 97.9220rd100-5 & 2513.41 & 643.81 & 15.04 & 0.00 & 628.22 & 0.55 & 97.5821eil101-2 & 2100.39 & 519.56 & 10.63 & 0.00 & 508.75 & 0.19 & 97.9221eil101-3 & 4245.95 & 1069.99 & 18.31 & 0.00 & 1051.25 & 0.43 & 98.2521eil101-4 & 906.82 & 227.88 & 7.48 & 0.00 & 220.15 & 0.25 & 96.6121eil101-5 & 682.82 & 172.40 & 4.07 & 0.00 & 168.13 & 0.19 & 97.5221lin105-2 & 86.33 & 21.14 & 20.93 & 0.00 & 0.03 & 0.18 & 0.1521lin105-3$^\dagger$ & 7200.00 & 2047.88 & 380.14 & 0.00 & 1566.72 & 101.02 & 76.5021lin105-4 & 3609.22 & 903.74 & 19.51 & 0.00 & 883.49 & 0.74 & 97.7621lin105-5$^\dagger$ & 7200.00 & 1890.67 & 45.87 & 0.00 & 1843.24 & 1.56 & 97.49
The results indicate that the proposed branch-and-cut algorithm can solve instances involving up to 105 targets with modest computation times. The preprocessing algorithm in Sec. \[subsec:preprocessing\] was applied to 53/116 instances. The time taken by the preprocessing algorithm is not included in the overall computation time. The preprocessing algorithm reduced the size of these instances by 6 targets on average and the maximum reduction obtained was 14 targets. We observe that the instances that have a larger number of violated path elimination constraints take considerably large amount of computation time. The last column in table \[tab:times\], whose average is 73%, indicates the percentage of separation time spent for finding violated path elimination constraints. This is not surprising because the time complexity for identifying violated path elimination constraints in and given a fractional solution, is $O(|T|^5)$ and $O(m|T|^4)$ respectively. The average number of T-comb inequalities that were generated in the enumeration tree were larger for some of the bigger instances (see table \[tab:results\]). They were effective, especially in tightening the lower bound for the instances that were not solved to optimality; for the instances where violated T-comb inequalities were separated out, the average linear programming relaxation gap improvement was 18%. They were also useful in reducing the computation times for larger instances despite increasing the computation times for smaller instances. Overall, we were able to solve 108/116 instances to the optimality with the largest instance involving 105 targets, 21 clusters and 5 depots. The “**opt**” column for the remaining 6/116 instances is the cost of the best feasible solution obtained by the branch-and-cut algorithm at the end of 7200 seconds. For the instances not solved to optimality within the time limit of 7200 seconds, the LP-rounding heuristic was effective in generating feasible solutions within 2.1% of the best feasible solution, on average.
Conclusion \[sec:conc\]
=======================
In summary, we have presented an exact algorithm for the GMDTSP, a problem that has several practical applications including maritime transportation, health-care logistics, survivable telecommunication network design, and routing unmanned vehicles to name a few. A mixed-integer linear programming formulation including several classes of valid inequalities was proposed the facial structure of the polytope of feasible solutions was studied in detail. All the results were used to develop a branch-and-cut algorithm whose performance was corroborated through extensive numerical experiments on a wide range of benchmark instances from the standard library. The largest solved instance involved 105 targets, 21 clusters and 4 depots. Future work can be directed towards development of branch-and-cut approaches accompanied with a polyhedral study to solve the asymmetric counterpart of the problem.
{#section .unnumbered}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We argue that the non-adiabatic evolution of spin states in Stern-Gerlach apparatus can blur the manifestation of path spin entanglement. This fact questions the usual practice of spin measurement even in a formally and operationally ideal situation. Considering azimuthal inhomogeneity, we have identified through quantitative calculation the specific reason behind the breakdown of adiabatic evolution to be the spatial inhomogeneity of applied magnetic field. The angle $\theta$ between the $z$ component of magnetic moment of a particle and direction of applied magnetic field is also an important factor in determining the category of evolution of spin states. Adiabaticity always can be restored by choosing a sufficiently small value of $\theta$.'
address:
- '$^1$Theoretical Physics Division, Variable Energy Cyclotron Centre, 1/AF Bidhan Nagar, Kolkata 700 064, India'
- '$^2$S. N. Bose National Centre For basic Sciences, Saltlake, Kolkata-700098, India'
author:
- 'Nirupam Dutta $^1$, Ansuman Dey $^2$'
bibliography:
- 'entanglement.bib'
title: |
Blurred path-spin entanglement in Stern-Gerlach apparatus:\
interplay between magnetic inhomogeneity and Larmor precession
---
Introduction
============
The Stern-Gerlach experiment (SGE)[@gerlach1922] plays a fundamental role in quantum mechanics because of its conceptual relevance. SGE first witnessed the quantum nature of intrinsic spin of a particle. Besides realizing the existence of spin angular momentum and its space quantization, the experiment became a formal ground to study decoherence and quantum non locality like entanglement [@reinisch1999; @englert1988; @einstein1922]. Different aspects of SGE are in general quite complicated and are still being studied by present day researchers [@hsu2011] . In the conventional description of the experiment, a beam of particles is passed through a Stern-Gerlach(SG) apparatus which consists of a magnet providing an inhomogeneous magnetic field, in a certain direction(z, say). When a screen is put at a distance from where the particles are emerging from the SG apparatus, two distinct peaks are observed in positive and negative $z$ directions corresponding to two different spin values along $z$ axis. This clearly demonstrates a path spin entanglement. More precisely, the eigenstates of $S_{z}$ of a charge neutral particle entering the SG apparatus are entangled with its spatial degrees of freedom. The observed peaks on the screen always lead to the detection of up spin in the upward ($z>0$) and down spin in the downward ($z<0$) direction. This statement is not only true for the measurement of spin along $z$ axis, but is also valid when the spin is measured along any arbitrary direction $\hat{u}$ along which the magnetic field is applied. The magnetic field spatially separates two different spin states $|+\rangle_{u}$ and $|-\rangle_{u}$ and they evolve being coupled with two different parts of the wave function. The final state becomes $|\psi \rangle =c_+\phi_{+}({\bf x}) \otimes |+\rangle_{u} + c_{-}\phi_{-}({\bf x})\otimes|-\rangle_{u}$ with $c_{+}$, $c_{-}$ complex constants. Here, $\phi_{+}({\bf x})$, $\phi_{-}({\bf x})$ are the spatial parts in upward and downward directions of $u=0$ plane while $|+\rangle_{u}$, $|-\rangle_{u}$ are the eigenstates of $\hat{S_{u}}$. ${\bf x}$ denotes the spatial degrees of freedom of the particle. In a ideal SG experiment, the up (down) spin can only be found in upward (downward) direction. This is possible only when the deflected beams are well separated such that $\phi_{+}({\bf x})$ and $\phi_{-}({\bf x})$ are orthogonal inside the SG apparatus and the orthogonality is also preserved during the free evolution of the particles as they leave the apparatus and travel to the screen. Hence, the overlap function $$I = \int_{-\infty}^{\infty} \phi_{+}(\bold{x}) \phi_{-}(\bold{x})d^3 \bold{x}=0. \label{orthogonality}$$ This is referred to as a formally ideal situation [@home2007]. An additional requirement of operational idealness [@home2007] may also be invoked where $$\begin{aligned}
\fl E=\int_{x=-\infty}^{\infty} \int_{y=-\infty}^{\infty} \int_{z=0}^\infty |\phi_{-}(\bold{x})|^2 dx dy dz=\int_{x=-\infty}^{\infty} \int_{y=-\infty}^{\infty} \int_{z=-\infty}^0 |\phi_+(\bold{x})|^2 dxdydz = 0~.\label{opid}\end{aligned}$$ Any exception to the above conditions is known to lead to a non ideal outcome [^1]. However, we shall argue that these criteria, though necessary, are not sufficient to ensure an ideal outcome. Till date, another important assumption has been implicitly incorporated in the analysis of this experiment – it is the adiabatic evolution of spin states. By adiabatic evolution [@griffithsqm; @messiahqm] we mean that the $|+\rangle$ ($|-\rangle$) state which is coupled to $\phi_+$ ($\phi_-$) remains in the same state during its travel through the apparatus. But it can not be true in every situation, especially in a weak magnetic environment. This non adiabatic evolution of spin states will be shown to lead to the emergence of non idealness even in a ‘formally and operationally ideal’ case ($I =0,E=0$).
Though most studies on conceptual and experimental aspects of SGE have been conducted within the confines of adiabatic regime, there also are instances in literature [@bliokh2007; @papanicolaou1988] where the concept of adiabaticity has been addressed for explicitly time varying magnetic fields. However, the spatial variation of static inhomogeneous magnetic field can also play a crucial role in the measurement of spin through path spin entanglement. Spatial inhomogeneity of the magnetic field lends an implicit time dependence to the interaction Hamiltonian which becomes explicit in the rest frame of the particle. A rapid spatial variation of interaction Hamiltonian in that case may cause a non adiabatic evolution for spin states. This gives rise to the possibility of detecting both up and down spins in either direction. Hence, one cannot infer a definite spin of the particle by observing its deflection although the split in the distribution pattern on the screen is still there (due to the force on the particles caused by the magnetic field). In other words, though the observed state on the screen is, in general, a path spin entangled one, a specific path does not necessarily correspond to a particular spin. This, we call [*blurred path spin entanglement*]{}. In fact, under certain conditions, instead of down spin the up spin can couple with $\phi_{-}$ and vice versa. This is absolutely opposite to the conventional wisdom and obtaining such a results especially from the analysis of a formally and operationally ideal SG experiment is a highly non trivial issue.
In order to demonstrate our finding, we have organised the article in the following way. After the introduction, we have discussed the relevance of adiabatic and non-adiabatic evolution of spin$\frac{1}{2}$ particles in the context of SG experiments. Section $3$ is devoted to examining the time evolution of spin states by using Schroedinger equation. In this context, we also have pointed out the parameters which decide the category (adiabatic or non-adiabatic) of the time evolution. This result is carried over to section $4$ for the interpretation of blurred path spin entanglement and its non trivial consequences. Finally, to summarise, a concise discussion is presented at the end of the article.
Adiabaticity and path spin entanglement
=======================================
Suppose a bunch of charge-neutral spin$\frac{1}{2}$ particles, prepared in a superposed state of up $|+\rangle _{u}$ and down $|-\rangle_{u}$ spins are collimated towards an SG apparatus. The state of a particle before it enters the device can be expressed as, $$|\psi_{i}\rangle = \big(c_{+} |+\rangle _{u}+ c_{-} |-\rangle _{u} \big)\otimes \phi_{i}(\bold{x}),$$ where $\phi_{i}(\bold{x})$ is the spatial part of the state. The moment the particle enters the inhomogeneous magnetic field along some direction $\hat{u}$, the state becomes path spin entangled due to its interaction with the magnetic field and can be written as, $$|\psi\rangle = c_{+} |+\rangle _{u}\otimes \phi_{+}({\bf x})+ c_{-} |-\rangle _{u} \otimes \phi _{-}({\bf x}).\label{ent}$$ As a result, a part of the particle beam will start to propagate along the upward direction and the rest will get directed downward. This fact is evident from the above equation which has already been derived many times in this context. In conventional treatments inspired by the original Stern-Gerlach paper [@gerlach1922], the inhomogeneous magnetic field is assumed to be directed along a fixed direction (say $z$). However, this does not make the field divergence free, thus contradicting Maxwell’s equations [@platt1992; @aharonov1988]. As a way out, the authors in [@platt1992; @oliveira2006] introduced a two component magnetic field and adopted a time averaged description of Pauli equation thereby obtaining an effective magnetic field along a fixed direction. This assumption is based on the consideration of very strong magnetic fields [@platt1992; @alstrom1982] observed over a time scale much larger than the characteristic time scale of the spin half system. Obviously, it is not impossible to design such a circumstance with sufficient control over laboratory conditions. But the same cannot be said of general practical scenarios. The Stern-Gerlach model is widely used to study path spin entanglement in diverse magnetic environments where this assumption may not always hold. Especially for a weak magnetic field, the above mentioned treatment is far from warranted. In a more general situation, the inhomogeneous magnetic field inside the apparatus changes spatially not only in magnitude but also in its direction. One can always think of a collection of tiny SG apparatus each providing magnetic fields in different directions so that they together mimic the effect of a single apparatus having a magnetic field varying along the length.
Due to such inhomogeneity, the particle will experience a varying magnetic field during its flight from one space point to another. So, it will see a magnetic field which has an implicit time dependence. We would like to examine the situation for upward and downward beams separately from the rest frames of the corresponding particles. Here it is worth mentioning that the implicit time dependence of the magnetic field $B$ becomes explicit in the rest frame of the particle. Let us call the interaction Hamiltonian in the rest frame of the particle $H_{part}(t)$. It is clear that the interaction Hamiltonian $H_{part}(t)$ can not affect the spatial parts $\phi_+$ and $\phi_-$ as it no longer contains any spatial degree of freedom. Therefore, we need to study the evolution of up $|+\rangle$ and $|-\rangle$ spin states only. Before proceeding, below we justify and explain our scheme of working in the particle rest frame.
Obviously, the particle rest frame is not ideally inertial. Note that in an SG apparatus, interaction with a weak magnetic field can cause only a small change $\Delta p $ in the momentum of an outgoing particle. This small change is negligible compared to the momentum $p$ of a high energy incoming beam. Thus, taking the particle frame to be inertial is practically a nice approximation in this context, unlike in a strong magnetic environment where $\frac{\Delta p}{P}$ is not likely to be much less than unity. But an exception occurs even in the latter case when a particle has a de Broglie wavelength $\lambda$ comparable to the length of the apparatus. This is because the particle does not experience a noticeable change in its momentum inside the apparatus as the dimension of the SG apparatus is not much bigger than the size of the wave packet. Except for these two special cases, one must take into account non-inertial modifications of the Hamiltonian while working in the particle rest frame. In the present article we restrict ourselves to the two special simple cases only.
Suppose at some instant, the direction of $\vec{B}_{part}(\tau)$, the magnetic field seen from the particle rest frame, inside the apparatus is along $\hat{u}(\tau)$. $H_{part}(\tau)$ and $S_{u(\tau)}$ have common set of eigenvectors $|+\rangle_{u(\tau)}$ and $|-\rangle_{u(\tau)}$ as, $$H_{part}(\tau)= -\gamma B_{part}(\tau) S_{u(\tau)},\label{gyro}$$ where $\gamma$ is the gyromagnetic ratio of the particle. Now, the task is to understand whether the spin states evolve adiabatically or not under the action of such a time dependent Hamiltonian. If the Hamiltonian changes sufficiently slowly with time such that $$\fl \Bigg |\frac{_{u(\tau)}\langle \pm|\frac{d}{d\tau}|\mp\rangle_{u(\tau)} }{E_{\pm}-E_{\mp}}\Bigg |=\bigg |\frac{_{u(\tau)}\langle \pm|\dot{H}_{part}|\mp \rangle_{u(\tau)}}{\Big(E_{\pm}-E_{\mp}\Big)^{2}}\bigg|
=\bigg |\frac{_{u(\tau)}\langle \pm|\vec{\mu}.\dot{\vec{B}}_{part}|\mp \rangle_{u(\tau)}}{\Big(E_{\pm}-E_{\mp}\Big)^{2}}\bigg|<<1~, \label{adiabatic}$$ then up (down) spin state will evolve to up(down) spin state of the corresponding instantaneous Hamiltonian $H_{part}(\tau)$ [@born1928; @aharonov1987; @amin2009] but with some geometrical (Berry phase) and dynamical phase factors [@berry1984]. $|\mp \rangle_{u(\tau)}$ is the eigenstate corresponding to the energy eigenvalue $E_{\mp}$ and $\vec{\mu}$ is the magnetic moment of the particle. This implies that the upward (downward) beam always contains the up (down) spin only. In those situations for which the above condition is not satisfied, the evolved state in each direction becomes a superposition of up and down spins. Therefore, in the upward direction we will have finite transition probability to $|-\rangle_{u(\tau)}$ from $|+\rangle_{u(0)}$ and similarly to the state $|+\rangle_{u(\tau)}$ from$|-\rangle_{u(0)}$ in the downward direction. As a result, one will find both up and down spins in either direction even though a formally and operationally ideal situation has been maintained by keeping $I=0=E$. This reveals the so far overlooked, salient role played by the adiabatic approximation in an SG experiment.
Interplay between Larmor precession and azimuthal inhomogeneity of magnetic field
=================================================================================
In this section we explicitly examine the evolution of spin states in an inhomogeneous magnetic field keeping aside the assumptions made in earlier literature [@platt1992; @potel2005]. We avoid any time averaged description of Pauli equation so that our analysis can be applied to a more general scenario. In the rest frame of the particle, the magnetic field $B_{part}$ appears to change with time. So, by considering an azimuthally inhomogeneous magnetic field interacting with the magnetic dipole moment $\vec{\mu}$, the explicit time dependence of the Hamiltonian can be expressed as,
![A schematic diagram which shows how the direction $\hat{u}$ of the magnetic field $\vec{B}_{part}$ appears to be changing with time from the rest frame of the particle as it moves through the SG apparatus.[]{data-label="dia"}](sgdia.pdf)
$$\begin{aligned}
H_{part}(t)
&=-\vec{\mu}.\vec{B}_{part}(\tau)=-\big(\vec{\mu}.\hat{u}(\tau)\big)B_{part}=\omega_{0}\hat{S}_{u(\tau)} \nonumber \\
&=\frac{\omega_{0} }{2}\Big(\sigma_{x}\sin \theta \cos\Phi(\tau) + \sigma_{y}\sin\theta \sin\Phi(\tau) +\sigma_{z}\cos\theta \Big) \label{Hamiltonian}.\end{aligned}$$
Here, $\hat{u}(\tau)$ is the direction of the magnetic field $\vec{B}_{part}$ and the azimuthal inhomogeneity makes the Hamiltonian a time dependent quantity through $\Phi(\tau)$. The Larmor frequency $\omega_{0}$ is taken to be a constant quantity by assuming that the magnitude of the magnetic field is not changing appreciably inside the apparatus. $\sigma_{x}$, $\sigma_{y}$, $\sigma_{z}$ are the Pauli spin matrices and $\theta$ is the angle which the magnetic field makes with $z$ axis. We have already mentioned in the last section that after entering SG apparatus, the particle beam gets split into two different parts, expressed through a path spin entangled state, $$|\Psi(0)\rangle=c_+|+\rangle_{u(0)} \otimes\phi_{+}({\bf x},0)+ c_-|-\rangle_{u(0)}\otimes \phi_{-}({\bf x},0). \label{initial}$$ We denote by $\tau=0$ the instant when the particle just becomes path spin entangled inside the apparatus. Here $\phi_{+}(\bold{x},0)$ and $\phi_{-}(\bold{x},0)$ are the up and down components of the spatial wave function at $\tau=0$ with respect to the plane whose normal makes an angle $\theta$ with $z$ axis. For $\theta=0$, this plane is nothing but the $z=0$ plane and we get the usual text book set up. Now, we analyse both the beams separately from the corresponding rest frame of the particles. Let us say that the particle leaves the apparatus at some instant $\tau=t$ and then moves freely up to the screen. A schematic diagram is presented in fig. \[dia\] to demonstrate the set up under consideration. So, the final state of the particle we are interested in is $|\Psi(t)\rangle$.
Whether the evolution of the spin states in both upward and downward directions is adiabatic or not is determined from eq.\[adiabatic\] by making use of instantaneous energy eigenstates of $H_{part}(\tau)$, $$\begin{aligned}
|-\rangle_{u(\tau)} =
\left(\begin{array}{c}
e^{-i \frac{\Phi(\tau)}{2}} \sin{\frac{\theta}{2}} \\
-e^{i\frac{\Phi(\tau)}{2}} \cos{\frac{\theta}{2}}
\end{array} \right)
\nonumber \\
|+\rangle_{u(\tau)} =
\left( \begin{array}{c}
e^{-i\frac{\Phi(\tau)}{2}}\cos \frac{\theta}{2}\\
e^{i\frac{\Phi(\tau)}{2}}\sin \frac{\theta}{2}
\end{array}\right)
\label{inst_phi}\end{aligned}$$ corresponding to the energy eigenvalues $E_{-}$ and $E_{+}$ respectively. Henceforth, we will denote $u(\tau)$ by the slightly more compact notation $u_{\tau}$. So, the ratio $$\begin{aligned}
\Bigg |\frac{_{u_{\tau}}\langle -|\frac{d}{d\tau}|+\rangle_{u_{\tau}} }{E_{+}-E_{-}}\Bigg |=\frac{\dot{\Phi}}{2\omega_{0}} \sin \theta \label{adbc} .\end{aligned}$$ The above equation shows that $\theta$, $\dot{\Phi}$ and $\omega_{0}$ are the determining factors for the adiabatic evolution of spin states. For non zero values of $\sin\theta$, $\omega_{0}$ should be much larger than $\dot{\Phi}$ in order to satisfy the adiabatic condition \[adiabatic\] . It implies that the Larmor precession should be rapid enough to cope with the instantaneous direction of the magnetic field.
In order to know the time evolution of spin states, one needs to solve the Schrödinger equation for the time dependent Hamiltonian $H_{part}(\tau)$ with a given initial state. The evolved state at any instant $\tau$ could be expressed in terms of instantaneous basis vectors, $$|\tau \rangle = a(\tau)|+\rangle_{u_{\tau}} + b(\tau)|-\rangle_{u_{\tau}}.$$ Plugging this into Schrödinger equation we get the following set of linear differential equations, $$\begin{aligned}
\dot{a}(\tau)+\frac{i}{2}\big[\omega_{0}-\dot{\Phi}(\tau)\cos\theta\big]a(\tau) -i\frac{\dot{\Phi}(\tau)}{2}b(\tau)\sin\theta = 0\nonumber\\
\dot{b}(\tau)-\frac{i}{2}\big[\omega_{0}-\dot{\Phi}(\tau)\cos\theta\big]b(\tau) -i\frac{\dot{\Phi}(\tau)}{2}a(\tau)\sin\theta =0 \label{gensoln}\end{aligned}$$ For any quantitative prediction we need to know the time dependence of $\Phi(\tau)$. To proceed further, we present a solution by considering terms up to 1st order in the Taylor series expansion of $\Phi(\tau)$. A general solution is always possible but we have considered this simple case to show the drastic role played by the azimuthal inhomogeneity. $$\Phi(\tau)= \Phi_{0}+ \frac{d\Phi(\tau)}{d\tau}\big |_{\tau=0}\tau+ \Or(2)\label{Taylor}$$ This is actually equivalent to assuming a small change in velocity of the particle as it travels through the apparatus (mentioned in the previous section). The instantaneous states look (by setting $\Phi_{0}=0$ without any loss of generality), $$\begin{aligned}
|-\rangle_{u_{\tau}} =
\left(\begin{array}{c}
e^{-i\frac{\omega \tau}{2}} \sin{\frac{\theta}{2}}\\
-e^{i\frac{\omega \tau}{2}} \cos{\frac{\theta}{2}}
\end{array}\right)
\nonumber \\
|+\rangle_{u_{\tau}} =
\left(\begin{array}{c}
e^{-i\frac{\omega \tau}{2}} \cos \frac{\theta}{2}\\
e^{i\frac{\omega \tau}{2}}\sin \frac{\theta}{2}
\end{array}\right).
\label{inststate}\end{aligned}$$ The adiabatic condition \[adiabatic\] becomes $\frac{\omega}{\omega_{0}} \sin \theta <<1$, where $\omega = \frac{d\Phi(\tau)}{d\tau}\big |_{\tau=0}$. The spin part $|+\rangle_{u_0}$ in \[initial\] which is coupled to the spatial part $\phi_{+}$ evolves to the final state $$|t\rangle_{+} = \alpha_{+}|+\rangle_{u_{t}} + \alpha_{-}|-\rangle_{u_{t}}\label{state2}.$$ at the instant $\tau=t$. The coefficients $\alpha_{+}$ and $\alpha_{-}$can be evaluated by solving the differential equation \[gensoln\] with the condition $a(0)=1$, $b(0)=0$ and $a(t)=\alpha_{+}$, $b(t)=\alpha_{-}$. They are $$\begin{aligned}
\alpha_{+} = \cos \frac{\bar{\omega}t}{2} -i\frac{\omega_{0}-\omega\cos\theta}{\bar{\omega}}\sin\frac{\bar{\omega} t}{2} \nonumber\\
\alpha_{-}= i\frac{\omega\sin \theta}{\bar{\omega}}\sin\frac{\bar{\omega}t}{2} .\label{coeff1}\end{aligned}$$ $\bar{\omega}$ is a combination of $\omega$ and $\omega_{0}$: $$\bar{\omega}= \sqrt{\omega^{2}_{0}+\omega^{2} -2\omega_{0}\omega \cos\theta}\label{omegabar}.$$ So, we witness that the up spins in the upward direction will generate down spins through their non-adiabatic evolution.
Similarly, the evolved state corresponding to the initial state $|-\rangle_{u_0}$ which is coupled with $\phi_{-}$ is given by, $$|t\rangle_{-} = \beta_{+}|+\rangle_{u_{t}} + \beta_{-}|-\rangle_{u_{t}},\label{state}$$ The coefficients $\beta_{+}$ and $\beta_{-}$ are determined in the same way by inserting the conditions $a(0)=0$, $b(0)=1$ and $a(t)=\beta_{+}$, $b(t)=\beta_{-}$ in Schrödinger equation [@tong2010]: $$\begin{aligned}
\beta_{-} = \cos \frac{\bar{\omega}t}{2} +i\frac{\omega_{0}-\omega\cos\theta}{\bar{\omega}}\sin\frac{\bar{\omega}t}{2} \nonumber\\
\beta_{+}= i\frac{\omega\sin \theta}{\bar{\omega}}\sin\frac{\bar{\omega}t}{2}~.\label{eq:coeff2}\end{aligned}$$ This again shows a finite transition probability to up spin state from the down spin in the downward direction. Therefore, we have up and down spins in both the directions. Hence, the usual perception of spin measurement through SG experiment can not be valid in general. In principle, for any nonzero $\sin\theta$, one has to consider such non-adiabatic evolution of states whenever the quantity $\omega_{0}$ is not very large compared to $\omega$.
![$P_{sr}$ as a function of time $t$ (in the unit of $\frac{1}{\bar{\omega}}$) and the angle $\theta$ for a chosen value of $\omega =1.2\omega_{0}$.[]{data-label="survival"}](trans.pdf){width="\linewidth"}
![$P_{sr}$ as a function of time $t$ (in the unit of $\frac{1}{\bar{\omega}}$) and the angle $\theta$ for a chosen value of $\omega =1.2\omega_{0}$.[]{data-label="survival"}](surv.pdf){width="\linewidth"}
Equation \[coeff1\] and \[eq:coeff2\] show that the probability of having down spin from up spin in the upward direction is $|\alpha_{-}|^{2}$ which is exactly equal to the transition probability $|\beta_{+}|^{2}$ from down spin to up spin in the downward direction. We call it $P_{tr}$ and present it graphically in Figure \[transition\] as a function of $\theta$ and time $t$. The survival probability of up(down) spin states in the upward(downward) direction is equal to $|\alpha_{+}|^{2}$ ($|\beta_{-}|^{2}$). And it is clear that $|\alpha_{+}|^{2}=|\beta_{-}|^{2}$. Lets denote this probability as $P_{sr}$. Figure \[survival\] shows its variation with $\theta$ and $t$. This is a very important consequence of the analysis. We see that the number of down (up) spins disappearing from the downward (upward) beam is exactly equal to the number of down (up) spins appearing in the the upward (downward) beam. Therefore, as a whole the total spin is always conserved.
The final spin state which we observe on the screen is the one at the instant $t$ when the particle just leaves the apparatus. That instant has to be determined from the time taken by the particle to complete its travel through the SG apparatus. Therefore, the longitudinal dimension of the apparatus is important to explain the observed pattern on the screen.\
It is also noticeable that the non adiabatically evolved spin states have a precession frequency which is different from the Larmor frequency $\omega_{0}$. The modification also depends on $\theta$ and $\omega$. This happens because the instantaneous spin eigenstates are now weighted by a periodic function ($\cos\bar{\omega}t$ or $\sin\bar{\omega}t$).
Blurred path-spin entanglement
==============================
In the previous section, we have derived the final spin states coupled to spatial up and down parts of $|\psi(0)\rangle$. Now, the final state $|\Psi(t)\rangle$ of the particle when it leaves the apparatus becomes, $$\begin{aligned}
\fl |\Psi(t)\rangle &= c_{+}\big(\alpha_{+}|+\rangle_{u_{t}}+\alpha_{-}|-\rangle_{u_{t}}\big)\otimes \phi_{+}^{\prime}({\bf x^{\prime}},t)
+ c_{-}\big(\beta_{+}|+\rangle_{u_{t}}+\beta_{-}|-\rangle_{u_{t}}\big)\otimes \phi_{-}^{\prime}({\bf x^{\prime}},t)\nonumber\\
\fl&= \big(d_{1}\phi_{+}^{\prime}(\bold{x^{\prime}},t)+d_{2}\phi_{-}^{\prime}({\bf x^{\prime}},t)\big)\otimes |+\rangle_{u_{t}} +\big(k_{1}\phi_{+}^{\prime}({\bf x^{\prime}},t)+k_{2}\phi_{-}^{\prime}({\bf x^{\prime}},t)\big)\otimes |-\rangle_{u_{t}} ~,\label{eq:soln}\end{aligned}$$ where $\phi_{+}^{\prime}({\bf x^{\prime}},t)$ and $\phi_{-}^{\prime}({\bf x^{\prime}},t)$ are the spatial up and down components of the state as seen from the rest (primed) frame of the particle. We have mentioned earlier that the Hamiltonian $H_{part}(t)$ does not affect the spatial part of the state. The spatial up (down) part remains as the up (down) part with respect to the plane whose normal is now $\hat{u}(t)$ even at the instant $\tau=t$ as long as the angle $\theta$ is constant. No wonder that the situation remains formally and operationally ideal if we start with $I=0, E=0$ at $\tau =0$. Equation \[eq:soln\] shows that the final state is still an entangled one as the spatial and spin parts are not product separable. We call it a path spin entangled state but a big difference from the conventional result is that the entangled state can no longer be used to draw inference on the nature of particle spin from its path. In other words, the evolved state has blurred the possibility of determining the spin of a particle from a knowledge of its path. Moreover, one can easily check that a breakdown of adiabaticity can not result in disentanglement of the initially path spin entangled state. Disentanglement occurs when $d_{1}=d_{2}=\frac{d}{\sqrt{2}}$ and $k_{1}=k_{2}=\frac{k}{\sqrt{2}}$ which can not be achieved for any value of $\theta$ or $t$.
However, at the instant $t=\frac{(2n+1)\pi}{\bar{\omega}}$, $n$ being any positive integer, when $\cos\theta = \frac{\omega_{0}}{\omega}$, $$\alpha_{+}=\beta_{-}=0 .$$ The up spin is now found in the downward direction whereas the down spin appears only in the upward direction, as is clear from the solution presented in eq. \[eq:soln\]. This goes absolutely against the dictates of conventional wisdom. It is evident that the angle $\theta$ plays an important role in determining the final state of the particle which leaves the apparatus and evolves freely to the screen. Furthermore, in the limit of a strong magnetic field, we can always get back the well known text book result of SG experiment where the Larmor frequency $\omega_{0}$ is much larger than $\omega$. Here, $$\begin{aligned}
\alpha_{+} = e^{-i\frac{\omega_{0}t}{2}};\qquad \alpha_{-} = 0\nonumber\\
\beta_{-} = e^{i\frac{\omega_{0}t}{2}} ; \qquad \beta_{+} = 0~, \label{largeomega}\end{aligned}$$ as $$\bar{\omega}= \omega_{0}\sqrt{1+\frac{\omega^{2}}{\omega_{0}}-2\frac{\omega \cos\theta}{\omega_{0}}}\approx \omega_{0}\label{large}.$$ Now, plugging this in equation \[eq:soln\], we have $$\Psi(t)
= c_{+}e^{-i\frac{\omega_{0}t}{2}}|+\rangle_{u_0}\otimes \phi_{+}^{\prime}({\bf x^{\prime}},t)
+c_{-} e^{i\frac{\omega_{0}t}{2}}|-\rangle_{u_0}\otimes\phi_{-}^{\prime}({\bf x^{\prime}},t) \label{last}$$
This is equivalent to the solution \[ent\] obtained in references [@platt1992; @oliveira2006] for a time averaged magnetic field with fixed direction. One can easily realize that in this case the up (down) spin is found only in the upward (downward) direction. Thus we arrive at the conventional result of path spin entanglement from our analysis in a limiting situation where the adiabatic evolution is permissible. Note that we have not introduced any averaging concept in Schrödinger equation thereby ensuring that the solution holds good at arbitrary time scales. It is clear from eq. \[last\] that the final state if observed above a time scale $t=\frac{2\pi}{\omega_{0}}$ reduces to a solution of Pauli equation given in [@platt1992]
Discussion and Conclusion
=========================
In this article, we have investigated the interplay between Larmor precession and azimuthal inhomogeneity of magnetic field in the context of path-spin entanglement of a spin half system in an SG experiment. Emphasis has been laid on determining the necessary and sufficient condition for adiabatic evolution of spin states owing to its importance in spin measurement using SG apparatus. To the best of our knowledge, the phenomenon predicted by us has not been explored in extant literature on this topic. Earlier studies on dynamical nature of entanglement also took into consideration the effect of inhomogeneous magnetic fields, a relevant example being the article by Caldeira et al. [@oliveira2006]. But the authors in [@oliveira2006] presented the analysis with the help of a magnetic field which is changing in magnitude only. It is obvious that the instantaneous states will not then change with time. A similar setup may also be incorporated in our work to get back the result of [@oliveira2006]. For the sake of simplicity and clarity, we have restricted ourselves to an azimuthally inhomogeneous magnetic field and postponed a study of the more general situation to a future article [@nirupamprep2014ent]. Our present analysis reveals that not only the ratio of $\dot{\Phi}(t)$ to $\omega_{0}$ but the angle $\theta$ is also a key factor in determining the category of evolution of spin states. Setting $\theta = 0$ takes us back to the standard text book result for SG experiment even if we choose a description in terms of instantaneous states. When $\theta$ is not sufficiently small, unless $\omega_0 >> \dot{\Phi}$, the spin states evolve non-adiabatically. Though entanglement can still persist, it will no longer allow one to infer any definite spin from the trajectory of a particle. It must be carefully noted that the observation of both spins in either direction, as discussed here, is completely independent of formal and operational non-idealness. Rather, we stress that the interplay between inhomogeneity of magnetic field and Larmor precession is quite generic and can lead to novel unexpected phenomena.
There is a practical applicability of this finding, as well. Till date, an ideal SG apparatus has often been used as a beam splitter in quantum information devices. We have seen in section $3$ that due to a non-adiabatic evolution, a part (some times all) of the up (down) spin can appear in the downward (upward) direction, keeping the total spin of the system conserved. This can provide a spin flipping tools in various quantum information protocols.
Our sincere thanks are due to Nicolas Borghini and Debashis Mukherjee for their insightful advice. We thank Archan. S. Majumder for suggestin reference [@home2007] which helped us to understand operational and formal non-idealness in SG apparatus. N.D acknowledges support from DAE, Govt. of India and A. D. acknowledges DST, Govt. of India for supporting the research.
References {#references .unnumbered}
==========
[^1]: Note that the definitions of both formal and operational idealness concern only the spatial parts of the state of the particle
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Determining the maximal separation between sensitivity and block sensitivity of Boolean functions is of interest for computational complexity theory. We construct a sequence of Boolean functions with $bs(f) = \frac{1}{2}{s(f)}^2+\frac{1}{2} s(f)$. The best known separation previously was $bs(f) = \frac{1}{2} {s(f)}^2$ due to Rubinstein. We also report results of computer search for functions with at most $12$ variables.'
address: 'University of Latvia, Rainis bld. 19, Riga, LV-1586, Latvia'
author:
- Madars Virza
bibliography:
- 'blocksens.bib'
title: Sensitivity versus block sensitivity of Boolean functions
---
sensitivity ,block sensitivity
Introduction
============
In his 1989 paper [@nisan] Noam Nisan gives tight bounds for computing the value of a Boolean function in CREW-PRAM model. These bounds are expressed in terms of two complexity measures, namely the sensitivity and block sensitivity.
Let $w$ be a Boolean string of length $n$ and let $S$ be any subset of indices. Following definitions in [@nisan], by $w^{S}$ we will mean $w$ with all bits in $S$ flipped. If $f(w) \neq f(w^{S})$, we will say that $f$ is *sensitive to $S$ on $w$*.
The *sensitivity* $s(f, w)$ of $f$ on the input $w$ is defined as number of indices $i$ such that $f(w) \neq f(w^{\{i\}})$, $s(f, w) = | \{ i : f(w) \neq f(w^{\{i\}}) \} |$. The *sensitivity* $s(f)$ of $f$ is $\displaystyle\max_w s(f, w)$.
The *block sensitivity* $bs(f, w)$ of $f$ on n inputs $w_1, \ldots, w_n$ ($w = w_1 \ldots w_n$) is defined as maximum number of disjoint subsets $B_1, \ldots, B_k$ of $\{ 1, 2, \ldots, n \}$ such that for each $B_i$, $f(w) \neq f(w^{B_i})$. The *block sensitivity* $bs(f)$ of $f$ is $\displaystyle\max_w bs(f, w)$.
Block sensitivity is polynomially related to many other complexity measures such as decision tree complexity, certificate complexity, and quantum query complexity, among others [@buhrman]. The prime open problem associated with block sensitivity is whether sensitivity and block sensitivity are polynomially related.
In 1992 Rubinstein proposed [@rubinstein] a Boolean function $f$ for which $bs(f) = \frac{1}{2} {s(f)}^2$, demonstrating a quadratic separation between these two complexity measures, which we quote in verbatim:
For every $n$ that is an even perfect square, we can construct a Boolean $f$ on $n$ variables with $2 bs(f)={s(f)}^2=n$.
Let $\Delta_i$ denote the interval $\Delta_i = \{ (i-1) \sqrt{n}+1, \ldots, i \sqrt{n} \}$ ($i = 1, \ldots, \sqrt{n}$).
Let $g_i$ denote the Boolean function defined as follows: $g_i(H) = 1$ exactly if $H \cap \Delta_i = \{ 2j-1, 2j \}$ for some $j$ such that $2j \in \Delta_i$.
We define $f$ to be join of all such $g_i$: $f(H) = g_1(H) \wedge \ldots \wedge g_{\sqrt{n}}(H)$.
It is believed that this separation is not far from optimal. However, the best known upper bound of block sensitivity in terms of sensitivity is exponential [@lblock].
Our result
==========
In this paper we give an improvement on Rubinstein’s example, in the case where the input length $n$ is an odd square. Our function family is a modification of Rubinstein’s, and our analysis is modelled on his.
For every non-negative integer $k$ there exists a Boolean function $f$ of $n = (2k+1)^2$ variables, for which $s(f) = 2k+1$ and $bs(f) = (2k+1)(k+1)$.
We will divide variables into $2k+1$ disjoint sections with $2k+1$ variables in each section. We define $f$ to be $1$ iff there is a section $x_1, x_2, \ldots, x_{2k+1}$ such that either:
(i) $x_{2i-1} = x_{2i} = 1$ for some $1 \leq i \leq k$ and all other $x_j$’s are $0$, or
(ii) $x_{2k+1} = 1$ and all other $x_j$’s are $0$.
We will call such section a “good” section. For any $i$, we will call $x_{2i-1}$ and $x_{2i}$ a “pair”.
We observe that for input $w = 0 \ldots 0$ we have $s(f, w) = 2k+1$ and $bs(f, w) = (2k+1)(k+1)$. We will now prove that these are extremal values for sensitivity and block sensitivity, respectively.
Suppose we have an arbitrary input $w$. We will consider two cases:
1. $f(w) = 0$. Then we claim that for each of $2k+1$ sections there is at most one bit whose change could flip the value of $f$.
We will call a pair “incomplete”, if exactly one of its bits is set to $1$. We consider three cases based on the number of incomplete pairs in a section:
1. there are at least two “incomplete” pairs. Then we can’t change the value of $f$ by flipping just one bit, because doing so will leave at least one “incomplete” pair.
2. there is exactly one “incomplete” pair $P$. We first note that we can’t make the section “good” without flipping a bit in $P$. If unsetting the $1$-bit in $P$ makes the section “good”, then $x_{2k+1} = 1$ or $x_{2i-1} = x_{2i} = 1$ for some $1 \leq i \leq k$ and setting the $0$-bit in $P$ will not make the section “good”. If setting the $0$-bit in $P$ makes the section “good”, then the only $1$-bit in $x$ is also in $P$ and unsetting it will not make the section “good”. In either case there is at most one choice for the bit to alter.
3. there are no “incomplete” pairs. Then flipping any of $x_i$ for $1 \leq i \leq 2k$ will introduce an “incomplete” pair and the section will not become “good”. Therefore there is at most one bit whose change could flip the value of $f$, namely, $x_{2k+1}$.
2. $f(w) = 1$. If there are two “good” sections, we can’t change the value of $f$ by flipping just one input bit. If there is only one “good” section, we have at most $2k+1$ choices for the bit to alter.
This proves that $s(f) \leq 2k+1$ and we indeed have $s(f) = 2k+1$.
We will now prove that $bs(f) \leq (2k+1)(k+1)$. Assume that the maximal block sensitivity is achieved using $u$ blocks of size $1$ and $v$ blocks of size at least $2$. Number of blocks of size $1$ can’t exceed the sensitivity of the function, therefore $u \leq 2k+1$. The total size of all blocks is at most $u + 2v$, furthermore, the total size of all blocks can’t exceed the total number of variables, therefore $u + 2v \leq {(2k+1)}^2$. Taking these two inequalities together we obtain $bs(f) = u + v = \frac{1}{2}(u + (u + 2v)) \leq \frac{1}{2} ((2k+1) + (2k+1)^2) = (2k+1)(k+1)$.
This implies that for our function we have $bs(f) = \frac{1}{2}{s(f)}^2+\frac{1}{2} s(f)$. Our function improves an inequality from an open problem in [@lblock].
Maximal separation for functions of few variables
=================================================
We discovered our function by performing exhaustive computer search of all Boolean functions with at most $12$ variables.
The number of $12$-variable Boolean functions is $2^{4096}$ and a brute-force approach is clearly not feasible. In our experiments we reduced the block sensitivity problem to the SAT problem and used a SAT solver on the resulting problem instances. We built our SAT instances by considering $2^{12}$ variables corresponding to the values of $f(x_1, x_2, \ldots x_{12})$ and additional variables and clauses to describe constraints $s(f) \leq s$ and $bs(f) \geq bs$, for arbitrary constants $s$, $bs$. The SAT solver we used was cryptominisat [@cryptominisat] and it took us about a week of computing time.
The constraint $bs(f) \geq bs$
------------------------------
In order to describe $bs(f) \geq bs$ we constructed one SAT instance for each partition of $n$ into $bs$ parts $p_1, \ldots, p_{bs}$. The semantic meaning for the constraint we will describe is “there is a function which is sensitive on blocks of sizes $p_1, \ldots, p_{bs}$”. The constraint $bs(f) \geq bs$ is feasible if and only if there exists a partition for which the constraint described below is satisfiable.
We note that for any $T \subseteq \{1, \ldots, n \}$ and $g(x) = f(x^{T})$ we have $s(f) = s(g)$ and $bs(f) = bs(g)$. This permits us to assume, without loss of generality, that $f$ attains the maximal block sensitivity on input $0\ldots0$, furthermore we can assume that $f(0\ldots0) = 0$. If $bs(f) \geq bs$ then by reordering variables we can enforce $\neg f(0\ldots0) \wedge \bigwedge_{i}{f({{0\ldots0}^{P_i}})}$ where $\displaystyle P_i = \{ \sum_{j=1}^{i-1} p_j + 1, \sum_{j=1}^{i-1} p_j + 2, \ldots, \sum_{j=1}^{i} p_j \}$. This was our constraint. We used this constraint to enforce $bs(f) \geq bs$.
The constraint $s(f) \leq s$
----------------------------
We constructed the constraint $s(f) \leq s$ as a conjunction of constraints of the form $s(f, w) \leq s$, where $w$ ranged over all $2^n$ inputs.
We will now describe the setup for the constraint $s(f, w) \leq s$. Let $b_i$ be the variable denoting $f(w^{\{i\}})$. We used auxiliary variables and clauses to implement “counting” in unary. That is, we introduced new Boolean variables $a_1, \ldots, a_n$ and defined a set of constraints that force $a_1+a_2+\ldots+a_n=b_1+b_2+\ldots+b_n$ and $a_1 \geq a_2 \geq \ldots \geq a_n$, essentially sorting the $b_i$’s.
If $\{ a_i \}$ is a permutation of $\{ b_i \}$ sorted in descending order, then the constraint $s(f, w) \leq s$ can be implemented as $(\neg f(w) \wedge \neg a_{s+1}) \vee (f(w) \wedge a_{n-s})$. The set of $a_i$’s was built incrementally using a dynamic programming table $\{ c_{i,j} \}$ ($0 \leq i, j \leq n$), where $c_{i,j} = 1$ iff at least $j$ of $b_1, \ldots, b_i$ are $1$. Following the semantic meaning of $c_{i,j}$ we have the base conditions of $c_{i,0} = 1$ and $c_{i, j} = 0$ ($i < j$). Furthermore, the following recurrence holds: $c_{i,j} = (c_{i-1, j-1} \wedge b_i) \vee c_{i-1,j}$ ($1 \leq j \leq i \leq n$). This can be proved by considering two cases: if $b_i = 1$ then $c_{i,j} = c_{i-1,j-1} \vee c_{i-1, j} = c_{i-1,j-1}$, because $c_{i-1,j} \to c_{i-1,j-1}$, if $b_i = 0$ then $c_{i,j} = c_{i-1,j}$; both of these equalities agree with the definition of $c_{i,j}$. The set of $a_i$’s is exactly the last row of the matrix: $a_i = c_{i,n}$.
Results
-------
We have summarized our results in Table \[senst\]. Our experiments show that for all functions with at most 12 variables $bs(f)$ does not exceed $\frac{1}{2} {s(f)}^2 + \frac{1}{2} s(f)$. This might suggest that our example is indeed optimal. Due to the results of Kenyon and Kutin, the best possible separation that can be achieved between the sensitivity and the block sensitivity for blocks of size at most 2 is quadratic, so the new result is not far from optimal in such sense.
$n$ $s$ $bs$ $n$ $s$ $bs$ $n$ $s$ $bs$ $n$ $s$ $bs$
---------- ----- ------ ---------- ----- ------ ----------- ----- ------ ----------- ------ ------
$\geq 4$ $2$ $3$ $\geq 8$ $4$ $6$ $\geq 10$ $4$ $7$ $\geq 11$ $9$ $10$
$\geq 5$ $3$ $4$ $\geq 8$ $6$ $7$ $\geq 10$ $6$ $8$ $12$ $4$ $8$
$\geq 6$ $4$ $5$ $\geq 9$ $3$ $6$ $\geq 10$ $8$ $9$ $12$ $6$ $9$
$\geq 7$ $3$ $5$ $\geq 9$ $5$ $7$ $\geq 11$ $5$ $8$ $12$ $8$ $10$
$\geq 7$ $5$ $6$ $\geq 9$ $7$ $8$ $\geq 11$ $7$ $9$ $12$ $10$ $11$
: All permissable $(n, s, bs)$ triples for $n \leq 12$[]{data-label="senst"}
Acknowledgements
================
I would like to thank Andris Ambainis for introducing me to this problem and subsequent helpful discussions. Computational resources were provided in part by AILab of Institute of Mathematics and Computer Science, University of Latvia. I am thankful to the anonymous reviewers for their comments, which substantially improved this paper (FIXME).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
Limitations in processing capabilities and memory of today’s computers make spiking neuron-based (human) whole-brain simulations inevitably characterized by a compromise between bio-plausibility and computational cost. It translates into brain models composed of a reduced number of neurons and a simplified neuron’s mathematical model, leading to the search for new simulation strategies.\
Taking advantage of the sparse character of brain-like computation, the *event-driven* technique could represent a way to carry out efficient simulation of large-scale *Spiking Neural Networks* (SNN). The recent *Leaky Integrate-and-Fire with Latency* (LIFL) spiking neuron model is event-driven compatible and exhibits some realistic neuronal features, opening new avenues for brain modelling. In this paper we introduce FNS, the first LIFL-based spiking neural network framework, which combines spiking/synaptic neural modelling with the event-driven approach, allowing us to define heterogeneous neuron modules and multi-scale connectivity with delayed connections and plastic synapses. In order to allow multi-thread implementations a novel parallelization strategy is also introduced. This paper presents mathematical models, software implementation and simulation routines on which FNS is based. Finally, a brain subnetwork is modeled on the basis of real brain structural data, and the resulting simulated activity is compared with associated brain functional (source-space MEG) data, demonstrating a good matching between the activity of the model and that of the experimetal data. This work aims to lay the groundwork for future event-driven based personalised brain models.\
address:
- 'Laboratory of Cognitive and Computational Neuroscience, Center for Biomedical Technology, Technical University of Madrid & Complutense University of Madrid, Spain.'
- 'Department of Civil Engineering and Computer Science, University of Rome ‘Tor Vergata’, Italy'
- 'Department of Electronic Engineering, University of Rome ‘Tor Vergata’, Italy;'
- 'Department of Experimental Psychology, Cognitive Processes and Logopedy, Faculty of Psychology, Complutense University of Madrid, Spain;'
- 'Department of Industrial Engineering & IUNE, University of La Laguna, Spain'
author:
- Gianluca Susi
- Pilar Garcés
- Alessandro Cristini
- Emanuele Paracone
- Mario Salerno
- Fernando Maestú
- Ernesto Pereda
title: |
FNS: an event-driven spiking neural network simulator based on the LIFL neuron model\
[preprint version]{}
---
Spiking Neural Network ,Event-driven Simulation ,Neuronal modeling ,Functional connectivity ,Magnetoencephalography
Introduction
============
Today’s advanced *magnetic resonance imaging* (MRI) -based techniques allow a thorough estimation of the structural *connectome* (i.e., the map of neural connections in the brain [@HagmannConn; @SpornsConn]), as well as volume and morphology of single brain areas. Through the application of graph theory, such data can be employed to synthesise brain dynamic models, which today are more and more able to appropriately reproduce brain oscillations revealed by functional imaging techniques such as *Multi-Unit Activity* (MUA) and *Local Field Potential* (LFP) [@Barardi2014], *functional* MRI [@Cabral2011; @Deco2012; @Bettinardi2017] and *Magnetoencephalography* (MEG) [@Nakagawa2014; @Cabral2014; @Deco2017], providing new information on the brain operation. In such approaches, *nodes* represent surrogates of brain regions (corresponding to *gray matter*), and *edges* represent the long-range connections, along fibre tracts, between them (corresponding to *white matter*), usually estimated using *diffusion tensor imaging* (DTI) (Fig. \[FIG:BrainModGrAppr\]).
![Synthesis of a computational brain model using the graph approach. White matter connections can be extracted by means of DTI. Brains of individual subjects can be coregistered to a parcellation template (*atlas*) in order to assign connections to specific brain areas. By assigning node local dynamics to the obtained *structural connectome*, it is possible to extract simulated activity. The number of nodes of the model depends on the template used, and each node can be represented at different levels of abstraction (e.g., ensemble of spiking neurons).[]{data-label="FIG:BrainModGrAppr"}](1.png){width="100.00000%"}
Among the different approaches used to represent brain regions [@Deco2008], *spiking/synaptic models* [@Vicente2008; @Gollo2010; @Nakagawa2014; @Maslennikov2014] present a very large number of degrees of freedom, which gives rise to highly complex and realistic behaviours on a broad frequency range of the related oscillations [@Barardi2014]. In addition, spiking/synaptic models offer the opportunity to relate to real-brain data transversely [*micro-*, *meso-*, and *macro-scale*, referring to the categorisation of @Bohland2009], as well as to implement *spike-timing dependent plasticity* (STDP), which is indispensable in many kinds of computational neuroscience studies.
On the other hand, spiking/synaptic-based brain simulations are extremely expensive from a computational point of view [@Izhikevich2004]. This issue translates to the use of simplified spiking neuron models, and nodes composed of a low number of elements, thereby reducing the realism of the overall brain model.
Spiking neuron models are described by differential equations and usually simulated with clock-driven (synchronous) algorithms, by means of proper integration methods [see @Brette2007 for an extensive review]. In this way the update is done at every tick of a clock **X$(t)$ $\rightarrow$ X$(t+dt)$**, and involves all network elements (neurons and possibly synapses). Such two characteristics cause a fast growth of simulation times/used memory when the network size increases, and can lead the simulation to entail missing spikes (although recent work is aimed to overcome the latter limitation, as in [@Hanuschkin2010] and [@Krishnan2017]).
Conversely, in the event-driven (or asynchronous) approach a network element is updated only when it receives or emits a spike. Then, such approach does not envisage a periodic update, neither a check of all network elements, producing simulations devoid of missed spikes, and exploits the sparseness of brain-like activity. Since the latter is irregular in time with low average [@Mouraud2009], this approach has the potential to reduce the computational cost for large-scale network simulations [@Ros2006]. Nevertheless, the need of an explicit solution for the neuron state between spikes, and the consideration of incoming and outgoing pulses as discrete events, make the event-driven simulation of classic bio-realistic models very challenging. This has stimulated a big interest in scientific community in developing both realistic and event-driven-compatible spiking neuron models [see @Brette2006; @Brettepa2007; @Tonnelier2007; @Salerno2011; @RudolphLilith2012], which led to the development of event-driven based SNN simulators [see @Pecevski2014; @Cristini2015], and hybrid *event*/*time-step* based simulation strategies [see @Morrison2006; @Hanuschkin2010; @DHaene2014; @NEST2007; @BrianDocumentation].
In particular, the *Leaky Integrate-and-Fire with Latency* (LIFL) model is a recent neuron model that can be simulated in event-driven fashion, preserving important computational features at the same time [@Susi2018; @Salerno2011; @Cardarilli2013; @Cristini2015; @Susi2015; @Susi2016; @Acciarito2017]. Differently from the *Leaky Integrate-and-Fire* (LIF), LIFL incorporates important neuronal features extracted from the bio-realistic *Hodgkin-Huxley* (HH) model, such as the *spike latency* [@FitzHugh1955; @Izhikevich2004]. The latter has been proved to be fundamental in many scenarios of neural computation, providing a large range of behaviors. Then, the LIFL may open new avenues for the efficient simulation of large scale brain models.
In this work we present FNS (literally, *Firnet NeuroScience*), a LIFL-based exact event-driven SNN framework oriented to brain simulations, implemented in *Java*. FNS allows us to generate brain network models on the basis of a versatile graph-based multi-scale neuroanatomical connectivity scheme, allowing for heterogeneous neuron modules and connections. In addition to the high-customizability of the network, proper input and output sections make it possible to relate model activity to real data, with the option to enable plasticity, then making the network parameters evolve depending on the network activity.
In section 2, we describe the neurobiological principles and mathematical models on which FNS is based: neuron model, region model, fibre tracts model, plasticity, input and output signals.\
In section 3, we present the possibilities that the framework offers for the synthesis of custom models and the design of specific simulations: generator section, neuroanatomical model section and output section.\
In section 4, we illustrate the technical aspects of the simulation framework itself: design principles, data structures and parallelization strategy.\
In section 5, we present an example to show how to conduct a simulation in FNS, and to evaluate the realism and performances of the framework itself. In short, we model a brain subnetwork using structural data of a real subject, and through FNS we simulate brain activity and synthesize electrophysiological-like output signals. Then, we compare such signals with those of the real subject.\
In the Discussion section, we summarize our work and envisage how to improve FNS in future works.
In this manuscript, a single neuron is indicated with $n$; an axonal connection between two whichever neurons with $e$; a neuron module (corresponding to a region or subregion in real case) with $N$, and called *network node*; the complete set of connections between two nodes (corresponding to fibre tracts in real case) with $E$, and called *network edge*.
The software can be freely downloaded at:\
[www.fnsneuralsimulator.org](www.fnsneuralsimulator.org).\
In the download link a user guide (including a short description of how to install and run it) and some network models are also provided with the software.
From neurobiology to mathematical models
========================================
Specificity and heterogeneity characterize the human brain at all scales. In this regard, recent works highlight crucial aspects that have to be taken into account in brain models to obtain realistic dynamics:
- *Region bioplausibility*: in spiking/synaptic models, an inappropriate choice of the spiking neuron model or the intra-module connectivity configuration may lead to results having nothing to do with the information processing of real brain [@Izhikevich2004]. Of course, also the cardinality of the nodes is important for achieving an appropriate network behaviour.
- *Region diversity*: diversity among and within regions specializes the behaviour of single parts of the network, enhancing the information content and coding performances and shaping properties of collective behavior such as synchronization [see @Thivierge2008; @Gollo2016].
- *Inter-region connection bioplausibility*: synchronization between network nodes is strongly sensitive to edge parameters (as weights, delays and connection number) and their distributions [@Brunel1999; @Brunel2003; @Vicente2008; @Viriyopase2012; @Gollo2014].
- *Inter-region connection diversity*: selective variations of edge parameters are able to reconfigure the network synchronization profile [@Abuhassan2014], including synchronization between nodes that are not directly connected to the modified edge [@Gollo2014].
FNS aims to guarantee the possibility to take into account such aspects in order to avoid the alteration of the network operation. In this section we present mathematical models used in FNS.
LIFL Neuron model
-----------------
Altough the classic LIF model is very fast to simulate, it has been regarded as unrealistically simple, thereby incapable of reproducing the dynamics exhibited by cortical neurons [@Izhikevich2003]. FNS is based on the LIFL, that besides being computationally simple it is also able to support a greater number of neuronal features than the LIF.
### A brief introduction to the spike latency neuro-computational feature {#SECT:SLat}
The *spike latency* is the membrane potential-dependent delay time between the overcoming of the “threshold” potential and the actual spike generation [@Izhikevich2004]. Among all the neuron features, it is of considerable importance because it extends the neuron computation capabilities over the “threshold”, giving rise to a range of new behaviors. Spike latency is ubiquitous in the nervous system, including the auditory, visual, and somatosensory systems [@Wang2013; @Trotta2013].
From a computational point of view it provides a spike-timing mechanism to encode the strength of the input [@Izhikevich2007] conferring many coding/decoding capabilities to the network [e.g., @Gollisch2008; @Fontaine2009; @Susi2015], whereas, from a statistical point of view it results in a desynchronizing effect [@Salerno2011; @Cardarilli2013], fostering the emergence of higher frequencies [@Susi2016] and providing robustness to noise to the network [@Izhikevich2007 chapter 7]. Spike latency has already been introduced in some variants of the LIF, as *QIF* [@Vilela2009] and *EIF* [@Forcaud2003]. In LIFL spike latency is embedded with a mechanism extracted from the realistic HH model [@Salerno2011], both simple and suitable to the event-driven simulation strategy.
### LIFL operation {#SECT:LIFLoperation}
In this section, we briefly describe the behaviour of the LIFL neuron model. For the sake of simplicity, we will refer to its basic configuration.
LIFL neuron model is characterized by a real non-negative quantity $S$ (the *inner state*, corresponding to the membrane potential of the biological neuron), which is defined from 0 (corresponding to the resting potential of the biological neuron) to $S_{max}$ (*maximum state*), a value much greater than one, at most $\infty$. Simple Dirac delta functions (representing the [action potentials]{}) are supposed to be exchanged between network’s neurons, in form of *pulse* trains. The model is able to operate in two different modes: *passive mode* when $S<S_{th}$, and *active mode* when $S\geq S_{th}$, where $S_{th}$ is the *firing threshold*, a value slightly greater than 1. In passive mode, $S$ is affected by a decay, whereas the active mode is characterized by a spontaneous growth of $S$. Assuming that neuron $n_j$ (i.e., the *post-synaptic neuron*) is receiving a pulse from neuron $n_i$ (i.e., the *pre-synaptic neuron*), its inner state is updated through one of the following equations, depending on whether $n_j$ was in passive or in active mode, respectively:
[S\_[\_j]{}= ]{} S\_[p\_[j]{}]{}+A\_[\_[i]{}]{}W\_[\_[i,j]{}]{}-T\_l , 0S\_[p\_[j]{}]{}<S\_[th]{} \[EQ:Pm\]\
S\_[p\_[j]{}]{}+A\_[\_[i]{}]{}W\_[\_[i,j]{}]{}+T\_r , S\_[th]{}S\_[p\_[j]{}]{}< S\_[max]{} \[EQ:Am\]
$S_{p\;_{j}}$ represents the post-synaptic neuron’s *previous state*, i.e., the inner state immediately before the new pulse arrives. $A_{_{i}}$ represents the *pre-synaptic amplitude*, which is related to the pre-synaptic neuron, and can be positive or negative depending on whether the neuron sends excitatory or inhibitory connections, respectively.
$W_{_{i,j}}$ represents the *post-synaptic weight* related to the pre-/post-synaptic neuron couple; if this quantity is equal to 0, the related connection is not present. The product $A_{i}\cdot W_{i,j}$ globally represents the amplitude of the pulse arriving to the post-synaptic neuron $n_j$ (i.e., the *synaptic pulse*) from the pre-synaptic neuron $n_i$. In this paper, $w$ or $\omega$ will be used instead of W, depending on the connection is intra- o inter- node, respectively.
$T_l$ (the *leakage term*) takes into account the behaviour of $S$ during two consecutive input pulses in passive mode. The user is allowed to select among two kinds of underthreshold decay: *linear* decay [as in @Mattia2000] or *exponential* decay [as in @Barranca2014], which behaviour is modulated by the *decay parameter* D, as explained in the *Appendix A*.
$T_r$ (the *rise term*) takes into account the overthreshold growth acting upon $S$ during two consecutive input pulses in active mode. Specifically, once the neuron’s inner state crosses the threshold, the neuron is ready to fire. The firing is not instantaneous, but it occurs after a continuous-time delay, representing the spike latency, that we call *time-to-fire* and indicate with $t_f$ in our model. This quantity can be affected by further inputs, making the neuron sensitive to changes in the network spiking activity for a certain time window, until the actual spike generation. $S$ and $t_f$ are related through the following bijective relationship, called the *firing equation*:
$$t_f ={\frac{a}{(S-1)}}-b \; \label{EQ:Fe}$$
where $a,b \ge 0$. Such rectangular hyperbola has been obtained through the simulation of a membrane patch stimulated by brief current pulses (i.e., 0.01 *ms* of duration), solving the Hodgkin-Huxley equations [@Hodgkin1952] in *NEURON* environment [@Neuron], as described in [@Salerno2011]. Then, if the inner state of a neuron is known, the related $t_{f}$ can be exactly calculated by means of Eq. \[EQ:Fe\]. As introduced in \[SECT:SLat\], this nonlinear trend has been observed in most cortical neurons [@Izhikevich2004]; similar behaviors have been also found by other authors, such as [@Wang2013] and [@Trotta2013], using DC inputs. Conversely to previous versions of LIFL [@Cristini2015; @Susi2018], positive constants $a$ and $b$ have been introduced in order to make the model able to encompass the latency curves of a greater number of neuron types; in particular, $a$ allows us to distance/approach the hyperbola to its centre, while $b$ allows us to define a $S_{max}$, conferring a bio-physical meaning to the inner state in active mode (note that if $b=0$, then $S_{max}=\infty$; nevertheless, the neuron will continue to show the spike latency feature).
The firing threshold can be equivalently written as:
$$S_{th} = 1+c \; \label{EQ:Sth}$$
where $c$ is a positive value called *threshold constant*, that fixes a bound for the maximum $t_{f}$. According to Eq. \[EQ:Sth\], when $S = S_{th}$, the $t_{f}$ is maximum, and equal to:
$$\label{EQ:TfM}
t_{f,max} = a/c - b$$
where $t_{f,max}$ represents the upper bound of the time-to-fire. As mentioned above, the latter consideration is crucial in order to have a finite maximum spike latency as in biological neurons [@FitzHugh1955]. From the last equation, we obtain the restriction $c<a/b$.
As described in *Appendix B*, using Eq. \[EQ:Fe\], it is possible to obtain $T_r$ (*rise term*), as follows:
$$T_r =\frac{(S_{p}-1)^{2} \Delta t}{a-(S_{p}-1)\Delta t} \; \label{EQ:Rt}$$
in which $S_{p}$ represents the previous state, whereas $\Delta t$ is the temporal distance between two consecutive incoming pre-synaptic spikes. The Eq. \[EQ:Rt\] allows us to determine the inner state of a neuron at the time that it receives further inputs during the $t_f$ time window. In Fig. \[FIG:LIFLOp\], the operation of LIFL is illustrated, while the effect of Eq. \[EQ:Rt\] is shown in Fig. \[FIG:TfModification\].
![Neural summation and spike generation in a LIFL neuron. (a) Input/output process scheme; (b) temporal diagram of LIFL operation (basic configuration), assuming the neuron starts from its resting potential. Each incoming excitatory (inhibitory) input causes an instantaneous increase (decrease) of the inner state. In passive mode the neuron is affected by a decay; when $S$ exceeds the threshold ($S=S^{+}$) the neuron is ready to spike; due to the latency effect, the firing is not instantaneous but it occurs after $t_{f}$. Once emitted, the pulse of amplitude $A_{i}$ (positive, if the neuron $i$ is excitatory as supposed to be in this case, without loss of generality) is routed to all the subsequent connections. In (c) is shown the firing equation, i.e., the latency curve for the determination of $t_{f}$ from $S^{+}$[see @Salerno2011]. The simplest case of firing equation curve has been chosen ($a=1$, $b=0$), and $c$ set to $0.04$[]{data-label="FIG:LIFLOp"}](2.png){width="70.00000%"}
Assuming that an input spike leads the inner state overthreshold at time $t_{A}$, the arrival of a contribution during the latency time (i.e., at time $t_{B}$) results in a new $t_f$ (i.e., a change of the firing time). Excitatory (inhibitory) inputs increase (decrease) the inner state of a post-synaptic neuron. Therefore, when a neuron is in active mode, excitatory (inhibitory) inputs decrease (increase) the related time-to-fire (*post-trigger anticipation/postponement* respectively). If the inhibitory effect is as strong as to pull the post-synaptic neuron state under the firing threshold, its $t_{f}$ will be suppressed and its state will come back to the passive mode (*post-trigger inhibition*)[@Salerno2011; @Cristini2015].
![Arrival of further inputs when the neuron is overthreshold. (a) The arrival of a new excitatory synaptic pulse at time $t_{B}$ anticipates the spike generation (post-trigger anticipation). The arrival of a new inhibitory synaptic pulse at time $t_{B}$ is able to (b) delay the spike generation (post-trigger postponement), or (c) to cancel the spike generation (post-trigger inhibition). In order to simplify the comparison, the state evolution in active mode in the simple case of no further inputs is reported in the same figure (grey). Neuron $i$ is supposed to be excitatory as in Fig. \[FIG:LIFLOp\].[]{data-label="FIG:TfModification"}](3.png){width="100.00000%"}
For a given neuron $j$ in active mode, the arrival of new input contributions provokes $t_{f}$ updating. Once the $t_{f}$ is reached, the output spike is generated and the inner state is reset. Note that if incoming spikes are such as to bring $S$ to a value $<0$ ($> S_{max}$), $S$ is automatically put to 0 (a spike is immediately generated). We emphasize the fact that spike latency enables a mechanism to encode neural information, supported from all the most plausible models. Thus, there is lack of information in models that do not exhibit this relevant property.
Hitherto we have discussed a basic configuration of LIFL, which defines an intrinsically *class 1 excitable*, *integrator* neuron, supporting *tonic spiking* and *spike latency*. Nevertheless, thanks to the simplicity of its mathematical model, it can be easily enriched with other neuro-computational features to reproduce different kinds of cortical neurons [see @Izhikevich2004] by introducing minimal modifications to the model equations, or by adding extrinsic properties at the programming level. This is the case of *refractory period* for which the neuron becomes insensitive, for a period $t_{arp}$, to further incoming spikes after the spike generation, and *tonic bursting* for which the neuron produces a train of $N_{b}$ spikes interspaced by an interval $IBI$, instead of a single one.
In addition to the spike latency, emerging from the pure computation of the neuron, in the next section another kind of delay will be introduced, independent from the activity, used to characterize the long-range connections between neurons belonging to different groups.
Connection between 2 neurons {#SECT:Connection}
----------------------------
In FNS the network nodes are composed of modules of spiking neurons to represent brain regions. Neurons of the same node interact instantaneously, whereas a settable time delay ($\ge 0$) is present between neurons of different nodes to reflect the remoteness between the regions to which they pertain.
A scheme of inter-node neuron connection ($e_{i,j}$) is illustrated in Fig. \[FIG:InterneuronLinkWithDelay\], where $\lambda_{i,j}$ represents the *axonal length* block and $\omega_{i,j}$ represents the *post-synaptic weight* block. Such two link elements (belonging to a directed connection) are able to introduce delay and amplification/attenuation of the passing pulse, respectively. As in [@Nakagawa2014; @Cabral2014] a global propagation speed $v$ is set for FNS simulations, so that inter-node connection delays are automatically defined from the axonal lengths, as $\tau_{i,j}=\lambda_{i,j} / v$. Connection delays are important since they allow to take into account the three-dimensionality (i.e., spatial embeddedness) of the real anatomical brain networks. For the motivations mentioned before, conversely to the inter-node connection (represented as $e_{i,j}$ in Fig. \[FIG:NodesBundle\]), intra-node connection (represented as $e_{j,k}$ in the same Figure) does not provide the axonal length block (although synaptic weight block continues to be defined).
![Neuron connection model and pulse transfer. (a) compact representation, (b) logical block representation, (c) temporal diagram: length block produces a translation of the output pulse along time axis. Note that in this example the neuron is supposed to be excitatory, otherwise all the amplitudes would be negative. Output pulses can be considered as a correlate of spiking activity, whereas synaptic pulses can be considered as correlate of synaptic currents. Note that in a) and b) the black dot represents the synaptic junction[]{data-label="FIG:InterneuronLinkWithDelay"}](4.png){width="70.00000%"}
For biological and mathematical reasons, it is desirable to keep the synaptic weights under a certain value, *$W_{max}$*, a global parameter of the model.
In the following sections we call *firing event* the pulse emission by a pre-synaptic neuron, and *burning event* the pulse delivery to a post-synaptic neuron.
From brain regions to graph nodes {#SECT:BrainRegionNode}
---------------------------------
FNS allows us to define regions constituted by one or more *nodes* where each node consists of a neuron module with specific properties. In order to reproduce heterogeneous nodes, a Watts-Strogatz based generative procedure is implemented [@Watts1998] as detailed below, allowing the generation of complex networks with structure properties of real neuron populations.
The implemented procedure allows us to model intra- and inter-node diversity: number of neurons and connectivity, percentage of inhibitory neurons, distribution of weights and type of neuron; in addition, it is possible to represent a region with more than one node to model intra-region neuronal pools of different connectivity and neuron types. In the extreme case, a module can be composed of a single neuron, e.g., for reproducing small and deterministic motifs. In the following sections we illustrate the procedure used by FNS for the generation of network nodes and the structure of intra- and inter- node connections.
### Watts-Strogatz-based module generation procedure {#SECT:SWWSproc}
The original Watts-Strogatz procedure is able to generate different types of complex networks (from regular to random), including networks with *small-world* properties (i.e., networks that present large *clustering coefficient* and small *average path length*), that has been demonstrated to reasonably approximate a mid-sized patch of cortex (in the order of $10 \mu m$) with its neighborhood [@Riecke2007]. The original Watts-Strogatz procedure is here adapted to generate a module including both inhibitory and excitatory, oriented, connections, analogously to [@Maslennikov2014]. Given the integer $n$ (i.e., *number of neurons*), $k$ (i.e., *mean degree*), $p$ (i.e., *rewiring probability*), and $R$ (i.e., *excitatory ratio*), with $0\leq p \leq 1$ and $n\gg k \gg ln(n)\gg 1$, the model generates an oriented graph with $n$ vertices and $nk$ single connections in the following way:
- a regular ring lattice of $n$ spiking neurons is created, of which $R\cdot n$ are able to send excitatory connections and the remaining $(1-R)\cdot n$ are able to send inhibitory connections;
- for each neuron an outgoing connection to the closest $k$ neurons is generated ($k/2$ connections for each side, with $k$ integer and even);
- for each neuron $i$, every link $e_{i,j}$ with $i<j$, is rewired with probability $p$; rewiring is done by exchanging $e_{i,j}$ and $e_{i,m}$ where $m$ is chosen with uniform probability from all possible (excitatory or inhibitory) neurons that avoid self-loops ($m\neq i$) and link duplication. This process is repeated $n$ times, each one considering a different neuron.
Note that the $p$ parameter allows to interpolate between a regular lattice ($p=0$) and a random graph ($p=1$): as $p$ increases, the graph becomes increasingly disordered. For intermediate values of $p$ the network presents small-world properties. The parameters $n$, $k$, $p$ allow the user to customize the network nodes on the basis of the real anatomy. For example, $n$ can be chosen in accord to the volume of the region that is intended to be represented (estimated from a specific subject through volumetry, or extracted from existing *atlases*).
### Characterization of intra-module connections
Once connections have been established, weights have to be assigned. Several authors have addressed this problem, setting intra-node weights in different manners. Depending on the specific study, weights have been chosen to have the same, static value [@Deco2012], or characterized by a specific distribution [@Abuhassan2014], or varying in a certain range by means of plasticity [@Izhikevich2004b]. In order to encompass the most of these possibilities, in FNS a set of Gaussian distributed values can be defined by the user for the initialization of the intra-module post-synaptic weights of each module.
From fibre tracts to graph edges
--------------------------------
In FSN an *edge* represents a monodirectional set of long-range axons that links a module to another. In the brain, inter-region connections are often characterized by non negligible delays, which are determined by axon length, diameter and myelination.
![Two nodes connected by an edge. While an intra-node connection is characterized by its weight, an inter-node connection is defined by weight and length; an edge is defined by number of axons and related distribution of weights and lengths. In order to represent the two modules, for illustrative purpose the following values are used: $n_{1}=12$, $ n_{2}=16$; $R_{1}=2/3$, $R_{2}=3/4$; $k_{1} = k_{2} =4$; $p_{1} =1/6$, $p_{2} =1/8$.[]{data-label="FIG:NodesBundle"}](5.png){width="90.00000%"}
### Characterization of inter-region connections
FNS allows the user to set proper number of connections $N_e$ and distributions of weights and lengths can for the network edges. The distribution of edge weights follows a Gaussian function [as in @Abuhassan2014], characterized by the parameters $\mu_{\omega}$ and $\sigma_{\omega}$. Differently, a gamma distribution is implemented for the edge lengths, characterized by mean parameter $\mu_{\lambda}$ and shape parameter $\alpha_{\lambda}$ (see *Appendix C*), since there is probably not a unique prototypical shape for edge delays [as discussed in @Vicente2008]. Indeed, this distribution allows the user to explore different shapes, to investigate the impact of different choices on the network activity, to mimic pathological states as the effect of structural inhomogeneity , or spatially-selective conduction speed decrease due to demyelination [@Smith1994].
When defining inter-region edges, the user can specify the kind of connections to set among the nodes (excitatory, inhibitory, mixed).
STDP {#SECT:STDP}
----
Synaptic plasticity consists of an unsupervised spike-based process able to modify weights on the basis of the network activity. The STDP, a well-known type of plasticity mechanism, is believed to underlie learning and information storage in the brain, and refine neuronal circuits during brain development [@Sjostrom2010]. Considering a synapse connecting two neurons, such mechanism is based on the precise timings of *pre-synaptic pulse* (i.e., the *synaptic pulse* arriving from the pre-synaptic neuron) and *post-synaptic pulse* (i.e., the *output pulse* generated by the post-synaptic neuron), influencing the magnitude and direction of change of the synaptic weight. In case of inter-node connection the pre-synaptic pulse is taken after the axonal delay block and not before, in order to not to alter information on causality between pulse arrival and pulse generation. The original STDP behaviour [@Bi1998] can be approximated by two exponential functions [@Abbott2000].
[W = ]{} \[EQ:STDP\] A\_+ \^[-]{} , T > 0\[EQ:stdpPot\]\
0 , T = 0\
A\_- \^ , T < 0\[EQ:stdpDep\]
where:
- $\Delta T$ is the difference between post-synaptic pulse generation (i.e., $t_{post}$) and pre-synaptic pulse arrival (i.e., $t_{pre}$) instants: $$\label{EQ:Delta_t}
\Delta T = t_{post} - t_{pre}$$ as illustrated in Fig. \[FIG:STDP\]
- $\tau_+$ and $\tau_-$ are positive time constants for *long-term potentiation* (LTP, \[EQ:stdpPot\]) and *long-term depression* (LTD, \[EQ:stdpDep\]), respectively;
- $A_{+}$ and $A_{-}$ are chosen in order to keep weight values bounded between minimum and maximum values (as discussed in Sect. \[SECT:Connection\]).
Then, weight is increased or decreased depending on the pulse order (*pre-*before *post-*, or *post-* before *pre-*, respectively). To make the weight change dependent also on the current weight value, *soft bounds* [@Sjostrom2010] are introduced in FNS, so that $A_+ (W_p) = (W_{max}-W_{p})\eta_+$ and $A_- (W_p)=W_{p} \eta_-$, where $W_{p}$ is the past value of the synaptic weight, $W_{max}$ the upper bound (see \[SECT:Connection\]), and $\eta_+$ and $\eta_-$ are positive learning constants, usually in the order of $\sim10^{-5}$. Therefore, the weight update relations implemented in FNS are:
[W = ]{} W\_[p]{} + (W\_[max]{}- W\_[p]{})\_+ \^[- ]{} , T 0 \[LTP\]\
W\_[p]{} - W\_[p]{}\_- \^ , T < 0 \[LTD\]
It is important to stress that the *soft-bounds* approach allows an increase of both the synaptic capacity and the memory lifetime, with respect to the alternative *hard-bounds* approach [@vanRossum2012].
In addition, to simplify the STDP event list management, exponential tails are suppressed after a certain time value ${TO} \cdot {max}(\tau_+ , \tau_-)$, where *TO* is the *STDP timeout constant*, defined by the user, and usually in the order of 100 ms. In this way, neuron couples whose interval exceeds such time limit are not considered for the STDP process (see Fig. \[FIG:STDP\]).
![STDP process in FNS: (a) $\Delta T$ calculation in relation to the synapse $w_{ij}$, considering an inter-module connection (without loss of generality); (b) Shapes of the learning windows (LTP in green, LTD in red) considering exponential tail suppression (dash dot). []{data-label="FIG:STDP"}](6.png){width="90.00000%"}
STDP varies tremendously across synapse types and brain regions [@Abbott2000; @Caporale2008]. Accordingly, in FNS it is possible to specify a different set of STDP parameters for each node, or to apply STDP uniquely for certain nodes.
Input stimuli {#sec:Input Stimuli}
-------------
Several types of stimuli can be of interest in brain simulation studies. Of these, two prototypical types of stimuli are:
- the noisy fluctuations tipically observed in vivo, which can be modeled by uncorrelated Poisson-distributed spike trains [see @Frohlich2008; @Abuhassan2014; @Nakagawa2014];
- the DC current used by neurophysiologists to test some neuron features [@Izhikevich2004] that can be modeled by constant spike trains [as in @Vicente2008].
In addition, in many simulation scenarios the possibility of giving arbitrary spike streams (e.g., sequences that mimic sensory-like processed data) can be of interest, in order to test the response of specific subnetworks.
In light of these observations, in FNS it is possible to stimulate brain nodes with three different types of inputs: *Poisson-distributed spike train*, *constant spike train*, and *arbitrary spike stream*. The user is allowed to stimulate all or only a part of the network nodes, choosing for each kind of input a customizable number of fictive excitatory *external neurons*, and the characteristics of the required stimuli.
### Poisson-distributed spike train {#subsec:Poisson Distributed Spikes}
This option provides the injection of spike trains distributed according to an homogeneous Poisson process, in which the underlying *instantaneous firing rate* $r_P$ [@Gerstner2014], is constant over time.
Given a long interval ($t_{A}$, $t_{B}$) we place a single spike in that interval at random. Then, considering the sub-interval ($t_1$, $t_2$) of length $\Delta t = t_2 - t_1$, the probability that the spike occurs during this sub-interval is equal to ${\Delta t}/({t_{B}-t_{A}})$. Now, placing $k$ spikes in the long interval, the probability that $n$ of them fall in the sub-interval is given by the following binomial formula:
$$P{\{ \mbox{$n$ spikes during $\Delta t$}\}}=\frac{k!}{(k-n)!n!}p^nq^{k-n} \; \label{bin1Formula}$$
where $p =\delta t/(t_{B}-t_{A})$ and $q = 1 - p$.\
Under proper conditions this expression can be rewritten removing the time dependence as:
$$P{\{ \mbox{1 spike during $\delta t$} \}} \approx r\delta t \; \label{bin2Formula}$$
This equation can be used to generate a Poisson spike train by first subdividing time into a set of short intervals, each one of duration $\delta t$. Then generate a sequence of random numbers **R**, uniformly distributed between 0 and 1. For each interval $i$, a spike is generated if $\textbf{R}(i) \leq r\delta t$. This procedure is appropriate only when $\delta t$ is very small, i.e, only when $r\delta t \ll 1$ [@Heeger00poissonmodel].
In FNS, a user-defined number of fictive *external neurons* $n_{ext P,k}$ is set for each stimulated node $N_k$. By defining a $t_{start P,k}$ and a $t_{end P,k}$ for the external stimuli, each external neuron can send spikes in a discrete number of instants $(t_{start P,k} - t_{end P,k})/ \delta t_P$. The target neurons receive pulses of amplitude $A_{P,k}$.
Pulses are injected from each external neuron to all neurons belonging to a set of nodes defined by the user, by specifying the following set of parameters for each chosen node $N_k$: $n_{ext P,k}$, $t_{start P,k}$, $t_{end P,k}$, $r_{P,k}$, $\delta t_{P,k}$ and $A_{P,k}$.
### Constant spike train {#subsec:DC input}
This option provides the injection of constant spike trains in order to emulate DC current stimulation. Note that since we simulate the network by means of an event-driven approach, the *DC* input is not continuous as in the real counterpart, but it is constantly sampled with an adequately small time step (i.e., smaller than the spike duration) called *interspike interval* and indicated with $int_{\,c}$.
In FNS, a user-defined number of fictive *external neurons* $n_{ext \,c,k}$ is set for each stimulated node $N_k$. Each external neuron can send spikes from time $t_{start \,c,k}$ to $t_{end \,c,k}$, with amplitude $A_{\,c,k}$. Such kind of input is injected from each external neuron to all neurons belonging to a set of nodes defined by the user, by specifying the following set of parameters for each chosen node $N_k$: $n_{ext \,c, k}$,$t_{start \,c,k}$, $t_{end \,c,k}$, $int_{\,c,k}$ and $A_{\,c,k}$.
Note that the situation $int_{\,c,k} < t_{arp,k}$ should be avoided because pulses would arrive during the refractory time.
### Arbitrary spike stream
Arbitrary spike streams can be injected to neurons belonging to a set of nodes defined by the user by specifying the following set of parameters for each chosen node $N_k$: the spike *amplitude* $A_{ss,k}$, and a couple ($n_{ss,k}$, $t_{ss,k}$) for each event to introduce (i.e., *external source number* and related *spike timing*, respectively). External sources are permanently associated to the neurons of the indicated node, using a random procedure.
Output signals
--------------
Depending on the type of contributions we are considering at the network level, i.e., output pulses (corresponding to *action potentials*) or synaptic pulses (corresponding to *post-synaptic currents*), the same network activity gives rise to different signals, due to the presence of connection delays and weights.
In particular, action potentials coincide with the activity emerging from *firing* events (see Sect. \[SECT:LIFLoperation\]), because they take place before the axon, thus they are spatially localized at the emitter node; whereas post-synaptic currents coincide with the post-synaptic activity (see Sect. \[SECT:LIFLoperation\]), because they take place downstream the axon, thus they are spatially localized to the receiver node, and are affected by the shifting effect introduced by (heterogeneous) fibre tract’s delays and post-synaptic weights.
Action potentials are of interest for some studies [see @Vicente2008], whereas post-synaptic currents can be useful for some others (see [@Mazzoni2008; @Nakagawa2014] for LFP and MEG signal reconstruction).
In order to give the user the possibility to recostruct such different types of signals, output section of FNS allows to store both pulse emission and arrival times ($t_F$ and $t_B$), transmitter and receiver neurons ($n_F$ and $n_B$) and related nodes ($N_F$ and $N_B$), as well as amplitude weights ($W_{ev}$) involved in each event occurring during the simulation interval, for some nodes indicated by the user before the simulation starts.
Simulation framework structure
==============================
On the basis of the modelling introduced in the previous section, here we describe the framework structure and the tools it offers to the user for implementing a custom network, stimulating it, and obtaining the outputs of interest.\
The framework is articulated in three main sections : *Generator section*, *Neuroanatomical model section* and *Output section* (see Fig. \[FIG:SIM\]). In order to design a simulation, the user interacts with such sections by means of proper configuration files, that are internally translated into configuration *vectors*, which are defined in table \[TAB:ParDef\].
![FNS framework overall structure. The reader can find the means of the abbreviations in Table \[TAB:ParDef\][]{data-label="FIG:SIM"}](7.png){width="95.00000%"}
[|\*4[c|]{}]{}
**Section**&**Vector**&**Components**&**Name**\
Generator&$\mathbf{PV}$&$n_{ext P}$&number of *Poisson spike train* external neurons\
section&&$t_{start P}$&Poisson input onset\
&&$t_{end P}$&Poisson input offset\
&&$r_P$&firing rate\
&&$\delta t_P$&delta\
&&$A_P$&Poisson input amplitude\
&$\mathbf{CV}$&$n_{ext \, c}$&number of *constant spike train* external neurons\
&&$t_{start{\,c}}$&constant input onset\
&&$t_{end\,c}$&constant input offset\
&&$int_{\,c}$&interspike interval\
&&$A_{\,c}$&constant input amplitude\
&$\mathbf{SV}$&$t_{ss 1}, t_{ss 2}, ...$& *input stream* spike timings\
&&$n_{ss 1}, n_{ss 2}, ...$&related neuron numbers\
&&$A_{ss}$&stream input amplitude\
Neuroanatomical&$\mathbf{LDV}$&$n$&number of neurons\
module&&$p$&rewiring probability\
section&&$k$&mean degree\
&&$R$&excitatory ratio\
&&$A_{exc}$&excitatory pre-synaptic amplitude\
&&$A_{inh}$&inhibitory pre-synaptic amplitude\
&&$\mu_{_w}$&intra-node post-synaptic weight distr.mean (Gaussian)\
&&$\sigma_{_w}$&intra-node post-synaptic weight distr.st.dev. (Gaussian)\
&&$a$&latency curve center distance\
&&$b$&latency curve x-axis intersection\
&&$c$&threshold constant\
&&$D$&decay parameter\
&&$t_{arp}$&absolute refractory period\
&&$N_b$&burst number\
&&$IBI$&inter-burst interval\
&$\mathbf{ICV}$&$N_{e}$&number of connections (edge cardinality)\
&&$\mu_{\omega}$&inter-node post-synaptic weight distr.mean (Gaussian)\
&&$\sigma_{\omega}$&inter-node post-synaptic weight distr.st.dev. (Gaussian)\
&&$\mu_{\lambda}$&inter-node length distr.mean (gamma)\
&&$\alpha_{\lambda}$&inter-node length distr.shape (gamma)\
&$\mathbf{STDPV}$&$\tau_{+}$<P time constant\
&&$\tau_{-}$<D time constant\
&&$\eta_{+}$<P learning constant\
&&$\eta_{-}$<D learning constant\
&&$TO$&STDP timeout constant\
&$\mathbf{GPV}$&$W_{max}$&maximum weight\
&&$v$&global conduction speed\
&&$t_{stop}$&simulation stop time\
Output&$\mathbf{NV}$&$NOI_1, NOI_2, ...$&list of *NOI*s\
section&$\mathbf{FV}$&$n_{F}$&pre-synaptic neuron number (if firing event)\
&&$N_{F}$&pre-synaptic node number (if firing event)\
&&$t_F$&firing event time (if firing event)\
&$\mathbf{BV}$&$n_{B}$&post-synaptic neuron number (if burning event)\
&&$N_{B}$&post-synaptic node number (if burning event)\
&&$t_B$&pulse arrival time (if burning event)\
&&$W_{B}$&synaptic weight (if burning event)\
Generator section
-----------------
This section allows the user to inject the desired input to some selected nodes. *Poisson spike train vectors* (i.e., $\mathbf{PV}$), *constant spike train vectors* (i.e., $\mathbf{CV}$) and *arbitrary spike stream vectors* (i.e., $\mathbf{SV}$) can be combined to send more than a kind of input to the same node simultaneously.
Neuroanatomical model section {#SECT:NaMS}
-----------------------------
This section allows the user to define the network model: local dynamics, structural parameters, plasticity constants and global parameters. Each node is fully characterized by a *local dynamics vector* (i.e., $\mathbf{LDV}$), consisting of *local topology parameters*, *intra-node connection parameters* and *neuron parameters*. From the definition of node’s weight distribution, the simulator computes all the single intra-node synaptic weights and stores them in proper data structures (see Sect. \[SECT:dataStr\]).\
Each edge is fully characterized by a *inter-connectivity vector* (i.e., $\mathbf{ICV}$), consisting of *edge cardinality* and *inter-node weight distribution* and *length distribution parameters*. From the definition of such parameters the simulator computes all the single inter-node lengths and weights and stores them in proper data structures (see Sect. \[SECT:dataStr\]).
Note that values arising from the distribution of inter-node lengths have to be positive to be stored in the internal matrices; in case they assume negative values FNS allows the user to consider the absolute value of such quantities, or to terminate the program execution. In addition, weight values are kept below the value $W_{max}$.\
The *STDP vector* (i.e., $\mathbf{STDPV}$) contains all STDP parameters discussed in Sect. \[SECT:STDP\] and defines the STDP to act on a specific node.\
The global parameters of the system are defined by the *global parameters vector* (i.e., $\mathbf{GPV}$). Among these, $t_{stop}$ specifies the neural activity time we want to simulate in biological time units ($ms$).
Output section {#SECT:OtpS}
--------------
This section allows the user to choose regions and type of contributions for which to extract information. Before the simulation starts, the user can specify the list of nodes for which to store all simulation data (*nodes of interests*, or NOIs) through the *NOI vector* (i.e., $\mathbf{NV}$). At the end of the simulation, data of all firing and burning events in which such NOIs are implicated are available to the user, in form of a vector for each event. Depending whether it is a firing or burning event, in the output we obtain different vectors: *firing event vector* (i.e., $\mathbf{FV}$) or *burning event vector* (i.e., $\mathbf{BV}$), respectively. Such vectors are collected and made available to the user through the two files *firing.CSV* and *burning.CSV*. The former reports exhaustive information on firing events and burning events, respectively, to extract simulated electrophysiological signal (firing activity in the first case and postsynaptic activity in the second case), see Fig. \[FIG:Output\].
![Two seconds of simulated electrophysiological signal extracted from *firing.CSV* and *burning.CSV* files (spike rasterplot and postsynaptic activity, respectively) of a simulation of $N_N$ nodes, composed of 100 neurons each. The figures have been obtained throught the Matlab script available on the FNS website.[]{data-label="FIG:Output"}](8.png){width="98.00000%"}
Implementation aspects
======================
When a simulation is launched in FNS, two phases are performed in sequence: the first phase consists in the *initialization* of the data structures needed by the simulation; the second phase consists in the actual *simulation* of events.
The first phase is carried out through the following steps:
- reading of the *Generator section* and *Neuroanatomical model section* configuration vectors, and the list of NOIs for which the event vectors *FV* and *BV* (of *Output section*) have to be stored;
- creation of the node-specific data structures and neuron clusters to be run concurrently;
- creation of the global data structures.
After those steps have been accomplished, the second phase begins. The parallelization strategy implemented allows the program to proceed through the sequential simulation of single slices of simulated time with a specific and constant duration, for multiple nodes at the same time. Each cycle terminates with the synchronization between nodes whose events affects each other.
Data structures {#SECT:dataStr}
---------------
During the initialization phase the program generates the sets of values of weights and delays of the network. Possible values arising from the distributions that are negative or $> W_{max}$ are notified and rectified (see Sect. \[SECT:NaMS\]). Below we briefly describe the main data structures used by the software, highlighting which of these act at the node level:
**inter-connection** *dictionary*: it is a map containing weight and length of each inter-node connection. The connection is identified by the pre-synaptic neuron (pertaining to node $A$) and the post-synaptic neuron (pertaining to node $B$, with $B \neq A$);
**intra-connection** *dictionary* (node-specific): this is the intra-node equivalent of the *inter-connection* dictionary, where each entry represents the weight of an intra-node connection. The connection is identified by the pre-synaptic neuron and the post-synaptic neuron (pertaining to the same node);
**state** *dictionary* (node-specific): it contains the inner states of the neurons pertaining to a specific node, and it is constantly updated through the node simulation;
**active neuron** *list* (node-specific): list of neurons in active mode pertaining to a specific node, sorted on their firing time; this list is constantly updated through the node simulation;
**outgoing spike** *list* (node-specific): lists of output pulses, including post-synaptic neuron, node and instant, generated from a specific node within a specific time slice;
**STDP timing** *list*: it temporarily stores event timings in order to compute the $\Delta W$. Such timings are automatically discarded after the *TO* value defined by the user.
Event-driven procedure
----------------------
An asynchronous or event-driven algorithm allows the simulation to “jump” from an event to the next one. If the SNN was characterized by identical transmission delays and absence of spike latency, the data structure would be just a *First In First Out* queue, which has fast implementations [@Cormen2001]. Although the neuronal and connectivity features of FNS require more complex data structures (as shown in Sect. \[SECT:dataStr\]), the simulation procedure is quite simple.\
At any instant, each network neuron is characterized by its inner state, and active neurons are also characterized by their proper $t_f$. When a firing event occurs, it propagates toward the target neurons taking into account the connection delays (if present). Such events modify the inner state of post-synaptic neurons (and their $t_f$, for the active neurons), on the basis of the amplitude and sign of the pulse, and the time elapsed from the last state update. Four different cases of state update can happen to the target neuron:
$\cdot$
: *passive-to-passive*. This does not have any effect on the event list
$\cdot$
: *passive-to-active*. This elicits the insertion of an event (and related firing time) orderly on the event list
$\cdot$
: *active-to-active* (i.e., post-trigger anticipation/postponement). This elicits the update (and reordering) of its firing time on the event list
$\cdot$
: *active-to-passive* (i.e., post-trigger inhibition). This elicits the elimination of an event (and related firing time) from the event list
In addition to the four cases listed, two “forbidden” cases can occur during the simulation: from *passive mode* to $S<0$ and from *active mode* to $S\ge S_{max}$; for such cases, specific actions are included in the procedure (state value correction and output spike production, respectively).
At the same time, weights for which plasticity is active are updated accordingly, taking in account the *STDP timing list*.
Parallelization
---------------
In a parallel computing scenario the problem is splitted in many sub-problems such that their solutions can be computed independently and then collected to provide the global solution. In simulations of brain networks it is not trivial to determine which events can be executed in parallel because of the intricate cause-effect relations between the network elements. This led to the development of specific strategies for parallelising event-driven SNNs avoiding causality errors [e.g. @DHaene2006; @Mouraud2009; @Lobb2005; @Grassmann1998; @Djurfeldt2005; @DelormeSPIKENET2003].\
The event-driven parallelization method on which FNS is based on can be defined as *adaptive* [@DHaene2006], since the algorithm chooses an appropriate network-specific interval of simulated time to be used for the synchronization of the parallel tasks, avoiding as much as possible the underuse of available hardware resources. Given a generic network, the *opaque period* (OP) is the minimum simulated time needed by a signal to travel from a network element to an adjacent one [@Lubachevsky1989]. Then, within any simulated time window smaller than the OP of the network, each event cannot be caused by (or cannot affect to) any other event happened during the same time window.
If $ev_i$ and $ev_j$ are two distinct events such as
$$\left|\tau(ev_j)-\tau(ev_i)\right| < OP, \quad i \neq j$$
\
then they can be computed in parallel without loss of cause-effect relationship. This allows us to parallelize the computation within the *time slice* $T_s< OP$. On the other hand, each unit must wait until each of the others has ended to simulate the events of the previous $T_s$ to process new information; then, a *sync step* is needed to “deliver” the events just calculated to the unit which should use them to produce new events during the next *OPs*.
In the case of neural computation, an event could affect another one in a very short time, leading to a short $OP$, then counteracting the benefit of the parallelization. In order to efficiently perform parallel computation, in FNS the following strategy is adopted:
- each node is assigned to a specific *thread* (i.e., a process that deals with a local problem);
- the $T_s$ duration is sized as the minimum among all the network inter-node connection delays (that we define *Bounded Opaque Period* (BOP)), since this is the shortest interval needed by a neuron of a node to affect the state of a neuron of another node.
- inter-node spikes in queue are delivered to the corresponding threads through the *synch step*, at intervals equals to the BOP. Once a thread gets a spike event from a node, it puts this event orderly to the internal node-specific *active neuron list* and updates the internal state of the post-synaptic neuron at the proper time.
If the network presents two or more inter-connected nodes with zero-delay, FNS considers such nodes as a single *group*, and the mechanism continues to be valid. The fact of considering the more ample concept of group instead that of node enables the possibility of representing heterogeneous regions without losing the parallelization feature (see Sect. \[SECT:BrainRegionNode\]).
Threads are executed in parallel by the *thread workers* (i.e., multi-threaded or hyperthreaded CPUs), each of which can execute at most a thread at a time. In order to minimize the processing times, each worker can serve queued threads in turns for a short time, and different workers can swap threads each other, to achieve a dynamic balancing of the computational load.
Once $\mathbf{ICV}s$ (and then, the groups) have been defined, we can simulate the neural activity within each group. Each *thread* can assume one of the three states *running*, *waiting* or *runnable*. When the whole simulation starts, the steps until the completion of the execution are the following:\
1. all threads are set as *runnable*;\
2. each of them simulates the generation of events within the (simulated) $BOP_1$ (i.e., the time window from time $t=0$ to time $t=\bar{B}$). If empty threads occur, for them the simulation jumps directly to point 3;\
3. once a thread has generated all the events within the current $BOP_1$ window, it sets its status to *waiting*;\
4. once all threads have completed to simulate the events of the $BOP_1$ window, all threads synchronize each other through the synch step;\
5. all threads get again to the state *runnable* and simulate the events generated within the $BOP_2$ (i.e., the time window from time $t=\bar{B}$ to time $t=2\bar{B}$), and so on.\
This algorithm cycles until the stop condition set on the overall simulated time $t_{stop}$.\
Obviously, a firing event generated within ${BOP_n}$ not necessarily will be delivered as burning event during ${BOP_{n+1}}$: it could be delivered in one of the following BOPs, depending on the connection delay involved.\
The concept of parallelization through BOP is summarized in Fig. \[FIG:BOP\], where for simplicity we consider the simple *resonance pair* motif [@Gollo2014; @Maslennikov2014] (i.e., two nodes bidirectionally connected with delay).
![BOP-based parallelization mechanism. (a) Example network: two interconnected nodes are generating spiking activity (under proper external stimulation, not shown in the figure). (b) Temporal diagram of the parallelization process. Inter-node pulses are represented by dashed arrows, firing events by circles, burning events by crosses. An inter-node firing event generates burning events in the target node. The delivery of the events to the burning node will happen at the end of the firing event’s BOP. Given a firing event, related burnings will happen in one (or more) following BOPs. Given a burning event, the related firing event belongs to a BOP that precedes the current one (not necessarily the previous one). The axis represents simulated time. (c) Event schedule grid.[]{data-label="FIG:BOP"}](9.png){width="100.00000%"}
In *Appendix D* we report a description in pseudo-code of the procedure implemented in FNS, supporting the synchronization mechanism between nodes.
Reproduction of spontaneous MEG functional connectivity {#SECT:ReprFC}
=======================================================
In order to reproduce subject specific brain’s spontaneous electrophysiological activity using FNS, both structural and functional *connectomes* (i.e., the maps of structural and functional connections within brain areas [@Fornito2016]) have been extracted from a healthy participant using DTI and source-space MEG, respectively. Connectomes have been estimated using 14 regions (7 regions per hemisphere, see table \[TAB:NodDef\]) composing the Default Mode Network [@Raichle2001; @DePasquale2010], a *task-negative* resting state network, which is more strongly active during idling states than during task performance. The participant’s structural connectivity was used to estimate a structural model in the simulator, and its functional connectivity was employed to fine-tune phase and model evaluation.
Simulation setup
----------------
For the simulation we used brain data of a 66 years old male, chosen by lot from the set of control participants of a previous study [@Garces2014].
[|\*4[c|]{}]{} **\#**&**Name**\
1&Left precuneus\
2&Right precuneus\
3&Left isthmuscingulate\
4&Right isthmuscingulate\
5&Left inferiorparietal\
6&Right inferiorparietal\
7&Left superiorfrontal\
8&Right superiorfrontal\
9&Left middletemporal\
10&Right middletemporal\
11&Left anteriorcingulate\
12&Right anteriorcingulate\
13&Left hippocampus\
14&Right hippocampus\
To model spontaneous activity, we adjusted the LIFL model to emulate the behaviour of real pyramidal neurons, with a maximum latency of $25 ms$, chosen on the basis of the data reported in the electrophysiology database *Neuroelectro* [@Neuroelectro2014]. Neuron’s parameters have been set to the following values: $a = 1$, $b = 0$, $c = 0.04$, $t_{arp} = 2$, $D = 0.07$. An external excitatory background input is predisposed consisting of spike trains representing the noisy fluctuations observed in vivo, with amplitude chosen in such a way that an isolated neuron displays predominant spiking activity in the alpha band (i.e., 7.5-12 Hz, the frequency range that characterizes real resting state data [@Abuhassan2014]). Each node is modeled with $n = 100$ neurons for a total of 1400 neurons. $R=0.8$ as revealed by experimental observations [@Izhikevich2004b] and $p = 0.5$ in order to obtain small-world properties. Remaining intra-node connectivity parameters are chosen in such a way that the post-synaptic activity of the nodes preserves its peak in the alpha band, ensuring in the meantime that there is no strict periodicity of individual oscillations, and under the condition $n>>k>>ln(n)>>1$ (as discussed in \[SECT:SWWSproc\]), obtaining $k=30$ and $\mu_{w_{e,i}} =0.04$. Each edge is initialized with a number of connections $N_e / N_{e,max}$ between the considered brain regions, where $N_e$ is equal to the number of streamlines connecting two NOIs reconstructed through DTI and the denominator operates the normalization of the values in the range \[0, 1\]. Distances between regions $\mu_{\lambda}$ have been obtained considering the Hagmann DSI data set [@Hagmann2008], available as a package of *The Virtual Brain* [@SanzLeon2013].
Final structural model counts 45000 connections, for which plasticity is not active since we are reproducing a static scenario. Simulations have been repeated with two free parameters with the goal to match the resulting FC profile with the real MEG data: conduction velocity $v$, which has been varied in the neighborhood of the best-matching value reported in literature [i.e., 5.1 m/s, see @Cabral2014; @Nakagawa2014], and interconnection weights $\mu_{\omega}$ , which we varied in a range that ensures interaction between the nodes but without altering significatively the power spectrum previously set (i.e., \[0.045, 0.09\]). To extract a source-space MEG comparable signal from the model, the *burning event vector* is subsequently imported in Matlab where the events are collected in contiguous bins of 1 ms of simulated time, and the simulated time-series are calculated by summing up all synaptic pulses, as in [@Nakagawa2014].
From the simulated activity, for each combination of $v$ and $\omega$ we discarded initial transients and extracted 5 segments of 6 *s* of activity. Then we processed the trials of simulated activity in the same way of real MEG source space signal, with the method described in [@Garces2014].
Finally, for both simulated and real signal we calculated the *amplitude envelope correlation* FC index [@Brookes2011] between all pairs of nodes considered.
The comparison between MEG and model FC matrices is calculated through the Pearson’s correlation coefficient $r$ between the strictly upper triangular parts of alpha band FC values across all links connecting the 14 regions of interest (as in [@Cabral2014] and [@Nakagawa2014]), after the application of the Fisher-Z transform to the FC measures, due to the non-additivity of correlation coefficients [@Zimmerman2003].
The model shows the best agreement with experimental data for the *optimal values* $\omega = 0.080, v=5.2$, reaching an average correlation of $r = 0.51$ between the empirical and simulated FC profiles in the alpha band (Fig. \[FIG:Results\]) considering the overall set of trials composing the time series, a result that is more than satisfactory if compared to similar studies. Interestingly, although the groups have been tuned such that isolated neurons display predominant spiking activity in the alpha band, the final model (where the groups are interconnected through weights and delays) shows notable correlation values between MEG and model FC values also in other bands. In facts, using the optimal values $\omega=0.085$ and $v=5.2$, we have that $r > 0.35$ both in theta and beta bands, indicating the presence of multi-rhythmic activity, in agreement with real data (the complete set of simulation results is published on the Github page of FNS).
For the synthesis of MEG-like signals from the model, two major simplifications have been made:
- regarding the spatial organization of cortical neurons, we consider that pyramidal cells contribute the most to signal generation, taking into account that it is sufficient to reproduce the post-synaptic currents correctly, as shown in previous work ;
- the resulting signal more directly corresponds to a simulated LFP then to MEG. Nevertheless, a good correspondence between them has been reported [@Nakagawa2014; @Zhu2009] due to the fact that both signals arise from the same process (i.e., post-synaptic currents) [@Buzsaki2012; @Susi2018Ch].
![One second of simulated electrophysiological post-synaptic signal extracted from one node of the model with related alpha-filtered counterpart (a); power spectrum obtained from one of the trials (b); (c) represents the Pearson correlation coefficient between the simulated and the empirical FC as a function of simulation parameters v and $\omega$. The figures have been obtained throught the Matlab script available on the FNS website.[]{data-label="FIG:Results"}](10.png){width="98.00000%"}
Although the comparison shows that the model is successful in reproducing the FC strength measured in real resting state MEG data, FNS offers many possibilities for improving the model in order to achieve a better comparison, i.e., to consider a larger brain network, to use of a higher parcellation resolution, or to enhance diversity among the modules on the basis of real data.
### Performance
Considering the example described above, the average time needed for generating a segment of $6s$ of simulated activity is about $100s$ of real time on an *Intel(R) Core(TM) i7-7700CPU @ 3.60GHz* (ram 16GB). The time needed for the initialization phase is about $2s$; the remaining time is for the simulation phase. While the duration of the first phase depends on the network size only (number of neurons and connections of the network), the duration of the second phase also depends on the value of minimum inter-node length of the network. This time can be reduced by increasing the *maximum heap size* in Java (set to $8GB$ in these simulations). As explained in Sect. \[SECT:OtpS\], during the simulation FNS transcribes $\mathbf{FV}$ and $\mathbf{BV}$ of the selected NOI on the two files *firing.CSV* and *burning.CSV*; the time required to do these operations depends on the number of the selected NOI (14 in this case).
Discussion
==========
Dynamic models of brain networks can help us to understand the fundamental mechanisms that underpin neural processes, and to relate these processes to neural data. Among the different existing approaches, SNN-based brain simulators allow the user to perform a structure-function mapping at the level of single neurons/synapses, offering a multi-level perspective of brain dynamics. Here we have presented *FNS*, the first neural simulation framework based on the LIFL model, which combines spiking/synaptic neural modelling with the event-driven simulation technique, able to support real neuroanatomical schemes. FNS allows us to generate models with heterogeneous regions and fibre tracts (initializable on the basis of real structural data), and synaptic plasticity; in addition, it enables the introduction of various types of stimuli and the extraction of outputs at different network stages, depending on the kind of activity to be reproduced. With the aim of showing to the reader a simulation example, we have synthesized a subject-specific brain model sized on real structural data, and analyzed the network spontaneous activity; the comparison with real MEG data shows that FNS is able to reproduce satisfactorily the patterns of neuromagnetic brain sources. In addition, although the nodes have been tuned to oscillate prominently in the alpha band, in the final model there are quite relevant correlation values between MEG and model FC also in other bands (as theta and beta bands), testifying the presence of multirhythmic activity, in agreement with experimental data. This indicates that the spike latency feature could be a key-aspect to reproduce cross-frequency interactions.\
FNS is downloadable for free and allows the researcher to face realistic simulations, with limited time and budget. The current version of the software is written in Java$ ^{\textregistered}$ and the parallelization is currently implemented as a *multi-threaded* Java standalone application, taking advantage that Java presents specific features oriented to the optimization of memory resources (i.e., the *Java Garbage Collector*), and native support for parallel computation [@JavaForkJoin]. Future improvements include:
- translation of our framework to hybrid CPU-GPU technologies (e.g., *CUDA*) or cloud computing scenarios (e.g., *MapReduce*);
- development of an user-friendly interface and improvement of the compatibility with existent FC estimation tools (e.g., *Hermes* [@Niso2013]);
- development of a version of FNS based on low level programming languages (e.g., *C* or *GO*) to improve performances and compatibility with *high performing computing* platforms;
- considering the use of look-up tables [@Reutimann2003; @Naveros2017] to characterize neuronal dynamics of the LIFL model, for further speeding up simulations. In this process attention must be paid to the insertion of timing errors [@Ros2006] to not to compromise some *temporal coding*-based LIFL applications, as [@Susi2018].
Although a simulation test has been conducted and commented, this document is not intended as user guide, but as an explanation of the mathematical operation and structure of the simulator. The reader can find the software package and technical documentation on the FNS website: [www.fnsneuralsimulator.org](www.fnsneuralsimulator.org)\
Supplementary material
======================
Appendix A: LIFL Features {#appendix-a-lifl-features .unnumbered}
-------------------------
LIFL neuron supports natively the following neurophysiological properties: *integrator*, *spike latency*, *tonic spiking* and *class 1 excitability*, obtained by the implementation of the neuron equations in MATLAB environment. Among these:
- [*Tonic Spiking* takes place when a neuron fires a continuous spike train when stimulated through a DC current input [@Izhikevich2004]. We show this property stimulating a single neuron with a constant spike train (i.e., a discretized DC current input) with amplitude $A_c$. The raster plot of the spiking neuron activity (i.e., the output neuron response) is reported in Fig. \[FIG:StimulatedLIFL\]. Note that the firing frequency of the neuron is constant.]{}
- [*Class 1 Excitability*, exhibited by some cortical neurons, allows the neuron to spike with a frequency that depends on the input strength (ranging from 2 Hz to 200 Hz or more), including the fire at low-frequency when the input is weak. This property allows neurons to encode the input strength into their firing rate [@Izhikevich2004]. In Fig. \[FIG:StimulatedLIFL\], we show this behavior stimulating a neuron with a ramp input. Of course, in this case the firing frequency of the neuron is not constant.]{}
In order to further improve the realism of the LIFL neuron, some adjustments can be made at the programming level (as done for the *refractory period* and the *tonic bursting*), obtaining other computational features as, for example *mixed mode*[@Izhikevich2004].\
With regards to the underthreshold decay, FNS gives the possibility to choose among:
- Linear decay:\
Assuming $T_l = D\Delta t$, in which $D$ is the *decay parameter*, and $\Delta t$ represents the temporal distance between a couple of consecutive incoming spikes (of course, $D \geq$ 0; for $D = 0$ no decay is applied in passive mode, and the neuron behaves as a *perfect integrator*).\
- Exponential decay:\
Assuming $T_l=S_{p\;_{j}}\cdot(1-e^{- \Delta t/D})$, obtaining for the overall update equation: $$S_{j} = A_{_{i}}\cdot W_{_{i,j}} + S_{p\;_{j}} \cdot e^{\frac{-\Delta t} {D}}$$ in which $D$ here represents the classic *time constant*.
Appendix B: $T_{r}$ calculation {#appendix-b-t_r-calculation .unnumbered}
-------------------------------
Referring to Fig. \[Fig:S1\_TrCalculation\], at the time the neurons inner state is altered from a second input (here excitatory, but non influencial to calculation purposes), the *intermediate state* $S_i$ is determined, and then $T_{r}$ is calculated.
![Representation of $T_r$. LIFL neuron in active mode is characterized by a spontaneous growth of $S$. If a pulse arrives before the actual spike generation, $S$ is modified and the $t_{f}$ will be recalculated. The recalculation considers the intermediate state $S_i$, i.e., the neuron state at the time the pulse arrives.[]{data-label="Fig:S1_TrCalculation"}](13.png){width="80.00000%"}
In event-driven, network elements are updated only when they receive or emit a spike. Once an input spike arrives in active mode, the $S_i$ is calculated on the basis of the time remaining to the spike generation.
Referring to the generic inner state $S_{i}$ the firing equation is:
$$\label{EQ:tfi}
t_{f,i}=\frac{a}{S_i-1}-b$$
We define:
$$\Delta t=t_2-t_1$$
where $t_1$ and $t_2$ represent the arrival instants of the synaptic pulses to the considered neuron. Then:
$$\label{EQ:tfiNEW}
t_{f,i}=t_{f,1}-\Delta t$$
Rearranging Eq. \[EQ:tfi\], we obtain: $$\label{EQ:SiNEW}
S_i=\frac{a}{t_{f,i}+b} +1$$
Now we combine Eq. \[EQ:tfiNEW\] with Eq. \[EQ:SiNEW\] $$\label{EQ:SiNEW2}
S_i=\frac{a}{t_{f,1}-\Delta t + b}+1$$
By defining $$\label{EQ:Treq}
T_r = S_i - S_1$$
where $$\label{EQ:S1}
S_1=\frac{a}{(t_{f,1} + b)} + 1$$
and putting Eq .\[EQ:SiNEW2\] and \[EQ:S1\] in \[EQ:Treq\], we obtain:
$$T_r=\frac{a}{t_{f,1}-\Delta t + b}-\frac {a}{t_{f,1} + b}$$
that can be rearranged as
$$\label{EQ:DeltaSNEW}
T_r=\frac{a \Delta t}{(t_{f,1} + b -\Delta t) (t_{f,1}+b) }$$
Note that we are interested in determining an intermediate state; this implies that we consider the second synaptic pulse only if its timing (i.e., $t_2$) falls before the spike occurs. This gives us: $$\label{EQ:condition}
\Delta t < t_{f,1}$$ thus we do not have restrictions from the denominator of \[EQ:DeltaSNEW\].
The relation \[EQ:DeltaSNEW\] can be generalized to the case as more input modify the firing time; then, we can write $$T_r =S_{ic}-S_{ip}=\frac{a \Delta t_i}{(t_{f,ip} + b -\Delta t_i)(t_{f,ip}+b) }$$
with
$$\Delta t_i=t_{ic}-t_{ip}$$
where the subscript $ip$ stays for *intermediate-previous* and $ic$ for *intermediate-current*.
We can also make explicit the dependence of $T_r$ from the previous state, by inverting $t_{f,ip}$ trough Eq. \[EQ:tfi\], obtaining:
$$\label{EQ:DeltaSNEW_doubleintermediate}
T_r=\frac{(S_{ip}-1)^2 \Delta t}{a-(S_{ip}-1)\Delta t}$$
Obviously, the same considerations on the arrival time of the second pulse remain valid, thus we do not have restrictions imposed by the denominator of \[EQ:DeltaSNEW\_doubleintermediate\].
Appendix C: types of distributions used in the model {#appendix-c-types-of-distributions-used-in-the-model .unnumbered}
----------------------------------------------------
In FNS Gaussian distributions are implemented both for the initialization of intra-module weights of modules (one set for excitatory and one set for inhibitory) and inter-module weights of each edge:
$$f(W) =\frac{1}{{\sqrt{2 \pi {\sigma_{W}} ^2 } } } exp{\frac {(W-\mu_{W})^2} {-2 {\sigma_{W}} ^2}} \; \label{EQ:Gw}$$
where $\mu_{W}$ is the mean, and $\sigma_{W} ^{2}$ is the variance of the distribution. In the formula, W is intended to represent $A$ (distribution of intra-module weights) or $\omega$ (distribution of intermodule weights).
In time, a gamma distribution is implemented for inter-module lengths, which reflects on a gamma distribution of delays. Such gamma distribution is characterized from parameters $\mu_{\lambda}$ (i.e., *mean parameter*) and $\alpha_{\lambda}$ (i.e., *shape parameter*). If we call $\lambda$ the (axonal) delay, the probability density function of a gamma distribution can be written as:
$$f(\lambda) = \lambda^{\alpha_{\lambda}-1} \frac {exp(-\lambda \alpha_{\lambda}/\mu_{\lambda})}{(\mu_{\lambda}/\alpha_{\lambda})^{\alpha_{\lambda}} \Gamma (\alpha_{\lambda})} \; \label{GammaFunct1}$$
Note that $\mu_{\lambda}$ can be defined as:
$$\mu_{\lambda} =\alpha_{\lambda} \theta\; \label{GammaFunct2}$$
where $\theta$ is known as the *scale parameter*.\
Note that with the parameter $\alpha_{\lambda}$ is possible to control the type of distribution (low $\alpha_{\lambda}$ values lead toward the exponential distribution; high $\alpha_{\lambda}$ values lead toward the Dirac distribution); the more $\alpha_{\lambda}$ is high, the more the distribution *mode* approaches $\mu_{\lambda}$ (from the left).
Appendix D: Pseudo-code procedure {#appendix-d-pseudo-code-procedure .unnumbered}
---------------------------------
$\cdot$
: let `spike_queue` be the list of all spikes generated by any pre-synaptic neuron sorted in ascending order of spike time (the first spike in the list is the one with the least spike time);
$\cdot$
: let `outer_burning_event`: the list of spikes to be delivered to post-synaptic neurons belonging to outer nodes. At each synchronization:\
1. each entry of this list is read and sent to the thread executing the routine for the right node;\
2. after each item in the list has been sent, the whole list is cleared.
$\cdot$
: let `current_time` be the current simulated time;
$\cdot$
: let `split_stop_time` be the stop time for the current BOP simulation;
$\cdot$
: let `final_stop_time` be the simulated time at which the whole simulation must stop;
$\cdot$
: let `run_burning_routine(s*)` be the routine which calculates all the burning events caused by the fire event hold by the spike `s*`: during this procedure, all the burning events involving post-synaptic neurons of outer nodes are stored to a special `outer_burning_event`;
$\cdot$
: let `send_fires_for_outer_nodes()` be the routine which sends the spikes stored in `outer_burning_event` to post-synaptic neurons in outer nodes;
$\cdot$
: let `update_incoming_spikes_queue()` be the routine which updates the `spike_queue` with the spikes coming from outer nodes and targeting post-synaptic neurons of the present node before beginning the next BOP simulation.
<!-- -->
while true:
while current_time < split_stop_time:
s* = spike_queue.pop()
run_burning_routine(s*)
if current_time >= split_stop_time: (1)
if split_stop_time >= final_stop_time: (2)
return
send_fires_for_outer_nodes()
wait_until(current_time < split_stop_time) (3)
update_spikes_queue()
(1) end of the BOP: send the fires to the outer nodes
(2) end of the last BOP: end of the node simulation
(3) stops the thread until split_stop_time is updated
with the next BOP stop time
Information Sharing Statement {#information-sharing-statement .unnumbered}
-----------------------------
Software, simulation results and new network models are available at the\
FNS official website \[www.fnsbrainsimulator.org\] and GitHub\
page \[https://github.com/fnsneuralsimulator\].\
References {#references .unnumbered}
==========
[50]{}
urlstyle \[1\][doi: \#1]{}
P. Hagmann. *From diffusion [MRI]{} to brain connectomics*. issertation, Lausanne: EPFL, 2005. URL <https://infoscience.epfl.ch/record/33696>.
C. Sporns, G. Tononi, and R. Kötter. The human connectome: A structural description of the human brain. *PLOS Computational Biology*, 10 (4), 09 2005. URL <https://doi.org/10.1371/journal.pcbi.0010042>.
A. Barardi, B. Sancristóbal, and J. Garcia-Ojalvo. Phase-coherence transitions and communication in the gamma range between delay-coupled neuronal populations. *PLOS Computational Biology*, 100 (7):0 e22561, 2014. [doi: ]{}[10.1371/journal.pcbi.1003723]{}. URL <https://doi.org/10.1371/journal.pcbi.1003723>.
J. Cabral, E. Hugues, O. Sporns, and G. Deco. Role of local network oscillations in resting-state functional connectivity. *Neuroimage*, 570 (1), 2011.
G. Deco and V.K. Jirsa. Ongoing cortical activity at rest: Criticality, multistability, and ghost attractors. *The Journal of Neuroscience*, 320 (10):0 3366–3375, 2012.
RG Bettinardi, G Deco, VM Karlaftis, TJ Van Hartevelt, HM Fernandes, Z Kourtzi, ML Kringelbach, and G. Zamora-López. How structure sculpts function: Unveiling the contribution of anatomical connectivity to the brain’s spontaneous correlation structure. *Chaos*, 270 (4), 2017.
T.T. Nakagawa, M. Woolrich, H. Luckhoo, M. Joensson, H. Mohseni, M.L. Kringelbach, V Jirsa, and G. Deco. How delays matter in an oscillatory whole-brain spiking-neuron network model for [MEG]{} alpha-rhythms at rest. *Neuroimage*, 87:0 383–394, 2014. URL <https://doi.org/10.1016/j.neuroimage.2013.11.009>.
J. Cabral, H. Luckhoo, M. Woolrich, M. Joensson, H. Mohseni, A. Baker, M.L. Kringelbach, and G. Deco. Exploring mechanisms of spontaneous functional connectivity in [MEG]{}: How delayed network interactions lead to structured amplitude envelopes of band-pass filtered oscillations. *Neuroimage*, 90:0 423–435, 2014. URL <https://doi.org/10.1016/j.neuroimage.2013.11.047>.
Gustavo Deco, Joana Cabral, Mark W. Woolrich, Angus B.A. Stevner, Tim J. van Hartevelt, and Morten L. Kringelbach. Single or multiple frequency generators in on-going brain activity: A mechanistic whole-brain model of empirical meg data. *Neuroimage*, 152:0 538–550, 2017.
G. Deco, V.K. Jirsa, P. A. Robinson, M. Breakspear, and K.l. Friston. The dynamic brain: From spiking neurons to neural masses and cortical fields. *PLOS Comput Biol*, 4:0 1–35, 2008. URL <https://doi.org/10.1371/journal.pcbi.1000092>.
R. Vicente, L. L. Gollo, C. R. Mirasso, I. Fischer, and G. Pipa. Dynamical relaying can yield zero time lag neuronal synchrony despite long conduction delays. *Proceedings of the National Academy of Sciences USA*, 1050 (44):0 17157–17162, 2008.
L Gollo, C Mirasso, and A Villa. Dynamic control for synchronization of separated cortical areas through thalamic relay. *NeuroImage*, 520 (3):0 947–955, 2010. [doi: ]{}[10.1016/j.neuroimage.2009.11.058]{}. URL <https://doi.org/10.1016/j.neuroimage.2009.11.058>.
Oleg V. Maslennikov and Vladimir I. Nekorkin. Modular networks with delayed coupling: Synchronization and frequency control. *Phys. Rev. E*, 90:0 012901, Jul 2014. [doi: ]{}[10.1103/PhysRevE.90.012901]{}.
JW Bohland, C Wu, H Barbas, H Bokil, M Bota, HC Breiter, HT Cline, JC Doyle, PJ Freed, RJ Greenspan, SN Haber, M Hawrylycz, DG Herrera, CC Hilgetag, ZJ Huang, A Jones, EG Jones, HJ Karten, D Kleinfeld, R Kötter, HA Lester, JM Lin, BD Mensh, S Mikula, J Panksepp, JL Price, J Safdieh, CB Saper, ND Schiff, JD Schmahmann, BW Stillman, K Svoboda, LW Swanson, AW Toga, DC Van Essen, JD Watson, and PP Mitra. A proposal for a coordinated effort for the determination of brainwide neuroanatomical connectivity in model organisms at a mesoscopic scale. *PLOS Computational Biology*, 50 (3):0 1–9, 2009. [doi: ]{}[10.1371/journal.pcbi.1000334]{}. URL <https://doi.org/10.1371/journal.pcbi.1000334>.
E. M. Izhikevich. Which model to use for cortical spiking neurons? *IEEE Transaction on Neural Networks*, 150 (5):0 1063–1070, 2004.
R. Brette, M. Rudolph, T. Carnevale, H. Hines, D. Beeman, J. M. Bower, M. Diesmann, A. Morrison, P. H. Goodman, F. C. Jr Harris, M. Zirpe, T. Natschläger, D. Pecevski, B. Ermentrout, M. Djurfeldt, A. Lansner, O. Rochel, T. Vieville, E. Muller, A. P. Davison, S. El Boustani, and A. Destexhe. Simulation of networks of spiking neurons: [A]{} review of tools and strategies. *Journal of Computational Neuroscience*, 230 (3):0 349–398, 2007.
A. Hanuschkin, S. Kunkel, M. Helias, A. Morrison, and M. Diesmann. A general and efficient method for incorporating precise spike times in globally time-driven simulations. *Frontiers in Neuroinformatics*, 40 (113), 2010. [doi: ]{}[10.3389/fninf.2010.00113]{}. URL <https://doi.org/10.3389/fninf.2010.00113>.
J Krishnan, P Porta Mana, M Helias, M Diesmann, and E. Di Napoli. Perfect detection of spikes in the linear sub-threshold dynamics of point neurons. *Frontiers in Neuroinformatics*, 110 (75), 2017.
Anthony Mouraud and Didier Puzenat. Simulation of large spiking neural networks on distributed architectures. the “[DAMNED]{}” simulator. In Dominic Palmer-Brown, Chrisina Draganova, Elias Pimenidis, and Haris Mouratidis, editors, *Engineering Applications of Neural Networks.*, pages 359–370. Springer International Publishing, 2009. ISBN 978-3-642-03968-3. [doi: ]{}[10.1007/978-3-642-03969-0\_33]{}.
E. Ros, R. Carrillo, E.M. Ortigosa, B. Barbour, and R. Agís. Event-driven simulation scheme for spiking neural networks using lookup tables to characterize neuronal dynamics. *Neural Computation*, 180 (12):0 2959–2993, 2006.
R. Brette. Exact simulation of integrate-and-fire models with synaptic conductances. *Neural Computation*, 180 (8):0 2004–2027, 2006.
R. Brette. Exact simulation of integrate-and-fire models with exponential currents. *Neural Computation*, 190 (10):0 2604–2609, 2007.
A Tonnelier, H Belmabrouk, and D Martinez. Event-driven simulations of nonlinear integrate-and-fire neurons. *Neural Comput.*, 190 (12):0 1426–1461, 2007. [doi: ]{}[10.1162/NECO\_a\_00278]{}.
M. Salerno, G. Susi, and A. Cristini. Accurate latency characterization for very large asynchronous spiking neural networks. In M. Pellegrini, A. L. N. Fred, J. Filipe, and H. Gamboa, editors, *BIOINFORMATICS 2011 - Proceedings of the International Conference on Bioinformatics Models, Methods and Algorithms*, pages 116–124. SciTePress, 2011.
M Rudolph-Lilith, M Dubois, and A Destexhe. Analytical integrate-and-fire neuron models with conductance-based dynamics and realistic postsynaptic potential time course for event-driven simulation strategies. *Neural Comput.*, 240 (6):0 1426–1461, 2012. [doi: ]{}[10.1162/NECO\_a\_00278]{}.
D. Pecevski, D. Kappel, and Z. Jonke. : event-driven neural simulation framework with a python interface. *Frontiers in Neuroinformatics*, 80 (70), 2014. ISSN 1662-5196. [doi: ]{}[10.3389/fninf.2014.00070]{}. URL <http://journal.frontiersin.org/article/10.3389/fninf.2014.00070>.
A. Cristini, M. Salerno, and G. Susi. A continuous-time spiking neural network paradigm. In S. Bassis, A. Esposito, and F. C. Morabito, editors, *Advances in Neural Networks: Computational and Theoretical Issues*, pages 49–60. Springer International Publishing, 2015. ISBN 978-3-319-18163-9. [doi: ]{}[10.1007/978-3-319-18164-6\_6]{}.
A Morrison, S Straube, H Plesser, and M Diesmann. Exact subthreshold integration with continuous spike times in discrete time neural network simulations. *Neural Computation*, 19:0 47–79, 2006.
M D’Haene, M Hermans, and B Schrauwen. Toward unified hybrid simulation techniques for spiking neural networks. *Neural Comput.*, 260 (6):0 1055–79, 2014. [doi: ]{}[10.1162/NECO\_a\_00587]{}.
M. Gewaltig and M. Diesmann. ([N]{}[E]{}ural [S]{}imulation [T]{}ool). *Scholarpedia*, 20 (4):0 1430, 2007.
R. Brette and D. Goodman. *Brian Documentation. Release 1.4.3*, 2016. URL <http://www.briansimulator.org>.
Gianluca Susi, Luis Antón Toro, Leonides Canuet, Maria Eugenia López, Fernando Maestú, Claudio R. Mirasso, and Ernesto Pereda. A neuro-inspired system for online learning and recognition of parallel spike trains, based on spike latency, and heterosynaptic stdp. *Frontiers in Neuroscience*, 12:0 780, 2018[[a]{}]{}. ISSN 1662-453X. [doi: ]{}[10.3389/fnins.2018.00780]{}.
G. C. Cardarilli, A. Cristini, L. Di Nunzio, M. Re, M. Salerno, and G. Susi. Spiking neural networks based on [LIF]{} with latency: Simulation and synchronization effects. In *2013 Asilomar Conference on Signals, Systems and Computers*, pages 1838–1842, Pacific Grove, CA, USA, 2013. IEEE.
G. Susi. Bio-inspired temporal-decoding network topologies for the accurate recognition of spike patterns. *Transactions on Machine Learning and Artificial Intelligence*, 30 (4):0 27–41, 2015. [doi: ]{}[10.14738/tmlai.34.1438]{}.
G. Susi, A. Cristini, and M. Salerno. Path multimodality in [F]{}eedforward [SNN]{} module, using [LIF]{} with latency model. *Neural Network World*, 260 (4):0 363–376, 2016.
S. Acciarito, G.C. Cardarilli, A. Cristini, L. Di Nunzio, R. Fazzolari, G.M. Khanal, M. Re, and G. Susi. Hardware design of [LIF]{} with latency neuron model with memristive [STDP]{} synapses. *Integration, the VLSI Journal*, 59:0 81–89, 2017. ISSN 0167-9260.
R. FitzHugh. Mathematical models of threshold phenomena in the nerve membrane. *Bulletin of Mathematical Biology*, 170 (4):0 257–278, 1955.
JP Thivierge. Neural diversity creates a rich repertoire of brain activity. *Communicative & Integrative Biology*, 1:0 188–189, 2008.
LL. Gollo, M. Copelli, and J. A. Roberts. Diversity improves performances in excitable networks. *PeerJ*, 4:0 e1912, 2016. URL <https://doi.org/10.7717/peerj.1912>.
N Brunel and V. Hakim. Fast global oscillations in networks of integrate-and-fire neurons with low firing rates. *Neural computation*, 110 (7):0 1621–71, 1999.
N Brunel and X.J. Wang. What determines the frequency of fast network oscillations with irregular neural discharges? i. synaptic dynamics and excitation-inhibition balance. *Journal of Neurophysiology*, 900 (1):0 415–30, 2003.
A Viriyopase, I Bojak, M Zeitler, and S Gielen. When long-range zero-lag synchronization is feasible in cortical networks. *Frontiers in computational neuroscience*, 6:0 49, 2012. URL <https://www.frontiersin.org/article/10.3389/fncom.2012.00049>.
LL Gollo, C Mirasso, O Sporns, and M Breakspear. Mechanisms of zero-lag synchronization in cortical motifs. *PLOS computational biology*, 100 (4):0 1–17, 2014. [doi: ]{}[10.1371/journal.pcbi.1003548]{}. URL <https://doi.org/10.1371/journal.pcbi.1003548>.
K Abuhassan, D Coyle, and L Maguire. Compensating for thalamocortical synaptic loss in [A]{}lzheimer’s disease. *Frontiers in computational neuroscience*, 80 (65), 2014. ISSN 1662-5188. [doi: ]{}[10.3389/fncom.2014.00065]{}. URL <https://doi.org/10.3389/fncom.2014.00065>.
E. M. Izhikevich. Simple model of spiking neurons. *IEEE Transaction on Neural Networks*, 140 (6):0 1569–1572, 2003.
H. Wang, Y. Chen, and Y. Chen. First-spike latency in hodgkin’s three classes of neurons. *Journal of Theoretical Biology*, 328:0 19–25, 2013.
L Trotta, A Franci, and R Sepulchre. First spike latency sensitivity of spiking neuron models. *BMC Neuroscience*, 140 (1):0 354, 2013. ISSN 1471-2202.
E. M. Izhikevich. *Dynamical systems in neuroscience: the geometry of excitability and bursting*. Computational neuroscience. MIT Press, Cambridge, Mass., London, 2007.
Tim Gollisch and Markus Meister. Rapid neural coding in the retina with relative spike latencies. *Science*, 3190 (5866):0 1108–1111, 2008. ISSN 0036-8075. [doi: ]{}[10.1126/science.1149639]{}. URL <http://science.sciencemag.org/content/319/5866/1108>.
B Fontaine and H Peremans. Bat echolocation processing using first-spike latency coding. *Neural Networks*, 220 (10):0 1372 – 1382, 2009. ISSN 0893-6080.
Rafael D. Vilela and Benjamin Lindner. Comparative study of different integrate-and-fire neurons: Spontaneous activity, dynamical response, and stimulus-induced correlation. *Phys. Rev. E*, 80:0 031909, Sep 2009. [doi: ]{}[10.1103/PhysRevE.80.031909]{}. URL <http://link.aps.org/doi/10.1103/PhysRevE.80.031909>.
N Fourcaud-Trocme, D Hansel, C van Vreeswijk, and N Brunel. How spike generation mechanisms determine the neuronal response to fluctuating inputs. *The Journal of Neuroscience*, 230 (5):0 11628–11640, 2003. ISSN 1529-2401.
M. Mattia and P. Del Giudice. Efficient event-driven simulation of large networks of spiking neurons and dynamical synapses. *Neural Computation*, 120 (10):0 2305–2329, 2000.
V.J. Barranca, D.C. Johnson, J.L. Moyher, J.P. Sauppe, M.S. Shkarayev, G. Kovacic, and D. Cai. Dynamics of the exponential integrate-and-fire model with slow currents and adaptation. *Journal of Computational Neuroscience*, 370 (1):0 161–180, 2014.
A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane current and application to conduction and excitation in nerve. *The Journal of Physiology*, 1170 (4):0 500–544, 1952.
M. L. Hines and N. T. Carnevale. The [NEURON]{} simulation environment. *Neural Computation*, 90 (6):0 1179–1209, 1997.
J. D. Watts and S. H. Strogatz. Collective dynamics of ’small-world’ networks. *Nature*, 3930 (1):0 440–442, 1998.
H Riecke, A Roxin, S Madruga, and SA Solla. Multiple attractors, long chaotic transients, and failure in small-world networks of excitable neurons. *Chaos*, 170 (2), 2007. URL <https://doi.org/10.1063/1.2743611>.
EM Izhikevich, JA Gally, and GM Edelman. Spike-timing dynamics of neuronal groups. *Cereb Cortex.*, 140 (8):0 933–944, 2004.
R Ton, G Deco, and A Daffertshofer. Structure-function discrepancy: Inhomogeneity and delays in synchronized neural networks. *PLOS Computational Biology*, 100 (7):0 1–15, 2014. [doi: ]{}[10.1371/journal.pcbi.1003736]{}.
KJ Smith. Conduction properties of central demyelinated and remyelinated axons, and their relation to symptom production in demyelinating disorders. *Nature (eye)*, 80 (2):0 224–237, 1994.
J. Sjöström and W. Gerstner. Spike-timing dependent plasticity. <http://www.scholarpedia.org/article/Spike-timing_dependent_plasticity>, 2010.
G.Q. Bi and M.M. Poo. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. *The Journal of Neuroscience*, 180 (24):0 10464–10472, 1998.
L. F. Abbott and S. B. Nelson. Synaptic plasticity: taming the beast. *Nature Neuroscience*, 3:0 1178–1183, 2000.
M. C. van Rossum, M. Shippi, and A. B. Barrett. Soft-bound synaptic plasticity increases storage capacity. *PLOS Computational Biology*, 80 (12):0 1–11, 2012. URL <https://doi.org/10.1371/journal.pcbi.1002836>.
N Caporale and Y Dan. Spike timing-dependent plasticity: a [H]{}ebbian learning rule. *Annu Rev Neurosci.*, 31:0 25–46, 2008.
F Frohlich, M Bazhenov, and T J Sejnowski. Pathological effect of homeostatic synaptic scaling on network dynamics in diseases of the cortex. *The Journal of Neuroscience The Official Journal of the Society for Neuroscience*, 280 (7):0 1709–1720, 2008. [doi: ]{}[10.1523/JNEUROSCI.4263-07.2008]{}.
W. Gerstner, W.M. Kistler, R. Naud, and L. Paninski. *Neuronal Dynamics. From single neurons to networks and models of cognition*. Cambridge University Press, 2014.
David Heeger and Professor David Heeger. Poisson model of spike generation, 2000. URL <http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.37.6580>.
A. Mazzoni, S. Panzeri, N.K. Logothetis, and N. Brunel. Encoding of naturalistic stimuli by local field potential spectra in networks of excitatory and inhibitory neurons. *PLOS Computational Biology*, 40 (12):0 1–20, 12 2008. URL <https://doi.org/10.1371/journal.pcbi.1000239>.
T. Cormen, C. Leiserson, R. Rivest, and C. Stein. *Introduction to algorithms, 2nd ed.* The MIT Press, 2001.
Michiel D’Haene. A framework for parallel event driven simulation of large spiking neural networks. In *7e FirW Doctoraatssymposium*, 2006.
CJ Lobb, Z Chao, Fujimoto, RM, and SM Potter. Parallel event-driven neural network simulations using the hodgkin-huxley neuron model. In *Principles of Advanced and Distributed Simulation (PADS)*, pages 16–25, 2005. [doi: ]{}[10.1109/PADS.2005.18]{}.
C Grassmann and JK Anlauf. Distributed, event driven simulation of spiking neural networks. In *Proceedings of the International [ICSC]{} - [IFAC]{} Symposium on Neural Computation*, pages 100–105, 1998.
M Djurfeldt, C Johansson, O Ekeberg, M Rehn, M Lundqvist, and A Lansner. Massively parallel simulation of brain-scale neural network models. Technical report, “KTH” School of Computer Science and Communication, Stockholm, Sweden, 2005.
A Delorme and SJ Thorpe. pike[NET]{}: an event-driven simulation package for modelling large networks of spiking neurons. *Network Computation in Neural Systems*, 140 (4):0 613–627, 2003.
B.D. Lubachevsky. Efficient distributed event-driven simulations of multiple-loop networks. *Communications of the ACM*, 320 (1):0 111–123, 1989.
A. Fornito, A. Zaleski, and E.T. Bullmore. *Fundamentals of brain network analysis*. Elsevier, 2016.
Marcus Raichle, Ann Mary MacLeod, Abraham Z. Snyder, William J. Powers, Debra A. Gusnard, and Gordon L. Shulman. A default mode of brain function. *Proceedings of the National Academy of Sciences USA*, 980 (2):0 676–682, 2001.
Francesco de Pasquale, Stefania Della Penna, Abraham Z. Snyder, Christofer Lewis, Dante Mantini, Laura Marzetti, Paolo Belardnelli, Luca Ciancetta, Vittorio Pizzella, Gian Luca Romani, and Maurizio Corbetta. Temporal dynamics of spontaneous [MEG]{} activity in brain networks. *Proceedings of the National Academy of Sciences USA*, 1070 (13):0 6040–6045, 2010.
Pilar Garces, Jose-Angel Pineda, Leonides Canuet, Sara Aurtenetxe, Maria Eugenia Lopez, Alberto Marcos, Miguel Yus, Marcos Llanero-Luque, Francisco del Pozo, Miguel Sancho, and Fernando Maest[ú]{}. The [D]{}efault [M]{}ode [N]{}etwork is functionally and structurally disrupted in amnestic mild cognitive impairment - a bimodal [MEG]{} [DTI]{} study. *Neuroimage Clin.*, 6:0 214–221, 2014. URL <https://doi.org/10.1016/j.nicl.2014.09.004>.
B Fischl. Free[s]{}urfer. *Neuroimage*, 620 (2):0 774–781, 2012. [doi: ]{}[10.1016/j.neuroimage.2012.01.021]{}. URL <https://doi.org/10.1016/j.neuroimage.2012.01.021>.
R.S. Desikan, F Segonne, B Fischl, B.T. Quinn, B.C. Dickerson, D. Blacker, R.L. Buckner, A.M. Dale, R.P. Maguire, B.T. Hyman, M.S. Albert, and R.J. Killiany. An automated labeling system for subdividing the human cerebral cortex on [MRI]{} scans into gyral based region of interests. *Neuroimage*, 31:0 968–980, 2006. URL <https://doi.org/10.1016/j.neuroimage.2006.01.021>.
P Hagmann, L Cammoun, X Gigandet, R Meuli, C Honey, VJ Wedeen, and O Sporns. Mapping the structural core of human cerebral cortex. *Plos Biology*, 60 (7):0 1–15, 2008. URL <https://doi.org/10.1371/journal.pbio.0060159>.
Shreejoy J. Tripathy, Judith Savitskaya, Shawn D. Burton, Nathaniel N. Urban, and Richard C. Gerkin. Neuro[E]{}lectro: a window to the world’s neuron electrophysiology data. *Frontiers in Neuroinformatics*, 8:0 40, 2014. ISSN 1662-5196. [doi: ]{}[10.3389/fninf.2014.00040]{}. URL <http://journal.frontiersin.org/article/10.3389/fninf.2014.00040>.
P Sanz Leon, SA Knock, M Marmaduke Woodman, L Domide, J Mersmann, AR McIntosh, and V Jirsa. The virtual brain: a simulator of primate brain network dynamics. *Frontiers in Neuroinformatics.*, 7:0 10, 2013. ISSN 1662-5196. URL <https://www.frontiersin.org/article/10.3389/fninf.2013.00010>.
M.J. Brookes, J.R. Hale, J.M. Zumer, C.M. Stevenson, S.T. Francis, G.R. Barnes, J.P. Owen, P.G. Morris, and S.S. Nagarajan. Measuring functional connectivity using [MEG]{}: Methodology and comparison with fc[MRI]{}. *Neuroimage*, 560 (3):0 1082–1104, 2011. URL <https://doi.org/10.1016/j.neuroimage.2011.02.054>.
DW Zimmerman, BD Zumbo, and RH Williams. Bias in estimation and hypothesis testing of correlation. *Psicologica*, 24:0 133–158, 2003.
Zhao Zhu, Johanna M. Zumer, Marianne E. Lowenthal, Jeff Padberg, Gregg H. Recanzone, Leah A. Krubitzer, Srikantan S. Nagarajan, and Elizabeth A. Disbrow. The relationship between magnetic and electrophysiological responses to complex tactile stimuli. *BMC Neuroscience*, 100 (1):0 4, 2009. ISSN 1471-2202. URL <https://doi.org/10.1186/1471-2202-10-4>.
G. Buzsaki, C.A. Anastassiou, and C. Koch. The origin of extracellular fields and currents - [EEG]{}, [EC]{}o[G]{}, [LFP]{} and spikes. *Nat Rev Neurosci*, 13:0 407–420, 2012.
G Susi, SM Ye-Chen, J de Frutos Lucas, G Niso, and F Maest[ú]{}. Neurocognitive aging and functional connectivity using magnetoencephalography. In *Oxford research encyclopedia of psychology and aging*. Oxford University press, Oxford, 2018[[b]{}]{}.
Julien Ponge. Fork and join: Java can excel at painless parallel programming too! [O]{}racle technology network, 2011. URL <http://www.oracle.com/technetwork/articles/java/fork-join-422606.html>.
G. Niso, R. Bru[ñ]{}a, E. Pereda, R. Guti[é]{}rrez, R. Bajo, F. Maest[ú]{}, and F. del Pozo. : Towards an integrated toolbox to characterize functional and effective brain connectivity. *Neuroinformatics*, 110 (4):0 405–434, 2013.
J. Reutimann, M. Giugliano, and S. Fusi. Event-driven simulation of spiking neurons with stochastic dynamics. *Neural Computation*, 15:0 811–830, 2003.
F. Naveros, J.A. Garrido, R.R. Carrillo, E. Ros, and N.R. Luque. Event- and time-driven techniques using parallel cpu-gpu co-processing for snn. *Frontiers in neuroinformatics*, 110 (7), 2017.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The maximum entropy method is shown to be a special limit of the stochastic analytic continuation method introduced by Sandvik \[Phys.Rev. B [**57**]{}, 10287 (1998)\]. We employ a mapping between the analytic continuation problem and a system of interacting classical fields. The Hamiltonian of this system is chosen such that the determination of its ground state field configuration corresponds to an unregularized inversion of the analytic continuation input data. The regularization is effected by performing a thermal average over the field configurations at a small fictitious temperature using Monte Carlo sampling. We prove that the maximum entropy method, the currently accepted state of the art, is simply the mean field limit of this fully dynamical procedure. We also describe a technical innovation: we suggest that a parallel tempering algorithm leads to better traversal of the phase space and makes it easy to identify the critical value of the regularization temperature.'
author:
- 'K. S. D. Beach'
bibliography:
- 'cont.bib'
date: ' March 1, 2004 '
title: |
Identifying the maximum entropy method as a special limit\
of stochastic analytic continuation
---
Introduction
============
Wick rotation transforms imaginary time correlation functions into real, measurable response functions. Analytical results, or numerical results fit to a known functional form, allow for a simple substitution of variables: *e.g.*, $-i\tau \mapsto t(1+i0^+)$. In general, however, this is not possible. To interpret the results of computer simulations such as quantum Monte Carlo and to make comparisons with experiment, we require a technique that reliably extracts spectral information from imaginary time data. At issue is how best to do this given that the input data is intrinsically noisy and incomplete.
The most widely used technique is the maximum entropy method (MEM), [@Gull84; @Silver90; @Gubernatis91] which selects the best candidate solution that is consistent with the data. Here, “best” means most likely in the Bayesian sense. There are several variations on the algorithm, but in general it plays out as a competition between the goodness-of-fit measure $\chi^2$ and the entropic prior $\mathcal{S}$. In practice, one minimizes the functional $\chi^2 -
\alpha^{-1}\mathcal{S}$ (for some $\alpha^{-1} \neq 0$). The presence of the entropic prior introduces a non-linearity that pulls the minimum away from the least squares solution. One of the key advantages to the method is that it is rigourously derived from statistical considerations and guarantees a unique solution.
Another strategy is to generate a sequence of possible solutions and then take their mean, with the hope that spurious features will be averaged out and legitimate features reinforced (as, *e.g.*, in Ref. ). Such methods, however, tend to be *ad hoc* and are not rigourously justified. There are no criteria for selecting which solutions to include or for assigning their relative weights in the sum. Moreover, how these schemes are related to the MEM solution is unclear. There is no reason *a priori* to believe that an average over several possible spectra will be closer to the true spectrum than the single most probable one.
Nonetheless, there is compelling evidence that averaging methods can produce better spectra than the MEM. In particular, Sandvik [@Sandvik] has shown that an unbiased thermal average of all possible spectra, Boltzmann weighted according to $\chi^2$, produces (in several test cases) an average spectrum that is in better agreement with the true spectrum (found via exact diagonalization) than is the MEM result. Indeed, our own experience suggests that the MEM is unduly biased toward smooth solutions: sharp spectral features tend to be washed out or obliterated.
In this paper, we show how the averaging approach can be made systematic. We relate the analytic continuation problem to a system of interacting classical fields living on the unit interval and prove that the MEM solution is realized as its mean field configuration. From that point of view, Sandvik’s method amounts to allowing thermal fluctuations about this mean field configuration. It is, in some sense, the most natural dynamical generalization of the MEM. Finally, we sketch out an improved algorithm for performing the stochastic sampling and provide test results for the two methods applied to the spectrum of a simple BCS superconductor.
Analytic Continuation
=====================
A dynamical correlation function of imaginary time, $G(\tau) = \langle
\operatorname{T}[\hat{O}(\tau)\hat{O}^{\dagger}(0)]\rangle$, satisfies the (anti-)periodicity relation $G(\tau + \beta) = \mp G(\tau)$, where the upper sign holds for fermionic operators and the lower sign for bosonic ones. Since it is uniquely determined by its values in the region $\tau \in [0,\beta)$, the function admits a discrete Fourier transform $$\begin{aligned}
\label{EQ:Gtau}
G(\tau) &= \frac{1}{\beta}\sum_{\omega_n}
e^{-i\omega_n\tau}G(\omega_n),\\ \label{EQ:Gomegan}
G(\omega_n) &= \int_0^\beta \!d\tau\, e^{i\omega_n\tau}G(\tau),\end{aligned}$$ where the sum is over the Matsubara frequencies $\omega_n =
(2n+1)\pi/\beta$ for fermions and $\omega_n = 2n\pi/\beta$ for bosons, with $n \in \mathbb{Z}$.
Provided that $\lvert G(\omega_n) \rvert$ falls off at least as fast as $1/\lvert \omega_n \rvert$ when $n \rightarrow \infty$ (which is guaranteed so long as the operator (anti-)commutator satisfies $\langle
\hat{O}\hat{O}^{\dagger}\pm\hat{O}^{\dagger}\hat{O}\rangle < \infty$), the Fourier components are representable in terms of a function of the form $$\label{EQ-Gz}
\mathcal{G}(z) = \mp \int_{-\infty}^\infty \!\frac{d\omega}{2\pi}
\frac{\rho(\omega)}{z-\omega}$$ with the identification $G(\omega_n) = \mathcal{G}(i\omega_n)$. The function $\rho(\omega)$ is real-valued and satisfies $\rho(\omega) \ge
0$ for fermions and $\operatorname{sgn}(\omega)\rho(\omega) \ge 0$ for bosons. Note that $\mathcal{G}(z)$ is analytic everywhere in the complex plane, with the possible exception of the real line. Wherever $\rho(\omega)$ is nonzero, there will be a corresponding jump in $\mathcal{G}(z)$: $$\label{EQ-jump}
\mathcal{G}(\omega + i0^+) - \mathcal{G}(\omega - i0^+) = \pm
\rho(\omega).$$
The principle of analytic continuation states that given the value of $\mathcal{G}(z)$ at a countably infinite number of points along the imaginary axis—by which we mean that $G(\omega_n)$ or, equivalently, $G(\tau)$ is known—we can *uniquely* extend $\mathcal{G}(z)$ from those points to the full complex plane. In particular, we can find its values just above and just below the real axis and hence, via Eq. (\[EQ-jump\]), extract $\rho(\omega)$.
According to Eq. (\[EQ-Gz\]), we can write $$G(\omega_n) = \mp \int \!\frac{d\omega}{2\pi} \frac{\rho(\omega)}
{i\omega_n-\omega}.$$ Transforming back to imaginary time, via Eq. , and performing the Matsubara frequency sum yields $$\label{EQ-lin-func}
\begin{split}
G(\tau) &= \mp \int \!\frac{d\omega}{2\pi}
\frac{1}{\beta}\sum_{\omega_n}
\frac{e^{-i\omega_n\tau}}{i\omega_n-\omega}\rho(\omega)\\
&= \int \!\frac{d\omega}{2\pi}
\frac{e^{-\omega\tau}\rho(\omega)}
{e^{-\beta\omega}\pm 1}\\
&= \int \!d\omega \, K(\tau,\omega)A(\omega).
\end{split}$$ In the last line, we have defined $$K(\tau,\omega) = \begin{cases}
e^{-\omega\tau}/(e^{-\beta\omega} + 1) & \text{fermions}\\
\omega e^{-\omega\tau}/(e^{-\beta\omega} - 1) & \text{bosons}
\end{cases}$$ and $$\label{EQ:Arho}
A(\omega) = \begin{cases}
\rho(\omega)/2\pi & \text{fermions}\\
\rho(\omega)/2\pi\omega & \text{bosons.}
\end{cases}$$ (For some applications it may be more appropriate to define $K(\tau,\omega) = e^{-\omega\tau}$ and $A(\omega) = \rho(\omega)/2\pi
(e^{-\beta\omega}-1)$ in the bosonic case.) The *spectral function* $A(\omega)$, which we shall view as the main quanitity of interest, is positive definite and satisfies a sum rule $\int
\!d\omega\, A(\omega) = \mathcal{N} < \infty$.
Equation (\[EQ-lin-func\]) tells us that we can interpret $G(\tau)$ as a linear functional of $A(\omega)$ with kernel $K(\tau,\omega)$. Hence, the analytic continutation is equivalent to the functional inversion $A(\omega) = \mathbf{K}^{-1}[G(\tau)]$. Only a finite inversion is practicable, however. If we discretize frequency and imaginary time using a uniform mesh (with spacings $\Delta\tau$ and $\Delta\omega$), then $A_j = A(\Delta\omega \cdot
j)\Delta
\omega$ and $G_k = G(\Delta\tau \cdot k)$ are related by $A_j = \sum_k K^{-1}_{jk} G_k$. The problem is thus reduced to a matrix inversion of $$K_{kj} = \frac{e^{\Delta\omega\Delta\tau\cdot j \cdot k}}
{e^{-\beta \Delta \omega \cdot j}\pm 1}.$$
This inversion is not an easy one to perform, however. The condition number of $K_{jk}$ is extremely large: the matrix will have eigenvalues both exponentially large and exponentially small in $\beta$. This means that computation of the inversion requires extremely high numerical precision. [@Beach] Worse, the inversion problem is ill-posed and responds badly to any measurement error in the input set $G_k$. The inversion typically overfits the noise with spurious high-frequency modes in $A_j$.
The history of practical analytic continuation methods is one of continual refinement of the procedures for regularization of the matrix inversion. The simplest example of regularization is to try $$A_j = \sum_k (K_{kj}+\lambda\delta_{kj})^{-1} G_k.$$ Since the high-frequency modes in $A_j$ are generated by the smallest eigenvalues of $K_{jk}$, a nonzero value of $\lambda$ will have the effect of suppressing those modes with eigenvalues on the order or $\lambda$ or smaller. To see this, note that for each eigenvalue $E$ of $K_{jk}$, there is an eigenvalue in the inverse matrix that is modified according to $1/E \rightarrow 1/(E + \lambda)$.
This naive scheme has two major flaws. First, filtering out the high frequency modes in this way has the effect of eliminating from the spectral function *all* fine structure below a certain frequency scale, whether spurious or real. Second, it does not ensure that $A_j
\ge 0$, as required. The MEM, which we describe briefly in the next section, is considerably more sophisticated about what to filter and has nonnegativity built in.
\[SECT:maxentmeth\]Maximum Entropy Method
=========================================
Suppose that to the exact function $G(\tau)$ we have a measured approximation $\bar{G}(\tau)$. In practice, this will usually have been generated from some Monte Carlo simulation, so that $$\bar{G}(\tau) = G(\tau) + \text{statistical noise.}$$
The goodness-of-fit functional $$\label{EQ:chi_squared}
\chi^2[A] = \int_0^\beta \!\frac{d\tau}{\sigma(\tau)^2}\, \left\lvert
\int \!d\omega\, K(\tau,\omega)A(\omega) - \bar{G}(\tau) \right\rvert^2$$ measures how closely the correlation function generated from $A(\omega)$ \[via Eq. (\[EQ-lin-func\]), the forward model\] matches $\bar{G}(\tau)$. Here, $\sigma(\tau)$ is the best-guess estimate of the total measurement error in $\bar{G}(\tau)$. (See Appendix \[APP:Discretization\].) There is also an entropy associated with each spectral function, $$\label{EQ:SA}
\mathcal{S}[A] = -\int \!d\omega\,A(\omega)\ln\left(A(\omega)/D(\omega)
\right),$$ which measures the information content of $A(\omega)$. Here, $D(\omega)$ is the so-called *default model*, a smooth function that serves as the zero (maximum) entropy configuration. Any features of the true spectral function known in advance can be encoded in $D(\omega$).
It can be shown that the likelihood of any $A(\omega)$ being the true spectral function is equal to $\mathcal{P}[A] \sim e^{-Q[A]}$ where $Q = \chi^2 -
\alpha^{-1}\mathcal{S}$ (and $\alpha^{-1}$ is a parameter that controls the degree of regularization). The MEM solution corresponds to the spectral function that minimizes $Q$. In practice, the minimization of $Q$ is treated as a numerical optimization problem and is typically performed using the Newton-Raphson algorithm or some other gradient search technique. Nonetheless, a formal solution can be found by identifying the spectral function for which $Q$ is stationary with respect to functional variation. The result, derived in Appendix \[APP:formalsol\], is $$\label{EQ:MEM-formal-solution}
\bar{A}(\omega) = e^{\alpha\mu} D(\omega)\exp\biggl[-2\alpha
\int_0^\beta
\!\frac{d\tau}{\sigma(\tau)^2}\,\psi(\tau)K(\tau,\omega)\biggr]$$ where $$\psi(\tau) = \int \!d\omega\,K(\tau,\omega)\bar{A}(\omega) -
\bar{G}(\tau)$$ and $\mu$ is a Lagrange multiplier chosen to enforce the normalization $\int \!d\omega\, A(\omega) = \mathcal{N}$.
In two trivial limits, this set of equations can be solved exactly. When $\alpha \rightarrow \infty$, Eq. demands that $\psi \rightarrow 0$. This yields the noisy, unregularized spectrum $\bar{A}(\omega) = \mathbf{K}^{-1}[\bar{G}(\tau)]$, which is the solution that minimizes $\chi^2[A]$ . When $\alpha \rightarrow 0$, $\bar{A}(\omega) = D(\omega)$, the smooth default function. This solution maximizes $\mathcal{S}[A]$. Note that these results come about because $Q \sim \chi^2[A]$ and $Q \sim -\mathcal{S}[A]$, respectively, in the two limits.
Over the full range of intermediate values ($0 < \alpha < \infty$), Eq. constitutes a one-parameter family of solutions interpolating between these two extremes. An additional condition must be imposed to remove this ambiguity, *i.e.*, to turn the family of solutions into a single final spectrum. In *classic* MEM, one takes the point of view that somewhere between over-fitting and over-smoothing lies an ideal intermediate range centred on some optimal value of $\alpha$. In other schemes, the final result is produced by averaging, $\bar{A}(\omega) = \int_0^\infty\!d\alpha \, w(\alpha)
\bar{A}(\alpha,\omega) / \int_0^\infty\!d\alpha \, w(\alpha)$, in which case the question becomes which weighting function $w(\alpha)$ to use. In their definitive review, [@Jarrell96] Jarrel and Gubernatis address these issues in greater detail.
\[SECT-Alternative-Approach\]The Stochastic Approach
====================================================
In this section and the next, we introduce the stochastic analytic continuation approach and demonstrate how it is related to the MEM. To start, consider a smooth mapping $\phi : \mathbb{R} \mapsto [0,1]$, which takes the frequency domain of the spectral function onto the unit interval. Such a function will be of the form $$\label{EQ:phimap}
\phi(\omega) = \frac{1}{\mathcal{N}}\int_{-\infty}^\omega \!d\nu\,
D(\nu)$$ where $D = \mathcal{N}\phi'$ is positive definite and (like $A$) normalized to $\mathcal{N}$ but otherwise arbitrary. (We use the notation $D$ for the mapping’s kernel in anticipation of identifying it with the default model of the MEM.) Then, $$\label{EQ:nxnorm}
1 = \frac{1}{\mathcal{N}}\int \!d\omega\, A(\omega)
= \int \!d\phi(\omega)
\frac{A(\omega)}{D(\omega)}
= \int_0^1 \!dx\, n(x).$$ In the last line, we have made the change of variables $x = \phi(\omega)$ and introduced the dimensionless field $$\label{EQ:n_field_def}
n(x) = \frac{A(\phi^{-1}(x))}{D(\phi^{-1}(x))}$$ which, according to Eq. , is normalized to unity.
Under this change of variables, Eq. (\[EQ:chi\_squared\]) becomes $$\label{EQ:Hnx}
H[n(x)] = \int_0^\beta \!\frac{d\tau}{\sigma(\tau)^2}\left\lvert
\int_0^1 \!dx\, \hat{K}(\tau,x)n(x) - \bar{G}(\tau) \right\rvert^2$$ with $\hat{K}(\tau,\phi(\omega)) = K(\tau,\omega)$. We take the point of view that Eq. is the Hamiltonian for the system of classical fields $\{n(x)\}$. Then, supposing the system is held fixed at a fictitious inverse temperature $\alpha$, it has a partition function $Z = \int \!\mathcal{D}n\, e^{-\alpha H[n]}$ with a measure of integration $$\label{EQ:measureint}
\int\!\mathcal{D}n = \int_0^\infty \!\biggl(\prod_x dn(x)\biggr)
\delta\biggl(\int_0^1\!dx\,n(x) - 1 \biggr).$$ The thermally averaged value of the field is $$\label{EQ:nalpha}
\langle n(x) \rangle = \frac{1}{Z}\int \!\mathcal{D}n\, n(x) e^{-\alpha
H[n]}.$$ The corresponding “thermally regulated” spectral function, $$\langle{A}(\omega)\rangle = \langle n(\phi(\omega)) \rangle D(\omega),$$ can be recovered using Eq. (\[EQ:n\_field\_def\]).
At zero temperature ($\alpha \rightarrow \infty$), Eq. simply picks out the ground-state field configuration; the corresponding spectral function is the unregularized analytic continuation result. In the high temperature limit ($\alpha \rightarrow 0$), Eq. represents an *unweighted* average over all possible field configurations. In that case, the average is completely independent of the input function $\bar{G}(\tau)$ and as such can only yield the zero-information result $\langle n(x) \rangle = 1$. From Eq. , it follows that $D(\omega)$ is the corresponding spectral function.
These limits are precisely those of the MEM, which we discussed at the end of Sect. \[SECT:maxentmeth\]. Note that the kernel of the mapping in Eq. plays the same role as the MEM’s default model and the fictitious temperature the same role as the MEM’s regularization parameter.
\[SECT:approx\]Approximate Solutions
====================================
Now let us extend our “interacting classical field” analogy a little further. Expanding the square in Eq. , we can cast the Hamiltonian in the familiar form $$\begin{gathered}
H[n(x)] = \int_0^1 \!dx\, \epsilon(x) n(x)\\ + \frac{1}{2}\int_0^1
\!dx\,dy\,V(x,y)n(x)n(y),\end{gathered}$$ with a free dispersion $$\epsilon(x) = -2\int_0^\beta \!\frac{d\tau}{\sigma(\tau)^2}\,
\bar{G}(\tau)\hat{K}(\tau,x)$$ and an interaction term $$V(x,y) = V(y,x) = 2\int_0^\beta
\!\frac{d\tau}{\sigma(\tau)^2}\,\hat{K}(\tau,x)
\hat{K}(\tau,y).$$
*Noninteracting system*—Let us ignore the interaction term for a moment and proceed by setting $V=0$. Then, if we represent the delta function constraint in Eq. with an integral representation $$\delta(X) = \int_{-\infty}^\infty\!d\zeta \, \exp\bigl(i\zeta X \bigr),$$ the partition function is simply $\mathcal{Z} = \int_{-\infty}^\infty\!d\zeta \, e^{-i\zeta}
\mathcal{Z}(\zeta)$, where $$\mathcal{Z}(\zeta) = \int_0^\infty \!\biggl(\prod_x dn(x)\biggr)
e^{-\int_0^1\!dx\,(\alpha\epsilon(x)
-i\zeta)n(x)}.$$ The saddle point solution for the field is $$\label{EQ:nonintsol}
\bar{n}(x) = \frac{\delta}{\delta \epsilon(x)}\biggl(-\frac{1}{\alpha}
\ln Z(\bar{\zeta})\biggr) = e^{-\alpha(\epsilon(x)-\mu)}.$$ This says that the fields are Maxwell-Boltzmann distributed according to their energy as measured with respect to a chemical potential $\mu
\equiv i\bar{\zeta}/\alpha$, which is chosen such that $\int_0^1 \!dx\,
\bar{n}(x) = 1$.
*Mean field treatment*—Now let us reintroduce $V$. Assuming that fluctuations of the $n(x)$ field about its mean value are negligible, $$\bigl(n(x) - \bar{n}(x)\bigr) \bigl(n(y) - \bar{n}(y)\bigr) \approx 0,$$ the Hamiltonian has a mean field form $$\label{EQ:HMF}
H_{\text{MF}} = \int_0^1 \!dx\, E(x)n(x) + \text{const.},$$ where $$\label{EQ:Ex}
E(x) = \frac{\delta H[n]}{\delta n(x)}\biggr\rvert_{n=\bar{n}}
= \epsilon(x) + \int \!dy\, V(x,y)\bar{n}(y).$$ Equation leads to the saddle point solution given by Eq. but now with $\epsilon(x)$ replaced by $E(x)$. Using the definition of $E(x)$ from Eq. , we arrive at the self-consistent equation $$\label{EQ:nxformal}
\bar{n}(x) = e^{\alpha\mu}\exp\left[-\alpha\Bigl(\epsilon(x)+\int
\!dy\,V(x,y)
\bar{n}(y) \Bigr)\right].$$ Again, $\mu$ is a chemical potential used to fix the normalization.
Now consider the reverse change of variables taking $n(x)$ back to $A(\omega)$. With only a little effort, one can show that Eq. is identical to Eq. . What this tells us is that the mean field treatment of the classical field system is *formally equivalent to the MEM*.
We can make this equivalence more explicit still. The free energy density of the system we have just described is $F = U - \alpha^{-1}S -
\mu$, where the internal energy is given by $U = H[\bar{n}(x)]$ and the entropy (see Appendix \[APP:Entropy\]) by $$\label{EQ:Snx}
S[\bar{n}] = -\int_0^1 \!dx\, \bar{n}(x) \ln \bar{n}(x).$$
As we saw earlier, Eqs. and are connected by a change of variables. Similarly, $$\label{EQ:Snx}
\begin{split}
S[\bar{n}] &= -\int_0^1 \!dx\, \bar{n}(x) \ln \bar{n}(x)\\
&= -\int \!d\phi(\omega) \frac{\bar{A}(\omega)}{D(\omega)}\ln\left(
\frac{\bar{A}(\omega)}{D(\omega)}\right)\\
&= -\int \!d\omega \, \bar{A}(\omega) \ln\left(
\frac{\bar{A}(\omega)}{D(\omega)}\right) = \mathcal{S}[\bar{A}],
\end{split}$$ where the final equality follows from comparison with Eq. . Thus, $\chi^2 = H[\bar{n}(x)]$ and $\mathcal{S} = S[\bar{n}(x)]$, which makes clear that $F\mathcal{N} = Q = \chi^2 - \alpha^{-1}\mathcal{S} -
\mu \mathcal{N}$. This means that the MEM solution is just the one that minimizes the free energy of the $\{n(x)\}$ system at the mean field level.
Monte Carlo Evalutaion
======================
Configurations and Update Scheme
--------------------------------
The energy of a given field configuration, given by Eq. , can be written in the form $$\label{EQ:Hh2}
H[n(x)] = \int_0^\beta \!d\tau\, h(\tau)^2,$$ where $$h(\tau) = \frac{1}{\sigma(\tau)}\int_0^1 \!dx\, \hat{K}(\tau,x)n(x) -
g(\tau)$$ and $g(\tau) = \bar{G}(\tau)/\sigma(\tau)$ is the input Green’s function rescaled by the variance.
Computing $\langle n(x) \rangle$ requires that we integrate over all possible field configurations. To accomplish this, we need some ansatz to render the measure $\mathcal{D}n$ finite. One choice is to represent each field configuration as a superposition of delta functions. In that case, we can parameterize each configuration by a set of residues and coordinates $C = \{r_\gamma, a_\gamma \}$ satisfying $r_\gamma > 0$, $0
\le a_\gamma \le 1$, and $\sum_\gamma r_\gamma = 1$. The corresponding field configuration is $$n_C(x) = \sum_\gamma r_\gamma\, \delta(x - a_\gamma ).$$ The partition function $\mathcal{Z} = \int dC\, \exp(-\alpha H_C)$ has a new computationally tractable measure $$\int\!dC = \prod_\gamma \int_0^\infty\!dr_\gamma
\int_{0}^{1}\!da_\gamma\,\delta\biggl(\sum_\gamma r_\gamma - 1 \biggr).$$
In order to calculate the energy $H_C$ of a given configuration via Eq. , we shall need the relation $$\begin{split}
g(\tau) + h_C(\tau) &= \frac{1}{\sigma(\tau)}\int_0^1\!dx\,
\hat{K}(\tau,x)n_C(x)\\
&= \frac{1}{\sigma(\tau)}\sum_\gamma r_\gamma \hat{K}(\tau,a_\gamma).
\end{split}$$ Now suppose that the configuration is modified ($C \mapsto C'$) by altering the parameters in some subset $\Lambda$ of the delta function walkers: $$\begin{split}
r_\gamma \mapsto r'_\gamma & = r_\gamma + \sum_{\lambda\in\Lambda}
\delta_{\gamma\lambda}\Delta r_\lambda,\\
a_\gamma \mapsto a'_\gamma & = a_\gamma + \sum_{\lambda\in\Lambda}
\delta_{\gamma\lambda} \Delta a_\lambda.
\end{split}$$ Accordingly, $h_C \mapsto h_{C'} = h_C + \Delta h$, where $$\Delta h(\tau) = \frac{1}{\sigma(\tau)}\sum_{\lambda\in\Lambda} \Bigl[
r'_{\lambda}
\hat{K}(\tau,a'_\lambda) - r_{\lambda}\hat{K}(\tau,a_\lambda) \Bigr].$$ The configuration energy changes to $$\label{EQ:Delta_H}
\begin{split}
H_{C'} &= \int_0^\beta \!d\tau\, \bigl(h_C(\tau) + \Delta
h(\tau)\bigr)^2\\
&= H_C + \int_0^\beta \!d\tau\,\Delta h(\tau)\bigl[2h_C(\tau) + \Delta
h(\tau)\bigr].
\end{split}$$
The Monte Carlo procedure is to calculate $H_C$ and $h_C(\tau)$ for some arbitrary starting configuation $C$ and then update them whenever a walk is accepted. Acceptance is determined according to the usual Metropolis algorithm: create a modified trial configuration and compute its energy shift $\Delta H = H_{C'} - H_C$ following Eq. (\[EQ:Delta\_H\]); choose a random real number $\xi \in [0,1]$; if $\exp(-\alpha \Delta H) > \xi$, accept the walk and update $$\begin{split}
H_C &\mapsto H_{C'} = H_C + \Delta H, \\
h_C & \mapsto h_{C'} = h_C + \Delta h.
\end{split}$$
The path of the delta function walkers through the configuration space must be normalization-conserving and must satisfy detailed balance. Moreover, the entire phase space must, in principle, be accessible. Only two types of moves are necessary to meet these criteria: (1) coordinate shifting moves, in which the walker $\lambda$ is translated by a distance $\Delta a_\lambda$, and (2) weight sharing moves, in which the total residue of a subset of walkers is reapportioned amongst themselves such that $\sum_\gamma r_\gamma = 1$ is preserved.
It is useful, however, to introduce additional weight sharing moves that also conserve higher moments $$\mathcal{M}^{(n)} = \int_0^1\!dx\,n(x)x^n = \sum_\gamma r_\gamma
E_\gamma^n.$$ Sandvik has shown that such moves dramatically improve the acceptance ratio of attempted walks at low temperature. At a minimum we want to consider walks that preserve the overall normalization $\mathcal{M}^{(0)} = 1$. But we also consider rearrangements of weight between $n > 2$ walkers that conserve the first $n-1$ moments. Such a move can be effected as follows. Let $\Lambda = \{\lambda_1, \lambda_2,
\ldots , \lambda_n\} = \{\lambda_1\} \cup \tilde{\Lambda}$. Defining the scale factors $$Q_{\lambda} = \begin{cases}
-1 & \text{if $\lambda = \lambda_1$}\\
\frac{\prod_{\mu \in \tilde{\Lambda}} (a_\mu - a_{\lambda_1}
)}{\prod_{\mu \in \tilde{\Lambda}} (a_\mu - a_\lambda)} & \text{if
$\lambda \in \tilde{\Lambda}$,}
\end{cases}$$ we can express the changes in residue as $$r'_{\lambda} = r_{\lambda} + \Delta r_\lambda = r_{\lambda}
-s\,Q_{\lambda},$$ where $s$ parameterizes the 1-dimensional line of constraint through the $n$-dimensional space of residues. In order to preserve the positivity of the residues, we must impose $r'_\lambda > 0$. Hence, we need to ensure that $r_\lambda > Q_\lambda \Delta r_\lambda$ for all $\lambda
\in \Lambda$. Accordingly, we take $s$ to be randomly distributed in the interval $$\max_{\lambda \in \Lambda^-} \bigl(r_\lambda/Q_\lambda\bigr) < s <
\min_{\lambda \in \Lambda^+} \bigl(r_\lambda/Q_\lambda\bigr),$$ where $\Lambda^- = \{ \lambda : Q_{\lambda} < 0 \}$ and $\Lambda^+ = \{
\lambda : Q_{\lambda} > 0 \}$.
Parallel Tempering
------------------
The Monte Carlo algorithm described above can be improved by introducing parallel tempering. [@Marinari] The idea is to allow multiple instantiations of the simulation to proceed simultaneously for a variety of parameters $\{\alpha_0, \alpha_1, \ldots, \alpha_N\}$ covering a large range of inverse temperatures. The temperature profile is arbitrary, but we shall find it convenient to choose a constant ratio $\alpha_{p+1} / \alpha_p = R$ between one temperature layer and the next.
Most important, the field configurations in each layer are made to evolve in parallel but not independently. Configurations are swapped between adjacent layers in such a way that preserves detailed balance and ensures that each layer $p$ will eventually settle into thermal equilibrium at inverse temperature $\alpha_p$. The update rule is quite simple: given two adjacent layers $p$ and $q=p\pm 1$, choose a random real number $\xi \in [0,1]$ and swap the $p$ and $q$ configurations if $$\exp\Bigl[\bigl(\alpha_p-\alpha_q\bigr)
\bigl(H_p-H_q\bigr)\Bigr] > \xi.$$
![\[FIG:swap\]The acceptance ratio of configuration swaps between adjacent levels ($\alpha_p \leftrightarrow \alpha_{p+1}$) evolves as a function of the number of updates performed. When the system is fully thermalized, the acceptance ratios stabilize to asymptotic values.](Swap.eps)
Parallel tempering eliminates the need for a separate, initial annealing stage. [@Sandvik] Because the simulation simultaneously samples over a large temperature range, there is no danger of getting trapped in false minima: the interlayer walks always provide a cheap pathway between configurations separated by large energy barriers. All that is required is to let the system thermalize for some time before sampling (*i.e.*, before actually beginning to bin and tabulate the field configurations). By tracking the average acceptance rates for swaps between layers, it is straightforward to determine when the system has equilibrated. Figure \[FIG:swap\] shows a sample run (for a test case to be described in Sect. \[SECT:bcs\]). We see that on a stochastic time-scale of several tens of thousands of moves, each temperature layer settles into thermal equilibrium.
An additional advantage of the parallel tempering algorithm is that it yields in one run a complete temperature profile of all the important thermodynamic variables. In the next section, we discuss how we can put that information to use.
Critical Temperature
====================
The Monte Carlo simulation yields a set of thermally averaged field configurations $\{ \langle n(x) \rangle_{\alpha_p} : p = 0,1,\ldots
N\}$. With little additional effort, we can also keep track of the internal energies $\{U(\alpha_p) = \langle H[n] \rangle_{\alpha_p} : p =
0,1,\ldots N\}$. In this section, we propose a final candidate spectral function constructed from only these quantities.
To start, note that the specific heat can be written as $$C(\alpha_p) = \frac{dU(\alpha)}{d(\alpha^{-1})} \biggr\rvert_{\alpha =
\alpha_p}
\approx \frac{a_p U(\alpha_p)}{\ln R} \frac{d\ln U(\alpha_p)}{dp}.$$ (See Appendix \[APP:logmesh\].) In Fig. \[FIG:knee\], $\ln
U(\alpha_p)$ is plotted for each temperature level. The knee in the function, occurring in the vicinity of the level $p = p^*$, indicates there is a jump in the specific heat. At low temperatures ($\alpha >
\alpha^* \equiv \alpha_{p*}$), the system freezes out and the correlations $\langle n(x)n(y) \rangle - \langle n(x) \rangle \langle
n(y) \rangle$ become short-ranged. There is a characteristic energy scale $E^* = U(\alpha^*)$ associated with this phase transition.
![\[FIG:knee\]The internal energy of the $\{n(x)\}$ system at each temperature layer is plotted. The knee at $p = p^*$, corresponding to a jump in the specific heat, signals a thermodynamic phase transition.](Knee.eps)
Recall that in the *microcanonical ensemble*, the average over all configurations having energy $E$ is given by $$\langle n(x) \rangle_E = \int \mathcal{D}n\, n(x)\delta(E-H[n]).$$ We propose that the final spectrum be defined as $$\label{EQ:nfinal}
\langle\langle n(x) \rangle\rangle = \frac{1}{E^*}\int_0^{E^*}\! dE\,
\langle n(x) \rangle_E,$$ which sums over all field configurations in the ordered phase (*i.e.*, configurations with energies $E$ satisfying $0 \le E <
E^*$). Roughly speaking, this amounts to performing an unbiased average over all spectral functions $A$ that surpass the fitting threshold $\chi^2[A] < E^*$.
Since the Monte Carlo simulation is performed at fixed temperature, however, we must make the change of variables $dE =
(dU/d\alpha)d\alpha$. Equation becomes $$\langle\langle n(x) \rangle\rangle =
\frac{1}{U(\alpha^*)}\int_{\alpha^*}^{\infty}\! d\alpha\,
\biggl(-\frac{dU}{d\alpha}\biggr)\langle n(x) \rangle_{\alpha}.$$ The discretized version of this integral is $$\label{EQ:finaln}
\langle\langle n(x) \rangle\rangle = \frac{\sum_{p=p^*}^{N-1}
\bigl(U(\alpha_p) - U(\alpha_{p+1})\bigr)\langle n(x)
\rangle_{\alpha_p}}{U(\alpha_{p^*})-U(\alpha_N)}.$$
\[SECT:bcs\]BCS Test Case
=========================
![\[FIG:BCS\]The stochastic analytic continuation method is used to extract the spectrum of a BCS superconductor (bandwidth $W=6$ and gap $2\Delta = 1$) from noisy data. The grey region indicates the statistical uncertainty of the computed spectrum. The inset shows the classic and Bryan MEM results.](SpectrumSum.eps)
We showed in Sect. \[SECT:approx\] that the stochastic analytic continuation method is a dynamical generalization of the MEM. The question remains, What is gained by going beyond the mean field calculation? Our contention is that the stochastic method is better able to resolve sharp spectral features buried in noisy data. To illustrate this point, we have taken the spectrum of a BCS superconductor—which contains flat regions, steep peaks, and sharp gap edges—as a test case. The exact spectral function is $$\label{EQ:Abcs}
A(\omega) = \begin{cases}
\frac{1}{W}\frac{\lvert \omega \rvert}{\sqrt{\omega^2 - \Delta^2}} &
\text{if $\Delta < |\omega| < W/2$}\\
0 & \text{otherwise},
\end{cases}$$ where $W$ is the bandwidth and $2\Delta$ the gap magnitude.
From Eq. we generated an exact $G(\tau)$ using the forward model. We then applied random error to the function to create an approximate $\bar{G}(\tau)$, which was made to serve as the input data for our stochastic algorithm and for two flavours of the MEM—the classic method and a method due to Bryan[@Bryan90] (both described in Ref. ). Figure \[FIG:BCS\] shows these computed spectra alongside the exact result.
The most striking aspect of the comparison is that the stochastically generated spectrum does a superior job of modelling the gap. It closely follows the trough of the gap and captures some of the sharpness of the peaks at the gap edges. The MEM spectra, on the other hand, are much too smooth. The classic MEM spectrum is especially poor. It is at best a caricature of the true BCS spectrum: the sharp features are completely washed out and the depression at $\omega = 0$ is not a fully developed gap.
Bryan’s algorithm does a somewhat better job of reproducing the gap and its adjacent peaks, but in doing so it also forms a second pair of spurious humps around $\omega = 2$. In our experience, this is typical behaviour. The MEM method has trouble making sudden transitions from regions of high to low curvature. What one tends to get is a smooth curve gently oscillating around the correct result. The stochastic method, in contrast, seems to have no trouble generating a flat region next to a sharp peak.
Conclusions
===========
In this paper, we have made the case that the MEM is not the best method for extracting spectral information from imaginary time data. Instead, we advocate the use of the stochastic analytic continuation method. Our claim is that the stochastic method is at least as good as the MEM and may even surpass it for a broad class of problems in which the spectrum to be extracted has very sharp features.
This is a difficult point to argue convincingly. New analytic continuation methods tend to face considerable resistance, and claims of superiority on their behalf are met (quite rightly) with a high degree of skepticism. The MEM has a record of years of successful use in a variety of settings; plus, it offers the comfort of a seemingly rock-solid mathematical rationale. Competing schemes tend to lack any clear justification other than a few tantalizing examples of their performance in a handful of test cases.
The prevailing opinion is that the MEM is the definitive “solution” to the analytic continuation problem. Some other method may produce better spectra in particular special cases, but as a general method, the MEM has to win out. The thinking goes: no other algorithm can outperform the MEM because its solution is, by construction, the unique, best candidate spectrum—a claim that rests on the firm foundation of Bayesian logic.
What this line of reasoning misses, however, is the possibility that an average of *many* likely candidates might better reproduce the true spectral function than does the *single* most likely spectrum. To give a path integral analogy, we would argue that including fluctuations about a saddle point solution (the single most likely field configuration) can yield a result closer to the full integral. This is how we go about justifying the stochastic analytic continuation method.
Let us be careful about what can be established rigourously. To be precise, the standard conditional probability analysis used to derive the MEM proves only that the most likely spectrum belongs to the family of solutions (parameterized by $\alpha^{-1}$) that minimizes $Q = \chi^2
- \alpha^{-1} \mathcal{S}$. From our point of view, then, what is required of an averaging method is that it produce at the mean field level a family of solutions that coincides with the MEM result. The stochastic method, as we have formulated it, does exactly this—under the guise of minimizing the free energy $F\mathcal{N}$ $(= Q)$ of a system of classical fields at a fictitious temperature $\alpha^{-1}$.
This correspondence gives us a new way of thinking about the MEM solution. We know that even though a path integral contains jagged, discontinuous field configurations, its saddle point solution is always a smooth, continuous function. This highlights the main deficiency of the MEM—that it fails to model well spectral functions that are not sufficiently smooth—and makes clear why the stochastic method does not suffer from the same limitation.
Another advantage of the stochastic approach is that it helps us to talk about the analytic continuation problem using a more physical language. Having identified the regularization parameter as a temperature, we can ask how the system behaves thermodynamically. The answer, we have suggested, is that the system exhibits ordered and disordered phases that can be interpreted as the good-fitting and ill-fitting regimes. We believe that this gives a much more intuitive picture than does the somewhat obscure probability analysis of the MEM.
We close with a recapitulation of the main results. We have presented a new variant of the stochastic analytic continuation method that differs from Sandvik’s original prescription as follows: as a matter of mathematical formulation, it includes an additional internal freedom that turns out to be equivalent to specifying a default model; as a matter of practical implementation, it is built on a delta function walker scheme and takes advantage of parallel tempering. We have proved that the mean field version of this stochastic method is equivalent to the MEM. Our tests suggest that it outperforms the MEM for spectra with sharp features and fine structure.
The author would like to thank Anders Sandvik for several helpful discussions and Philippe Monthoux for generously making his maximum entropy code available. Support for this work was provided by the Department of Energy under grant DE-FG02-03ER46076.
\[APP:Discretization\]Statistical Error and Discretization
==========================================================
In Eq. , we have used notational shorthand to gloss over two subtle issues. First, we have ignored the fact that the statistical errors between $\bar{G}(\tau)$ and $\bar{G}(\tau')$ are not independent for $\tau\neq \tau'$. In general, the errors will be positively correlated whenever $\lvert \tau - \tau' \rvert$ is sufficiently small. There is also a tendency for them to be negatively (positively) correlated over long-separated times since $G(0^-) = \mp
G(\beta)$ is built in to the definition of the correlation function. Thus, one should more properly write the goodness-of-fit measure as $$\label{EQ:chi-cov}
\chi^2[A] = \frac{1}{\beta^2} \!\int d\tau d\tau'\, \Delta(\tau)
C^{-1}(\tau,\tau') \Delta(\tau'),$$ where $\Delta(\tau) = \int d\omega\, K(\tau,\omega)A(\omega) -
\bar{G}(\tau)$ and $C(\tau,\tau')$ is the covariance function for $\bar{G}(\tau)$.
Second, we have ignored the discrete nature of the known input data. A Quantum Monte Carlo algorithm, for example, is used to generate stochastically a sequence of independent measurements $\{ G^{(1)},
G^{(2)}, \ldots, G^{(M)} \}$, where each $G^{(m)}$ is an $(L\!\!+\!\!1)$-vector holding the values of the single-particle propagator at imaginary times $\tau_l = \beta l/L$ for $l = 0, 1,
\ldots, L$.
The numerical measurement of the Green’s function is accomplished by taking the average $$\label{EQ:Gbar-sumM}
\begin{split}
\overline{G}_l &= \frac{1}{M}\sum_{m=1}^M G^{(m)}_l.\\
\overline{G_l G_{l'}} &= \frac{1}{M}\sum_{m=1}^M G^{(m)}_l G^{(m)}_{l'}.
\end{split}$$ The corresponding covariance matrix is given by $$\label{EQ:Cov-sumM}
\begin{split}
C_{ll'} &= \frac{1}{M\bigl(M-1\bigr)}\sum_{m=1}^M \bigl(G^{(m)}_l -
\bar{G}_l \bigr)\bigl(G^{(m)}_{l'} - \bar{G}_{l'} \bigr).\\
&= \frac{1}{M-1}\sum_{m=1}^M \biggl[ \overline{G_l G_{l'}} - \bar{G}_l
\bar{G}_{l'} \biggr]
\end{split}$$ Equation must now be discretized in order to make use of Eqs. and . The imaginary time integrals are carried out numerically on a uniform mesh of $L$ time slices (spaced by $\Delta\tau = \beta / L$) according to the formula $$\int_0^\beta d\tau f(\tau) \approx \sum_{l=0}^L \Delta\tau \, w_l \,
f_l,$$ where $f_l = f(\Delta\tau\cdot l)$ and the Bode’s rule weights $w_l$ satisfy $\sum_{l=0}^L w_l = L$. Equaton becomes $$\label{EQ:chi-cov-discrete}
\chi^2 = \frac{1}{L^2} \sum_{l,l'=0}^L w_l \Delta_l C^{-1}_{ll'} w_{l'}
\Delta_{l'}.$$ Since $\Delta(0) = \pm \Delta(\beta)$, $$\label{EQ:chi-cov-discrete}
\chi^2 = \frac{1}{L^2} \sum_{l,l'=0}^{L-1} \tilde{w}_l \Delta_l
C^{-1}_{ll'} \tilde{w}_{l'} \Delta_{l'}.$$ Here, $\tilde{w}_l = w_l + \delta_{l,0} w_L$ for $l = 0, 1, \ldots L-1$.
We now want to solve for the unitary transformation $U$ that diagonalizes the covariance matrix. This allows us to write $C =
U^t\Sigma U$ in terms of a set of *statistically independent* variances $\Sigma = \operatorname{diag}(\sigma^2_1, \sigma^2_2, \ldots, \sigma^2_L)$. The inverse matrix is $C^{-1} = U^t \Sigma^{-1} U$.
Putting $C^{-1}_{ll'} = \sum_{k=0}^L \frac{1}{\sigma^2_k} U_{kl}
U_{kl'}$ into Eq. yields $$\begin{split}
\chi^2 &= \sum_{k=0}^L \frac{1}{\sigma^2_k} \biggl( \frac{1}{L}
\sum_{l=0}^L U_{kl} w_l \Delta_l \biggr)^2\\
&= \sum_{k=0}^L \frac{1}{\sigma^2_k} \Bigl\lvert (VK)[A]_k - (VG)_k
\Bigr\rvert^2
\end{split}$$ where we have defined the matrix $V_{kl} = U_{kl} w_l / L$.
To recapitulate, the discretization of the $\tau$ integration is implicit in Eq. ; it also presumes that we are working in the $V$ basis in which the covariance matrix is diagonal.
\[APP:formalsol\]Maximum Entropy Formal Solution
================================================
We want to examine the changes in $\mathcal{S}$ with variations in $A(\omega)$. Since the spectral function is subject the the normalization constraint $\int d\omega\, A(\omega) = \mathcal{N}$, variations in $A(x)$ and $A(y)$ for $x\neq y$ are not independent. We can enforce the constraint by introducing a lagrange multiplier $\Gamma = 1 +
\alpha\mu$. Let us define $$\label{EQ:extendS}
\mathcal{S}[A,\Gamma] \equiv -\int d\omega\,A\ln\left(A/D\right)
+\Gamma\int d\omega\,\left(A-D\right).$$ We have assumed here that $D(\omega)$ and $A(\omega)$ have the same normalization.
Variations of the extended functional, Eq. , look like $$\begin{split}
\frac{\delta \mathcal{S}}{\delta A(x)} &=
-\ln\left(\frac{A(x)}{D(x)}\right) +
\alpha \mu\\
\frac{d \mathcal{S}}{d\mu} &= \alpha\int\!
d\omega\,\left(A(\omega)-D(\omega)\right)
\end{split}$$ There is a unique solution that causes these two equations to vanish: $A(x) = D(x)$, $\mu = 0$. This implies that $\mathcal{S} = 0$ and $\delta \mathcal{S} = 0$.
Also, since $$\frac{\delta^2 \mathcal{S}}{\delta A(x)\delta A(y)} =
-\frac{\delta(x-y)}{A(x)}$$ we find that $\delta^2 \mathcal{S} \le 0$. This means that the entropy functional is strictly non-positive and takes its maximum $\mathcal{S}=0$ when $A$ is equal to the default model.
Similar considerations for $\chi^2[A]$ allow us to construct the total variation in $Q = \chi^2 - \alpha^{-1} \mathcal{S}$. We find that $$\begin{split}
0 = \frac{\delta Q[A,\mu]}{\delta A(\omega)} &= 2\int_0^\beta d\tau\,
K(\tau,x)\psi(\tau)\\
&\quad - \alpha^{-1} \left[- \ln \left( \frac{A(\omega)}{D(\omega)}
\right) + \alpha\mu\right]
\end{split}$$ where $$\psi(\tau) = \int d\nu\,K(\tau,\nu)A(\nu) - \bar{G}(\tau).$$
\[APP:Entropy\]Configurational Entropy
======================================
Consider a system of $N$ energy levels with degeneracies $m_p$ ($p=1,2,\ldots,N$). Suppose that each level is filled with $n_p$ indistinguishable particles. The state of the system is unchanged by the rearrangement of particles within a given level. Thus, given a set of occupancies $\{0 \le n_p \le m_p\}$, the number of equivalent configurations is $\Omega(\{n_p\}) = \prod_p\begin{pmatrix}
m_p\\
n_p
\end{pmatrix}$ and the entropy due to this configuration is $$\ln \Omega(\{n_p\}) = \frac{1}{N}\sum_p\ln
\begin{pmatrix}
m_p\\
n_p
\end{pmatrix}.$$
The binomial coefficient $\begin{pmatrix}
m\\
n
\end{pmatrix}
= m!/(m-n)!/n!$ can be approximated using Stirling’s formula $m! \approx m\ln m$. In the limit of small relative occupancy, this gives $$\begin{split}
\ln
\begin{pmatrix}
m\\
n
\end{pmatrix} &= m\ln m - (m-n)\ln (m-n) - n \ln n\\
&\xrightarrow{m>>n} - n \ln n.
\end{split}$$
Going over to the continuum, we make the identification $$\begin{split}
\frac{1}{N}\sum_p &\rightarrow \int \!dx\\
m_p &\rightarrow \infty\\
n_p &\rightarrow n(x)
\end{split}$$ and use the counting arguments above to write the entropy associated with each field configuration: $$\ln \Omega[n] = -\int_0^1 \!dx\, n(x)\ln n(x).$$ The total entropy is $$S = \int\!\mathcal{D}n\,\ln \Omega[n] \approx \ln \Omega[\bar{n}].$$
\[APP:logmesh\]Discretization over a logarithmic mesh
=====================================================
Suppose that we want to integrate a function $f(\alpha)$ known only at the points $\alpha_n = R^n \alpha_0$ for $n = 0, 1, \ldots N$. The integral identity $$\int \! d\alpha\, f(\alpha) = \int \! d\tilde{\alpha}\,
e^{\tilde{\alpha}} f(e^{\tilde{\alpha}})$$ follows from the change of variables $\alpha = \exp(\tilde{\alpha})$. In this basis, the known points describe a uniform mesh $$\tilde{\alpha}_n = \ln \alpha_n = \ln \alpha_0 + n \ln R$$ with spacing $\Delta\tilde{\alpha} = \tilde{\alpha}_{n+1} -
\tilde{\alpha}_{n} = \ln R$. Accordingly, $$\begin{split}
\int \! d\alpha\, f(\alpha) &\approx \sum_{n=0}^N \Delta\tilde{\alpha}\,
e^{\tilde{\alpha_n}} f(e^{\tilde{\alpha}_n})\\
&= \sum_{n=0}^N \bigl(\ln R\bigr) \alpha_n f(\alpha_n).
\end{split}$$
When the integrand is of the form $$\frac{dU}{d\alpha} = \frac{1}{e^{\tilde{\alpha}}}
\frac{dU}{d\tilde{\alpha}}$$ we must first discretize the derivative $$\begin{split}
\frac{dU}{d\alpha}\biggr\rvert_{\alpha = \alpha_n} &\approx
\frac{1}{e^{\tilde{\alpha}_n}} \frac{U(\alpha_{n+1}) -
U(\alpha_n)}{\Delta\tilde{\alpha}}\\
&= \frac{U(\alpha_{n+1}) - U(\alpha_n)}{\alpha_n \ln R},
\end{split}$$ which leads to the integrals $$\begin{gathered}
\int_{\alpha_{p}}^{\alpha_{N-1}}\! d\alpha\,
\biggl(-\frac{dU}{d\alpha}\biggr) \langle n(x) \rangle_{\alpha}\\
\approx \sum_{n=p}^{N-1} \bigl[U(\alpha_n)-U(\alpha_{n+1})\bigr]\langle
n(x) \rangle_{\alpha_n}\end{gathered}$$ and $$\begin{split}
\int_{\alpha_{p}}^{\alpha_{N-1}}\! d\alpha\,
\biggl(-\frac{dU}{d\alpha}\biggr)
&\approx \sum_{n=p}^{N-1} \bigl[U(\alpha_n)-U(\alpha_{n+1})\bigr]\\
&= U(\alpha_p) - U(\alpha_N).
\end{split}$$ Equation is simply the ratio of these two results.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
title: Review Comments
---
{#section .unnumbered}
1. Reviewer 2 of 2019 ACC submission 8
Comments to the author ======================
Review of the paper “A Closed-Form Analytical Solution for Optimal Coordination of Connected and Automated Vehicles” by Andreas A. Malikopoulos and Liuhui Zhao
The paper “A Closed-Form Analytical Solution for Optimal Coordination of Connected and Automated Vehicles” by Andreas A. Malikopoulos and Liuhui Zhao presents a closed-form analytical solution for an optimal control problem that is related to connected automated vehicles. The vehicles are modeled by simple double integrators, but the authors discuss how non-trivial rear-end collision avoidance constraints can be taken into account.
In the reviewer’s opinion the paper is generally well-written, although there are number of issues that need to be improved. Major comments can be found below.
Major comments:
1\) Analysis: The presented analysis methods use indirect optimal control ideas, which is fine for the case that no constraint or only a few constraints are active such that the solution can be worked out explicitly. However, the current framework seems not suited for larger-scale problems. For example, if more complex traffic scenarios are considered, with additional collision avoidance constraint inside the intersection area, it may become very cumbersome to work out all possible active set combinations, especially if these constraints are enforced point-wise in time (exponential complexity). The authors might consider to mention that this is a limitation of their approach.
2\) Literature: The current manuscript is related to a number of existing papers that are not cited. This includes
Theory: there are tons of articles on Pontryagin’s maximum principle, indirect optimal control approaches, etc.. In my opinion, the basic methods used in this paper are not new, even if the particular application to rear-end collision avoidance constraint satisfaction may be new. Thus, it is relevant to cite existing indirect control methods and explain more clearly what the contribution of the current manuscript relative to all these existing works are.
Methods: there exist a number of more advanced control methods, which are based on active set enumeration, e.g., Explicit Model Predictive Control. Unfortunately, the corresponding links to Explicit MPC are not made in the paper. I recommend to review the related literature carefully and try to find out whether there are any relevant new ideas, perhaps, in the context of the considered constrained double integrator optimization problem.
Application: there exists many articles on traffic control at intersections. For example, the authors are missing the work by Robert Hult (Chalmers), who has many articles on control of autonomous vehicles at intersections. Please also search for related work by Mario Zanon. If I remember correctly, some recent papers by Mario Zanon and co-workers also consider double integrator (as well as more advanced Dubin car models) in combination with rear-end collision constraints. In my opinion, this work is so strikingly closely related to the submitted manuscript that is should not be missed in the literature review.
2. Reviewer 3 of 2019 ACC submission 8
Comments to the author ======================
The paper analyzes optimal control of sequenced vehicles subject to point mass model with Force as the input and force square as the objective for each vehicle. The paper based on Pontryagin maximum principle computes parametric polynomial velocity and position trajectories, calculates parameters based on boundary and jump conditions and pieces optimal arcs to generate complete optimal trajectory.
The paper is in right direction to construct analytic optimal trajectory when it can be and reduce real-time computation bourdon. It is well understood that deriving analytic optimal solution demands significant model reductions. However, the models in this paper is over simplified and the reviewer suggest that authors provide the path forward to more realistic models and assumptions so that the future results are applicable in fleet optimization. The following are major concerns that can be addressed in future studies:
1- The objective is force square with distance constraint. No time constraint is included. In reality, the constraint is a mixture of time and fuel consumption. Saving fuel at the cost of arriving late and arriving too early at the cost of fuel is not acceptable. Adding time cost index would have helped without complicating analytic analysis. The cost index in the cost is a common practice in aviation systems and is used in all flight management systems since 70s. 2-Applying the derived analysis on one single car would lead to acceleration or deceleration to maximum or minimum speed and cruise in that condition until the final distance constraint is met. Namely, the car would for instance accelerate to the maximum speed no matter how large it is. Clearly this is not a realistic optimal trajectory. Because the fuel consumption is a function of force (thrust) as well as speed and higher speed is not necessarily optimal. 3- taking (2) into account would differentiate the cooperative optimal structure where all cars cooperate to minimize aggregated fuel and the case where each car acts as a selfish rational player. Each car would have and optimal cruise condition and velocity. Hence, if a specific car goes too much inefficient with non-cruise velocity, then other cars would accommodate the care so that it is not forced to switch to other optimal arcs. A scenario that is different from the presented results. 4-The reviewer strongly encourages the authors to study the flight optimization papers form 70s and early 80s. There pontryagin principle is extensively used to characterize different optimal arcs including cruise conditions with complete engine model taken into account. More interestingly it is shown that the optimal trajectory in some cases is indeed periodic. The models are all point mass so mathematically is the same as the models studied in this paper.
The reviewer concludes that even though the model and assumptions are simplistic, the results are certainly right steps in the right direction and recommends the paper for publications.
3. Reviewer 4 of 2019 ACC submission 8
Comments to the author ======================
Authors propose a method to obtain optimal control for a platooning of autonomous vehicles. The application is the problem of crossing intersections with flows of vehicles. A centralised system is responsible of the scheduling of the vehicles and assigns an exit time for each vehicle in order to avoid collisions between vehicles in different flows. This system is out of the scope of the paper. Indeed in the paper the problem of ensuring the correct exiting time from the intersection while preserving a safety distance from a front vehicle is considered. The vehicle dynamics is a double integrator and an additional state is considered as the distance from the front vehicle. The proposed optimal control problem considers the final state (distance to be covered is the intersection length) and the final time. Moreover, it considers also constraints on states (speed limits and minimum safety distance) and control (acceleration limits). The problem is solved in closed form by considering separately the unconstrained and the constrained cased with the junction conditions. All the possible constraint activation cases are considered and linear systems in the multipliers variable are obtained. The solution is found by considering the unconstrained optimal solution until a constraint is activated and then by switching to the relative case. This is repeated until a the solution does not violate any constraint.
The paper is clearly written and results are correct. In my opinion, however, the part of the paper dedicated to the computation of the overall solution (by cases glued together) is not sufficiently detailed. Indeed, what is not completely clear is how this approach guarantees that the final conditions are met. In particular I am referring to the conditions on the final position. This condition does not appear to be taken into account in the glueing procedure when the solution is found by switching from an unconstrained case to a constrained one or between constrained cases. Indeed this is a boundary condition problem but it does not seem to be correctly solved with the proposed procedure. Although the aspect is not sufficiently tackled in the paper, solutions obtained by simulations seem to verify the condition.
Please fix the numbering of references of figures 2.a and 2.b that are swapped.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider an extremely broad class of M/G/1 scheduling policies called SOAP: Schedule Ordered by Age-based Priority. The SOAP policies include almost all scheduling policies in the literature as well as an infinite number of variants which have never been analyzed, or maybe not even conceived. SOAP policies range from classic policies, like first-come, first-serve (FCFS), foreground-background (FB), class-based priority, and shortest remaining processing time (SRPT); to much more complicated scheduling rules, such as the famously complex Gittins index policy and other policies in which a job’s priority changes arbitrarily with its age. While the response time of policies in the former category is well understood, policies in the latter category have resisted response time analysis. We present a universal analysis of all SOAP policies, deriving the mean and Laplace-Stieltjes transform of response time.'
author:
- Ziv Scully
- 'Mor Harchol-Balter'
- 'Alan Scheller-Wolf'
bibliography:
- 'refs.bib'
title: ' SOAP: One Clean Analysis of All Age-Based Scheduling Policies'
---
<ccs2012> <concept> <concept\_id>10002944.10011123.10011674</concept\_id> <concept\_desc>General and reference Performance</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10002950.10003648.10003688.10003689</concept\_id> <concept\_desc>Mathematics of computing Queueing theory</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10011007.10010940.10010941.10010949.10010957.10010688</concept\_id> <concept\_desc>Software and its engineering Scheduling</concept\_desc> <concept\_significance>500</concept\_significance> </concept> <concept> <concept\_id>10010147.10010341.10010342</concept\_id> <concept\_desc>Computing methodologies Model development and analysis</concept\_desc> <concept\_significance>300</concept\_significance> </concept> <concept> <concept\_id>10003752.10003809.10003636.10003808</concept\_id> <concept\_desc>Theory of computation Scheduling algorithms</concept\_desc> <concept\_significance>100</concept\_significance> </concept> </ccs2012>
Introduction
============
Analyzing the response time of scheduling policies in the M/G/1 setting has been the focus of thousands of papers over the past half century, from classic early works [@book_kleinrock; @fb_analysis_schrage; @m/g/1_kendall; @nonpreemptive_takacs; @ps_analysis_kleinrock; @srpt_analysis_schrage; @vacations_fuhrmann] to more recent works in the SIGMETRICS community [@fb_optimality_misra; @fairness_wierman; @mlps_delay_aalto; @multiclass_ayesta; @ps_asymptotics_borst; @ps_heavy_robert; @ps_beyond_aalto; @smart_insensitive_wierman; @smart_preventing_wierman; @smart_epsilon_wierman; @srpt_heavy_zwart; @vacations_ayesta]. Examples of common scheduling policies include
- *first-come, first-served* (FCFS), which serves jobs nonpreemptively in the order they arrive;
- *class-based priority*, which serves the job of highest priority class, possibly preemptively and possibly nonpreemptively;
- *shortest remaining processing time* (SRPT), which preemptively serves the job with the least remaining time;
- *foreground-background* (FB), which preemptively serves the job that has received the least service so far; and
- *processor sharing* (PS), which concurrently serves all jobs in the system at the same rate.
In just these few examples we see a variety of features represented: preemptible jobs, nonpreemptible jobs, prioritizing by class, prioritizing by job size, and prioritizing by service received so far, or *age*. Each policy requires a custom response time analysis that takes into account its particular combination of features.
Although there has been much success in analyzing the response time of specific scheduling policies in the M/G/1 setting, such as those listed above, results are ad-hoc and *limited to relatively simple policies*. Analyzing variants of the above simple policies, let alone fundamentally different policies, is an open problem. For instance, none of the following scenarios have been analyzed before.
- Suppose we have exact size information for some “sized” jobs but not other “unsized” jobs. We run SRPT on sized jobs and FB on unsized jobs, meaning that we serve the sized job of minimum remaining time or unsized job of minimum age, whichever measurement is smaller.
- Suppose we have jobs that are neither fully preemptible nor fully nonpreemptible but instead preemptible only at specific “checkpoint” ages. We run a preemptive policy, such as SRPT or FB, but only preempt jobs when they reach checkpoint ages.
- The *Gittins index policy* [@book_gittins; @m/g/1_gittins_aalto], long known to be optimal for minimizing mean response time in the M/G/1 queue[^1], has only been analyzed in certain special cases [@multiclass_ayesta; @rs_slowdown_hyytia]. In general, the Gittins index policy can have a complex priority scheme [@mlps_gittins_aalto] which, while known to perform optimally, has not been analyzed before in its general form.
Approaching the above examples with state-of-the-art techniques, if possible at all, would require an ad-hoc analysis for each scenario. We seek *general principles and techniques* for response time analysis that apply to not just the above examples but to as many scheduling policies as possible, even those not yet imagined.
Contributions
-------------
We introduce *SOAP*, a *universal framework* for defining and analyzing M/G/1 scheduling policies. The SOAP framework can analyze *any SOAP scheduling policy*, which includes nearly any policy where a job’s priority depends on its own characteristics: class, size, age, and so on. Specifically, we make the following contributions.
- We *define* the class of SOAP policies (), a broad class of policies that includes the three unsolved examples above as well as many other policies, from practical scenarios to policies not yet imagined. We *encode* many policies old and new as SOAP policies ().
- We give a *universal response time analysis* that works for any SOAP policy (), obtaining closed forms for the mean () and Laplace-Stieltjes transform (). In particular, we apply our results to *previously intractable analyses* (), such as the response time of the Gittins index policy.
In defining and analyzing SOAP policies, there are two major technical challenges. The first major challenge is that to have a single analysis apply to many scheduling policies at once, we need to *express all such policies within a single framework*. The SOAP framework encodes a scheduling policy as a *rank function*, which maps each job to a priority level, or *rank*. All SOAP policies are based on a single rule: always serve the job of *minimal rank*. For example, in a preemptive class-based priority system, a job’s rank is its class (), whereas in SRPT, a job’s rank is its remaining time (). Rank functions can express a huge variety of policies, from virtually all classic policies () to complex policies which have never been analyzed before (). A notable exception is PS, which does not fit into the SOAP framework.
The second major challenge is to *analyze policies with arbitrary rank functions*. In particular, nearly all previously analyzed scheduling policies, when expressed as SOAP policies, have rank functions that are *monotonic* in age. For example, under SRPT, a job’s rank decreases with age, making it less and less likely to be preempted by another job, while under FB, a job’s rank increases with age, making it more and more likely to be preempted by another job. Unfortunately, the techniques used in the past to analyze policies with monotonic rank functions *break down for arbitrary nonmonotonic rank functions*, which appear, for instance, when studying the Gittins index policy () and jobs that are preemptible only at certain checkpoints (). We develop *new analytical tools* that work for arbitrary rank functions ().
Related Work
------------
Our work on SOAP policies follows in the tradition of analyses that address an entire class of policies at once. Two such classes are *SMART* [@smart_insensitive_wierman] and *multilevel processor sharing* (MLPS) [@book_kleinrock].
- The SMART class includes all policies that satisfy certain criteria that ensure they prioritize small jobs over large ones, such as SRPT and PSJF (). Some recent work on SMART policies includes analyzing the tail behavior of response time [@smart_preventing_wierman] and characterizing the tradeoff between accuracy of size estimates and response time [@smart_epsilon_wierman].
- The MLPS class consists of policies that divide all jobs in the system into echelons based on age, then serves jobs in the youngest echelon according to FCFS, FB, or PS. Some recent work on MLPS policies includes optimally choosing the age echelon cutoffs [@mlps_threshold_osipova] and connecting MLPS to the Gittins index policy [@mlps_gittins_aalto].
While the SMART and MLPS classes have nearly no overlap, the SOAP class includes many policies from *both* classes. Specifically, the SMART\* subclass of SMART [@smart_insensitive_wierman] and MLPS policies which do not use PS are all SOAP policies.
A particularly important SOAP policy is the Gittins index policy [@book_gittins; @m/g/1_gittins_aalto], which minimizes mean response time in the M/G/1 queue when job sizes are not known. The Gittins index policy has a rather complex definition, but recent work [@m/g/1_gittins_aalto; @mlps_gittins_aalto] has revealed some of its structural properties. analyze a specific case of the Gittins index policy for a multiclass M/G/1 queue where each class’s job size distribution has the *decreasing hazard rate* (DHR) property. Using the SOAP framework, we can analyze the Gittins index policy for *arbitrary size distributions*. show that for jobs with known sizes, a weighted version of the Gittins index yields the *shortest processing time product* (SPTP) policy and that this policy minimizes mean *slowdown*, which is the ratio of a job’s response time to its size. SPTP is a SOAP policy, so the SOAP framework can obtain the Laplace-Stieltjes transform of slowdown for SPTP, extending the previous mean analysis.
System Model and SOAP Policies {#sec:model}
==============================
We consider work-conserving scheduling policies for the M/G/1 queue. We write $\lambda$ for the total arrival rate and $X$ for the overall job size distribution. We assume a stable system, meaning $\lambda\E{X} < 1$, and a preempt-resume model, meaning preemption and processor sharing are permitted without penalty or loss of work.
Descriptors {#sub:descriptors}
-----------
Scheduling algorithms use information about jobs in the system when deciding which job to serve. We can divide this information into two types: *static* and *dynamic*.
- *Static* information about a job is revealed when it enters the system and never changes. For example, in a system with multiple job classes, a job’s class would be static information, and in a system where exact job sizes are known, a job’s exact size would be static information. We call a job’s static information its *descriptor* and write $\mc{D}$ for the set of descriptors. A job’s descriptor $d$ determines its size distribution $X_d = (X \mid \text{job has descriptor } d)$.
- *Dynamic* information about a job changes as a job is served. In this paper, the only dynamic information about a job is its *age*, the amount of time it has been served. The set of possible ages is $\R_{\geq 0}$.
Descriptors are often tuples. To distinguish descriptors from ranks, another type of tuple introduced in , we write descriptors in \[square brackets\] and ranks in $\langle$angle brackets$\rangle$.
\[ex:descriptors\] Consider a system with a set of job classes $\mc{K}$, where $X_k$ is the size distribution of class $k \in \mc{K}$. Depending on what information is known to the scheduler, the set of descriptors $\mc{D}$ may be one of several options.
- If jobs do not reveal their exact size upon entering the system, then $\mc{D} = \mc{K}$, because the only static information we have about each job is its class. The size distribution of jobs with descriptor $k$ is simply $X_k$.
- If jobs reveal their exact size upon entering the system, then $\mc{D} = \mc{K} \times \R_{\geq 0}$, because we know each job’s class and size. The size distribution of jobs with descriptor $[k, x]$ is $X_{[k, x]} = x$, the deterministic distribution with value $x$.
- If only some jobs reveal their exact size, then $$\mc{D} = \mc{K} \times (\R_{\geq 0} \cup \{?\}),$$ because some jobs have known exact size $x \in \R_{\geq 0}$ while others have unknown size, which we denote by $?$. The size distributions are $X_{[k, x]} = x$ for $x \in \R_{\geq 0}$ and $X_{[k, ?]} = X_k$.
We require that the descriptors of jobs must be chosen i.i.d. according to a fixed distribution. For instance, in , each job’s class must be chosen i.i.d., and in the third scenario, having each job independently reveal its size with probability $1/2$ is permitted, but having alternating arrivals reveal their sizes is not.
SOAP Policies and Rank Functions {#sub:soap_policies}
--------------------------------
A *SOAP scheduling policy* is a preemptive priority policy where a job’s descriptor and age determine its priority. SOAP is an acronym for *Schedule Ordered by Age-based Priority*. Specifically, a SOAP policy is specified by the following ingredients:
- a set $\mc{R}$ of *ranks*,
- a strict total order $\prec$ on $\mc{R}$, and
- a *rank function* assigning a rank $r(d, a)$ to each pair of descriptor $d$ and age $a$, $$r : \mc{D} \times \R_{\geq 0} \to \mc{R}.$$
The defining property of SOAP policies is the following.
> **Every moment in time, a SOAP policy serves the job of *minimum rank*.**
Ties between jobs of the same rank are broken using a tiebreaking rule. For simplicity of exposition, we focus on *first-come, first-served* (FCFS) tiebreaking, which, among the jobs of minimal rank, serves the job that arrived to the system first. Our results also apply to SOAP policies that use *last-come, first-served* (LCFS) tiebreaking, which we defer to [^2].
Suppose job $J$ has descriptor $d_J$ and age $a_J$ and job $K$ has descriptor $d_K$ and age $a_K$. We say $J$ *outranks* $K$ if $r(d_J, a_J) \prec r(d_K, a_K)$ or both $r(d_J, a_J) = r(d_K, a_K)$ and $J$ arrived before $K$. SOAP policies always serve the job that outranks all other jobs in the system.
A great many SOAP policies can be expressed using $\mc{R} = \R$, in which case $\prec$ is the usual ordering $<$ on $\R$. However, it is often convenient to have a nested rank structure in which jobs are first prioritized by a *primary rank*, then by a *secondary rank*, and only then by the tiebreaking rule. We express such nested ranks using $\mc{R} = \R^2$. Each rank is a pair $\angle{r_1, r_2}$ of primary rank $r_1$ and secondary rank $r_2$, and $\prec$ is the lexicographic ordering: $\angle{r_1, r_2} \prec \angle{r_1', r_2'}$ if $r_1 < r_1'$ or both $r_1 = r_1'$ and $r_2 < r_2'$. We write primary and secondary ranks of a state $(d, a)$ as $r_1(d, a)$ and $r_2(d, a)$, respectively, so $$r(d, a) = \angle{r_1(d, a), r_2(d, a)}.$$
When specifying a SOAP policy, we usually leave the choice of $\mc{R}$ unstated, as it is implied from the formula for the rank function. Our results apply to $\mc{R} = \R^n$ ordered lexicographically for any $n \geq 1$, and they easily generalize to other choices of $\mc{R}$.
We will devote much time to discussing how jobs’ ranks change with age. Thus, when we call a rank function “monotonic” or similar, we mean that it is so with respect to age, not descriptor.
For a SOAP policy to be well-defined, its rank function must satisfy some technical conditions, which are given in .
SOAP Policies Are Everywhere {#sec:examples}
============================
Previously Analyzed SOAP Policies {#sub:examples_previous}
---------------------------------
\[ex:soap\_fb\] The *foreground-background* (FB) policy is a SOAP policy. It uses no static information, so $\mc{D} = \{{\emptyset}\}$, where ${\emptyset}$ is a “placeholder” descriptor assigned to every job. FB always serves the job of least age, so it has rank function $r({\emptyset}, a) = a$. It is likely that many jobs are tied for minimum rank under FB, but whichever job is served immediately loses minimum status, resulting in a processor-sharing effect.
There are always many rank functions that encode the same SOAP policy. For instance, any rank function monotonically increasing in age, such as $r({\emptyset}, a) = a^2$, also describes FB.
\[ex:soap\_fcfs\] The *first-come, first-served* (FCFS) policy is a SOAP policy. It uses no static information, so $\mc{D} = \{{\emptyset}\}$. FCFS is nonpreemptive, which is equivalent to always serving the job of maximal age, so it has rank function $r({\emptyset}, a) = -a$. FCFS tiebreaking plays a crucial role by breaking ties between jobs of age $0$.
Once again, there are multiple rank functions that describe FCFS. In particular, a constant rank function yields FCFS due to the tiebreaking rule, but we prefer the given encoding because it makes it clear that FCFS is a *nonpreemptive* policy. As the following examples demonstrate, using primary rank $-a$ is a general way to indicate nonpreemptiveness in a rank function.
\[ex:soap\_classes\] Consider a system with classes $\mc{K} = \{1, \dots, n\}$ where jobs within each class are served in FCFS order but class $1$ has highest priority, class $2$ has next-highest priority, and so on. The *nonpreemptive priority* and *preemptive priority* policies are SOAP policies. Both policies use job class as static information, so $\mc{D} = \mc{K}$.
- Nonpreemptive priority has rank function $r(k, a) = \angle{-a, k}$: the primary rank prevents preemption, and the secondary rank prioritizes the classes when starting a new job.
- Preemptive priority has rank function $r(k, a) = \angle{k, -a}$: because $k$ is the primary rank, jobs from high-priority classes preempt those in low priority classes.
\[ex:soap\_srpt\] The *shortest job first* (SJF), *preemptive shortest job first* (PSJF), and *shortest remaining processing time* (SRPT) policies are SOAP policies. All three policies assume exact size information is known and use it when scheduling, so all use $\mc{D} = \R_{\geq 0}$.
- SJF has rank function $r(x, a) = \angle{-a, x}$: it is a *nonpreemptive* priority policy with size as priority.
- PSJF has rank function $r(x, a) = \angle{x, -a}$: it is a *preemptive* priority policy with size as priority.
- SRPT has rank function $r(x, a) = x - a$: a job’s rank is its remaining size.
Newly Analyzed SOAP Policies {#sub:examples_new}
----------------------------
\[ex:soap\_serpt\] The *shortest expected processing time* (SEPT), *preemptive shortest expected processing time* (PSEPT) and *shortest expected remaining processing time* (SERPT) policies are SOAP policies. The policies are respective analogues of SJF, PSJF, and SRPT, but they do not have access to exact size information.
- SEPT has rank function $r(d, a) = \angle{-a, \E{X_d}}$: it is a *nonpreemptive* priority policy with expected size as priority.
- PSEPT has rank function $r(d, a) = \angle{\E{X_d}, -a}$: it is a *preemptive* priority policy with expected size as priority.
- SERPT has rank function $r(d, a) = \E[X_d > a]{X_d - a}$: a job’s rank is its expected remaining size.
While SEPT and PSEPT have analyses similar to those of SJF and PSJF, respectively, SERPT has never been analyzed before in full generality. We have left the set of descriptors $\mc{D}$ unspecified because the definitions above work for any set of descriptors.
For concreteness, consider a system where all jobs have the same descriptor ${\emptyset}$ and size either $2$ or $14$, each with probability $1/2$. The resulting rank function for SERPT, shown in , is *nonmonotonic* with respect to age. This differs from *every policy described in* , all of which have monotonic rank functions. The potential nonmonotonicity of SERPT’s rank function has prevented previous techniques from analyzing SERPT in full generality. We give the first response time analysis of SERPT using our general analysis of all SOAP policies ().
[\
]{} The rank function for SERPT using the distribution described in : jobs have size either $2$ or $14$, each with probability $1/2$. The rank is the expected remaining size of a job given it has reached its age $a$. In this case, the initial expected size is $8$, but if the job does not finish at age $2$, then we know it must be size $14$, so its expected remaining size jumps up to $12$.
\[ex:soap\_gittins\] The *Gittins index* of a job with descriptor $d$ and age $a$ is [@book_gittins; @m/g/1_gittins_aalto] $$G(d, a)
= \sup_{\Delta > 0} \frac{ \P[X_d > a]{X_d - a \leq \Delta}}{ \E[X_d > a]{\min\{X_d - a, \Delta\}}}
= \sup_{\Delta > 0} \frac{ \int_a^{a + \Delta} \density[d]{t} \, dt}{ \int_a^{a + \Delta} \tail[d]{t} \, dt},$$ where $\density[d]{}$ and $\tail[d]{}$ are the density and tail functions of $X_d$, respectively. The *Gittins index policy* is the scheduling policy that always serves the job of maximal Gittins index, and it is known to minimize mean response time in the M/G/1 queue [@book_gittins]. Although optimality of the Gittins index policy has long been known, only a few special cases have been analyzed in the past [@multiclass_ayesta; @rs_slowdown_hyytia]. The Gittins index policy is a SOAP policy with rank function $$r(d, a) = \frac{1}{G(d, a)}.$$ Like the policies in , the Gittins index policy can be defined with any set of descriptors.
The Gittins index policy is in general not the same as SERPT, as shown in , but, like SERPT, the Gittins index policy often uses a nonmonotonic rank function, making it impossible to analyze in general using previous techniques. We give the first response time analysis of the Gittins index policy using our general analysis of all SOAP policies ().
[\
]{} The rank function for the Gittins Index Policy using the same distribution as in : jobs have size either $2$ or $14$, each with probability $1/2$. Compared to SERPT, the Gittins index policy gives more priority to jobs before they reach age $2$. For instance, while SERPT ranks a job with age $1.99$ on par with a hypothetical job that deterministically has remaining size $6.01$, the Gittins index policy ranks such a job on par with a hypothetical job that deterministically has remaining size $0.02$. This reflects the fact that it is almost free to run such a job to age $2$, just in case it is about to finish. We show in that the Gittins index policy achieves lower mean response time than SERPT due to its prioritizing potentially short jobs.
\[ex:soap\_checkpoints\] Consider a system in which jobs, rather than being completely nonpreemptible or preemptible, are *preemptible at specific checkpoints*, say every $1$ time unit. The *discretized FB* policy is a variant of FB for jobs with checkpoints: when possible, it serves the job of minimal age, but it does not preempt jobs between checkpoints[^3]. Discretized FB is a SOAP policy. It uses no static information, so $\mc{D} = \{{\emptyset}\}$, and it has rank function $$r({\emptyset}, a) = \angle{\floor{a} - a, a},$$ This rank function is illustrated in . Roughly speaking, the primary rank encodes the “discretized” aspect, preempting a job only at integer ages $a$ when $\floor{a} - a = 0$, and the secondary rank encodes the “FB” aspect.
Discretized FB is just one example of a policy for jobs preemptible only at specific checkpoints, but we can “discretize” any other SOAP policy by using primary rank $\floor{a} - a$. For instance, *discretized SRPT* has rank function $r(x, a) = \angle{\floor{a} - a, x - a}$.
[\
]{} The rank function for the discretized FB policy described in . The primary rank (top) ensures that jobs are only preempted at integer ages. This is because new jobs enter the system with primary rank $0$, and jobs only have primary rank $0$ at integer ages. The secondary rank (bottom) prioritizes the job of lowest age, as in traditional FB.
We have seen a variety of features that SOAP policies can model:
- jobs that are nonpreemptible, preemptible, or preemptible at checkpoints;
- jobs with known or unknown exact size;
- priority based on a job’s exact size or expected size;
- class-based priority in multiclass systems; and
- priority that changes nonmonotonically as a job ages.
As the following examples show, SOAP policies go even further: they can *combine many such features* as part of a single policy.
\[ex:soap\_humans\_and\_robots\] Consider a system with two customer classes, humans ($H$) and robots ($R$).
- Humans, unpredictable and easily offended, have unknown service time, are nonpreemptible, and are served according to FCFS relative to other humans.
- Robots, precise and ruthlessly efficient, have known service time, are preemptible, and are served according to SRPT relative to other robots.
We can model the system using $\mc{D} = \{[H, ?]\} \cup \{[R, x] \mid x \in \R_{\geq 0}\}$, where $?$ denotes unknown size. A reasonable policy might have humans outrank most robots but let short robots, say those with remaining size less than some threshold $x_H$, outrank humans that have not yet started service. This results in rank function $$\begin{aligned}
r([H, ?], a) &= \angle{-a, x_H} \\
r([R, x], a) &= \angle{0, x - a},
\end{aligned}$$ which uses primary rank to encode preemptibility and secondary rank to encode priority. We analyze this system in .
Consider a system as in with $n = 3$ classes. Suppose that, in addition to class-based priority,
- jobs in class $1$ are preemptible, have known size, and are served according to SRPT;
- jobs in class $2$ are nonpreemptible, have known size, and are served according to SJF; and
- jobs in class $3$ are preemptible at specific checkpoints, have unknown size, and are served according to discretized FB, as in .
The static information of a job is its class and, if known, its size, so $\mc{D} = \{1, 2\} \times \R_{\geq 0} \cup \{[3, ?]\}$, where $?$ denotes unknown size. The rank function, which uses $\mc{R} = \R^3$ as the set of ranks, is $$\begin{aligned}
r([1, x], a) &= \angle{0, 1, x - a} \\
r([2, x], a) &= \angle{-a, 2, x} \\
r([3, ?], a) &= \angle{\floor{a} - a, 3, a}.
\end{aligned}$$ The components of the rank respectively encode preemptibility, class-based priority, and the policy used within each class.
SOAP Policies with LCFS Tiebreaking
-----------------------------------
There are two common SOAP policies that use LCFS tiebreaking instead of FCFS. The *last-come, first-serve* (LCFS) policy, which has the same rank function as FCFS in but uses LCFS tiebreaking. The rank function still ensures nonpreemption, but now ties between jobs of age $0$ are broken by LCFS. Similarly, the *preemptive last-come, first-serve* (PLCFS) policy is a SOAP policy with constant rank function and LCFS tiebreaking. SOAP policies that use LCFS tiebreaking admit essentially the same analysis as those that use FCFS tiebreaking ().
Non-SOAP Policies
-----------------
As our examples have demonstrated, there is an extremely wide variety of SOAP policies. However, there are some policies which are not SOAP policies, many of which fit into three broad categories.
First, some policies *cannot be expressed using descriptors that are distributed i.i.d. for each arriving job*. For example, the *earliest deadline first* (EDF) policy could be a SOAP policy if each job’s descriptor were its deadline, but deadlines cannot be i.i.d. because later arrivals need later deadlines.
Second, some policies *require a tiebreaking rule other than FCFS or LCFS*. For example, the *random order of service* (ROS) policy could be a SOAP policy using the rank function in if it could break ties between jobs of age $0$ differently. A future generalization of the SOAP class might allow for ROS tiebreaking because, like FCFS and LCFS, it serves one job at a time. In contrast, the *processor sharing* (PS) policy, which is also not a SOAP policy, requires a fundamentally different tiebreaking rule.
Third, some policies have *job priorities that are context-dependent*. For example, a nonpreemptive policy in a multiclass system that tries to alternate between serving jobs of class $1$ and class $2$ is not a SOAP policy, because the priority of a job depends on external context, namely the class of the previously served job. The rank function approach used by SOAP policies inherently considers each job individually, so there is no way to capture such context.
How to Handle Any Rank Function {#sec:key_ideas}
===============================
We have seen how to express a vast space of policies in the SOAP framework by careful choice of rank function. However, as demonstrated by , the rank function that encodes a SOAP policy can be very complicated. This leaves us with a difficult technical challenge: how do we analyze SOAP policies with arbitrary rank functions?
Before tackling arbitrary rank functions, let us recall how classic response time analyses work. Though there are of course many approaches, all of the policies in can be analyzed with the “tagged job” approach, which follows a particular job through the system to analyze its response time. For instance, consider tagging a job of size $x$ in a system using PSJF (). There are two types of jobs that outrank the tagged job:
- jobs of size at most $x$ that are present in the system when the tagged job arrives and
- jobs of size less than $x$ that arrive at the system after the tagged job arrives but before it completes.
One way to think about PSJF is to view the tagged job as seeing the system through “transformer glasses” [@book_harchol-balter] which transform the system by *hiding jobs that the tagged job outranks*. For PSJF, this transformation is simple because each job’s rank is essentially constant[^4]. A similar approach still works for policies with increasing or decreasing rank functions, but the hiding transformation becomes more complicated. For instance, under SRPT (), which has a decreasing rank function, a tagged job of size $x$ sees a system transformed as follows.
- Jobs that arrive after the tagged job are hidden if their size is at least $x - a$, where $a$ is the tagged job’s age when the other job arrives.
- Jobs that arrive before the tagged job are hidden if their *remaining* size is greater than $x$. It may be that a job of *initial* size greater than $x$ remains visible.
In more general terms: because jobs’ ranks change with age, the tagged job’s hiding criterion changes with its age, and whether or not other jobs satisfy that criterion changes with their ages. Handling these changes in rank is already tricky for SRPT, where a job’s rank only decreases with age. The situation becomes even more complex when working with *nonmonotonic* rank functions.
Conventions
-----------
For the remainder of this section, we examine the journey of a tagged job $J$ through the system. The tagged job $J$ has descriptor $d_J$ and size $x_J$. Note that we may use $J$’s size as part of our analysis even if the scheduler does not have access to exact job sizes.
Throughout, we call jobs other than $J$ *old* if they arrive before $J$ and *new* if they arrive after $J$. As a mnemonic, we name old jobs $I$ and name new jobs $K$, using subscripts when there are multiple such jobs. In examples, old jobs have descriptor $d_I$ and new jobs have descriptor $d_K$, though in a real system it may of course happen that different old or new jobs have different descriptors. When it is unspecified whether a job is new or old, we name the job $L$.
Nonmonotonicity Difficulties {#sub:nonmonotonicity}
----------------------------
There are two major obstacles to analyzing policies with arbitrary nonmonotonic rank functions that do not occur when analyzing SRPT, which has a monotonic rank function. We illustrate these obstacles below and in . The first obstacle concerns the nonmonotonicity of $J$’s rank.
- In SRPT, we *permanently* hide some other jobs based on $J$’s current rank.
- In general, another job $L$ might be only *temporarily* hidden. If $J$’s rank starts below $L$’s rank but later exceeds it, the initially hidden $L$ becomes visible again.
The second obstacle concerns the nonmonotonicity of the ranks of old jobs.
- In SRPT, an old job *permanently* outranks $J$ if served for long enough before $J$ arrives.
- In general, an old job $I$ might only *temporarily* outrank $J$. Furthermore, if $I$’s rank oscillates above and below $J$’s initial rank, whether $I$ gets hidden or stays visible depends on *when during $I$’s service $J$ arrives*.
[\
]{} We show two obstacles to the “transformer glasses” approach by examining the interaction between the tagged job $J$ and an old job $I$ of age $a_I$. The cyan curve shows $I$’s rank as a function of its age. First (top), suppose a tagged job $J$ has rank $r_J$ upon entering the system, and suppose that old job $I$ of age $a_I$ is already in the system. Consider how $J$ views $I$ if, as pictured, $r_J \prec r(d_I, a_I)$. Then $J$ outranks $I$, so $I$ is initially hidden. However, $I$ may be only *temporarily* hidden, because $J$’s rank may later increase to $r_J' \succ r(d_I, a_I)$. Second (bottom), consider the same jobs $J$ and $I$ and suppose $J$ has not entered the system yet. Whether or not $I$ will be hidden from $J$ depends on *$I$’s age when $J$ arrives*. For instance, as $I$ advances in age from $a_I$ to $a_I'$ and later $a_I''$, it switches back and forth between being visible to and hidden from $J$.
Dealing with such arbitrarily varying rank functions appears intractable. We need *two key insights* in order to handle nonmonotonic rank functions: the *Pessimism Principle* () and the *Vacation Transformation* ().
The Pessimism Principle {#sub:pessimism_principle}
-----------------------
We call the amount of time any job $L$ is served before $J$ completes *$J$’s delay due to $L\esub$*. The response time of $J$ is its size $x_J$ plus its delays due to all jobs that are in the system with $J$ at some point.
Suppose that a new job $K$ arrives when $J$ has age $a_J$. To analyze $J$’s response time, we have to know its delay due to $K$. As we saw in , deciding whether or not to hide $K$ based only on $J$’s current rank $r(d_J, a_J)$ will not work. Instead, we need to examine $J$’s *current and future ranks*.
\[def:rworst\] The *worst future rank* of a job with descriptor $d$, size $x$, and age $a$ is $${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}} = \sup_{\mathclap{a \leq b < x}} r(d, b).$$ Note that the worst future rank only considers ages up to the job’s size $x$. See for an illustration.
[\
]{} The relationship between rank $r(d, a)$ (solid cyan), and worst future rank ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ (dashed magenta) for a job with descriptor $d$ and size $x$.
The reason we care about $J$’s worst future rank is that for the purposes of computing $J$’s delay delay due to $K$, we can essentially pretend that $J$ has its worst future rank $${r_J}= {{\appopt{\subopt{r^{\mathrm{worst}}}{d_J, x_J}}{a_J}}},$$ as illustrated in . This is because before $J$ reaches its worst future rank, $K$ must either complete or surpass[^5] rank ${r_J}$.
[\
]{} A new job $K$ arrives when $J$, whose rank as a function of age is shown in cyan, has age $a_J$. Suppose for simplicity that $K$’s rank is constant at $r_K$. An instant after $J$ reaches age $a_J'$, it will be outranked by $K$, so $K$ will complete before $J$ reaches age ${a_J''}$. In practice, $K$’s rank may change, but $K$ will still complete before $J$ reaches age ${a_J''}$ unless $K$ surpasses rank ${r_J}$, in which case $K$ never outranks $J$ again. Thus, for the purposes of finding $J$’s delay due to $K$, we can pretend that $J$’s rank is ${r_J}$ for all of $J$’s ages before ${a_J''}$.
We have so far focused on a new job $K$, but the story is very similar for old jobs. The result is the Pessimism Principle, so named for its pessimistic focus on the worst future rank.
[Pessimism Principle]{} The tagged job $J$’s delay due to any other job $L$ is the amount of time $L$ is served until it either *completes* or *surpasses $J$’s worst future rank*. To be precise, letting $${r_J}(a) = {{\appopt{\subopt{r^{\mathrm{worst}}}{d_J, x_J}}{a}}},$$ this means the following.
- Each old job $I$ is served until it completes or first has rank $r_I \succ {r_J}(0)$. In particular, if $r_I = {r_J}(0)$, then $I$ outranks $J$ due to FCFS tiebreaking, thus the strict inequality[^6].
- Each new job $K$ is served until it completes or first has rank $r_K \succeq {r_J}(a_J)$, where $a_J$ is the age of $J$ when $K$ arrives.
Furthermore, in both cases, the service occurs *before $J$ is served while at its worst future rank*.
To clarify, the discussion above addresses the *total amount of service another job receives* while $J$ is in the system. The service need not be contiguous but might be interleaved with that of $J$ and other jobs.
The Pessimism Principle gives an implicit description of $J$’s delay due to any other job $L$, but it remains to explicitly find this delay’s distribution. We do this now for the case where $L$ is new, treating the case where $L$ is old in .
\[def:new\_work\] Let $r$ be a rank. The *new $r$-work* is a random variable, written ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$, representing how long a job that just arrived to the system is served until it completes or surpasses rank $r$. Specifically, we define ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}} = {{\sqappopt{\subopt{X^{\mathrm{new}}}{D}}{r}}}$, where $D$ is the random descriptor assigned to a new job and, for any specific descriptor $d$, $$\begin{aligned}
c_d[r] &= \inf\{a \geq 0 \mid r(d, a) \succeq r\} \\
{{\sqappopt{\subopt{X^{\mathrm{new}}}{d}}{r}}} &= \min\{X_d, c_d[r]\}.
\end{aligned}$$ That is, $c_d[r]$ is the *cutoff age* at which a new job with descriptor $d$ surpasses rank $r$. See for an illustration.
[\
]{} The new $r$-work for a job with descriptor $d$, whose rank as a function of age is shown in cyan, is the amount of service the job requires until it either completes or surpasses rank $r$ at the cutoff age $c_d[r]$. Pictorially, the new $r$-work is the amount of service the job requires while its age is in the green region. The new $r$-work is at most $c_d[r]$ but may be less if the job completes. Note that the new $r$-work is not impacted by the job’s rank at ages $a > c_d[r]$, even if $r(d, a) \prec r$.
Together, the Pessimism Principle and say that if a new job $K$ has random descriptor and arrives when $J$ has age $a_J$, then $J$’s delay due to $K$ is ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d_J, x_J}}{a_J}}}}}}$.
\[ex:new\_work\_srpt\] Consider SRPT as described in , in which a job’s descriptor is its size and its rank is its remaining size. Suppose that a new job $K$ of size $x_K$ arrives when $J$ has age $a_J$. Under SRPT, $J$’s worst future rank is its current rank $r(x_J, a_J) = x_J - a_J$, so the cutoff age for $K$ is $$c_{x_K}[x_J - a_J] =
\begin{cases}
\infty & \text{if } x_K < x_J - a_J \\
0 & \text{if } x_K \geq x_J - a_J.
\end{cases}$$ That is, $K$ will always outrank $J$ if $x_K < x_J - a_J$, and $K$ will never outrank $J$ if $x_K \geq x_J - a_J$. This means $J$’s delay due to $K$ is $${{\sqappopt{\subopt{X^{\mathrm{new}}}{x_K}}{x_J - a_J}}} = x_K \1(x_K < x_J - a_J),$$ where $\1$ is the indicator function.
When analyzing the response time of SRPT, we need to know $J$’s delay due to a *random* new job, meaning one with size drawn from the size distribution $X$. This is the new $(x_J - a_J)$-work, $${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{x_J - a_J}}} = {{\sqappopt{\subopt{X^{\mathrm{new}}}{X}}{x_J - a_J}}} = X \1(X < x_J - a_J).$$
\[ex:new\_work\_serpt\] Consider the SERPT system described in , in which *all* jobs have the same descriptor ${\emptyset}$ and the same two-point size distribution: jobs are size $2$ with probability $1/2$ and size $14$ otherwise. The rank function is $$r({\emptyset}, a) = \E[X > a]{X - a} =
\begin{cases}
8 - a & \text{if } a < 2 \\
14 - a & \text{if } a \geq 2,
\end{cases}$$ as shown in .
Suppose that a new job $K$ arrives when $J$ has age $a_J < 2$. The worst future rank of $J$ depends crucially on $J$’s size $x_J$.
- If $x_J = 2$, then $J$’s worst future rank is its current rank, $8 - a_J$. In this case, the cutoff age for $K$ is $c_{\emptyset}[8 - a_J] = 0$ because $K$’s initial rank is $8$, which is at least $J$’s worst future rank $8 - a_J$. This means $J$’s delay due to $K$ is $${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{8 - a_J}}} = {{\sqappopt{\subopt{X^{\mathrm{new}}}{{\emptyset}}}{8 - a_J}}} = 0.$$
- If instead $x_J = 14$, then $J$’s worst future rank is $12$. In this case, the cutoff age for $K$ is $c_{\emptyset}[12] = 2$ because $K$ will either complete or jump up to rank $12$ when it reaches age $2$. This means $J$’s delay due to $K$ is $${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{12}}} = {{\sqappopt{\subopt{X^{\mathrm{new}}}{{\emptyset}}}{12}}} = 2.$$
The Vacation Transformation {#sub:vacation_transformation}
---------------------------
We have seen how the Pessimism Principle shows us how long each new job delays the tagged job $J$. The question remains: how long does each old job delay $J$? This is much harder than the corresponding question for new jobs because *old jobs can have any age*, whereas new jobs always start at age $0$.
Fortunately, we are actually not directly concerned with the delay due to individual old jobs. What ultimately matters is the delay due to *all old jobs together*. It turns out that we can view this total delay as the *queueing time of a carefully transformed system*. The careful transformation in question is the Vacation Transformation, but we are still a few definitions away from presenting it.
Our goal is to find the $J$’s total delay due to old jobs. We can think of this delay as the amount of “relevant” work in the system at the moment $J$ arrives. Because Poisson arrivals see time averages [@pasta_wolff], the distribution of the amount relevant work seen by $J$ upon arrival is the *stationary distribution* of the amount of relevant work. Thus, in this section, we imagine $J$ as a *witness* to the system, watching other jobs enter, receive service, and exit. For this purpose, the most important fact about $J$ is its worst future rank upon arrival, $${r_J}= {{\appopt{\subopt{r^{\mathrm{worst}}}{d_J, x_J}}{0}}}.$$
So far, we have considered old jobs as a monolithic category, but it is useful to consider three subcategories. At any moment in time, we can classify old jobs as follows.
- *Discarded* old jobs currently have rank greater than ${r_J}$.
- *Original* old jobs currently have rank at most ${r_J}$ and *have always had* rank at most ${r_J}$ since arriving themselves.
- *Recycled* old jobs currently have rank at most ${r_J}$ but had rank greater than ${r_J}$ at some point in the past.
More generally, we may call a job discarded, original, or recycled *with respect to rank $r\esub$*, in which case we replace ${r_J}$ with $r$.
An old job $I$ goes through the following transitions between these categories, as shown in Figures \[fig:transformer\_glasses\] and \[fig:old\_work\].
- When $I$ arrives in the system, if its initial rank is at most ${r_J}$, it starts out original. Otherwise, it starts out discarded.
- As $I$ ages, it may become discarded if it is not already.
- As $I$ ages further, it may become recycled, then discarded again, then recycled again, and so on until it completes.
Eventually, $J$ will arrive, and each of the old jobs will delay $J$ by some amount of time based on their category at the moment when $J$ arrives. By the Pessimism Principle, $J$’s delay due to discarded jobs is $0$, so such jobs do not concern us further. Original jobs are similar to new jobs but, due to the FCFS tiebreaking rule, not quite the same. Recycled jobs are the most difficult type of old job to handle, but fortunately, as we will soon see, $J$ sees *at most one* recycled job in the system when it arrives.
For the purposes of analyzing $J$’s response time, we view the system through “transformer glasses” [@book_harchol-balter] through which *only original and recycled jobs are visible*, as illustrated in . In the transformed system, jobs are transformed such that when they become discarded, they complete. Thus, the total work in the transformed system is exactly the “relevant” work in the untransformed system, which would be $J$’s total delay due to old jobs were $J$ to arrive immediately. Therefore, our goal is to find the *stationary distribution of work in the transformed system*.
[\
]{} A system contains jobs $I_1$, $I_2$, and $I_3$ with respective ages $a_1$, $a_2$, and $a_3$. All three jobs have the same descriptor $d_I$ and therefore the same rank as a function of age, drawn in cyan. We view the system from the transformed perspective of witness $J$, which has worst future rank ${r_J}$. By the Pessimism Principle, if $J$ were to arrive at the system now, it would only be delayed by original job $I_1$ and recycled job $I_2$, so discarded job $I_3$ is completely hidden from $J$. Furthermore, jobs $I_1$ and $I_2$ only delay $J$ until they either complete or exceed rank ${r_J}$, so there are upper bounds on the delays due to each, shown as magenta bars.
Both original and recycled jobs arrive in the transformed system. Arrivals of original jobs correspond to arrivals to the untransformed system, but arrivals of recycled jobs occur seemingly arbitrarily, as they are really caused by discarded jobs transitioning to recycled in the untransformed system. A busy period in the transformed system always starts with the arrival of an original or recycled job. Arrivals of original jobs continue during the busy period, but *no more recycled jobs arrive* for the rest of the busy period. This is because for a recycled job to arrive in the transformed system, a discarded job has to become recycled by receiving service in the untransformed system. But discarded jobs never outrank original or recycled jobs, which are present for the entire busy period in the transformed system, so such transitions never occur.
To analyze the amount of work in the transformed system, we need to know how long each old job spends as an original or recycled job. We call the amount of time an old job spends as original with respect to rank $r$ its *$0$-old $r$-work*, and we call the amount of time it spends as recycled for the $i$th time with respect to rank $r$ its *$i$-old $r$-work*, both of which we now define formally.
\[def:old\_work\] Let $r$ be a rank and $d$ be a descriptor. The *$0$-old $r$-interval* for descriptor $d$ is the interval of ages during which a job of descriptor $d$ is original with respect to rank $r$. Specifically, the interval is $[b_{0, d}[r], c_{0, d}[r]]$, where[^7] $$\begin{aligned}
b_{0, d}[r] &= 0 \\
c_{0, d}[r] &= \inf\{a \geq b_{0, d}[r] \mid r(d, a) \succ r\}.
\end{aligned}$$ For $i \geq 1$, the *$i$-old $r$-interval* for descriptor $d$ is the interval of ages during which the job is recycled with respect to rank $r$ for the $i$th time. Specifically, the interval is $[b_{i, d}[r], c_{i, d}[r]]$, where $$\begin{aligned}
b_{i, d}[r] &= \inf\{a > c_{i - 1, d}[r] \mid r(d, a) \preceq r\} \\
c_{i, d}[r] &= \inf\{a > b_{i, d}[r] \mid r(d, a) \succ r\}.
\end{aligned}$$ If $b_{i, d}[r] = c_{i, d}[r] = \infty$, the interval is the empty set. See for an illustration.
For $i \geq 0$, the *$i$-old $r$-work* is a random variable, written ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}$, representing how long a job will be served while its age is in its $i$-old $r$-interval. Specifically, we define ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} = {{\sqappopt{\subopt{X^{\mathrm{old}}}{i, D}}{r}}}$, where $D$ is the random descriptor assigned to a new job and, for any specific descriptor $d$, $${{\sqappopt{\subopt{X^{\mathrm{old}}}{i, d}}{r}}} =
\begin{cases}
0 & \text{if } X_d < b_{i, d}[r] \\
X_d - b_{i, d}[r] & \text{if } b_{i, d}[r] \leq X_d < c_{i, d}[r] \\
c_{i, d}[r] - b_{i, d}[r] & \text{if } c_{i, d}[r] \leq X_d.
\end{cases}$$ If $b_{i, d}[r] = c_{i, d}[r] = \infty$, we define ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i, d}}{r}}} = 0$.
[\
]{} As a job ages, it can transition repeatedly between being discarded and recycled with respect to a rank $r$. The $i$-old $r$-interval for descriptor $d$ is the interval $[b_{i, d}[r], c_{i, d}[r]]$ during which a job of descriptor $d$ is original ($i = 0$) or recycled for the $i$th time ($i \geq 1$). We highlight the $i$-old $r$-intervals in green. The $i$-old $r$-work is the amount of service the job requires while its age is in its $i$-old $r$-interval.
Suppose old job $I$ has a random descriptor. In the transformed system, $I$ receives service
- for time ${{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{{r_J}}}}$ as an original job and,
- for all $i \geq 1$, for time ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{r_J}}}}$ as a job being recycled for the $i$th time.
Note that ${{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{{r_J}}}}$ may be $0$, representing $I$ starting out discarded, and ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{r_J}}}}$ for $i \geq 1$ may be $0$, representing $I$ completing before being recycled for the $i$th time.
\[ex:old\_work\_srpt\] Consider SRPT as described in , in which a job’s descriptor is its size and its rank is its remaining size. Suppose that $J$ witnesses an old job $I$ of initial size $x_I$. Under SRPT, every job’s rank is strictly decreasing with age, so $J$’s worst future rank is its initial size $x_J$. The amount of time $I$ spends as an original or recycled job depends on its size relative to $J$’s.
- If $x_I \leq x_J$, then $I$ is original until its completion because its rank never exceeds $x_J$, so $${{\sqappopt{\subopt{X^{\mathrm{old}}}{0, x_I}}{x_J}}} = x_I.$$ $I$ is never recycled, so ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{x_J}}} = 0$ for $i \geq 1$.
- If $x_I > x_J$, then $I$ is starts out discarded but becomes recycled at age $x_I - x_J$, at which point it has remaining size $x_J$, so $$\begin{aligned}
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0, x_I}}{x_J}}} &= 0 \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{1, x_I}}{x_J}}} &= x_J.
\end{aligned}$$ $I$ is recycled only once, so ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i, x_I}}{x_J}}} = 0$ for $i \geq 2$.
When analyzing the response time of SRPT, we need to know the amount of time a *random* old job, meaning one with size drawn from the size distribution $X$, spends as an original or recycled job. From the above casework, we obtain $$\begin{aligned}
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{x_J}}} = {{\sqappopt{\subopt{X^{\mathrm{old}}}{0, X}}{x_J}}} &= X\1(X \leq x_J) \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{1}}{x_J}}} = {{\sqappopt{\subopt{X^{\mathrm{old}}}{1, X}}{x_J}}} &= x_J\1(X > x_J),
\end{aligned}$$ where $\1$ is the indicator function, and ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{x_J}}} = 0$ for $i \geq 2$.
The Vacation Transformation gives a simple specification of the amount of work the witness $J$ sees in the transformed system. It follows from three observations.
- The amount of work in the transformed system is independent of the scheduling policy used on transformed jobs, provided it is work-conserving, so we may assume FCFS among original and recycled jobs[^8].
- Because Poisson arrivals see time averages [@pasta_wolff], the stationary amount of work in the transformed system is the same as the stationary FCFS queueing time of an original job in the transformed system.
- In the transformed system, the arrival process of recycled jobs is not Poisson, but they only appear at the starts of busy periods, so it is convenient to view them as server vacations.
[Vacation Transformation]{} Consider the tagged job $J$ with descriptor $d_J$ and size $x_J$ arriving to the system, and let $${r_J}= {{\appopt{\subopt{r^{\mathrm{worst}}}{d_J, x_J}}{0}}}.$$ $J$’s total delay due to old jobs has the same distribution as *queueing time in a transformed M/G/1/FCFS system with “sparse” server vacations*. We call the transformed system the *Vacation Transformation System*, or simply *VT System*. In the VT System, jobs arrive at rate $\lambda$ and have size distribution ${{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{{r_J}}}}$, and several types of vacations occasionally occur, as described below.
The server in the VT System is always in one of three states:
- *busy*, meaning a job is in service;
- *idle*, meaning the system is empty but the server is ready to start a job immediately should one arrive; or
- *on vacation*, meaning the server will not serve jobs until the vacation finishes, even if there are jobs in the system.
The server only starts vacations when the system is empty. Unlike job arrivals, vacation start times are not a Poisson process. Each vacation has a *type* $i \geq 1$ determining its length. Specifically, type $i$ vacations have i.i.d. lengths drawn from distribution $$V_i = ({{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{r_J}}}} \mid {{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{r_J}}}} > 0).$$ The stationary probability that the the server is on a type $i$ vacation is $\lambda\E{{{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{r_J}}}}}$. The exact process determining when a type $i$ vacation starts is intractable, but to analyze queueing time in the VT System, as we do in , it fortunately suffices to know just this stationary probability.
Response Time of SOAP Policies {#sec:response_time}
==============================
Having spent the previous section understanding the perspective of a tagged job in a system using a SOAP policy, we are now ready to apply our insights to analyze the response time of SOAP policies. Specifically, we analyze $T_{d, x}$, the response time of a tagged job $J$ with descriptor $d$ and size $x$.
When analyzing traditional policies like SRPT or PSJF (), it often helps to think of $J$’s response time as the sum of two independent random variables [@srpt_analysis_schrage; @book_harchol-balter]:
- *waiting time*, written ${T^{\mathrm{wait}}}_{d, x}$, the time from when $J$ arrives to when it first enters service; and
- *residence time*, written ${T^{\mathrm{res}}}_{d, x}$, the time from when $J$ first enters service to when it exits.
Unfortunately, the usual strategy for analyzing waiting and residence time relies on $J$’s rank never increasing, which holds for SRPT and PSJF but does not hold for a great many SOAP policies. To overcome this obstacle, we *replace $J$’s rank with its worst future rank*, which never increases. This rank substitution, fully justified in , is made possible by the Pessimism Principle (). We call $J$ with its adjusted rank the *rank-substituted* tagged job.
After replacing $J$’s rank with its worst future rank, it remains to analyze its waiting and residence times. Even though $J$’s worst future rank is monotonic, other jobs’ ranks may both increase and decrease with age, making these analyses challenging.
- We can think of *waiting time* as a transformed busy period: the initial transformed work is $J$’s total delay due to old jobs, and each arriving new job’s transformed size is the amount it delays $J$. The main challenge is finding the distribution of initial transformed work, which requires using the Vacation Transformation (). We analyze waiting time in .
- We would also like to think of *residence time* as a transformed busy period. The initial work is simply $J$’s size $x$, but now the arriving new jobs present a challenge: because $J$’s worst future rank may decrease with time, not all new jobs have the same transformed size distribution. We analyze residence time in .
Waiting Time {#sub:waiting_time}
------------
Hereafter, “transform” means Laplace-Stieltjes transform. We write the transform of a random variable $V$ as $\tl{V}(s)$ and the transform of a parametrized random variable $V[r]$ as $\tl{V}[r](s)$.
Recall that the waiting time of the rank-substituted tagged job $J$ is the amount of time between $J$’s arrival and when $J$ first enters service. As mentioned previously, we can think of waiting time as a transformed busy period: the initial work is $J$’s total delay due to old jobs, and each arriving new job’s size is the amount it delays $J$. This type of busy period is formalized in the following definition.
Let $r$ be a rank. The *new $r$-work busy period*, written ${{\sqappopt{\subopt{B^{\mathrm{new}}}{}}{r}}}$, is the length of a busy period in an M/G/1 system with arrival rate $\lambda$ and job size ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$. Its transform satisfies [@book_harchol-balter] $${\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{r}}}{s}} = {\appopt{{\sqappopt{\subopt{\tl{X}^{\mathrm{new}}}{}}{r}}}{s + \lambda(1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{r}}}{s}})}}.$$ More generally, the new $r$-work busy period *started by work $W$*, written ${{\sqappopt{\subopt{B^{\mathrm{new}}}{W}}{r}}}$, is the length of a busy period in the same M/G/1 system with a random initial amount of work $W$. It has transform [@book_harchol-balter] $$\label{eq:busy_work}
{\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{W}}{r}}}{s}} = \tl{W}(s + \lambda(1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{r}}}{s}})).$$
Let $r = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{0}}}$ be $J$’s worst future rank upon arrival. Because $J$ is not served until its residence time, $r$ remains $J$’s worst future rank for the entirety of $J$’s waiting time. This means $J$’s delay due to each new job that arrives during its waiting time is ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$, so $J$’s waiting time is a new $r$-work busy period started by initial work $W$, where $W$ is $J$’s total delay due to old jobs.
All that remains is to determine $W$, for which it is convenient to define two new notations. First, let $$\begin{aligned}
{{\sqappopt{\subopt{\rho^{\mathrm{new}}}{}}{r}}} &= \lambda\E{{{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}} \\
{{\sqappopt{\subopt{\rho^{\mathrm{old}}}{i}}{r}}} &= \lambda\E{{{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}} \\
{{\sqappopt{\subopt{\rho^{\mathrm{old}}}{\Sigma}}{r}}} &= \sum_{i = 0}^\infty {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{i}}{r}}}\end{aligned}$$ be the “loads contributed by” new $r$-work, $i$-old $r$-work, and all old $r$-work, respectively. Second, let ${{\sqappopt{\subopt{Y^{\mathrm{old}}}{i}}{r}}}$ be the *equilibrium distribution* [@book_harchol-balter], or length-biased sample, of ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}$. It has transform $${\appopt{{\sqappopt{\subopt{\tl{Y}^{\mathrm{old}}}{i}}{r}}}{s}} = \frac{1 - {\appopt{{\sqappopt{\subopt{\tl{X}^{\mathrm{old}}}{i}}{r}}}{s}}}{s\E{{{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}}}.$$ Note that ${{\sqappopt{\subopt{Y^{\mathrm{old}}}{i}}{r}}}$ is also the equilibrium distribution of the length of a type $i$ vacation in the VT System, $V_i = ({{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} \mid {{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} > 0)$, because samples of length $0$ are never encountered in a length-biased sample.
\[lem:waiting\_time\] Under any SOAP policy, the Laplace-Stieltjes transform of waiting time for a rank-substituted tagged job with descriptor $d$ and size $x$ is $${\tl{T}^{\mathrm{wait}}}_{d, x}(s)
= \frac{ 1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{\Sigma}}{r}}}
+ \sum_{i = 1}^\infty {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{i}}{r}}}{\appopt{{\sqappopt{\subopt{\tl{Y}^{\mathrm{old}}}{i}}{r}}}{\sigma}}}{ 1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}}{\appopt{{\sqappopt{\subopt{\tl{Y}^{\mathrm{old}}}{0}}{r}}}{\sigma}}},$$ where $r = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{0}}}$ and $\sigma = s + \lambda(1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{r}}}{s}})$.
Call the rank-substituted tagged job $J$. As previously mentioned, ${T^{\mathrm{wait}}}_{d, x}$ is a new $r$-work busy period started by initial work $W$, where $W$ is $J$’s total delay due to old jobs. This means ${\tl{T}^{\mathrm{wait}}}_{d, x}(s) = \tl{W}(\sigma)$ by , so it remains only to compute $\tl{W}(\sigma)$.
The Vacation Transformation states that $W$ has the same distribution as the queueing time in the VT System, which is a particular M/G/1/FCFS system with vacations. A decomposition result of @vacations_fuhrmann [Equation (4)] states that the number of jobs in an M/G/1/FCFS system with vacations, such as the VT System, is distributed as the sum of two independent random variables:
- the number $N_Q$ of jobs in the queue of a vacation-free M/G/1/FCFS system; and
- the number $N_V$ of jobs in the queue observed by an arriving job conditional on observing a non-busy server.
Recalling the Vacation Transformation and a standard result for the M/G/1 queue [@book_harchol-balter], we immediately obtain the probability generating function for $N_Q$, $$\label{eq:pgf_nq}
\hat{N}_Q(z)
= \frac{1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}}}{1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}}{\appopt{{\sqappopt{\subopt{\tl{Y}^{\mathrm{old}}}{0}}{r}}}{\lambda(1 - z)}}}.$$
In the VT System, a job arriving to a non-busy server observes $N_V$ as one of the following.
- If the server is idle, which happens with probability $1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{\Sigma}}{r}}}$, there are $0$ jobs in the queue.
- If the server is in the middle of a type $i$ vacation, which happens with probability ${{\sqappopt{\subopt{\rho^{\mathrm{old}}}{i}}{r}}}$, the number of jobs in the queue is the number of Poisson arrivals at rate $\lambda$ during time ${{\sqappopt{\subopt{Y^{\mathrm{old}}}{i}}{r}}}$, which is the amount of time since the start of the type $i$ vacation.
Accounting for the fact that we measure $N_V$ only when a job arrives to a non-busy server, which happens with probability $1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}}$, we obtain the probability generating function of $N_V$, $$\label{eq:pgf_a}
\hat{N}_V(z)
= \frac{ 1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{\Sigma}}{r}}}
+ \sum_{i = 1}^\infty {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{i}}{r}}}{\appopt{{\sqappopt{\subopt{\tl{Y}^{\mathrm{old}}}{i}}{r}}}{\lambda(1-z)}}}{ 1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}}}.$$ Multiplying and , applying the distributional version of Little’s Law[^9] [@distributional_little_keilson], and substituting $\sigma = \lambda(1 - z)$ gives the transform of queueing time in the VT System, which matches the desired transform $\tl{W}(\sigma)$.
Residence Time {#sub:residence_time}
--------------
\[lem:residence\_time\] Under any SOAP policy, the Laplace-Stieltjes transform of residence time for a rank-substituted tagged job with descriptor $d$ and size $x$ is $${\tl{T}^{\mathrm{res}}}_{d, x}(s)
= \exp\biggl(
-\lambda \int_0^x (1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}}}}{s}}) \, da
\biggr).$$
We view residence time as a *sum of many small busy periods*, each started by a small amount of work $\delta$. We then take the $\delta \to 0$ limit, which exists thanks to conditions in . This is very similar to the argument used to compute the transform of residence time under SRPT [@srpt_analysis_schrage]. Specifically, we divide $[0, x]$ into chunks of size $\delta$ and consider each small busy period started by the work needed to bring the job from age $a$ to age $a + \delta$. In the $\delta \to 0$ limit, for the entirety of the small busy period starting at age $a$, we can assume the job has rank ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$.
By the Pessimism Principle, the amount of work in the small busy period starting at age $a$ is the length of a new ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$-work busy period started by work $\delta$, which by has transform $${\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{\delta}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}}}}{s}}
= \exp(-\delta\lambda(1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}}}}{s}})).$$ The lengths of the small busy periods are independent because the arrival process is Poisson, so the residence time, which is the total amount of work in all such busy periods, has transform $$\prod_{i = 0}^{\mathclap{x/\delta}}
{\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{\delta}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{\delta i}}}}}}{s}}
= \exp\biggl(
-\delta\lambda \sum_{i = 0}^{\mathclap{x/\delta}}
(1 - {\appopt{{\sqappopt{\subopt{\tl{B}^{\mathrm{new}}}{}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{\delta i}}}}}}{s}})
\biggr).$$ Taking the $\delta \to 0$ limit yields the desired expression.
Total Response Time
-------------------
\[thm:response\_time\_transform\] Under any SOAP policy, the Laplace-Stieltjes transform of response time of jobs with descriptor $d$ and size $x$ is $$\tl{T}_{d, x}(s) = {\tl{T}^{\mathrm{wait}}}_{d, x}(s){\tl{T}^{\mathrm{res}}}_{d, x}(s),$$ where ${\tl{T}^{\mathrm{res}}}_{d, x}(s)$ and ${\tl{T}^{\mathrm{wait}}}_{d, x}(s)$ are as in Lemmas \[lem:waiting\_time\] and \[lem:residence\_time\], respectively.
Because the arrival process is Poisson, ${T^{\mathrm{wait}}}_{d, x}$ and ${T^{\mathrm{res}}}_{d, x}$ are independent, so the result follows from Lemmas \[lem:waiting\_time\] and \[lem:residence\_time\]
\[thm:response\_time\_mean\] Under any SOAP policy, the mean response time of jobs with descriptor $d$ and size $x$ is $$\E{T_{d, x}}
= \frac{ \lambda \sum_{i = 0}^\infty \E{({{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}})^2}}{ 2(1 - {{\sqappopt{\subopt{\rho^{\mathrm{old}}}{0}}{r}}})(1 - {{\sqappopt{\subopt{\rho^{\mathrm{new}}}{}}{r}}})}
+ \int_0^x \frac{1}{1 - {{\sqappopt{\subopt{\rho^{\mathrm{new}}}{}}{r(a)}}}} \, da,$$ where $r(a) = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ and $r = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{0}}}$.
This follows from $\E{T_{d, x}} = -\tl{T}_{d, x}'(0)$ after straightforward computation.
We can use $T_{d, x}$ to analyze $T_d$, the response time of jobs with descriptor $d$ and any size, and $T$, the overall response time of all jobs. Specifically, $\tl{T}_d(s) = \E{\tl{T}_{d, X_d}(s)}$, where $X_d$ is the size distribution of jobs with descriptor $d$, and $\tl{T}(s) = \E{\tl{T}_D(s)}$, where $D$ is the random descriptor assigned to a new job. Analyzing response time of jobs with size $x$ and any descriptor is similar, but it requires computing $D_x$, the distribution of descriptors of jobs with size $x$.
New Analyses for Specific Policies {#sec:applications}
==================================
In this section, we analyze the response time of several policies discussed in . The main challenge to analyzing SOAP policies is determining ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$ and ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}$. Throughout, we give expressions for these focusing only on ranks $r$ that are important for the final result. Specifically, for the possible descriptors $d$ and sizes $x$, we only need to find
- ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}}}}$ for all ages $a$ and
- ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{0}}}}}}$.
For simplicity, we give final formulas only for mean response time.
Discretized FB {#sub:discretized_fb}
--------------
Consider the discretized FB policy (), which can only preempt jobs when their age is at specific checkpoints spaced $1$ time unit apart. We represent this using the rank function $$r({\emptyset}, a) = \angle{\floor{a} - a, a},$$ shown in . As in the analysis of traditional FB, it is convenient to work in terms of the *capped size distribution* and its associated load [@book_harchol-balter Section 30.3], $$\label{eq:capped}
\begin{aligned}
X_{\ol{x}} &= \min\{X, x\} \\
\rho_{\ol{x}} &= \lambda\E{X_{\ol{x}}}.
\end{aligned}$$
In discretized FB, a job of size $x$ and age $a$ has worst future rank $${{\appopt{\subopt{r^{\mathrm{worst}}}{{\emptyset}, x}}{a}}} =
\begin{cases}
\angle{0, \floor{x}} & \text{if } a \leq \floor{x} \\
\angle{\floor{a} - a, a} & \text{if } a > \floor{x}.
\end{cases}$$ The new $r$-work and $0$-old $r$-work relevant for obtaining response time are therefore $$\begin{aligned}
{{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{\angle{0, \floor{x}}}}} &= X_{{ \mkern 4mu\overline{\mkern-4mu\floor{x}\mkern-2.9mu}\mkern 2.9mu}} \\
{{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{\angle{\floor{a} - a, a}}}} &= 0 \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{\angle{0, \floor{x}}}}} &= X_{{ \mkern 4mu\overline{\mkern-4mu\floor{x} + 1\mkern-2.2mu}\mkern 2.2mu}},\end{aligned}$$ where $\floor{x} < a < x$ in the second equation. From , we conclude the following.
\[prop:discretized\_fb\] Under discretized FB (), the mean response time of jobs with size $x$ is[^10] $$\E{T_x}
= \frac{ \lambda\mathbf{E}\bigl[X_{{ \mkern 4mu\overline{\mkern-4mu\ceil{x}\mkern-2.9mu}\mkern 2.9mu}}^2\bigr]}{ 2\bigl(1 - \rho_{{ \mkern 4mu\overline{\mkern-4mu\ceil{x}\mkern-2.9mu}\mkern 2.9mu}}\bigr)\bigl(1 - \rho_{{ \mkern 4mu\overline{\mkern-4mu\floor{x}\mkern-2.9mu}\mkern 2.9mu}}\bigr)}
+ \frac{\floor{x}}{1 - \rho_{{ \mkern 4mu\overline{\mkern-4mu\floor{x}\mkern-2.9mu}\mkern 2.9mu}}} + x - \floor{x}.$$
Mixture of Known and Unknown Job Sizes {#sub:humans_and_robots}
--------------------------------------
Consider the “humans and robots” system (). This scenario has two features that were previously difficult to analyze.
- Some jobs have known size, namely robots, while others have unknown size, namely humans.
- Some jobs are preemptible, namely robots, while others are nonpreemptible, namely humans.
Let
- $X_H$ and $X_R$ be the respective size distributions of humans and robots,
- $p_H$ and $p_R = 1 - p_H$ be the respective probabilities that a given arrival is a human or a robot, and
- $\lambda_H = \lambda p_H$ and $\lambda_R = \lambda p_R$ be the respective arrival rates of humans and robots.
Recall that humans all have descriptor $[H, ?]$, indicating that their size is unknown, and robots each have a descriptor of the form $[R, x]$, indicating that their exact size is $x$. The rank function is $$\begin{aligned}
r([H, ?], a) &= \angle{-a, x_H} \\
r([R, x], a) &= \angle{0, x - a},\end{aligned}$$ where $x_H$ is a constant.
Both humans and robots have maximal primary rank, namely $0$, upon entering the system, so the only nonzero new $r$-work and $i$-old $r$-work occur for ranks of the form $r = \angle{0, x}$: $$\begin{aligned}
{{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{\angle{0, x}}}} &=
\begin{cases}
X_H \1(x_H < x) & \text{with probability } p_H \\
X_R \1(X_R < x) & \text{with probability } p_R
\end{cases} \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{\angle{0, x}}}} &=
\begin{cases}
X_H \1(x_H \leq x) & \text{with probability } p_H \\
X_R \1(X_R \leq x) & \text{with probability } p_R
\end{cases} \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{1}}{\angle{0, x}}}} &=
\begin{cases}
X_H \1(x_H > x) & \text{with probability } p_H \\
x \1(X_R > x) & \text{with probability } p_R,
\end{cases}\end{aligned}$$ where $\1$ is the indicator function. These follow from arguments similar to those given for SRPT in Examples \[ex:new\_work\_srpt\] and \[ex:old\_work\_srpt\]. Finally, ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} = 0$ for $i \geq 2$, so yields the following.
In the humans and robots system (), the mean response time of humans is $$\E{T_H}
= \frac{ \lambda_H\E{X_H^2} + \lambda_R\E{(X_R)_{\ol{x_H}}^2}}{ 2(1 - \rho_H - \rho_{R \leq x_H})(1 - \rho_{R < x_H})}
+ \E{X_H},$$ and the mean response time of robots with size $x$ is = + \_0\^x dt, where $(X_R)_{\ol{x_H}}\esub$ and $(X_R)_{\ol{x}}\esub$ are capped distributions as defined in and $$\begin{aligned}
\rho_H &= \lambda_H \E{X_H} \\
\rho_{R < x} &= \lambda_R \E{X_R \1(X_R < x)} \\
\rho_{R \leq x} &= \lambda_R \E{X_R \1(X_R \leq x)}.
\end{aligned}$$
The Gittins Index Policy
------------------------
As discussed in , the Gittins index policy can have a nonmonotonic rank function, and thus only special cases of it have been analyzed in the past [@multiclass_ayesta; @rs_slowdown_hyytia]. Theorems \[thm:response\_time\_transform\] and \[thm:response\_time\_mean\] give us exactly the framework we need to analyze the Gittins index policy used with *any* set of descriptors and job size distributions, though ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$ and ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}}$ do not have a general closed form and thus require details of the system to derive. We start in this section by using the SOAP framework to analyze the model considered by @multiclass_ayesta, which is relatively simple thanks to the fact that the rank functions involved are monotonic. In , we move to a more difficult setting in which the Gittins index policy has a nonmonotonic rank function.
We consider a system with two job classes $A$ and $B$, which serve as our descriptors, with respective arrival rates $\lambda_A = \lambda p_A$ and $\lambda_B = \lambda p_B$ and respective Pareto size distributions $X_A$ and $X_B$. Specifically, $$\P{X_A > t} = \biggl(1 + \frac{t}{\beta_A}\biggr)^{-\alpha_A}$$ and symmetrically for $B$, where $\alpha_A, \alpha_B > 1$ and $\beta_A, \beta_B > 0$ are parameters of the distributions. The rank function is [@m/g/1_gittins_aalto; @multiclass_ayesta] $$r(A, a) = \frac{1}{G(A, a)} = \frac{\beta_A + a}{\alpha_A}$$ and symmetrically for $B$. The rank is strictly increasing in $a$, so ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}} = r(d, x)$ for all ages $a < x$. Strictly increasing rank also means jobs are never recycled, so ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} = 0$ for all $i \geq 1$.
It remains only to compute ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$ and ${{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{r}}}$. As in , let $$c_A[r] = \max\{\alpha_A r - \beta_A, 0\}$$ be the age at which a class $A$ job first surpasses rank $r$, and symmetrically for $B$. Note that $c_A[r(A, x)] = x$. Because the rank function is strictly increasing, $${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}} = {{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{r}}} =
\begin{cases}
(X_A)_{\ol{c_A[r]}} & \text{with probability } p_A \\
(X_B)_{\ol{c_B[r]}} & \text{with probability } p_B,
\end{cases}$$ where $(X_A)_{\ol{c_A[r]}}$ and $(X_B)_{\ol{c_B[r]}}$ are capped distributions as defined in . By , $$\E{T_{A,x}}
= \frac{ \lambda_A\E{(X_A)_{\ol{x}}^2}
+ \lambda_B\E{(X_B)_{\ol{y}}^2}}{ 2(1 - \rho_{\ol{x}, \ol{y}})^2}
+ \frac{x}{1 - \rho_{\ol{x}, \ol{y}}},$$ where $$\begin{aligned}
y &= c_B[r(A, x)] \\
\rho_{\ol{x}, \ol{y}}
&= \lambda_A\E{(X_A)_{\ol{x}}}
+ \lambda_B\E{(X_B)_{\ol{y}}}.\end{aligned}$$ Of course, $\E{T_{B, x}}$ is symmetrical. It is simple to verify that, modulo notation, this matches the results of @multiclass_ayesta.
Case Study: SERPT vs. the Gittins Index {#sub:gittins_serpt}
---------------------------------------
We now analyze both SERPT and the Gittins index policy on the size distribution introduced in . Both policies have nonmonotonic rank functions in this case, so we need the full power of to compute mean response times. All jobs have descriptor ${\emptyset}$ and the same two-point size distribution. We write the size distribution as $X = {\mathbf{CoinFlip}}\{2, 14\}$, meaning jobs have size $2$ with probability $1/2$ and size $14$ otherwise.
We begin by analyzing SERPT. We have computed ${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}}$ for SERPT with this size distribution in , $${{\sqappopt{\subopt{X^{\mathrm{new}}}{}}{r}}} =
\begin{cases}
0 & r \leq 8 \\
2 & r > 8
\end{cases}$$ We only need to compute $i$-old $r$-work for ranks ${{\appopt{\subopt{r^{\mathrm{worst}}}{{\emptyset}, 2}}{0}}} = 8$ and ${{\appopt{\subopt{r^{\mathrm{worst}}}{{\emptyset}, 14}}{0}}} = 12$. From the rank function plot in , we see $$\begin{aligned}
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{8}}} &= 2 \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{12}}} &= {\mathbf{CoinFlip}}\{2, 14\} \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{1}}{8}}} &= {\mathbf{CoinFlip}}\{0, 8\} \\
{{\sqappopt{\subopt{X^{\mathrm{old}}}{1}}{12}}} &= 0.\end{aligned}$$ The most subtle of these is ${{\sqappopt{\subopt{X^{\mathrm{old}}}{1}}{8}}}$: the total time an old job $I$ spends as recycled with respect to rank $8$ is either $0$, if $I$ has size $2$, or $8$, if $I$ has size $14$. Finally, ${{\sqappopt{\subopt{X^{\mathrm{old}}}{i}}{r}}} = 0$ for $i \geq 2$. Applying yields size-specific mean response times $$\begin{aligned}
\E{T_2^{\mathrm{SERPT}}}
&= \frac{18\lambda}{1 - 2\lambda} + 2 \\
\E{T_{14}^{\mathrm{SERPT}}}
&= \frac{50\lambda}{(1 - 8\lambda)(1 - 2\lambda)}
+ \frac{6}{1 - 2\lambda} + 8.\end{aligned}$$
[\
]{} The rank function for SERPT where jobs have two-point size distribution ${\mathbf{CoinFlip}}\{2,14\}$, meaning size $2$ with probability $1/2$ and size $14$ otherwise. The same rank function appears in . The $0$-old (original) and $1$-old ($1$-recycled) intervals are highlighted in green. A job is original with respect to rank $8$ until age $2$, when its rank jumps up if it does not complete. Upon reaching age $6$, the job has remaining size $8$, so it becomes recycled.
The Gittins index policy for the same system has rank function $r({\emptyset}, a) = 1/G({\emptyset}, a)$, where, by the definition in , $$\frac{1}{G({\emptyset}, a)} =
\begin{cases}
4 - 2a & \text{if } a < 2 \\
14 - a & \text{if } a \geq 2.
\end{cases}$$ This rank function is illustrated in . Broadly speaking, the Gittins index policy places higher priority on jobs of age $a < 2$ than SERPT does. Like that of SERPT, the rank function is piecewise linear with negative slopes, so we omit the very similar analysis and simply state the size-specific mean response times: $$\begin{aligned}
\E{T_2^{\mathrm{Gittins}}}
&= \frac{6\lambda}{1 - 2\lambda} + 2 \\
\E{T_{14}^{\mathrm{Gittins}}}
&= \frac{50\lambda}{(1 - 8\lambda)(1 - 2\lambda)}
+ \frac{10}{1 - 2\lambda} + 4.\end{aligned}$$ As expected due to its prioritization of jobs of age $a < 2$, the Gittins index policy has shorter mean response time for jobs of size $2$ but longer mean response time for jobs of size $14$. The Gittins index policy is known to minimize overall mean response time, and it performs as promised: $$\begin{aligned}
\E{T_2^{\mathrm{SERPT}}} - \E{T_2^{\mathrm{Gittins}}}
&= \frac{12\lambda}{1 - 2\lambda} \\
\E{T_{14}^{\mathrm{SERPT}}} - \E{T_{14}^{\mathrm{Gittins}}}
&= \frac{-8\lambda}{1 - 2\lambda}.\end{aligned}$$ Because the two job sizes are equally likely, the Gittins index policy has lower overall mean response time than SERPT.
Conclusion
==========
We introduce *SOAP policies*, a very broad class of scheduling policies for the M/G/1 queue. The characteristic feature of a SOAP policy is its *rank function*, which maps each possible state a job could be in to a *rank*, meaning priority level. The SOAP class includes many policies old and new. While the mean response times of some relatively simple SOAP policies have been analyzed previously, the vast majority of SOAP policies, in particular those with *nonmonotonic* rank functions, have resisted analysis. Using two key technical insights, the *Pessimism Principle* and the *Vacation Transformation*, we overcome the obstacles presented by nonmonotonic rank functions to present a *universal response time analysis* that applies to any SOAP policy.
Our universal analysis applies to some notable policies. Among these is the *Gittins index policy*, which has long been known to minimize mean response time in settings where exact job sizes are not known. While prior work [@multiclass_ayesta; @rs_slowdown_hyytia] was restricted to the case of known job sizes or distributions with the decreasing hazard rate property, our analysis can handle the Gittins index policy with *arbitrary size distributions*, which was previously intractable. Our universal analysis also applies to several *practically motivated systems*, such as those in which jobs are only preemptible at certain checkpoints or only some jobs’ exact sizes are known. More broadly, we are optimistic that techniques similar to our Pessimism Principle and Vacation Transformation could help analyze the response times of scheduling policies in more complex M/G/1 settings, such as systems with setup times or server vacations.
We thank Peter van de Ven and the anonymous referees for their helpful comments. Ziv Scully was supported by an scholarship and the Graduate Research Fellowship Program under Grant No. . Mor Harchol-Balter was supported by NSF-, NSF-, NSF-, and a Faculty Award from .
Extension to LCFS Tiebreaking {#app:lcfs}
=============================
SOAP policies that use LCFS tiebreaking admit almost exactly the same analysis as those that use FCFS tiebreaking. As explained in detail below, the entire analysis is unchanged except for *reversing the strictness of rank comparisons* in Definitions \[def:new\_work\] and \[def:old\_work\], meaning swapping $\prec$ and $\preceq$ and swapping $\succ$ and $\succeq$.
Throughout Sections \[sub:pessimism\_principle\] and \[sub:vacation\_transformation\], which follow a tagged job $J$ through the system, we distinguish between *new* jobs, which arrive after $J$, and *old* jobs, which arrive before $J$. When there are multiple jobs of minimal rank, FCFS tiebreaking prioritizes old jobs, then $J$, and then new jobs. This prioritization affects the *strictness of rank comparisons* when defining new $r$-work and old $r$-work. For instance, defines $$c_d[r] = \inf\{a \geq 0 \mid r(d, a) \succeq r\},$$ whereas defines $$c_{0, d}[r] = \inf\{a \geq 0 \mid r(d, a) \succ r\},$$ which is the same but with $\succ$ in place of $\succeq$. When $J$’s worst future rank is $r$, the above values each represent a “cutoff age” before which a job of descriptor $d$ outranks $J$. Under FCFS tiebreaking, a new job $K$ outranks $J$ until $K$’s rank is *at least* $r$, whereas an old job $I$ outranks $J$ until $I$’s rank *strictly exceeds* $r$. Under LCFS tiebreaking, this situation is reversed, which manifests as reversing the strictness of rank comparisons.
Rank Function Details {#app:rank_function}
=====================
In order to ensure that a SOAP policy is well-defined, its rank function $r$ must satisfy the following conditions.
- With respect to descriptor, $r$ must be *piecewise continuous* to ensure that certain expectations are well-defined.
- With respect to age, $r$ must be *piecewise monotonic* and *piecewise differentiable* to determine when and how to share the processor between multiple jobs. Any compact region of $\R_{\geq 0}$ must contain only finitely many boundary points between pieces. Furthermore, upwards jump discontinuities must be continuous from the right[^11].
These conditions allow us to define a SOAP policy as the limit of discrete-time priority policies, with the limit taken as the discretization increment approaches $0$. When $\mc{R} = \R^2$ ordered lexicographically, the limiting policy is , which has a clear generalization to $\mc{R} = \R^n$. In , we say a job is “in state $(d, a)$” to mean it has descriptor $d$ and age $a$.
Let $\mc{J}$ be the set of jobs in states $(d, a)$ of minimal $r_1(d, a)$.
- Within $\mc{J}$, consider jobs such that $r$ is strictly decreasing in age. If there are any, schedule the job of minimal $r_2$, using FCFS tiebreaking if there are multiple such jobs.
- Otherwise, within $\mc{J}$, consider jobs such that $r_1$ is constant and $r_2$ is strictly increasing in age. If there are any, share the processor between all such jobs, giving a job in state $(d, a)$ share proportional to $1 / \partial_ar_2(d, a)$.
- Otherwise, $\mc{J}$ must only contain jobs such that $r_1$ is strictly increasing. Share the processor between jobs in $\mc{J}$, giving a job in state $(d, a)$ share proportional to $1 / \partial_ar_1(d, a)$.
Worst Future Rank Details {#app:rworst}
=========================
For simplicity of exposition, throughout , we assumed that a job’s worst future rank ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ was actually attained by that job in the future. However, there are two cases where the supremum in is *not* be attained by some age $a$: when there is a jump discontinuity or when the maximum is at the open boundary $a = x$. For example, in , if a job has integer size $x$, then the job never attains rank $\angle{0, x}$.
There are multiple ways to remedy the situation. The most intuitive is to say that ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ is not a rank but a *rank bound*. The set of rank bounds is $\mc{R} \times \{-1, 0\}$ ordered lexicographically. An ordinary rank $r$ corresponds to the pair $(r, 0)$ representing the *closed* upper bound $r' \preceq r$ over other ranks $r'$, but the pair $(r, -1)$ is “just below” $(r, 0)$, representing the *open* upper bound $r' \prec r$.
The corrections to definitions are as follows. In , we define the worst future rank to be the rank bound $${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}
= (\sup_{\mathclap{a \leq b < x}} r(d, b),
-\1(\text{the supremum is not attained}))$$ instead of just a rank. In Definitions \[def:new\_work\] and \[def:old\_work\], instead of defining $r$-work for a rank $r$, we define $(r, q)$-work for rank bounds $(r, q)$. When we compare a rank $r'$ against a rank bound $(r, q)$, we compare rank bound $(r', 0)$ against $(r, q)$. Concretely, in , we define $$c_d[(r, q)] = \inf\{a \geq 0 \mid (r(d, a), 0) \succeq (r, q)\},$$ and similarly for $b_{i, d}[(r, q)]$ and $c_{i, d}[(r, q)]$ in .
For example, consider the analysis of discretized FB in . When a job has integer size $x$, the supremum in ${{\appopt{\subopt{r^{\mathrm{worst}}}{{\emptyset}, x}}{a}}}$ is attained only in the $b \to x$ limit. Thus, ${{\sqappopt{\subopt{X^{\mathrm{old}}}{0}}{{{\appopt{\subopt{r^{\mathrm{worst}}}{\emptyset,x}}{0}}}}}} = X_{\ol{x}}$ when $x$ is an integer, not $X_{\ol{x + 1}}$ as would follow from the uncorrected . To correct for this, uses $\ceil{x}$ instead of $\floor{x} + 1$.
The above discussion assumes FCFS tiebreaking. As discussed in , the strictness of rank comparisons in Definitions \[def:new\_work\] and \[def:old\_work\] is reversed under LCFS tiebreaking, but the same changes described above apply without issue.
The Rank-Substituted Tagged Job {#app:rank_substitution}
===============================
In this section, we show that for the purposes of analyzing $T_{d, x}$, we can use a *rank-substituted* tagged job.
[Worst Future Rank Substitution]{} Consider an arbitrary arrival sequence that includes the arrival of a tagged job $J$ with descriptor $d$ and size $x$. The response time of $J$ is unaffected if we *schedule $J$ as if its rank were ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ instead of $r(d, a)\esub$* at every age $a$, without otherwise changing the arrival sequence or scheduling policy. We call this process *rank substitution*.
Worst Future Rank Substitution is a direct consequence of the Pessimism Principle (). To see why it holds, consider two systems experiencing identical job arrivals, including tagged job $J$.
- *System A* is unmodified, so $J$ is scheduled according to its current rank.
- *System B* uses rank substitution, so $J$ is scheduled according to its worst future rank.
We say the two systems *synchronize* at time $t$ if they contain the same jobs at the same ages at $t$. The systems clearly synchronize at $J$’s arrival time. We show below that the systems also synchronize at many other points in time, one of which is $J$’s exit time, so $J$’s response time is the same in each system.
The Pessimism Principle states that all of $J$’s delay due to another job $L$ occurs *before $J$ is served while at its worst future rank*. This suggests we should focus on $J$’s *worst future age*, which when $J$ has age $a$ is $${{\appopt{\subopt{a^{\mathrm{worst}}}{d, x}}{a}}} = \inf\{b \geq a \mid r(d, b) = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{b}}}\},$$ namely the earliest age at which $J$ attains its worst future rank[^12]. See for an illustration. Note that $a = {{\appopt{\subopt{a^{\mathrm{worst}}}{d, x}}{a}}}$ if and only if $r(d, a) = {{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$.
[\
]{} The relationship between rank $r(d, a)$ (solid cyan), worst future rank ${{\appopt{\subopt{r^{\mathrm{worst}}}{d, x}}{a}}}$ (dashed magenta), and worst future age ${{\appopt{\subopt{a^{\mathrm{worst}}}{d, x}}{a}}}$ (solid green) for a job with descriptor $d$ and size $x$. Age and worst future age coincide for ages in the intervals $[v_1, w_1]$ and $[v_2, w_2]$.
As $J$ ages, its worst future age alternates between being a constant future age and its current age. Let $[v_i, w_i]$ be the $i$th interval of ages $a$ such that $a = {{\appopt{\subopt{a^{\mathrm{worst}}}{d, x}}{a}}}$. It is convenient to set $w_0 = 0$ and $v_{n + 1} = x$, where $n$ is the number of $[v_i, w_i]$ intervals. We will show that System A and System B synchronize when, for some $i$, either
- $J$ has age $v_i$ and is in service or
- $J$’s age is in $(v_i, w_i]$.
It is clear that if the systems synchronize when $J$ is served at age $v_i$, then the systems remain synchronized until $J$ reaches age $w_i$, because $J$’s rank in the two systems is identical until $J$ reaches age $w_i$. Thus, it suffices to show that if the systems synchronize when $J$ has age $w_{i - 1}$, then the systems synchronize when $J$ is served at age $v_i$.
Consider how the two systems change during the interval between their synchronization when $J$ has age $w_{i - 1}$. Let $t_{\mathrm{A}}$ be the moment when $J$ is served at age $v_i$ in System A, and symmetrically for $t_{\mathrm{B}}$ in System B.
- In System A, by the Pessimism Principle, each other job $L$ in the system during the interval is served until it either completes or surpasses $J$’s worst future rank $r(d, v_i)$, after which $L$ is never served again. This is because ${{\appopt{\subopt{a^{\mathrm{worst}}}{d, x}}{a}}} = v_i$ for all $a \in (w_{i - 1}, v_i]$.
- In System B, by the Pessimism Principle, each other job $L$ is served for the same amount of time as in System A, because rank substitution does not change $J$’s worst future rank.
Thus, if System A experiences the same arrivals before $t_{\mathrm{A}}$ that System B experiences before $t_{\mathrm{B}}$, then the systems synchronize at $t_{\mathrm{A}} = t_{\mathrm{B}}$, as desired. Suppose for contradiction that one system, say System A, experiences an extra arrival. Because the arrival sequence is the same for the two systems, this occurs only if $t_{\mathrm{A}} > t_{\mathrm{B}}$. But by work conservation and the observations above, at $t_{\mathrm{B}}$, System A must serve $J$ at age $v_i$, so $t_{\mathrm{A}} = t_{\mathrm{B}}$ after all, contradicting $t_{\mathrm{A}} > t_{\mathrm{B}}$.
[^1]: While SRPT is optimal when exact job sizes are known, the Gittins index policy, of which SRPT is a special case, is optimal even when only size distributions are known.
[^2]: Sometimes ties for minimum rank lead to processor sharing, and these ties do not require tiebreaking. See and for details.
[^3]: We note that it is possible to model discretized FB as an MLPS policy with infinitely many thresholds.
[^4]: Due to FCFS tiebreaking, we could encode PSJF without its secondary rank.
[^5]: Unless otherwise noted, we mean “surpass” in a weak sense. That is, to surpass a rank $r$ means to attain rank at least $r$.
[^6]: See for discussion of corner cases where we need $\succeq$ instead of $\succ$.
[^7]: See for discussion of corner cases where we need $\succeq$ instead of $\succ$.
[^8]: Recall that an original or recycled job completes the transformed system if it either completes *or becomes discarded* in the untransformed sysytem.
[^9]: Here we apply the distributional version of Little’s Law to the VT System, not the original SOAP system, which is crucial because the law applies only to FCFS systems.
[^10]: See for discussion of why we use $\ceil{x}$ instead of $\floor{x} + 1$.
[^11]: That is, if a job jumps from low rank to high rank at age $a$, its rank exactly at age $a$ should be the high rank.
[^12]: As discussed in , a job’s worst future rank is sometimes only attained in a limit due to the job’s completion or a jump in the rank function. Accounting for this changes only minor details in the following discussion.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present a sampling-based framework for multi-robot motion planning which combines an implicit representation of a roadmap with a novel approach for pathfinding in geometrically embedded graphs tailored for our setting. Our pathfinding algorithm, *discrete-RRT* (dRRT), is an adaptation of the celebrated RRT algorithm for the discrete case of a graph, and it enables a rapid exploration of the high-dimensional configuration space by carefully walking through an implicit representation of a tensor product of roadmaps for the individual robots. We demonstrate our approach experimentally on scenarios of up to 60 degrees of freedom where our algorithm is faster by a factor of at least ten when compared to existing algorithms that we are aware of.'
author:
- 'Kiril Solovey[^1]'
- Oren Salzman
- Dan Halperin
title: 'Finding a needle in an exponential haystack: Discrete RRT for exploration of implicit roadmaps in multi-robot motion planning[^2]'
---
Introduction
============
*Multi-robot motion planning* is a fundamental problem in robotics and has been extensively studied. In this work we are concerned with finding paths for a group of robots, operating in the same workspace, moving from start to target positions while avoiding collisions with obstacles as well as with each other. We consider the continuous formulation of the problem, where the robots and obstacles are geometric entities and the robots operate in a configuration space, e.g., $\dR^d$ (as opposed to the discrete variant, sometimes called the where the robots move on a graph). Moreover, we assume that each robot has its own start and target positions, as opposed to the unlabeled case (see, e.g., [@abhs-emrmp13; @kh-pim05; @sh-kcmr; @TMK13]).
Previous work
-------------
We assume familiarity with the basic terminology of motion planning. For background, see, e.g., [@clhbkt-prmp; @l-pa]. Initial work on motion planning aimed to develop *complete* algorithms, which guarantee to find a solution when one exists or report that none exists otherwise. Such algorithms for the multi-robot case exist [@ss-pm3; @ss-cmp91; @y-cms84] yet are exponential in the number of robots. The exponential running time, can be attributed to the high number of *degrees of freedom* (*dof*)—the sum of the dofs of the individual robots.
The introduction of *sampling-based* algorithms such as the *probabilistic roadmap method* (PRM) [@kslo-prm], the *rapidly-exploring random trees* (RRT) [@kl-rrtc] and their many variants, had a significant impact on the field of [motion planning]{}due to their efficiency, simplicity and applicability to a wide range of problems. Sampling-based algorithms attempt to capture the connectivity of the *configuration space* ([C-space]{}) by sampling collision-free configurations and constructing a *roadmap*—a graph data structure where the free configurations are vertices and the edges represent collision-free paths between nearby configurations. Although these algorithms are not complete, most of them are *probabilistically complete*, that is, they are guaranteed to find a solution, if one exists, given a sufficient amount of time. Recently, Karaman and Frazzoli [@KF11] introduced several variants of these algorithms such that, with high probability they produce paths that are *asymptotically optimal* with respect to some quality measure. Sampling-based algorithms can be easily extended to the multi-robot case by considering the fleet of robots as one composite robot [@sl-upp]. Such a naive approach suffers from inefficiency as it overlooks aspects that are unique to the multi-robot problem. More tailor-made sampling-based techniques have been proposed for the multi-robot case [@hh-hmp; @shh-mms2; @sh-kcmr]. Particularly relevant to our efforts is the work of Švestka and Overmars [@so-cppmr] who suggested to construct a composite roadmap which is a Cartesian product of roadmaps of the individual robots. Due to the exponential nature of the resulting roadmap, this technique is only applicable to problems that involve a modest number of robots. A recent work by Wagner et al. [@wkh-ppp] suggests that the composite roadmap does not necessarily have to be explicitly represented. Instead, they maintain an implicitly represented composite roadmap, and apply their M\* algorithm [@wc-mstar] to efficiently retrieve paths, while minimizing the explored portion of the roadmap. The resulting technique is able to cope with a large number of robots, for certain types of scenarios. Additional information on these two approaches is provided in Section \[sec:c\_roadmaps\] below.
Contribution
------------
We present a sampling-based algorithm for the multi-robot motion-planning problem called *multi-robot discrete RRT* (MRdRRT). Similar to the approach of Wagner et al. [@wkh-ppp], we maintain an implicit representation of the composite roadmap. We propose an alternative, highly efficient, technique for pathfinding in the roadmap, which can cope with scenarios that involve tight coupling of the robots. Our new approach, which we call dRRT, is an adaptation of the celebrated RRT algorithm [@kl-rrtc] for the discrete case of a graph, embedded in Euclidean space[^3]. dRRT traverses a composite roadmap that may have exponentially many neighbors (exponential in the number of robots that need to be coordinated). The efficient traversal is achieved by retrieving only partial information of the explored roadmap. Specifically, it considers a single neighbor of a visited vertex at each step. dRRT rapidly explores the [C-space]{}represented by the implicit graph. Integrating the implicit representation of the roadmap allows us to solve multi-robot problems while exploring only a small portion of the [C-space]{}.
We demonstrate the capabilities of MRdRRT on the setting of polyhedral robots translating and rotating in space amidst polyhedral obstacles. We provide experimental results on several challenging scenarios, where MRdRRT is faster by a factor of at least ten when compared to existing algorithms that we are aware of. We show that we manage to solve problems of up to 60 dofs for highly coupled scenarios (Figure \[fig:3d\_scenarios\]).
The organization of this paper is as follows. In Section \[sec:c\_roadmaps\] we elaborate on two sampling-based multi-robot motion planning algorithms, namely the composite roadmap approach by Švestka and Overmars [@so-cppmr] and the work on subdimensional expansion and M\* by Wagner et al. [@wc-mstar; @wkh-ppp]. In Section \[sec:d\_rrt\] we introduce the dRRT algorithm. For clarity of exposition, we first describe it as a general pathfinding algorithm for geometrically embedded graphs. In the following section (Section \[sec:implementation\]) we describe the MRdRRT method where dRRT is used in the setting of multi-robot motion-planning problem for the exploration of the implicitly represented composite roadmaps. We show in Section \[sec:experimental\_results\] experimental results for the algorithm on different scenarios and conclude the paper in Section \[sec:discussion\] with possible future research directions.
\[Twisty\] [ ![3D environments with robots that are allowed to rotate and translate (6DOFs). In scenarios (a),(b),(c) robots of the same color need to exchange positions. (a) Twisty scenario with 8 corkscrew-shaped robots, in a room with a barrier. (b) Abstract scenario with 8 L-shaped robots. (c) Cubicles scenario with 10 L-shaped robots. (d) Home scenario with 5 table-shaped robots that are placed in different rooms. The goal is to change rooms in a clockwise order. The scenario were constructed using meshes that are provided by the Open Motion Planning Library [@ompl] (OMPL 0.10.2) distribution.](twisty.png "fig:"){width="45.00000%"} ]{} \[Abstract \] [ ![3D environments with robots that are allowed to rotate and translate (6DOFs). In scenarios (a),(b),(c) robots of the same color need to exchange positions. (a) Twisty scenario with 8 corkscrew-shaped robots, in a room with a barrier. (b) Abstract scenario with 8 L-shaped robots. (c) Cubicles scenario with 10 L-shaped robots. (d) Home scenario with 5 table-shaped robots that are placed in different rooms. The goal is to change rooms in a clockwise order. The scenario were constructed using meshes that are provided by the Open Motion Planning Library [@ompl] (OMPL 0.10.2) distribution.](abstract.png "fig:"){width="45.00000%"} ]{} \[Cubicles \] [ ![3D environments with robots that are allowed to rotate and translate (6DOFs). In scenarios (a),(b),(c) robots of the same color need to exchange positions. (a) Twisty scenario with 8 corkscrew-shaped robots, in a room with a barrier. (b) Abstract scenario with 8 L-shaped robots. (c) Cubicles scenario with 10 L-shaped robots. (d) Home scenario with 5 table-shaped robots that are placed in different rooms. The goal is to change rooms in a clockwise order. The scenario were constructed using meshes that are provided by the Open Motion Planning Library [@ompl] (OMPL 0.10.2) distribution.](cubicles.png "fig:"){width="45.00000%"} ]{} \[Home \] [ ![3D environments with robots that are allowed to rotate and translate (6DOFs). In scenarios (a),(b),(c) robots of the same color need to exchange positions. (a) Twisty scenario with 8 corkscrew-shaped robots, in a room with a barrier. (b) Abstract scenario with 8 L-shaped robots. (c) Cubicles scenario with 10 L-shaped robots. (d) Home scenario with 5 table-shaped robots that are placed in different rooms. The goal is to change rooms in a clockwise order. The scenario were constructed using meshes that are provided by the Open Motion Planning Library [@ompl] (OMPL 0.10.2) distribution.](home.png "fig:"){width="45.00000%"} ]{}
\[fig:3d\_scenarios\]
Composite roadmaps for multi-robot motion planning {#sec:c_roadmaps}
==================================================
We describe the composite roadmap approach introduced by Švestka and Overmars [@so-cppmr]. Here, a Cartesian product of PRM roadmaps of individual robots is considered as a means of devising a roadmap for the entire fleet of robots. However, since they consider an explicit construction of this roadmap, their technique is applicable to scenarios that involve only a small number of robots. To overcome this, Wagner et al. suggest [@wc-mstar; @wkh-ppp] to represent the roadmap *implicitly* and describe a novel algorithm to find paths on this implicit graph.
Let $r_1,\ldots,r_m$ be $m$ robots operating in a workspace $W$ with start and target configurations $s_i,t_i$. We wish to find paths for every robot from start to target, while avoiding collision with obstacles as well as with the other robots. Let $G_i=(V_i,E_i)$ be a PRM roadmap for $r_i$, $|V_i|=n$, and let $k$ denote the maximal degree of a vertex in any $G_i$. In addition, assume that $s_i,t_i \in V_i$, and that $s_i,t_i$ reside in the same connected component of $G_i$. Given such a collection of roadmaps $G_1,\ldots, G_m$ a composite roadmap can be defined in two different ways—one is the result of a *Cartesian product* of the individual roadmaps while in the other a *tensor product* is used [@wiki-gp].
The *composite roadmap* $\dG=(\dV, \dE)$ is defined as follows. The vertices $\dV$ represent all combinations of collision-free placements of the $m$ robots. Formally, a set of $m$ robot configurations $C=(v_1,\ldots,v_m)$ is a vertex of $\dG$ if for every $i$, $v_i\in V_i$, and in addition, when every robot $r_i$ is placed in $v_i$ the robots are pairwise collision-free. The Cartesian and tensor products differ in the type of edges in the resulting roadmap. If the Cartesian product is used, then $(C,C')\in \dE$, where $C=(v_1,\ldots,v_m), C'\in(v'_1,\ldots,v'_m)$, if there exists $i$ such that $(v_i,v'_i)\in E_i$, for every $j\neq i$ it holds that $v_j=v'_j$, and $r_i$ does not collide with the other robots stationed at $v_j=v'_j$ while moving from $v_i$ to $v'_i$. A tensor product generates many more edges. Specifically, $(C,C')\in\dE$ if $(v_i,v'_i)\in E_i$ for every $i$, and the robots remain collision-free while moving on the respective single-graph edges.
**Remark.**
Wagner et al. propose an adaptation of A\* to the case of a composite roadmap called M\* [@wc-mstar]. Their approach exploits the observation that only the motion of some robots has to be coupled in typical scenarios. Thus, planning in the joint [C-space]{}is only required for robots that have to be coupled, while the motion of the rest of the robots can be planned individually. Hence, their method dynamically explores low-dimensional search spaces embedded in the full [C-space]{}, instead of the joint high-dimensional [C-space]{}. This technique is highly effective for scenarios with a low degree of coupling, and can cope with large fleets of robots in such settings. However, when the degree of coupling increases, we observed sharp increase in the running time of this algorithm, as it has to consider many neighbors of a visited vertex.
Discrete RRT {#sec:d_rrt}
============
We describe a technique which we call *discrete RRT* (dRRT) for pathfinding in implicit graphs that are embedded in a Euclidean space. For clarity of exposition, we first describe dRRT without the technicalities related to motion planning. We add these details in the subsequent section. As the name suggests, dRRT is an adaptation of the RRT algorithm [@kl-rrtc] for the purpose of exploring discrete geometrically-embedded graphs, instead of a continuous space.
Since the graph serves as an approximation of some relevant portion of the Euclidean space, traversal of the graph can be viewed as a process of exploring the subspace. The dRRT algorithm rapidly explores the graph by biasing the search towards vertices embedded in unexplored regions of the space.
Let $G=(V,E)$ be a graph where every $v\in V$ is embedded in a point in Euclidean space $\dR^d$ and every edge $(v,v')\in E$ is a line segment connecting the points. Given two vertices $s,t\in V$, dRRT searches for a path in $G$ from $s$ to $t$. For simplicity, assume that the graph is embedded in $[0,1]^d$.
Similarly to its continuous counterpart, dRRT grows a tree rooted in $s$ and attempts to connect it to $t$ to form a path from $s$ to $t$. As in RRT, the growth of the tree is achieved by extending it towards random samples in $[0,1]^d$. In our case though, vertices and edges that are added to the trees are taken from $G$, and we do not generate new vertices and edges along the way.
As $G$ is represented implicitly, the algorithm uses an oracle to retrieve information regarding neighbors of visited vertices. We first describe this oracle and then proceed with a full description of the dRRT algorithm. Finally, we show that this technique is *probabilistically complete*.
Oracle to query the implicit graph
----------------------------------
In order to retrieve partial information regarding the neighbors of visited vertices, dRRT consults an oracle described below. We start with several basic definitions.
Given two points $v,v'\in [0,1]^d$, denote by $\rho(v,v')$ the ray that starts in $v$ and goes through $v'$. Given three points $v,v',v''\in [0,1]^d$, denote by $\angle_v(v',v'')$ the (smaller) angle between $\rho(v,v')$ and $\rho(v,v'')$.
Given a vertex $v\in V$, and a point $u\in [0,1]^d$ we define $$\orad(v,u):={\operatornamewithlimits{argmin}}_{v'}\left\{\angle_v(u,v')|(v,v')\in E\right\}.$$
In other words, the direction oracle returns the neighbor $v'$ of $v$ such that the direction from $v$ to $v'$ is closest to the direction from $v$ to $u$.
Description of dRRT
-------------------
At a high level, dRRT proceeds similar to the RRT algorithm, and we repeat it here for completeness. The dRRT algorithm (Algorithm \[alg:planner\]) grows a trees $\T$ which is a subgraphs of $G$ and is rooted in $s$ (line 1). The growth of $\T$ (line 3) is achieved by an expansion towards random samples. Additionally, an attempt to connect $\T$ with $t$ is made (line 4). The algorithm terminates when this operation succeeds and a solution path is generated (line 6), otherwise the algorithm repeats line 2.
Expansion of $\T$ is performed by the EXPAND operation (Algorithm \[alg:expand\]) which performs $N$ iterations that consist of the following steps: A point $q_{\text{rand}}$ is sampled uniformly from $[0,1]^d$ (line 2). Then, a node $q_{\text{near}}$ that is the closest to the sample (in Euclidean distance), is selected (line 3). $q_{\text{near}}$ is extended towards the sample by locating the vertex $q_{\text{new}}\in V$, that is the neighbor of $q_{\text{near}}$ in $G$ in the direction of $q_{\text{rand}}$ (by the direction oracle $\orad$). Once $q_{\text{new}}$ is found (line 4), it is added to the tree (line 6) with the edge $(q_{\text{near}},q_{\text{new}})$ (line 7). See an illustration of this process in Figure \[fig:dRRT\_alg\]. This is already different from the standard RRT as we cannot necessarily proceed exactly in the direction of the random point.
After the expansion, dRRT attempts to connect the tree $\T$ with $t$ using the CONNECT\_TO\_TARGET operation (Algorithm \[alg:connect\]). For every vertex $q$ of $\T$, which one of the $K$ nearest neighbors of $t$ in $\T$ (line 1), an attempt is made to connect $q$ to $t$ using the method LOCAL\_CONNECTOR (line 2) which is a crucial part of the dRRT algorithm (see Subsection \[sub:local\]).
Finally, given a path from some node $q$ of $\T$ to $t$ the method RETRIEVE\_PATH (Algorithm \[alg:planner\], line 6) returns the concatenation of the path from $s$ to $q$, with $\Pi$.
$\T$.init($s$) EXPAND($\T$) $\Pi \leftarrow$ CONNECT\_TO\_TARGET($\T,t$) RETRIEVE\_PATH($\T,\Pi$)
$q_{\text{rand}}\leftarrow$ RANDOM\_SAMPLE() $q_{\text{near}}\leftarrow$ NEAREST\_NEIGHBOR($\T,q_{\text{rand}}$) $q_{\text{new}}\leftarrow \orad(q_{\text{near}},q_{\text{rand}})$ $\T$.add\_vertex($q_{\text{new}}$) $\T$.add\_edge($q_{\text{near}} ,q_{\text{new}}$)
$\Pi \leftarrow$ LOCAL\_CONNECTOR($q,t$) $\emptyset$
Local connector {#sub:local}
---------------
We show in the following subsection that it is possible that $\T$ will eventually reach $t$ during the EXPAND stage, and therefore an application of LOCAL\_CONNECTOR will not be necessary. However, in practice this is unlikely to occur within a short time frame, especially when $G$ is large. Thus, we employ a heavy-duty technique, which given two vertices $q_0,q_1$ of $G$ tries to find a path between them. We mention that it is common to assume in sampling-based algorithms that connecting nearby samples will require less effort than solving the initial problem and here we make a similar assumption. We assume that a [local connector]{}is effective only on *restricted* pathfinding problems, thus in the general case it cannot be applied directly on $s,t$, as it may be highly costly (unless the problem is easy). A concrete example of a [local connector]{}is provided in the next section.
Probabilistic completeness of dRRT {#sec:complete}
----------------------------------
Recall that an algorithm is *probabilistically complete* if the probability it finds a solution tends to one as the run-time of the algorithm tends to infinity (when such a solution exists). For simplicity, we show that dRRT possesses a stronger property and with high probability will reveal all the vertices of the traversed graph, assuming this graph is connected.
The proof relies on the assumption that the vertices of the traversed graph $G$ are in *general position*, that is, every pair of distinct vertices are embedded in two distinct points in $\dR^d$, and for every triplet of distinct vertices the points in which they are embedded are non-collinear. This issue will be addressed in the following section, where we consider the application of dRRT on a specific type of graphs. The proof does not need to take into consideration the [local connector]{}.
\[thm:complete\] Let $G=(V,E)$ be a connected graph embedded in $[0,1]^d$ where the vertices are in general position. Then ,with high probability, every vertex of $G$ will be revealed by the dRRT algorithm, given sufficient amount of time.
Denote by $U$ the set of vertices of $\T$ after the completion of an iteration of the algorithm. Let $v^*\in V\setminus U$ be an unvisited vertex such that there exists $(v,v^*)\in E$, where $v\in U$. We wish to show that the probability that $\T$ will be expanded on the edge $(v,v^*)$, and thus $v^*$ will be added to $U$, is bounded away from zero. For simplicity we assume that there exists a single vertex $v\in U$ that has an edge to $v^*$.
Denote by $\vor(v)$ the *Voronoi cell* [@bkos-cg08] of the site $v$, in the Euclidean (standard) Voronoi diagram of point sites, where the sites are the vertices of $U$ (Figure \[fig:dRRT\_alg\](b)). In addition, denote by $\vor'(v,v^*)$ the Voronoi cell of $\rho(v,v^*)$, in a Voronoi diagram of the ray sites $\rho(v,v^*),\rho(v,u_1),\ldots,\rho(v,u_j)$, where $u_1,\ldots,u_j$ are the neighbors of $v$ in $\T$, not including $v^*$ (Figure \[fig:dRRT\_alg\](c)).
Notice that in order to extend $\T$ from $v$ to $v^*$ the random sample $q_{\text{rand}}$ in EXPAND (Algorithm \[alg:expand\]) has to fall inside $\vor(v)\cap\vor'(v,v^*)$. Thus, in order to guarantee that $v^*$ will be added to $\T$, with non-zero probability, we show that the shared region between these two cells has non-zero measure, namely $|\vor(v)\cap\vor'(v,v^*)|>0$, where $|\Gamma|$ denotes the volume of $\Gamma$.
By the general position assumption we can deduce that $|\vor(v)|>0$ and $|\vor'(v,v^*)|>0$. In addition, the intersection between the two cells is clearly non-empty: There is a ball with radius $r>0$ whose center is $v$ and is completely contained in $\vor(v)$; similarly, there is a cone of solid angle $\alpha >0$ with apex at $v$ fully contained in $\vor'(v,v^*)$. Hence, it holds that $|\vor(v)\cap\vor'(v,v^*)|>0$, otherwise $v$ and $v^*$ are embedded in the same point.
We note that a more careful analysis can yield an explicit bound on the convergence rate of dRRT. Such a bound may be computed using the size of the smallest cell in the Voronoi diagram of all nodes of $G$.
\[\] [ {width="30.00000%"} ]{} \[\] [ {width="30.00000%"} ]{}\
\[fig:drrt\_2\] \[\] [ {width="30.00000%"} ]{} \[fig:drrt\_3\] \[\] [ {width="30.00000%"} ]{}
\[fig:dRRT\_alg\]
Multi-robot motion planning with dRRT {#sec:implementation}
=====================================
In this section we describe the MRdRRT algorithm. Specifically, we discuss the adaptation of dRRT for pathfinding in a composite roadmap $\dG$, which is embedded in the joint [C-space]{}of $m$ robots. In particular, we show an implementation of the oracle $\orad$, which relies solely on the representation of $G_1,\ldots, G_m$. Additionally, we discuss an implementation of the [local connector]{}component that takes advantage of the fact that $\dG$ represents a set of valid positions and movements of multiple robots. Finally, we discuss the probabilistic completeness of our entire approach to multi-robot motion planning.
Oracle $\orad$
--------------
Recall that given $C\in \dV$ and a random sample $q$, $\O_D(C,q)$ returns $C'$ such that $C'$ is a neighbor of $C$ in $\dG$, and for every other neighbor $C''$ of $C$, $\rho(C,q)$ forms a smaller angle with $\rho(C,C')$ than with $\rho(C,C'')$, where $\rho$ is as defined in Section \[sec:complete\].
Denote by $\C(r_i)$ the [C-space]{}of $r_i$. Let $q=(q_1,\ldots,q_m)$ where $q_i\in\C(r_i)$, and let $C=(c_1,\ldots,c_m)$ where $c_i\in V_i$. To find a suitable neighbor for $C$ we first find the most suitable neighbor for every individual robot and combine the $m$ single-robot neighbors into a candidate neighbor for $C$. We denote by $c'_i=\orad(c_i,q_i)$ the neighbor of $c_i$ in $G_i$ that is in the direction of $q_i$. Notice that the implementation of the oracle for individual roadmaps is trivial—for example, by traversing all the neighbors of $c_i$ in $G_i$. Let $C'=(c'_1,\ldots,c'_m)$ be a candidate for the result of $\O_D(C,q)$. If $(C,C')$ represents a valid edge in $\dG$, i.e., no robot-robot collision occurs, we return $C'$. Otherwise, $\O_D(C,q)$ returns $\emptyset$. In this case, the new sample is ignored and another sample is drawn in the EXPAND phase (Algorithm \[alg:expand\]).
The completeness proof of the dRRT (Theorem \[thm:complete\]) for this specific implementation of $\orad$, is straightforward. Notice that in order to extend $C=(c_1,\ldots,c_m)$ to $C'=(c'_1,\ldots,c'_m)$ the sample $q=(q_1,\ldots,q_m)$ must obey the following restriction: For every robot $r_i$, $q_i$ must lie in $\vor(c_i)\cap\vor'(c_i,c'_i)$ (where in the original proof we required that $q$ will lie in $\vor(C)\cap\vor'(C,C')$). Also note that the points in $\C(r_i)$ are in general position, as required by Theorem \[thm:complete\], since they were uniformly sampled by PRM.
Local connector implementation
------------------------------
Recall that in the general dRRT algorithm the [local connector]{}is used for connecting two given vertices of a graph. As our [local connector]{}we rely on a framework described by van den Berg et al. [@bslm-cppmr]. Given two vertices $\dV=(v_1,\ldots,v_m), \dV'=(v'_1,\ldots,v'_m)$ of $\dG$ we find for each robot $i$ a path $\pi_i$ on $G_i$ from $v_i$ to $v'_i$. The connector attempts to find an ordering of the robots such that robot $i$ does not leave its start position on $\pi_i$ until robots with higher priority reached their target positions on their respective path, and of course that it also avoids collisions. When these robots reach their destination robot $i$ moves along $\pi_i$ from $\pi_i(0)$ to $\pi_i(1)$. During the movement of this robot the other robots stay put.
The priorities are assigned according to the following rule: if moving robot $i$ along $\pi_i$ causes a collision with robot $j$ that is placed in $v_j$ then robot $i$ should move *after* robot $j$. Similarly, if $i$ collides with robot $j$ that is placed in $v'_j$ then robot $i$ should move *before* robot $j$. This prioritization induces a directed graph $\I$. In case this graph is acyclic we generate a solution according to the prioritization of the robots. Otherwise, we report failure.
We decided to use this simple technique in our experiments due to its low cost, in terms of running time, regardless of whether it succeeds finding a solution or not. We wish to mention that we also experimented with M\* with a bounded degree of coupling (to avoid considering exponentially many neighbors) as the local connector in our algorithm. However, the ordering algorithm of [@bslm-cppmr] turned out to be considerably more efficient.
Probabilistic completeness of MRdRRT
------------------------------------
In order for the motion-planning framework to be probabilistically complete, we still need to show that (i) as the number of samples used for each single-robot roadmap tends to infinity, the composite roadmap will contain a path (if such a path exists) and (ii) that the proof of Theorem \[thm:complete\] still holds when the size of the graph tends to infinity. Indeed, Švestka and Overmars [@so-cppmr] show that the composite roadmap approach is probabilistically complete when the graph-search algorithm is complete. However, in our setting, the graph-search algorithm is only probabilistically complete and the proof may need to be refined as the size of each Voronoi cell tends to zero.
We note that as the composite roadmap is finite, it is easy to modify the dRRT algorithm such that it will be complete. This may be done by keeping a list of exposed nodes that still have unexposed edges. At the end of every iteration of the main loop of dRRT (Algorithm \[alg:planner\], line 2) one node is picked from the list and one of its unexposed edges is exposed (finding an unexposed edge is done in a brute force manner). Although the above modification ensures completeness of dRRT and hence probabilistic completeness of MRdRRT, we are currently looking for an alternative proof that does not require altering the dRRT algorithm.
Experimental results {#sec:experimental_results}
====================
We implemented MRdRRT for the case of polyhedral robots translating and rotating among polyhedral obstacles (see Figure \[fig:3d\_scenarios\]). We compared the performance of MRdRRT with RRT and an improved (recursive) version of M\* that appears in [@wc-mstar]. To make the comparison as equitable as possible, as dRRT does not take into consideration the quality of the solution, we use the *inflated* version of M\* [@wc-mstar] with relaxed optimality guarantees.
**Implementation details.** The algorithms were implemented in [C.08ex]{}. The experiments were conducted on a laptop with an Intel i5-3230M 2.60GHz processor with 16GB of memory, running 64-bit Windows 7. We implemented a generic framework for multi-robot motion planning based on composite roadmaps. The implementation relies on PQP [@pqp] for collision detection, and performs nearest-neighbor queries using the Fast Library for Approximate Nearest Neighbors (FLANN) [@muja_flann_2009]. Metrics, sampling and interpolation in the 3D environments followed the guidelines presented by Kuffner [@Kuffner04]. To eliminate the dependence of dRRT on parameters we assigned them according to the number of iterations the algorithm performed so far, i.e., the number of times that the main loop has been repeated. Specifically, in the $i$’th iteration each EXPAND (Algorithm \[alg:expand\]) call performs $2^i$ iterations ($N=2^i$), while CONNECT\_TO\_TARGET uses $K=i$ candidates that are connected with $t$.
**Test scenarios.** We report in Table \[tbl:coupled\_scenarios\] the running times of M\* and dRRT for the scenarios. The first three scenarios are especially challenging as they consist of a large number of robots, and require a substantial amount of coordination between them. The fourth scenario (“Home”) is more relaxed and consists of only five robots and requires little coordination.
We ran each of the three algorithms 10 times on each scenario. RRT proved incapable of solving any of the test scenarios, running for several tens of minutes until terminating due to exceeding the memory limits. We believe that RRT as-is is not suitable for high-dimensional, coupled, multi-robot motion planning. M\* exhibited slightly better performance. For the first three scenraios, which involve multiple robots and require a substantial amount of coordination, it never exceeded a success rate of 40%. In particular, it often ran out of memory or ran for a very long duration (we terminated it if its running time exceeded ten times the running time of MRdRRT). On the other hand, MRdRRT was stable in its results and managed to solve all the scenarios for each of the 10 attempts. When M\* did manage to solve one of the first three scenarios, it explored between 2.5 to 10 times the number of vertices that dRRT explored. For the fourth scenario the results of MRdRRT and M\* were comparable and in general we found M\* more suitable for situations where only a small number of robots have to interact at any given time. We mention that MRdRRT was unable to solve scenarios that consist of a substantially larger number of robot than we used in our experiments. We believe that it would be beneficial to consider a stronger *local connector* in such cases.
[c||c|c||c|c|c||c|c|c|c|c|]{} & & &\
& & & visited & total & success & visited & connect & expand & total & success\
& & & vertices & time & rate & vertices & time & time & time & rate\
& 8k & 10s & DNF & DNF & 0% & 8k & 3.3s & 6.7s & 11s & 100%\
& 10k & 24.8s & 300k & 267s & 30% & 34k & 30.4s & 25.5s & 55.9s & 100%\
& 10k & 16.2s & 27k & 31s & 40% & 12k & 16.3s & 36.8s & 53.1s & 100%\
& 5k & 10.1s & 2k & 3.9s & 100% & 8k & 1.5s & 2.9s & 4.4s & 100%\
\[tbl:coupled\_scenarios\]
Discussion {#sec:discussion}
==========
In this section we state the benefits of MRdRRT, which consists of an implicitly represented roadmaps for multi-robot motion planning combined with an efficient approach for pathfinding for such roadmaps.
Recall that the implicitly-represented composite roadmap $\dG$ results from a tensor product of $m$ PRM roadmaps $G_1,\ldots,G_m$. The reliance on the precomputed individual roadmaps eliminates the need to perform additional collision checking between robots and obstacles while querying $\dG$. This has a substantial impact on the performance of MRdRRT as it is often the case that checking whether $m$ robots collide with obstacles is much more costly than checking whether the $m$ robots collide between themselves. This is in contrast with more naive approaches, such as RRT which consider the group of robots as one large robot. In such cases, checking whether a configuration (or an edge) is collision free requires checking for the two types of collisions simultaneously.
The M\* algorithm, which also uses the underlying structure of $\dG$, performs very well in situations where only a small subset of the robots need to coordinate. In these situations it can cope, almost effortlessly, with several tens of robots while outperforming our framework. However, in scenarios where a substantial amount of coordination is required between the robots M\* suffers from a disadvantage, since it is forced to consider exponentially many neighbors when performing the search on $\dG$. In contrast, dRRT performs a “minimalistic” search and advances in small steps, little by little, regardless of the difficulty of the problem at hand. Moreover, dRRT strives to reach unknown regions in $\dG$ while avoiding spending too much time in the exploration of regions that are in the vicinity of explored vertices. This is done via the Voronoi bias, as shown in the proof of Theorem \[thm:complete\]. This is extremely beneficial when working on $\dG$ since it contains vertices which represent essentially the same conformation of the robots, and thus considering many vertices within a small region would not lead to a better understanding of the problem at hand. To justify this claim, consider the following example. Suppose that for every robot $i$, $v_i$ is a vertex of $V_i$ that has $k$ neighbors in $G_i$ at distance at most $\varepsilon$. Then the vertex $(v_1,...,v_m)\in \dV$ might have as much as $k^m$ neighbors that are at distance at most $\varepsilon\sqrt{m}$ in $\dG$.
Future work {#sec:future}
===========
**Towards optimality.** Currently, our algorithmic framework is concerned with finding *some* solution. Our immediate future goal is to modify it to provide a solution with quality guarantees, possibly by taking an approach similar to the continuous RRT\* algorithm [@KF11], which is known to be asymptotically optimal. A fundamental difference between RRT\* and the original formulation of RRT is in a rewiring step, where the structure of the tree is revised to improve previously examined paths. Specifically, when a new node is added to the tree, it is checked as to whether it will be more beneficial for some of the existing nodes to point to the new vertex instead of their current parent in the tree. This can be adapted, to some extent, to the discrete case, although it is not clear whether this indeed will lead to optimal paths.
**dRRT in other settings of motion planning.** In this paper we combined the dRRT algorithm with implicit composite roadmaps to provide an efficient algorithm for multi-robot motion planning. One of the benefits of our framework comes from the fact that it reuses some of the already computed information to avoid performing costly operations. In particular, it refrains from checking collisions between robots with obstacles by forcing the individual robots to move on precalculated individual roadmaps (i.e., $G_i$). We believe that a similar approach can be used in other settings of motion planning. In particular, we are currently working on a dRRT-based approach for motion planning of a multi-linked robot. The new approach generates an implicitly-represented roadmap, which encapsulates information on configurations and paths between configuration that do not induce self-intersections of the robot, while ignoring the existence of obstacles. Then, we overlay this roadmap on the workspace, an operation which invalidates some of the nodes and edges of the roadmap. Thus, we know only which configurations are self-collision free, but not obstacles collision-free. Then we use dRRT for pathfinding on the new roadmap, while avoiding self-collision tests and while exploring a small portion of the infinite roadmap.
Acknowledgements {#sec:ack}
================
We wish to thank Glenn Wagner for advising on the M\* algorithm and Ariel Felner for advice regarding pathfinding algorithms on graphs. We note that the title “Finding a Needle in an Exponential Haystack” has been previously used in a talk by Joel Spencer in a different context.
[10]{} \[1\][`#1`]{}
- A Proximity Query Package. <http://gamma.cs.unc.edu/SSV/>
Graph [P]{}roduct — [W]{}ikipedia[,]{} [T]{}he [F]{}ree [E]{}ncyclopedia (2013), <http://en.wikipedia.org/wiki/Graph_product>
Adler, A., de Berg, M., Halperin, D., Solovey, K.: Efficient multi-robot motion planning for unlabeled discs in simple polygons. CoRR abs/1312.1038 (2013)
Aronov, B., de Berg, M., van der Stappen, A.F., Švestka, P., Vleugels, J.: [Motion Planning for Multiple Robots](http://link.springer.com/article/10.1007%2FPL00009476?LI=true). Discrete [&]{} Computational Geometry 22(4), 505–525 (1999)
Auletta, V., Monti, A., Parente, M., Persiano, P.: A linear time algorithm for the feasibility of pebble motion on trees. In: SWAT. pp. 259–270 (1996)
van den Berg, J., Overmars, M.: Prioritized motion planning for multiple robots. In: IROS. pp. 430 – 435 (2005)
van den Berg, J., Snoeyink, J., Lin, M., Manocha, D.: [Centralized Path Planning for Multiple Robots: Optimal Decoupling into Sequential Plans](http://www.roboticsproceedings.org/rss05/p18.html). In: RSS (2009)
de Berg, M., van Kreveld, M., Overmars, M., Schwarzkopf, O.: Computational Geometry: Algorithms and Applications. Springer-Verlag, third edn. (2008)
Branicky, M.S., Curtiss, M.M., Levine, J.A., Morgan, S.B.: [[RRT]{}s for Nonlinear, Discrete, and Hybrid Planning and Control](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1272639&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1272639). In: Decision and Control. pp. 9–12 (2003)
Choset, H., Lynch, K., Hutchinson, S., Kantor, G., Burgard, G., Kavraki, L., Thrun, S.: Principles of Robot Motion: Theory, Algorithms, and Implementations. MIT Press (2005)
Şucan, I.A., Moll, M., Kavraki, L.E.: [The [O]{}pen [M]{}otion [P]{}lanning [L]{}ibrary](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6377468&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6377468). [IEEE]{} Robotics & Automation Magazine 19(4), 72–82 (2012)
Goraly, G., Hassin, R.: Multi-color pebble motion on graphs. Algorithmica 58(3), 610–636 (2010)
Hirsch, S., Halperin, D.: [Hybrid Motion Planning: Coordinating Two Discs Moving among Polygonal Obstacles in the Plane](http://link.springer.com/chapter/10.1007%2F978-3-540-45058-0_15?LI=true). In: WAFR. pp. 239–255. Springer (2002)
Hopcroft, J., Schwartz, J., Sharir, M.: [On the Complexity of Motion Planning for Multiple Independent Objects; [PSPACE]{}-Hardness of the “[W]{}arehouseman’s Problem”](http://ijr.sagepub.com/content/3/4/76.short). IJRR 3(4), 76–88 (1984)
Karaman, S., Frazzoli, E.: [Sampling-Based Algorithms for Optimal Motion Planning](http://ijr.sagepub.com/content/30/7/846.short). IJRR 30(7), 846–894 (2011)
Kavraki, L.E., Švestka, P., Latombe, J.C., Overmars, M.: . IEEE Transactions on Robotics and Automation 12(4), 566–580 (1996)
Kloder, S., Hutchinson, S.: [Path Planning for Permutation-Invariant Multi-Robot Formations](http://dx.doi.org/10.1109/ROBOT.2005.1570374). In: ICRA. pp. 1797–1802 (2005)
Kornhauser, D.: Coordinating Pebble Motion on Graphs, The Diameter of Permutation Groups, and Applications. M.[Sc]{}. thesis, Department of Electrical Engineering and Computer Scienec, Massachusetts Institute of Technology (1984)
Kuffner, J.J., La[V]{}alle, S.M.: [[RRT]{}-[C]{}onnect: An Efficient Approach to Single-Query Path Planning](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=844730&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D844730). In: ICRA. pp. 995–1001 (2000)
Kuffner, J.J.: Effective sampling and distance metrics for [3D]{} rigid body path planning. In: ICRA. pp. 3993–3998 (2004)
La[V]{}alle, S.M.: Planning Algorithms. Cambridge University Press (2006)
Leroy, S., Laumond, J.P., Simeon, T.: [Multiple Path Coordination for Mobile Robots: A Geometric Algorithm](http://ijcai.org/Past%20Proceedings/IJCAI-99%20VOL-2/PDF/064.pdf). In: IJCAI. pp. 1118–1123 (1999)
Luna, R., Bekris, K.E.: [Push and Swap: Fast Cooperative Path-Finding with Completeness Guarantees](http://dl.acm.org/citation.cfm?id=2283446). In: IJCAI. pp. 294–300 (2011)
Muja, M., Lowe, D.G.: [Fast Approximate Nearest Neighbors with Automatic Algorithm Configuration]{}. In: VISSAPP. pp. 331–340. INSTICC Press (2009)
Pearl, J.: Heuristics: Intelligent Search Strategies for Computer Problem Solving. Addison-Wesley (1984)
Salzman, O., Hemmer, M., Halperin, D.: [On the Power of Manifold Samples in Exploring Configuration Spaces and the Dimensionality of Narrow Passages](http://link.springer.com/chapter/10.1007%2F978-3-642-36279-8_19). WAFR pp. 313–329 (2012)
Sanchez, G., Latombe, J.C.: [Using a [PRM]{} Planner to Compare Centralized and Decoupled Planning for Multi-Robot Systems](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=1014852&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D1014852). In: ICRA. pp. 2112–2119 (2002)
Schwartz, J.T., Sharir, M.: [On the Piano Movers’ Problem: [III]{}. [C]{}oordinating the Motion of Several Independent Bodies](http://ijr.sagepub.com/content/2/3/46.short). IJRR 2(3), 46–75 (1983)
Sharir, M., Sifrony, S.: Coordinated motion planning for two independent robots. Annals of Mathematics and Artificial Intelligence 3(1), 107–130 (1991)
Solovey, K., Halperin, D.: [$k$-[C]{}olor [M]{}ulti-[R]{}obot [M]{}otion [P]{}lanning](http://link.springer.com/chapter/10.1007%2F978-3-642-36279-8_12). WAFR pp. 191–207 (2012)
Spirakis, P.G., Yap, C.K.: [[S]{}trong [NP]{}-[H]{}ardness of [M]{}oving [M]{}any [D]{}iscs](http://www.sciencedirect.com/science/article/pii/0020019084901303). Inf. Process. Lett. 19(1), 55–59 (1984)
Turpin, M., Michael, N., Kumar, V.: [Computationally Efficient Trajectory Planning and Task Assignment for Large Teams of Unlabeled Robots]{}. In: ICRA. pp. 834–840 (2013)
Švestka, P., Overmars, M.: [Coordinated Path Planning for Multiple Robots](http://www.sciencedirect.com/science/article/pii/S092188909700033X). Robotics and Autonomous Systems 23, 125–152 (1998)
Wagner, G., Choset, H.: [M\*: A Complete Multirobot Path Planning Algorithm with Performance Bounds](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6095022&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6095022). In: IROS. pp. 3260–3267. IEEE (2011)
Wagner, G., Kang, M., Choset, H.: [Probabilistic Path Planning for Multiple Robots with Subdimensional Expansion](http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=6225297&url=http%3A%2F%2Fieeexplore.ieee.org%2Fxpls%2Fabs_all.jsp%3Farnumber%3D6225297). In: ICRA. pp. 2886–2892 (2012)
Yap, C.: Coordinating the motion of several discs. Tech. rep., Courant Institute of Mathematical Sciences, Michigan State University, New York (1984)
[^1]: K. Solovey and O. Salzman contributed equally to this paper.
[^2]: This work has been supported in part by the 7th Framework Programme for Research of the European Commission, under FET-Open grant number 255827 (CGL—Computational Geometry Learning), by the Israel Science Foundation (grant no. 1102/11), by the German-Israeli Foundation (grant no. 1150-82.6/2011), and by the Hermann Minkowski–Minerva Center for Geometry at Tel Aviv University.
[^3]: We mention that we are not the first to consider RRTs in discrete domains. Branicky et al. [@BCLM03] applied the RRT algorithm to a discrete graph. However, a key difference between the approaches is that we assume that the graph is *geometrically embedded*, hence we use *random points* as samples while they use nodes of the graph as samples. Additionally, their technique requires that all the neighbors of a visited vertex will be considered—a costly operation in our setting, as mentioned above.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study the problem of the damping of collective modes close to a Pomeranchuk quantum critical point in a Fermi liquid. In analogy with problems in dissipative open quantum systems, we derive the Landau damping of a Fermi liquid by integrating out a macroscopic number of degrees of freedom from a generating functional. Being a reformulation of the linearized Boltzmann equation this approach reproduces well-known results from the theory of Fermi liquids. We also study the Bethe-Salpeter equations within the Landau theory and discuss the implications of these results on quantum phase transitions of the Pomeranchuk type and its dynamical exponent, $z$. We apply our results to the electronic nematic instability and find $z=3$ in the collisionless limit.'
author:
- Johan Nilsson
- 'A. H. Castro Neto'
bibliography:
- 'damping.bib'
title: Heat Bath Approach to Landau Damping and Pomeranchuk Quantum Critical Points
---
Introduction
============
Recently there has been interest in the so-called quantum critical Pomeranchuk instability where the interactions in a fermion system are strong enough to make the Fermi surface soft and ultimately unstable toward a spontaneous deformation [@pomeranchuk]. These instabilities have been proposed to be behind exotic electronic phases such as electronic liquid crystals in cuprate superconductors [@eduardonematic; @barci] and fractional quantum Hall effect [@qhe_1; @qhe_2], metamagnetic transitions in electronic systems such as Sr$_3$Ru$_2$O$_7$ [@honerkamp], electronic systems close to van Hove singularities [@wegner], “hidden order” in URu$_2$Si$_2$ [@varma; @varma_s], a mechanism for generation of spin-orbit coupling [@congjun], and even phase transitions in interacting quantum dots [@murthy]. The Pomeranchuk instability can take place both on a lattice and in the continuum [@eduardonematic; @metznersoft; @kunyang] and is driven by attractive quasiparticle-quasiparticle interactions. These interactions in Fermi liquid theory are parameterized by the Landau parameters, $F^{s,a}_m$, where $s$ ($a$) stands for the symmetric (anti-symmetric) spin channel and $m = 0,1,...$ for the angular momentum (we focus on the case of the two-dimensional isotropic liquid although the three-dimensional case is completely analogous). Pomeranchuk showed that if some $F^{s,a}_m$ becomes too attractive ($F^{s,a}_m < -1$ in two dimensions) at $T=0$ then the Fermi surface becomes unstable and cannot sustain collective oscillations. The simplest example of such a transition is a ferromagnetic (or Stoner) instability that occurs when $F^a_0<-1$, which is characterized by a diverging magnetic susceptibility, $\chi$ ($\chi \propto 1/(1+F^a_0)$) A similar instability with a diverging compressibility, ${\cal K}$, can be achieved by pulling on a liquid to put it into a negative pressure region when $F^s_0 < -1$ (${\cal K} \propto 1/(1+F^s_0)$) [@balibarreview]. There is also the possibility that the instability takes place in a higher angular momentum channel. For example, a deformation into an elliptical Fermi surface can be achieved by having a strong enough attraction in the $d$-channel, that is, $F^s_2 < -1$. The broken symmetry phase in this case has been proposed to be an *electronic nematic* because of the nomenclature in the classical theory of liquid crystals where an analogous phase exists. The properties of the broken symmetry phase has been found to be very peculiar with marked non-Fermi liquid properties [@eduardonematic]. Although Fermi liquid theory cannot predict what is the nature of the broken symmetry phase, it can be used to study the approach to the instability from the Fermi liquid (quantum disordered) side of the transition (see Fig.\[phase\_diagram\]).
![Phase diagram of a quantum critical Pomeranchuk instability of order $m$. Vertical axis: temperature $T$; Horizontal axis: Landau parameter $F_m$. The dashed arrow line shows the direction of approach to the quantum critical point used in this work. []{data-label="phase_diagram"}](phase_diagram.eps)
One of the main characteristics of a Pomeranchuk quantum critical point is that in its proximity the excitations are not of the simple quasiparticle type but rather become heavily damped because of the soft nature of the Fermi surface. This is so because the Fermi surface allows for a diverging number of low energy particle-hole pairs to accompany an excited quasiparticle. Presumably, a similar situation holds for most of the single-particle excitations in the broken symmetry phase [@eduardonematic] although the Fermi liquid theory is not able to access that phase directly.
A natural way to study the collective modes of an interacting electron fluid is via Fermi surface bosonization. The method of bosonization has been very successful in providing a framework for describing the breakdown of the normal Fermi liquid in one-dimensional (1D) systems, where instead the Luttinger liquid emerges as the “standard” model [@voitlong]. In Luttinger liquids there are no well defined quasiparticles and the electron breaks down into charge and spin collective modes. It is now well understood that the anomalous behavior in one dimension is mainly due to the restricted phase space. Indeed, there exists Pomeranchuk instabilities also in 1D which lead to a breakdown of the Luttinger liquid. What happens is that either the charge- or spin- velocity vanishes at the transition. Since the charge(spin) susceptibility goes as the inverse of the charge(spin) velocity [@voitlong] the signature of the instability is a diverging susceptibility, just as in the higher dimensional case. Thus, the Pomeranchuk instability is a general concept, independent of the dimensionality of the system under interest. The higher dimensional bosonization approach was developed many years ago [@haldane; @marston_1; @marstonreview; @antonioboson1; @antonioboson2; @kopietz] and reproduces the known results from the theory of Fermi liquids, especially the physics of collective modes and particle-hole excitations which are described in terms of boson coherent states. The basic reason for this is that both theories describes fermion systems dominated by forward scattering correctly, and this is what the Landau Fermi liquid is all about [@shankarrg].
For problems such as those involving classical limits it is a good idea to use a description in terms of bosons since they can have classical analogs whereas fermions cannot. In particular, when one is studying phase transitions in the language of a Landau-Ginzburg expansion the order parameter is a classical construction which in the quantum language should be described by bosons. One can generally trade a fermion system with density-density interactions to one involving bosons by using a Hubbard-Stratanovich transformation and then integrate out the fermions from the path integral that describes the system. This is the standard device to study quantum critical theories in Fermi systems [@Hertz; @Continentino; @Millis]. In such approaches to quantum critical phenomena dissipative terms are generated in the order parameter effective action, the so-called Landau damping. However, there have been some arguments about whether it is viable to integrate out all of the fermions from the system in this way, or if some properties of the system is lost due to the gapless nature of the electrons in a Fermi liquid. In particular there are particle-hole-like excitations with arbitrarily low energies which must be treated with care since they may control the critical behavior of a Fermi system. Nevertheless, the integration of high energy degrees of freedom as one usually does in renormalization group (RG) approaches should give sensible results. Obviously, the particle-hole continuum type excitations do not disappear in a faithful bosonic reformulation of the system, although their nature might be disguised. When some of these modes are in resonance with the collective excitations the result is Landau damping. This occurs only for modes with a velocity smaller than the Fermi velocity.
In this paper we show that the Landau damping problem can be recast in terms of a trace over bosonic modes. In particular the modes associated with the high angular momentum oscillations of the Fermi surface are traced out since the collective excitations we are interested are concentrated in the low angular momentum modes. Furthermore, using the Bethe-Salpeter equation in the context of the collision integral, we can study the lifetime of quasiparticles and collective modes in a Fermi liquid close to a Pomeranchuk quantum critical point. Moreover, we discuss the dynamical critical exponent of the same quantum critical point. The paper is organized as follows. In Section \[landaudamp\] we show how Landau damping as described above can be reproduced within a simple purely bosonic reformulation of the Fermi liquid theory. Here the damping comes about from the coupling to a bath of harmonic oscillators similarly to what happens in dissipative quantum mechanics. In Section \[collisionintegral\] we study the collision integral in the language of Fermi liquid theory, focusing on the behavior near the critical point. This turns out to be important since collisions sets the lifetime of the quasiparticles, and as one approaches the critical point the lifetime decays rapidly. We conclude in Section \[conclusions\] with a discussion of the implications of these phenomena on the theory close to the critical point. In an Appendix we present some mathematical details.
Landau damping from a Boson bath {#landaudamp}
================================
The phenomenon of Landau damping is conventionally thought of as the damping of a collective mode due to ‘a *coherent* interaction of the collective mode with those particles which “surf-ride” on the crests of the running wave.’[@pinesnoz] In this section we show explicitly how Landau damping can also be viewed as the dissipation of a mode due to its coupling to a set of independent oscillators, i.e. a “boson bath”. Here the origin of these bosons are particle-hole excitations of the fermion system itself. This approach is similar in spirit to theories of dissipation in quantum mechanics.[@feynman; @caldeiralegget]
Let us take as our starting point the Landau kinetic equation that describes the evolution of the quasiparticle distribution function in the spin symmetric channel, $n_{\vec{p}}(\vec{r},t)$ [@pinesnoz]: $$\begin{gathered}
\label{landaukin_n}
{\frac{\partialn_{\vec{p}}(\vec{r},t)}{\partialt}} +
\nabla_p \epsilon_{\vec{p}}(\vec{r},t) \cdot \nabla_r
n_{\vec{p}}(\vec{r},t)
\\
-
\nabla_r \epsilon_{\vec{p}}(\vec{r},t) \cdot \nabla_p n_{\vec{p}}(\vec{r},t)
= I[n_{\vec{p}'}] \, ,\end{gathered}$$ where $\epsilon_{\vec{p}}(\vec{r},t)$ is the quasiparticle energy and $I[n_{\vec{p}'}]$ is the collision integral ($\vec{p}$ is the quasiparticle momentum, $\vec{r}$ is the position in real space, and $t$ is time) [@baymp]. Since the physics goes on close to the Fermi surface it is convenient to change variables to $\nu_{\vec{p}}(\vec{r},t)$ which we define by writing the deviation from the ground state values of quasiparticle distribution as $\delta n_{\vec{p}}(\vec{r},t) =
-\bigl( {\frac{\partialn_{p}^0}{\partial\epsilon_p}} \bigr) \nu_{p}(\vec{r},t)$. Physically $\nu_{p}(\vec{r},t)$ measure the local deformation of the Fermi surface. Upon a Fourier transform in space the linearized kinetic equation for $\nu_{p}(\vec{r},t)$ becomes $$\label{landaukin_nu}
{\frac{\partial\nu_{\vec{p}}(\vec{q},t)}{\partialt}} +
i \vec{v}_{\vec{p}} \cdot \vec{q} \;
\bigl[ \nu_{\vec{p}}(\vec{q},t) + \delta \epsilon_{\vec{p}}(\vec{q},t) \bigr]
= I'[\nu_{\vec{p}'}],$$ where $\vec{v}_{\vec{p}}$ is the Fermi velocity at the momentum $\vec{p}$. Again the notation $$\begin{aligned}
\delta \epsilon_{\vec{p}} =
\sum_{\vec{p}'}f_{\vec{p}\vec{p}'}\delta n_{\vec{p}'} \, ,\end{aligned}$$ is standard and describes the interaction (parametrized by $f_{\vec{p}\vec{p}'}$) among excited quasiparticles. For an isotropic liquid in two dimensions, $\vec{v}_{\vec{p}}$ is parallel to $\vec{p}$, and since the magnitude of $p$ is nearly equal to the Fermi momentum $p_F$ we can label $\vec{p}$ by an angle $\theta_{p}$ (measured from the direction of $\vec{q}$). Then, the interaction $f_{\vec{p}\vec{p}'}$ only depends on the angle between the momenta and is convenient to expand in Fourier components because of the circular symmetry: $$\begin{aligned}
\label{nudef}
& \nu (\theta) = \sum_{m} \nu_m e^{i m \theta } \, ,
\\
& f(\theta,\theta') = \sum_m f_m e^{i m (\theta -\theta') },\end{aligned}$$ where $f_m = f_{-m}$. The sum over $m$ should be cut-off at some integer of order $E_F / \Lambda \equiv N_{\Lambda}\gg 1$, where $E_F$ is the Fermi energy and $\Lambda$ is some energy scale below which the Landau theory is valid. Projecting onto the Fourier components and setting the collision integral to zero for the time being we arrive at $$\begin{gathered}
\label{eomlin}
{\frac{\partial\nu_m(\vec{q},t)}{\partialt}} +\frac{i v_F q}{2} \Bigl[
(1+F_{m-1}) \nu_{m-1}(\vec{q},t) \\ +
(1+F_{m+1}) \nu_{m+1}(\vec{q},t)
\Bigr] = 0.\end{gathered}$$ Here $v_F$ is the Fermi velocity and the $F_n = N(0)f_n$ are the interaction parameters normalized with the density of states at the Fermi energy, $N(0)$. Note that these equations are linear in the time derivatives and couple time derivatives of odd components to the even components and vice versa. Thus the equations respect time reversal invariance implying non-dissipative dynamics.
To make connection with the more familiar simple oscillator we eliminate the odd components in favor of a second time derivative [@note]. With the notations: $$\begin{aligned}
a_m (\vec{q},t) &=& \sqrt{1 +F_m} \, \, \nu_m (\vec{q},t) \, ,
\\
A_m &=& \frac{1}{4} (1+F_m) (2 + F_{m-1} + F_{m+1}) \, ,
\\
B_m &=& \frac{1}{4} \sqrt{1+F_{m-1}} (1+ F_{m}) \sqrt{1 + F_{m+1}} \, ,\end{aligned}$$ the equations of motion become: $$\begin{gathered}
\label{eomsec}
{\frac{\partial^{2}{a_m(\vec{q},t)}}{\partial{t}^{2}}} +(v_F q)^2 \Bigl[
A_m a_{m}(\vec{q},t) \\ +
B_{m-1} a_{m-2}(\vec{q},t) +
B_{m+1} a_{m+2}(\vec{q},t)
\Bigr] =0.\end{gathered}$$ These equations of motion are also generated from Euler-Lagrange equations and the classical Lagrangian density: $$\begin{gathered}
\label{lagrangian1}
\mathcal{L} = \sum_m {\frac{\partiala_{2 m}^* (\vec{q},t)}{\partialt}}
{\frac{\partiala_{2 m} (\vec{q},t)}{\partialt}} \\
- (v_F q)^2 \sum_{m,n} a_{2 m}^* (\vec{q},t) V_{m n} a_{2 n} (\vec{q},t),\end{gathered}$$ where $$\begin{aligned}
V_{m n} = \delta_{m,n} A_{2 m} + \delta_{m-1,n}B_{2 m -1} +
\delta_{m+1,n} B_{2 m +1} \, .
\label{vmn}\end{aligned}$$ Notice that (\[lagrangian1\]) describes a set of coupled harmonic oscillators coupled by “springs” with spring constants given by $V_{m,n}$. This result shows the clear collective nature of the Fermi surface modes.
We can also encode the evolution of the system with the methods of path integrals[@negeleo] in the real time partition function: $$\label{Zdef}
Z = \int Da^* Da \, \, e^{i S},$$ where we also perform a Fourier transform in time $$\label{action1}
S = \sum_{\vec{q},m,n} \int d\omega \, a_{2m}^{*}(\vec{q},\omega) \,
\tilde{V}_{mn} \, a_{2n}(\vec{q},\omega),$$ with, $$\begin{aligned}
\tilde{V}_{mn} &=& (v_F q)^2 \left[s^2 \delta_{mn} -V_{mn}\right] \, ,
\label{tvmn}
\\
s&=& \omega/(v_F q) \, .\end{aligned}$$ It is clear from these results that the propagator of the collective modes is essentially given by the inverse of the coupling matrix $\tilde{V}_{m,n}$, that is, $\tilde{V}_{m,n}^{-1}$. Therefore, the eigenmodes of $\tilde{V}$ are the collective excitations of the system and its eigenenergies determine the frequency of oscillations of these modes.
We can gain further insight into the problem by splitting the action (\[action1\]) into two decoupled sectors ($S =S^+ +S^-$) by using the symmetric and antisymmetric combinations $$\begin{aligned}
a_{m}^{\pm} = \frac{1}{\sqrt{2}} (a_{m} \pm a_{-m} ) \, , \end{aligned}$$ defined for $m \geq 0$. These corresponds to the longitudinal ($+$) and transverse ($-$) modes in the sense of having even/odd parity with respect to the line defined by $\vec{q}$. Notice that the transverse modes have no density (i.e. $m=0$) component.
Formally, there can be an infinite number of Landau parameters $F_m$ that parameterize the interactions in a Fermi liquid. In practice, however, only a few Landau parameters are taken into account [@baymp]. In the case of the Pomeranchuk instability we can concentrate on a single Landau parameter, that is, the most singular, since all the other parameters only provide trivial renormalizations of the Fermi liquid properties. Without lack of generality let us assume that there are $M$ Landau parameters in the description of a quantum liquid, that is, the $F_m$ are zero for $m \geq M+1$. In this case the matrix $\tilde{V}_{m,n}$ in (\[tvmn\]) for $m,n \geq M+1$ (we call it the high-$M$ block) can be written in the longitudinal and transverse sectors in terms of blocks of the form: $$\label{blockmatrix}
\begin{pmatrix}
s^2 - \frac{1}{2} & -\frac{1}{4} & 0 & 0 & \cdots \\
-\frac{1}{4} & s^2 -\frac{1}{2} & -\frac{1}{4} & 0 & \\
0 & -\frac{1}{4} & s^2 -\frac{1}{2} & -\frac{1}{4} & \\
0 & 0 & -\frac{1}{4} & s^2 -\frac{1}{2} & \\
\vdots & & & & \ddots
\end{pmatrix}.$$ The matrix (\[blockmatrix\]) (assumed to have size $N \times N$) have orthonormal eigenvectors of the form: $$\label{eigenvectors1}
\vec{\varphi}(k) = \sqrt{\frac{2}{N+1}} \begin{pmatrix}
\sin(k), \sin(2k) , \sin(3k) ,\ldots , \sin(N k)
\end{pmatrix}^T,$$ with respective eigenvalues $\lambda(s,k) = s^2 - (1+\cos(k))/2$ (the allowed values for the $k$’s are: $k = \pi n / (N+1)$ where $n=1 \ldots N$). The block (\[blockmatrix\]) is coupled to the mode $M$ via two off-diagonal terms in $V_{m,n}$. Schematically we have: $$\begin{pmatrix}
\mbox{low-M block} &
\begin{matrix}
\vdots & \vdots & \\
0 & 0 & \cdots \\
B_{2M+1} & 0 & \cdots
\end{matrix}
\\
\begin{matrix}
\cdots & 0 & B_{2M+1} \\
\cdots & 0 & 0 \\
& \vdots & \vdots
\end{matrix}
& \mbox{high-M block}
\end{pmatrix}.$$ Using the diagonalized form of the block matrix it is not hard to integrate out the high-$M$ modes. The end result is to change $$\label{eq:dA}
A_M \rightarrow A_M + \delta A_M,$$ in the low-$M$ block of $\tilde{V}_{mn}$, where $$\begin{aligned}
\delta A_M = 2 B_{2 M +1}^2 G(s)
\end{aligned}$$ and $$\label{Gdef}
G(s) = \frac{1}{N+1} \sum_k \frac{\sin^2(k)}{s^2 - (1+\cos(k))/2}.$$ From this result we see that $G(s)$ as a function of complex $s$ is analytic in the upper and lower half plane. It has a set of poles on the real axis between $s=-1$ and $s=1$. In the limit of $N \rightarrow \infty$ the poles become dense resulting in a branch cut. By taking $N\rightarrow \infty$, $G(s)$ can be expressed as an integral: $$\label{Geval}
G(s) = \int_{-\pi}^{\pi}\frac{dk}{\pi} \frac{\sin^2(k)}{2 s^2 - 1 - \cos(k)}.$$ Taking $s$ to lie away from the branch cut one can turn the integral into a contour integral around the unit circle centered at the origin. The resulting integral is easily evaluated for $|s|>1$ by the residue theorem and can be analytically continued to $|s|<1$ with the result: $$\label{Gval}
G(s)= -2 + 4 s^2 - 4 \sqrt{s^4 -s^2}.$$ So that for real $s$ and $|s|>1$ $$\label{Glarge}
G(s) = -2 + 4s^2 -4 |s|\sqrt{s^2-1},$$ while for real $s$ and $|s|<1$ $$\label{Gsmall}
G(s \pm i0^+) = -2+ 4s^2 \mp 4 i s\sqrt{1 - s^2}.$$ Hence, the final result of integrating out the high angular momentum modes is the introduction of a shift in the Landau parameters with the introduction of an imaginary part when $|s|=|\omega|/(q v_F)<1$, which is nothing but the Landau damping. By only keeping the low-$M$ modes the time reversal invariance has been broken.
Since what we have done is a reformulation of the Landau theory it should give the same result as the conventional approach. A simple check amounts to taking only $F_0 \neq 0$ and integrate out all modes except the zeroth one in the longitudinal sector, that is, $M=0$. The well-known zero sound mode should then come out as the zero of the only remaining matrix element (which is equivalent to having a pole in the propagator). Explicitly we obtain: $$\label{ZScheck}
\frac{s^2}{2}-\frac{A_0+2 \delta A_0}{2} = 0.$$ Substituting the expression for $A_0$, $B_1$ and $G(s)$ we get an undamped solution for $F_0 > 0$ with $$\label{ZScheck2}
|s| = \frac{1+F_0}{\sqrt{(1+F_0)^2 - F_{0}^2}},$$ which is the dispersion of the zero mode in two dimensions, see e.g. Ref.\[\].
Lifetime effects close to a Pomeranchuck quantum critical point {#collisionintegral}
===============================================================
In this section we investigate the effect of the proximity of a Pomeranchuk quantum critical point on the single particle and collective modes of a Fermi liquid. Using the collision integral formalism we study the quasiparticle lifetime via the Bethe-Salpeter equation close to the quantum critical point. We show that the critical point associated with the quantum fluctuations change substantially the quasiparticle lifetime without affecting the damping of the collective modes at low angular momentum.
Single particle states
----------------------
We follow the usual procedure and study the lowest order process, i.e. quasiparticle-quasiparticle scattering. Then the expression for the collision integral becomes [@baymp]: $$\begin{aligned}
\label{CIB}
I[n_{\vec{p}}] = 2\pi \sum_{\vec{p}_2 \sigma_2}
\sum_{\stackrel{\vec{p}_3 \sigma_3}{\vec{p}_4 \sigma_4}}'
|<3 4|t|1 2>|^2 \nonumber \\
\delta_{\vec{p}_1+\vec{p}_2 ,\vec{p}_3+\vec{p}_4}
\delta_{\sigma_1+\sigma_2, \sigma_3 +\sigma_4}
\delta(\epsilon_1 + \epsilon_2 -\epsilon_3 -\epsilon_4) \nonumber \\
\Bigl[ n_3 n_4(1-n_2)(1-n_1) - n_1 n_2 (1-n_3) (1-n_4) \Bigr],\end{aligned}$$ which describe scattering processes in and out of the state $\vec{p}$ which is labeled $1$. The spin information can be encoded into the the definition of the matrix element $|<t>|^2$ as described in Ref. \[\], so we drop the spin information from now on (or assume a spin-independent scattering amplitude).
To make progress we linearize the collision integral in the deviations $\delta n_i$ from (local) equilibrium $n_{i}^{0}$. Moreover, we employ the relaxation time approximation which amounts to keeping only the term proportional to $\delta n_1$: $$\label{eq:4}
I = -\frac{1}{\tau_{\vec{p}}} \delta n_{\vec{p}},$$ where $$\begin{aligned}
\label{eq:5}
\frac{1}{\tau_{\vec{p}}} = 2\pi \sum_{\vec{p}_2}
\sum_{\vec{p}_3,\vec{p}_4}'
|<3 4|t|1 2>|^2 \nonumber \\
\delta_{\vec{p}_1+\vec{p}_2 ,\vec{p}_3+\vec{p}_4}
\delta(\epsilon_1 + \epsilon_2 -\epsilon_3 -\epsilon_4) \nonumber \\
\Bigl[ n_{3}^0 n_{4}^0(1-n_{2}^0) + n_{2}^0 (1-n_{3}^0) (1-n_{4}^0) \Bigr].\end{aligned}$$ One can do better than the relaxation time approximation but it is not necessary for our purposes [@baymp] .
Now we introduce the momentum transfer $\vec{q}$, and the energy transfer $\omega$ via the identity $\delta(\epsilon_1 + \epsilon_2 -\epsilon_3 -\epsilon_4) =
\int d\omega \delta(\epsilon_1 - \epsilon_3 - \omega)
\delta(\epsilon_2 - \epsilon_4 + \omega)$. We also specialize to nearly forward scattering. This is an approximation in three dimensions where it amounts to neglecting out of plane scattering. In two dimensions it is essentially exact because of momentum conservation and the reduced phase space. This means that we restrict the sum over $\vec{q}$ to $q < q_c$, where $q_c \ll p_F$ is some cut-off momentum. Dropping the superscripts on the occupation numbers for clarity we then get $$\begin{aligned}
\label{lt3}
\frac{1}{\tau} =
\frac{2 \pi}{\left[1-n(\epsilon)\right]} \int d\omega
\sum_{\vec{q}} \sum_{\vec{p}'}
|<3 4|t|1 2>|^2 \nonumber \\
\delta(v_F q \cos(\theta_p) - \omega)
\delta (v_F q \cos(\theta_{p'}) - \omega) \nonumber \\
\left[1-n(\epsilon -\omega \right)] n(\epsilon')
\left[1-n(\epsilon'+\omega)\right]. \end{aligned}$$ This form is convenient when $\epsilon >0$, for $\epsilon <0$ one can instead exchange $n \leftrightarrow 1-n$ everywhere. The end result is even in $\epsilon$.
To proceed we need an expression for the matrix element. This one can get from the solution of the Bethe-Salpeter equation, which describes repeated scattering of a quasiparticle-quasihole pair: $$\begin{gathered}
\label{bethesalpeter}
t_{\vec{p} \vec{p}'}(\vec{q},\omega +i\eta) = f_{\vec{p} \vec{p}'}
\nonumber \\ -
\sum_{\vec{p}'' \neq \vec{p}'} f_{\vec{p} \vec{p}''}
\frac{\vec{q} \cdot \nabla_{p''}n_{p''}^0}{\omega + i\eta - \vec{q}
\cdot \vec{v}_{\vec{p}''}}
t_{\vec{p}'' \vec{p}'}(\vec{q},\omega +i\eta).\end{gathered}$$ The formalism for solving the Bethe-Salpeter equation in three dimensions can be found in e.g. Ref. \[\]. Here we outline the procedure in two dimensions.
Firstly one expands $t$ in Fourier components $$t_{\vec{p} \vec{p}'}(\vec{q},\omega +i\eta) =
\sum_{m, m'} t_{m,m'} (\vec{q},\omega +i\eta) e^{i (m \theta_{p} - m'
\theta_{p'})},$$ where the angles $\theta$ are measured with respect to $\vec{q}$. A matrix equation for $t_{m,m'}$ can be derived from the orthogonality between the Fourier components $$\label{teq}
(1+F_m) t_{m,m'} = \delta_{m,m'} f_m +F_m \sum_{n} K(n-m) t_{n,m'},$$ where $$K(n) = \int_{0}^{2\pi}\frac{d\theta}{2 \pi} \frac{s}{s-\cos(\theta)}e^{i n \theta},$$ and $$s=\frac{\omega + i\eta}{v_F q}.$$ Explicit expressions for $K(n)$ can be found in Ref. \[\]. Given the values of $\{f_n\}$ one can solve (\[teq\]) to get $t_{m,m'}$.
Close to a Pomeranchuk quantum critical point when one of the Landau parameters reaches a critical value we can keep only the relevant Landau parameter (see Fig.\[phase\_diagram\]). In this paper we study the “nematic” critical point and choose to keep only $F_2=F_{-2}$ nonzero in which case it is straightforward to solve for $t$. The only nonzero components are $$\begin{aligned}
t_{2,2} &=& t_{-2,-2} = \frac{ 1+F_2 - F_2 K(0) }{D} f_2 \, ,
\nonumber
\\
t_{2,-2} &=& t_{-2,2} = \frac{F_2 K(4)}{D} f_2 \, ,\end{aligned}$$ where $$\begin{aligned}
D=[1+F_2 -F_2 K(0)]^2 - [F_2 K(4)]^2 \, .\end{aligned}$$
Now that we have an explicit expression for $<t>$ we can go back to the expression for the lifetime in (\[lt3\]), turn the sums into integrals and perform the angular integrals with the result $$\begin{aligned}
\label{lt4}
\frac{1}{\tau} =
\frac{N(0) L^2}{(2\pi v_F)^2 \left[1-n(\epsilon)\right]} \int d\omega
\left[1-n(\epsilon -\omega \right)] \nonumber
\\ \times \int d\epsilon' n(\epsilon')
\left[1-n(\epsilon'+\omega)\right]
\int_{|\omega|/ v_F}^{q_c} \frac{dq}{q} |t(s,q)|^2,\end{aligned}$$ where $$\begin{aligned}
\label{t2}
|t(s,q)|^2 = \frac{16}{1-s^2}\bigl[1-8 s^2(1-s^2)\bigr]|t_{2,2}+t_{2,-2}|^2
\nonumber \\
+512 s^4 (1-s^2) \bigl( |t_{2,2}|^2 + |t_{2,-2}|^2 \bigr).\end{aligned}$$ The first term in (\[t2\]) gives the dominating contribution (from small $s$) so we only keep it in what follows. Notice that due to the integration limits on $q$ in (\[lt4\]) we have $s \leq 1$ and the singularity near $s=1$ is removed by the behavior of $|t_{2,2}+t_{2,-2}|^2$.
Usually, i.e. for not too strong attractive interaction, it is a good approximation to take $t$ to only depend on $s$. Close to the critical point however, one must also include a $q$-dependence in $t$, explicitly we take $$\begin{aligned}
1+F_2 = \delta +\kappa (q / q_c)^2 \, ,\end{aligned}$$ where $\kappa \propto \partial^2 F_2/\partial q^2$. Now we can trade the integral over $q$ to one over $s$. The integral is then dominated by the small-$s$ contribution and it is possible to approximate the integral by a scaling argument. For small $s$ we have: $$\begin{aligned}
\label{tapprox}
t_{2,2}+t_{2,-2} &=& \frac{f_2}{1+F_2 -F_2 [K(0)+K(4)]} \, ,
\nonumber
\\
&\approx& \frac{f_2}{\delta +\kappa (q/q_c)^2 +2 i F_2 s} \, ,
\nonumber
\\
&\approx&
\frac{f_ 2 s^2}{\delta s^2 + (\omega / \omega_c)^2 -2i s^3}.\end{aligned}$$ Using (\[tapprox\]) the $q$-integral in (\[lt4\]) can be written as: $$\label{eq:8}
16 |f_2|^2 \int_{|\omega|/v_F q_c}^1 ds
\frac{s^3}{[\delta s^2 + (\omega / \omega_c)^2]^2 + 4 s^6}.$$ For $\delta \to 0$ (that is, at the quantum critical point) we scale $s^3 \propto \omega^2$ so that the integral becomes $\propto \omega^{-4/3}$. Performing the remaining integrals at $T=0$ one get $$\label{eq:tau1}
\frac{1}{\tau} \propto |\epsilon|^{2/3} \, ,$$ leading to a very anomalous energy dependence of the quasiparticle lifetime [@eduardonematic; @metznersoft]. For finite $\delta$ (away from the quantum critical point, inside of the quantum disordered regime) we should recover the standard Fermi liquid result. By scaling $s \propto \omega$, we get: $$\label{eq:tau2}
\frac{1}{\tau} \propto \epsilon^2 \ln (\omega_c / \epsilon)^2 +
\mbox{const} \times \epsilon^2,$$ which is the Fermi liquid result [@metznersoft]. The above results can be summarized by the scaling form of the quasiparticle lifetime that can be obtained using the contour integral trick of Pethick [@pethickfinite] that we give in the appendix. The general result for the quasiparticle lifetime is: $$\label{eq:tau3}
\frac{1}{\tau} = \mbox{const} \times \epsilon^2 + 2 \omega_c^2 (\epsilon/\omega_c)^{2/3}
F\Bigl[\Bigl( \frac{3}{\delta}\Bigl)^3 \Bigl(\frac{\epsilon}{\omega_c}\Bigr)^2\Bigr],$$ where $F(\xi)$ is a scaling function such that $F(\xi) \approx 1$ for $\xi \gg 1$ and $F(\xi) \approx -\xi^{2/3}\log(\xi)$ for $\xi \ll 1$, reproducing the results of eqs. (\[eq:tau1\]) and (\[eq:tau2\]). A similar crossover formula (for the imaginary part of the self-energy) in the case of a ferromagnetic quantum critical point was found by Chubukov *et al*.[@chubukovself; @chubukov3]. Notice that our results can be generalized for any kind of Pomeranchuk quantum critical point given the critical Landau parameter. The result of this lowest order approximation is thus that close to the “nematic” critical point in 2D the lifetime of the quasiparticle dominates over the energy and the concept of the usual quasiparticle is not well-defined. To see the real fate of the quasiparticle one must include higher order scattering processes or use a more microscopic theory. Nevertheless, the calculation in this section signals the breakdown of normal Fermi liquid behavior close to a Pomeranchuk quantum critical point.
Studies by other groups of similar models also find a breakdown of Fermi liquid theory at the critical point. The $\epsilon^{2/3}$ behavior in two dimensions has also been found before in related systems, see e.g. Ref. \[\].
Collective modes
----------------
To calculate the relaxation time for collective modes one must keep in mind that inside the linearized collision integral there are terms of the form $\delta n_{\vec{p}} - \delta n_{\vec{p}+\vec{q}}$ where the second term is neglected in the usual relaxation time approximation. The existence of the second term makes collisions less effective in the damping of collective modes. More explicitly, linearizing the collision integral (\[CIB\]) for the $\nu$’s in (\[landaukin\_nu\]) we get: $$\begin{aligned}
\label{Iprime}
I'[\nu] = -\frac{2\pi}{1-n(\epsilon_1)} \sum_{\vec{p}_2}
\sum_{\vec{p}_3,\vec{p}_4}'
|<3 4|t|1 2>|^2 \nonumber \\
\delta_{\vec{p}_1+\vec{p}_2 ,\vec{p}_3+\vec{p}_4}
\delta(\epsilon_1 + \epsilon_2 -\epsilon_3 -\epsilon_4) \nonumber \\
n_{2}(1-n_{3})(1-n_4)\Bigl[\nu_1 - \nu_3 +\nu_2 -\nu_4 \Bigr].\end{aligned}$$ Expanding the $\nu$’s in Fourier components as in Section \[landaudamp\] and projecting the collision integral we get $$\label{eq:14}
\frac{1}{\tau_m} = m^2 \frac{1}{\tau'}.$$ where $1/\tau'$ is given by (\[lt4\]) with the substitution $\int
dq \rightarrow \int dq \; q^2 /2 p_{F}^2$. To get this result we have neglected the term $\nu_2 - \nu_4$ since this term is typically of the same size as the first (i.e. $\nu_1 - \nu_3$) and just gives a different prefactor in the calculation of $1/\tau$ of order $1$. More important is the expansion of $\cos[m (\theta + q/ p_F)]$ in powers of $m q /p_F$, which is only valid for $m \ll p_F /q_c \approx N_{\Lambda}$. The terms with odd powers of $q$ vanish upon performing the angle averages.
For these modes the simple scaling evaluation of the integrals gives: $$\label{eq:15}
\frac{1}{\tau_m} \propto \Bigl(\frac{m}{N_{\Lambda}}\Bigr)^2
\epsilon^{4/3},$$ for $\delta \to 0$ and $$\label{eq:16}
\frac{1}{\tau_m} \propto \Bigl(\frac{m}{N_{\Lambda}}\Bigr)^2
\epsilon^{2},$$ for finite $\delta$. From this we can conclude that the collisionless approximation is good for self-driven modes ($\omega \approx \epsilon$) when $(m/N_{\Lambda})^2 \ll 1$. Note that this implies that it makes sense to talk about (e.g.) the $M=2$ mode also in the critical regime. The high-$M$ modes are more sensitive to low-$q$ scattering. For these it is not possible to relate $\cos[m(\theta +q / p_F)]$ in any simple way by $\cos(m\theta)$ and hence the single particle result above should be a good approximation. This means that the high-$M$ modes are heavily damped in the critical regime. For intermediate $M$ we expect a crossover between the two limiting cases. Also, note that in all cases (single-particle or collective) there is a prefactor of $1/(k_F v_F) \approx 1/\epsilon_F$ and the correct powers of $\omega_c$ to make the dimensions correct.
Quantum Critical Regime - Discussion {#conclusions}
====================================
Recently Yang[@kunyang] has argued in favor of a dynamical critical exponent of $z=2$ for the Pomeranchuk quantum critical point in 2D. This is in disagreement with e.g. the results of Ref. \[\]. The claim that the other non-critical modes just renormalize the couplings in the critical theory is not obvious. In particular, the low-energy particle-hole fermion excitations which are reproduced within the bosonic formulation as linear combinations of the high angular components are neglected in Ref.\[\]. The calculation in Section \[landaudamp\] illustrates the effect of these modes in the collisionless regime, and in a simpler setting than that of multidimensional bosonization which requires a complicated Bogoliubov transformation to achieve similar results.[@antonioboson2]
Taking only $F_0$, $F_1$ and $F_2 \neq 0$ we can get the equations for the transverse and longitudinal modes from the result of Section \[landaudamp\]. If we integrate out all the modes that we are not interested in (i.e. $M \geq 3$), the resulting equations for the modes are: $$\label{mode-}
s_-^2-A_2-\delta A_2 =0,$$ for the transverse ($-$) mode and $$\label{mode+}
\begin{vmatrix}
(s_+^2-A_0)/{2} & -B_1 \\
-B_1 & s^2_+-A_2-\delta A_2
\end{vmatrix}
= 0,$$ for the longitudinal ($+$) mode. From these equations one can extract the leading scaling behavior to be $s_{-}^2 \sim [\delta + \kappa (q/q_c)^2]/4$ and $s_{+} \sim -i [\delta+\kappa (q/q_c)^2 /2]$, assuming that only the $d$-channel goes critical. So the transverse mode has a dynamical critical exponent of $z=2$ and the longitudinal $z=3$ in agreement with Ref. \[\]. The longitudinal mode, being slower, dominates the critical behavior.
As we found in Section \[collisionintegral\], the collisionless behavior is only a good approximation for low-$M$ modes in the critical regime. The question that arises is what would be the effect of collisions on the modes that participate in the Landau damping. The coupling of the $M=2$ mode to the rest of the system still results in damped motion. Generally one can integrate out the modes with $M \geq 3$ with the same result as above in (\[mode-\]) and (\[mode+\]). The difference is that we no longer have an explicit expression for $\delta A_2$ since the approximations of the previous section break down. However, we still must have $\delta A_2 = 2B_{3}^2 U(q,\omega)$, where $U(q,\omega)$ is a function such that $U(q,0)= U(s=0) = -2$. The calculation of $U(q,\omega)$ is beyond of the scope of this work. Nevertheless, if $U(q,\omega)$ is known, the longitudinal mode equation becomes: $$\label{eq:2}
s_+^2 -\frac{1}{8} \Bigl[\delta + \kappa
\Bigl(\frac{q}{q_c}\Bigr)^2\Bigr]
\times
\Bigl[ U(q,\omega) - U(q,0) \Bigr] = 0.$$ If we assume that $U(q,\omega)-U(q,0) \sim -i \omega^{\alpha} / q$ in the $\omega \to 0$ limit, the dynamical critical exponent of the mode is $z=3/(2-\alpha)$. This result allows us to put bonds on the value of the dynamical exponent since we expect $2/3 \leq \alpha \leq 1$ in such a way that $9/4 \leq z \leq 3$. One can speculate that the modes that are responsible for the damping of the critical mode, having very low energy, are weakly damped since there is not much for them to decay into because of the small phase space. This would hint towards $z=3$ even when collisions are taken into account. Another argument in favor of $z=3$ is that the collisionless limit is a good approximation when the mean free path is much longer than the wavelength, which translates into $1/\tau
\ll q v_F $.[@pinesnoz] Taking $1 / \tau \sim \omega^{2/3}$ and $\omega \sim q^3$ the inequality is clearly satisfied as $q\rightarrow 0$ indicating the validity of the collisionless limit.
A questionable thing with the whole approach close to the critical point is whether the calculations are self-consistent. The simplest particle-particle scattering process gave a lifetime that is proportional to $\epsilon^{2/3}$. The same result can be obtained from a 1-loop calculation of the imaginary part of the fermion self energy, which describes similar physics. Taking this seriously would invalidate the quasiparticle picture, and hence make the whole ground for the calculation questionable. Notice, however, that although the quasiparticle description seems to break down, the low angular momentum collective spectrum remains essentially unchanged (except by the damping of the collective modes given by the subleading contributions (\[eq:15\]) and (\[eq:16\])).
One further circumstantial evidence for a breakdown of the quasiparticle near the quantum critical point comes from the theory of multi-dimensional bosonization[@antonioboson2]. The calculations are tractable for an isotropic interaction, the connection to the notations in this paper is $Ng = F_0$. The Bogoliubov transformation used to diagonalize the system connects different patches $i$ and $l$ of the Fermi surface according to the matrices: $$\begin{aligned}
\label{eq:1}
\mathcal{M}_{il} = \delta_{il} + \frac{1}{N} \frac{Ng}{1+Ng}
\frac{\sqrt{s_i s_l}}{s_l - s_i} \, ,
\nonumber \\
\mathcal{N}_{il} = - \frac{1}{N} \frac{Ng}{1+Ng}
\frac{\sqrt{s_i s_l}}{s_l + s_i} \, ,\end{aligned}$$ where $s_i = \cos(\theta_i)$. Away from the critical point the second term in $\mathcal{M}$ represent the dressing of the particle by interactions, and $\mathcal{N}$ is always small. Close to the critical point however, the factor $Ng/(1+Ng) = F_0/(1+F_0)$ blows up when $F_0 \to -1$ and a broad dressing cloud dominates the excitations, also $\mathcal{N}$ is no longer negligible and is an important part of the excitation. Moreover the expression for the quasiparticle residue $Z= \exp\left\{-\frac{1}{4}(\frac{F_0}{1+F_0})^2\right\}$ shows that it vanishes exponentially as the transition is approached. The calculation for the $d$-channel Pomeranchuk transition is much more complicated because the kernel of the integral equation is no longer a product kernel. We expect similar conclusions however. Because the Fermi liquid theory breaks down close to the critical point a calculation within the Fermi liquid theory itself doesn’t make much sense there, except for indicating its own breakdown. The correct description is not in terms of the standard Fermi liquid theory and it would be very interesting to know what it is.
In summary, we have presented an alternative approach for the study of Landau damping in Fermi liquids based on the trace over high angular momentum modes. We find that this approach reproduces well-known results in the literature and allows for the study of the quasiparticles and collective modes close to a Pomeranchuk quantum critical point. Although our approach is general and can be used to study any instability of the Pomeranchuk type, we focused on the problem of the electronic nematic that has been discussed in the literature recently [@eduardonematic; @kunyang; @kee; @khavkine; @eduardonew]. We found that the quasiparticle lifetime becomes anomalously short at the quantum critical point leading to a breakdown of the Fermi liquid description.
This work was supported through NSF grant DMR-0343790.
Contour integral
================
The formula for the inverse lifetime can be written (at $T=0$ and $\epsilon>0$ for simplicity) $$\begin{gathered}
\label{eq:13}
\frac{1}{\tau} =
\frac{1}{2\pi v_F k_F} \int_0^{\epsilon} d\omega \omega
\Bigl[-c_{s}^{<} + N^2(0) \int_{0}^{1} \frac{ds}{s} |t(s,q)|^2 \Bigr]
\\
\equiv
\frac{1}{2\pi v_F k_F} \int_0^{\epsilon} d\omega \omega
(-c_{s}^{<} + I_s).\end{gathered}$$ To get this we have added and subtracted the contribution of the $N^2(0)$ times the $s$-integral down to zero, this is approximately $$\begin{aligned}
\label{eq:10}
c_{s}^{<} = 16 \int_0^{|\omega|/v_F q_c} ds
\frac{s^3}{[\delta s^2 + (\omega / \omega_c)^2]^2 + 4 s^6}
\nonumber \\
= 16 \int_0^1 dx
\frac{x^3}{[\delta x^2 + \kappa]^2 + 4 x^6
(\frac{\omega}{v_F q_c})^2},\end{aligned}$$ which to leading order is a constant (for finite $\kappa$), note that we set $F_2 =-1$ everywhere where it is not dangerous to do so.
Now we can follow Pethick[@pethickfinite] and turn the integral in $I_s$ into two contour integrals, the difference being that in our case the two contours have different orientation $$\label{eq:6}
I_s = - \frac{2 N(0)}{i}\oint_{c}
\frac{ds}{s^2}\frac{1}{\sqrt{1-s^2}}
\frac{1}{1+4s^2(s^2 -1)} (t_{2,2}+t_{2,-2}),$$ using the fact that the integral is dominated by small values of $s$ one can rewrite it as $$\label{eq:7}
I_s = -8 \mathcal{P}\int_{0}^{\infty} dz
\frac{1}{\delta z^2 - (\omega/\omega_c)^2+2z^3},$$ after a deformation of the contour to the imaginary axis. Scaling $z = \delta x /6$ we get $$\label{eq:11}
I_s = -16 \bigl(\frac{3}{\delta}\bigr)^2
\mathcal{P} \int_0^{\infty} dx \frac{1}{x^3 + 3 x^2 - 4 Y},$$ where $Y = (3 /\delta )^3 (\omega / \omega_c)^2$. Finally performing the remaining integral we obtain $$\begin{gathered}
\label{eq:12}
\frac{1}{\tau} = \frac{1}{2\pi v_F k_F}
\Bigl[ - c_{s}^< \frac{\epsilon^2}{2}
\\+ 2 \epsilon^2 (\frac{\omega_c}{\epsilon})^{4/3}
\frac{1}{\xi^{1/3}}
\mathcal{P}\int_{1}^{\infty} dx
\log \left[\frac{x^3 -3 x + 2(1-2 \xi)}{x^3 - 3x +2} \right] \, , \end{gathered}$$ which is of the desired scaling form with $\xi = (3 /\delta )^3 (\epsilon / \omega_c)^2$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider Hamiltonian Floer cohomology groups associated to a Lefschetz fibration, and the structure of operations on them. As an application, we will (under an important additional assumption) equip those groups with connections, which differentiate with respect to the Novikov variable.'
author:
- Paul Seidel
title: |
Fukaya $A_\infty$-structures associated to\
Lefschetz fibrations. IV
---
Introduction\[sec:intro\]
=========================
This paper continues a discussion [@seidel12b; @seidel14b; @seidel15; @seidel16] of pseudo-holomorphic curve theory as applied to Lefschetz fibrations. Ultimately, our sights are set on Fukaya categories; but here, we remain in the “closed string” context of Hamiltonian Floer cohomology.
Given a symplectic Lefschetz fibration (over the disc), one can single out certain classes of time-dependent Hamiltonians on the total space, which give rise to Floer cohomology groups that are invariants of the fibration. More concretely, the outcome is an infinite family of such groups, indexed by the “rotation number at infinity” $r \in {{\mathbb Z}}$. All of them vanish if the Lefschetz fibration is trivial (has no critical points). Based on intuition from mirror symmetry, one expects these groups to carry a rich algebraic structure. The initial breakthrough in constructing the desired operations is due to Abouzaid-Ganatra (unpublished), who equipped the $r = 1$ group with a ring structure. Here, we streamline and generalize their insight, by introducing a suitable TQFT framework. Eventually, one wants to compare Hamiltonian Floer cohomology with certain twisted Hochschild cohomology groups of the associated Fukaya category, and thereby relate our operations to ones in homological algebra; however, to carry that out, one needs a specially adapted version of closed-open string maps, which is beyond the scope of the present paper (the literature contains several constructions relating the open and closed string sectors for Lefschetz fibrations, but none of them seems really practical for this purpose).
Our main motivation for pursuing this direction comes from [@seidel16]. There, symplectic cohomology was shown, under a fundamental additional assumption, to carry a one-parameter family of connections, which differentiate with respect to the Novikov formal parameter. That construction relied on the structure of symplectic cohomology as a BV algebra. Using some of the operations we have made available, a version of the same construction can be carried out for Lefschetz fibrations. The outcome, under a similar assumption as in [@seidel16], is that each of our Floer cohomology groups carries a connection; for different $r$, these correspond to different instances of the family of connections on symplectic cohomology. Ultimately, one wants to use these connections, and closed-open string maps, to carry over the enumerative results from [@seidel16] to Fukaya categories of Lefschetz fibrations; we refer the reader to [@seidel15 Section 4] for the conjectural outcome and its context.
The structure of this paper may deserve some comment. Floer-theoretic constructions and identities have often been thought of as realizations of abstract TQFT arguments, and that is also true here. Usually, the formal TQFT part is straightforward, and the main work goes into its Floer-theoretic implementation. In our situation, the weight shifts: the TQFT arguments are far from obvious, whereas the Floer theory is quite standard (except possibly for some compactness arguments). Therefore, after outlining our results in Section \[sec:results\], the bulk of the paper (Sections \[sec:sl2\]–\[sec:elliptic\]) is devoted to TQFT considerations and their geometric prerequisites. After that, we explain the necessary basic Floer theory (Sections \[sec:maps-to-the-disc\]–\[sec:floer\]). Putting together the two parts is a routine process, and we will only describe it briefly (Section \[sec:end\]).
[*Acknowledgments.*]{} This work was supported by the Simons Foundation, through a Simons Investigator award; by NSF grant DMS-1500954; by the Institute for Advanced Study, through a visiting appointment supported by grants from the Ambrose Monell Foundation and the Simonyi Endowment Fund; and by Columbia University, through an Eilenberg Visiting Professorship. I would like to thank Mohammed Abouzaid and Sheel Ganatra for generously sharing their seminal insights; Nick Sheridan for illuminating conversations on further developments; and Alexander Sukhov for patiently answering my questions concerning pseudoconvexity (see Remark \[th:sukhov\]).
Main constructions\[sec:results\]
=================================
This section summarizes the structures that will arise from our analysis of Floer-theoretic operations, including connections. This will be done with only the minimal amount of technical details, and without any real attempt at explaining the underlying geometry. At the end, we relate part of our constructions to more familiar ones in symplectic cohomology.
Background\[subsec:background\]
-------------------------------
Let $$\label{eq:lefschetz}
\pi: E \longrightarrow B$$ be a proper symplectic Lefschetz fibration whose base is the open unit disc $B \subset {{\mathbb C}}$, and whose total space is a $2n$-dimensional symplectic manifold. For simplicity, we assume that $$\label{eq:cy}
c_1(E) = 0,$$ and choose a trivialization of the anticanonical bundle (the complex line bundle representing the first Chern class), up to homotopy. The smooth fibres are then closed $(2n-2)$-dimensional symplectic manifolds, again with trivialized anticanonical bundle.
All Floer cohomology groups under discussion will be finite-dimensional ${{\mathbb Z}}$-graded vector spaces over the Novikov field ${{\mathbb K}}$ with real coefficients. Elements of ${{\mathbb K}}$ are of the form $$\label{eq:novikov}
f(q) = c_0 q^{d_0} + c_1 q^{d_1} + \cdots, \quad c_i \in {{\mathbb R}}, \;\; d_i \in {{\mathbb R}}, \;\; \textstyle \lim_i d_i = +\infty.$$ Concretely, to we associate a family of such Floer cohomology groups, denoted by $$\label{eq:floer-cohomology}
\mathit{HF}^*(E,r), \;\; r \in {{\mathbb Z}}.$$ The $r = 0$ group admits a description in purely topological terms, as the cohomology of $E$ “relative to a fibre at $\infty$”. If we compactify to $\bar\pi: \bar{E} \rightarrow \bar{B}$, then the statement is that $$\label{eq:floer-0}
\mathit{HF}^*(E,0) {\cong}H^*(\bar{E},\bar{E}_w) \quad \text{for some $w \in \partial \bar{B}$.}$$ The other groups , $r \neq 0$, are not classical topological invariants; they encode symplectic information related to the monodromy around $\partial \bar{B}$. In the simplest case $r = 1$, if $\mu$ is that monodromy (a symplectic automorphism of the fibre), with its fixed point Floer cohomology $\mathit{HF}^*(\mu)$, there is a long exact sequence $$\label{eq:b-sequence}
\cdots \rightarrow \mathit{HF}^*(\mu) \longrightarrow H^*(E;{{\mathbb K}}) \stackrel{}{\longrightarrow} \mathit{HF}^*(E,1) \longrightarrow \mathit{HF}^{*+1}(\mu) \rightarrow \cdots$$ This idea can be generalized to other $r>0$, leading to a version of the spectral sequence from [@mclean12], involving powers of $\mu$ (see [@seidel14b Lemma 6.14] for the statement). The situation for $r<0$ is dual: there are nondegenerate pairings $$\label{eq:pairing}
\mathit{HF}^*(E,-r) \otimes \mathit{HF}^{2n-*}(E,r) \longrightarrow {{\mathbb K}}.$$
Operations\[subsec:operations\]
-------------------------------
Our main topic is the structure of operations on . We start with linear maps. Each group $\mathit{HF}^*(E,r)$ has a canonical automorphism $\Sigma$, which generates an action of ${{\mathbb Z}}/r$. For $r = 0$, this automorphism admits a classical interpretation as the relative monodromy map, moving $w$ around $\partial \bar{B}$ in . For $r \neq 0$, one can think of $\mathit{HF}^*(E,r)$ as the Floer cohomology of the $r$-th power of a Hamiltonian automorphism, and then $\Sigma$ is induced by $1/r$ rotation of loops. We also have continuation maps which increase $r$, and which are invariant under the action of $\Sigma$ (on either the left or right): $$\begin{aligned}
\label{eq:continuation}
& C: \mathit{HF}^*(E,r) \longrightarrow \mathit{HF}^*(E,r+1), \\
\label{eq:continuation-2}
& D: \mathit{HF}^*(E,r) \longrightarrow \mathit{HF}^{*-1}(E,r+1).\end{aligned}$$ The structure of multilinear operations is more interesting, and does not seem to admit a single concise description. Two simple consequences of the general theory are the following:
\[th:operations-1\] $\mathit{HF}^*(E,1)$ carries the structure of a Gerstenhaber algebra (a commutative product and a bracket of degree $-1$, with suitable relations between them).
\[th:operations-1b\] Each $\mathit{HF}^*(E,r)$ is a Gerstenhaber module over $\mathit{HF}^*(E,1)$, in a ${{\mathbb Z}}/r$-equivariant way (and reducing to the diagonal module for $r = 1$).
The Gerstenhaber algebra structure includes the Abouzaid-Ganatra product. Propositions \[th:operations-1\] and \[th:operations-1b\] assign a distingushed role to $\mathit{HF}^*(E,1)$. Operations of this kind are the most important ones for our applications, but they by no means exhaust the structures that are present. Here are two other easily stated results that can be extracted from the general theory.
\[th:operations-2\] There is the structure of a bigraded associative algebra on $$\label{eq:bigraded-1}
\bigoplus_{s \geq 0} \mathit{HF}^*(E,s+1).$$
\[th:operations-3\] There is the structure of a bigraded Lie algebra on $$\label{eq:bigraded-lie}
\bigoplus_{s \geq 0} \mathit{HF}^{*+1}(E,s+1)^{{{\mathbb Z}}/(s+1)}.$$
The product from Proposition \[th:operations-2\] contains the algebra structure from Proposition \[th:operations-1\] (by restricting to $s = 0$), and also the module structure over that algebra from Proposition \[th:operations-1b\], for $r>0$. It may be worth while spelling out what Proposition \[th:operations-3\] says, in view of the various degree shifts involved. $\mathit{HF}^i(E,r)^{{{\mathbb Z}}/r}$ appears in our graded Lie algebra in (cohomological) degree $i-1$, and this determines the signs in the antisymmetry and Jacobi relations. In addition, the bracket is homogeneous with respect to $s = r-1$, hence is given by maps $$\mathit{HF}^{i_1}(E,r_1)^{{{\mathbb Z}}/r_1} \otimes \mathit{HF}^{i_2}(E,r_2)^{{{\mathbb Z}}/r_2} \longrightarrow
\mathit{HF}^{i_1+i_2-1}(E,r_1+r_2-1)^{{{\mathbb Z}}/(r_1+r_2-1)}.$$ The Lie structure from Proposition \[th:operations-1\] is contained in that of Proposition \[th:operations-3\], and so is the ${{\mathbb Z}}/r$-invariant part of the Lie module structure from Proposition \[th:operations-1b\], for $r>0$.
The product structures from Propositions \[th:operations-1\]–\[th:operations-2\] have a unit $$\label{eq:e-}
e \in \mathit{HF}^0(E,1),$$ which can be thought of as the image of the unit in ordinary cohomology under the map from . Another important ingredient for us will be the image of the symplectic class, or more precisely of $q^{-1}[\omega_E] \in H^2(E;{{\mathbb K}})$, under that map. We call this image the Kodaira-Spencer class, and write it as $$\label{eq:k-k}
k \in \mathit{HF}^2(E,1).$$ Following the model of [@seidel16], vanishing of the Kodaira-Spencer class gives rise to another operation, which is not ${{\mathbb K}}$-linear, but instead differentiates with respect to the Novikov parameter $q$:
\[th:diff-q\] Suppose that vanishes. Choose a bounding cochain for its cocycle representative, in the Floer cochain complex underlying $\mathit{HF}^*(E,1)$. That choice determines a connection on each group , which means a map $$\label{eq:nabla-r}
\begin{aligned}
& \nabla: \mathit{HF}^*(E,r) \longrightarrow \mathit{HF}^*(E,r), \\
& \nabla(f x) = f \nabla x + (\partial_q f) x \quad \text{for $f \in {{\mathbb K}}$.}
\end{aligned}$$ Changing the bounding cochain by adding a cocycle representing some class in $\mathit{HF}^*(E,1)$ has the effect of subtracting the Lie action of that class (in the sense of Proposition \[th:operations-1b\]) from .
The theory producing all these results is modelled on an abstract notion of TQFT for surfaces equipped with framed $\mathit{PSL}_2({{\mathbb R}})$-connections (“framed” means that we consider only connections on the trivial bundle, and do not quotient out by the gauge group). Here, $\mathit{PSL}_2({{\mathbb R}})$ is thought of as acting on the base of by hyperbolic isometries. Of course, this action cannot be lifted to symplectic automorphisms of the total space, but lifts exist infinitesimally if one only considers the Lefschetz fibration near infinity, and that is sufficient for our purpose. More precisely, the products and Lie brackets mentioned above are all constructed using flat $\mathit{PSL}_2({{\mathbb R}})$-connections. More generally, one can allow connections whose curvature belongs to the nonnegative cone inside $\mathfrak{sl}_2({{\mathbb R}})$. Nonnegatively curved connections are required to produce the maps , , as well as the classes , (and, naturally in view of the latter, play an important role in Proposition \[th:diff-q\]).
At this point, we want to make up for a slight lack of precision in the discussion above. There are many choices involved in setting up $\mathit{HF}^*(E,r)$, but the essential one (other than $r$, of course) is that of a hyperbolic element $g \in \mathit{PSL}_2({{\mathbb R}})$. One uses the action of $g$ on the disc to determine the structure at infinity of a Hamiltonian automorphism of $E$ (and then, $r$ is used to choose an extension of that automorphism over all of $E$). Different choices of $g$ give rise to isomorphic Floer cohomology groups. However, these isomorphisms are generally unique only up to composition with powers of $\Sigma$ (it doesn’t matter which side the composition is taken on, since the isomorphisms are $\Sigma$-equivariant). As an elementary illustration, let’s take a second look at the $r = 0$ case. A more precise statement of would require that $w$ should lie in the interval in $\partial \bar{B}$ bounded by the repelling fixed point of $g$ (on the left) and its attracting fixed point (on the right). For two choices of $g$ within the same one-parameter semigroup (or equivalently, which have the same attracting and repelling fixed points), the resulting relative cohomology groups $H^*(\bar{E},\bar{E}_w)$ can be identified canonically; otherwise, one needs to make a choice, of how to move one $w$ point to the other (or more intrinsically, of a path inside the hyperbolic locus connecting the two choices of $g$). In fact, this is also the situation for general $r$. We will prove that the algebraic structures from Proposition \[th:operations-1\] and \[th:operations-1b\] are compatible with those isomorphisms, and a similar argument would apply to Proposition \[th:diff-q\]. In the interest of brevity, we will not address the parallel uniqueness issue for the structures from Propositions \[th:operations-2\] and \[th:operations-3\], since those are less central for our intended applications.
\[th:open-q\] The way in which the various pieces fit together (together with the parallel situation in homological algebra [@seidel14b Section 2f]) suggests the following possible further developments (none of which will be attempted in this paper).
\(i) The relation between Proposition \[th:operations-2\] and the module structures from Proposition \[th:operations-1b\] seems to indicate that the algebra structure ought to be extended to $s \in {{\mathbb Z}}$.
\(ii) There might be an extension of which would add an $s = -1$ term of the form $\mathit{HF}^*(E,0)[[u]]$, for $u$ a formal variable of degree $2$. (iii) Take $r>0$. By using a suitably chosen chain level representative of $\Sigma$, one can define the algebraic mapping cone of $\mathit{id}-\Sigma$, which is a graded space $S^*(E,r)$ fitting into a long exact sequence $$\cdots \rightarrow S^*(E,r) \longrightarrow \mathit{HF}^*(E,r) \xrightarrow{\mathit{id}-\Sigma} \mathit{HF}^*(E,r) \rightarrow \cdots$$ Extend this to $r = 0$ by setting $S^*(E,0) = \mathit{HF}^*(E,0)$. Then, the expectation is that $$\label{eq:k-space}
\bigoplus_{s \geq -1} S^*(E,s+1)$$ carries the structure of a bigraded Batalin-Vilkovisky (BV) algebra. The Lie structure from Proposition \[th:operations-3\], with its extension from (ii), would be thought of as the “cyclic” counterpart of the bracket on (compare [@seidel14b Remark 2.10]). The maps $S^*(E,r) \rightarrow \mathit{HF}^*(E,r)$ would form an algebra homomorphism from to (an extended version of) .
Relation with symplectic cohomology
-----------------------------------
Symplectic cohomology is a well-known invariant of noncompact symplectic manifolds, which can be easily adapted to our situation .
\[th:limit\] In the limit $r \rightarrow \infty$, the maps recover the symplectic cohomology of $E$: $$\label{eq:direct-lim}
\underrightarrow{\lim}_r\,\mathit{HF}^*(E,r) {\cong}\mathit{SH}^*(E).$$
We can describe the situation a little more precisely, using versions of symplectic cohomology with “finite slope”. These are groups $$\label{eq:floer-cohomology-2}
\mathit{HF}^*(E,r), \;\; r \in {{\mathbb R}}\setminus {{\mathbb Z}}.$$ As the notation indicates, they fit in with , in the sense that there are continuation maps which increase $r$, and can mix the integer and non-integer cases (we will denote all such maps by $C$). By definition, symplectic cohomology is the direct limit of $\mathit{HF}^*(E,r)$ using only non-integer values of $r$. One proves Proposition \[th:limit\] by interleaving this with the integer values. Another familiar property of is that it comes with a natural map $H^*(E;{{\mathbb K}}) \rightarrow \mathit{HF}^*(E,r)$ for positive $r$. If we take $r \in (0,1)$, then composing this map with a continuation map $\mathit{HF}^*(E,r) \rightarrow \mathit{HF}^*(E,1)$ recovers the map from .
\[th:relate-operations\] It is well-known (see e.g. [@seidel07; @ritter10]) that symplectic cohomology carries the structure of a BV algebra (this notion has already been mentioned in Remark \[th:open-q\]). The basic operations are: the (graded commutative) product; and the BV operator $\Delta$, of degree $-1$. From them, one constructs a bracket of degree $-1$, $$[x_1,x_2] = \Delta(x_1 \cdot x_2) - (\Delta x_1) \cdot x_2 - (-1)^{|x_1|} x_1 \cdot (\Delta x_2),$$ which yields a Gerstenhaber algebra structure. It is natural to ask how relates the operations on with those in $\mathit{SH}^*(E)$. For instance, it is easy to show that turns into $\Delta$ in the limit (see Proposition \[th:mixed-continuation-2\](iv) for a sketch of the geometry underlying that argument). While we will not pursue such questions further here, it may be worthwhile summarizing what one can reasonably expect:
\(i) The map $\mathit{HF}^*(E,1) \rightarrow \mathit{SH}^*(E)$ should be a homomorphism of Gerstenhaber algebras.
\(ii) For the Gerstenhaber module structure from Proposition \[th:operations-1b\], the expected behaviour is slightly more complicated. The product is still compatible with that on symplectic cohomology, but the bracket should fit into a commutative diagram $$\label{eq:lie-lie}
\xymatrix{
\mathit{HF}^*(E,1) \otimes \mathit{HF}^*(E,r) \ar[rr]^-{[\cdot,\cdot]}
\ar[d]
&& \mathit{HF}^{*-1}(E,r) \ar[d] \\
\mathit{SH}^*(E) \otimes \mathit{SH}^*(E) \ar[rr]^-{[\cdot,\cdot]_c} && \mathit{SH}^*(E)
}$$ where the bottom $\rightarrow$ is the following operation on symplectic cohomology, for $c = r-1$: $$\label{eq:two-brackets}
[x_1,x_2]_c = [x_1,x_2] - c (\Delta x_1) \cdot x_2 = \Delta(x_1 \cdot x_2) - (c+1) (\Delta x_1) \cdot x_2 - (-1)^{|x_1|} x_1 \cdot (\Delta x_2).$$ We will see the same phenomenon in Proposition \[th:two-connections\] below (which is indeed closely related).
\(iii) One expects similar compatibility statements for Propositions \[th:operations-2\] and (less obviously) \[th:operations-3\].
For our intended applications, the most important aspect of the relation between and concerns connections. Let’s first summarize the outcome of the construction from [@seidel16]:
\[th:elliptic-connection\] Fix $r_1 \in {{\mathbb R}}^{>0} \setminus {{\mathbb Z}}$ and $r_2 \in {{\mathbb R}}\setminus {{\mathbb Z}}$, such that $r_1 + r_2 \notin {{\mathbb Z}}$. Suppose that the image of $q^{-1}[\omega_E]$ in $\mathit{HF}^2(E,r_1)$ vanishes, and choose a bounding cochain. This gives rise to a family of “connections” parametrized by $c \in {{\mathbb K}}$, which are maps $$\label{eq:nabla-c}
\begin{aligned}
& \nabla_c: \mathit{HF}^*(E,r_2) \longrightarrow \mathit{HF}^*(E,r_1+r_2), \\
& \nabla_c( fx) = f \nabla_c x + (\partial_q f) C(x),
\end{aligned}$$ related to each other as follows. Associated to the bounding cochain is an $a \in \mathit{HF}^0(E,r_1)$, and $$\nabla_c x = \nabla_1 x + (c-1)\, a \cdot x,$$ where $a \cdot x$ is the pair-of-pants product $\mathit{HF}^*(E,r_1) \otimes \mathit{HF}^*(E,r_2) \rightarrow \mathit{HF}^*(E,r_1+r_2)$. The maps commute with the continuation maps that increase $r_2$. Hence, in the limit, they induce a family of actual connections on $\mathit{SH}^*(E)$.
To relate this to Proposition \[th:diff-q\], one needs to restrict attention to integer values of the parameter $c$. We will show the following (which is only one of the possible ways of formulating the relationship, and not necessarily the most conceptually satisfying one, but sufficient for our purpose):
\[th:two-connections\] Suppose that the assumption of Proposition \[th:diff-q\] holds; then, so does that of Proposition \[th:elliptic-connection\], for any $r_1>1$. Choose $r_1 \in (1,2)$ and $r_2 \notin {{\mathbb Z}}$, such that $r_1+r_2 \notin {{\mathbb Z}}$ and the interval $(r_2,r_1+r_2)$ contains exactly one integer $r \in {{\mathbb Z}}$. Then, for suitably correlated choices of bounding cochains, for $c = r-1$ factors through , as follows: $$\label{eq:two-connections}
\xymatrix{
\ar@/_1.5pc/[rrr]_-{\nabla_c}
\mathit{HF}^*(E,r_2) \ar[r]^-{C} & \mathit{HF}^*(E,r)
\ar[r]^-{\nabla}
& \ar[r]^-{C} \mathit{HF}^*(E,r) &
\mathit{HF}^*(E,r_1+r_2).
}$$
$\mathfrak{sl}_2({{\mathbb R}})$\[sec:sl2\]
===========================================
The Lie algebras which occur naturally in symplectic geometry are those of Hamiltonian vector fields, which are complicated even in two dimensions. Fortunately, to see the structures emerging in our particular application, it is enough consider a finite-dimensional Lie subalgebra. This section contains the necessary elementary material, much of it borrowed from [@lalonde-mcduff97].
Nonnegativity and the rotation number
-------------------------------------
Set $G = \mathit{PSL}_2({{\mathbb R}})$ and ${\mathfrak{g}}= \mathfrak{sl}_2({{\mathbb R}})$. Call $\gamma \in {\mathfrak{g}}$ nonnegative if the associated quadratic form is positive semidefinite: $$\label{eq:quadratic-form}
\mathrm{det}(v, \gamma v) \geq 0 \quad \text{for all $v \in {{\mathbb R}}^2$,}$$ where $(v, \gamma v)$ is the matrix with columns $v$ and $\gamma v$; an equivalent notation would be $v \wedge \gamma v \geq 0$. Concretely, nonnegative $\gamma$ are of the form $$\label{eq:abc}
\gamma = \left( \begin{smallmatrix} \epsilon & \delta-\alpha \\ \delta+\alpha & -\epsilon \end{smallmatrix} \right), \quad
\alpha \geq \sqrt{\delta^2+\epsilon^2}.$$ If we think of $G$ as acting on ${{\mathbb R}}P^1$, an element of ${\mathfrak{g}}$ is nonnegative iff the associated vector field does not point in clockwise direction anywhere. Write ${\mathfrak{g}}_{\geq 0} \subset {\mathfrak{g}}$ for the cone of nonnegative elements, which is invariant under the adjoint action. Similarly, we call $\gamma$ positive if is positive definite; those elements constitute the interior ${\mathfrak{g}}_{>0} \subset {\mathfrak{g}}_{\geq 0}$. Correspondingly, there is the opposite cone ${\mathfrak{g}}_{\geq 0} = -{\mathfrak{g}}_{\leq 0}$ of nonpositive elements, whose interior consists of the negative ones.
There is a related notion for parabolic elements $g \in G$ (the terminology in this paper is that a nontrivial element of $G$ is either elliptic, parabolic, or hyperbolic; the identity ${\mathbbm{1}}$ is considered to be neither). Namely, given such an element, choose the preimage in $\mathit{SL}_2({{\mathbb R}})$ with $\mathrm{tr}(g) = 2$. We say that $g$ is a nonnegative parabolic if $\mathrm{det}(v,gv) \geq 0$ for all $v$; otherwise, it is a nonpositive parabolic (and those are precisely the two parabolic conjugacy classes in $G$). Equivalently, a parabolic $g$ is nonnegative if it can be written as $g = e^{\gamma}$ for some nonzero nilpotent $\gamma \in {\mathfrak{g}}_{\geq 0}$, and similarly for the other sign.
A path $g(t) \in G$ is called nonnegative if $$\label{eq:nonnegative-path}
g'(t) g(t)^{-1} \in {\mathfrak{g}}_{\geq 0}, \quad \text{or equivalently} \quad g(t)^{-1} g'(t) \in {\mathfrak{g}}_{\geq 0}.$$ Geometrically, a path in $G$ is nonnegative iff along it, points on ${{\mathbb R}}P^1$ never move clockwise. If $g_1(t)$ and $g_2(t)$ are nonnegative paths, so is their pointwise product, since $$\label{eq:make-positive}
g_0(t) = g_1(t)g_2(t) \;\; \Longrightarrow \;\;
g_0'(t) g_0(t)^{-1} = g_1'(t) g_1(t)^{-1} + g_1(t) \big( g_2'(t) g_2(t)^{-1} \big) g_1(t)^{-1}.$$ In parallel, we have notions of positive, nonpositive, and negative paths. If $g(t)$ is nonnegative, both $g(t)^{-1}$ and $g(-t)$ are nonpositive, while $g(-t)^{-1}$ is again nonnegative. If $g_1(t)$ is nonnegative, and $g_2(t)$ positive, their product is positive, by . In particular, any nonnegative path can be perturbed to a positive one, by multiplying it with $e^{t\gamma}$ for some small positive $\gamma$.
Let $\tilde{G} \rightarrow G$ be the universal cover. A preimage in $\tilde{G}$ of a given $g \in G$ is the same as a lift of the action of $g$ to ${{\mathbb R}}\rightarrow {{\mathbb R}}/\pi {{\mathbb Z}}= {{\mathbb R}}P^1$. Any $\tilde{g} \in \tilde{G}$ has a rotation number, defined in terms of the action on ${{\mathbb R}}$ by $$\label{eq:rotation}
\mathrm{rot}(\tilde{g}) = {\textstyle \lim_{k \rightarrow \infty}}\; \frac{\tilde{g}^k(x) - x}{\pi k}
\quad \text{for any $x \in {{\mathbb R}}$.}$$ If one composes $\tilde{g}$ with an element in the central subgroup $\mathit{ker}(\tilde{G} \rightarrow G) {\cong}{{\mathbb Z}}$, the rotation number changes by adding that same integer. The other general properties of rotation numbers are: they are continuous; conjugation invariant; homogeneous, meaning that $$\label{eq:homogeneity}
\mathrm{rot}(\tilde{g}^k) = k\, \mathrm{rot}(\tilde{g}), \quad k \in {{\mathbb Z}};$$ they can’t decrease along nonnegative paths; and they satisfy the quasimorphism property $$\label{eq:quasimorphism}
|\mathrm{rot}(\tilde{g}_1\tilde{g}_2) - \mathrm{rot}(\tilde{g}_1) - \mathrm{rot}(\tilde{g}_2)| \leq 1.$$ One can break down the situation into conjugacy types, as follows. Suppose that $\tilde{g}$ is the lift of an elliptic $g \in G$. Then, $g$ is a rotation with angle $\theta$, where $\theta \in ({{\mathbb R}}/\pi{{\mathbb Z}}) \setminus \{0\}$ (in our terminology, “rotations” can be with respect to any oriented basis of ${{\mathbb R}}^2$, hence are not necessarily orthogonal). The choice of $\tilde{g}$ corresponds to a lift $\tilde{\theta} \in {{\mathbb R}}$, and the rotation number is $\tilde{\theta}/\pi \in {{\mathbb R}}\setminus {{\mathbb Z}}$. In the other (hyperbolic or parabolic) case, the rotation number is an integer, and can be computed as follows. Fix an eigenvector of $g$. The choice of $\tilde{g}$ determines a homotopy class of paths from the identity to $g$. The image of the eigenvector along such a path yields a loop in ${{\mathbb R}}P^1$, whose degree is the rotation number. We would like to state two consequences of this discussion. For the first one, note that the map $\tilde{G} \rightarrow G$ factors through $\mathit{SL}_2({{\mathbb R}})$. Case-by-case analysis then directly shows that:
If $g \in \mathit{SL}_2({{\mathbb R}})$ is the image of $\tilde{g} \in \tilde{G}$, then $$\label{eq:parity}
\left\{
\begin{aligned}
& \mathrm{tr}(g) > 0\; \Longleftrightarrow\; \mathrm{rot}(\tilde{g}) \in 2{{\mathbb Z}}+ (-{{\textstyle\frac{1}{2}}},{{\textstyle\frac{1}{2}}}), \\
& \mathrm{tr}(g) < 0\; \Longleftrightarrow\; \mathrm{rot}(\tilde{g}) \in 2{{\mathbb Z}}+ 1 + (-{{\textstyle\frac{1}{2}}},{{\textstyle\frac{1}{2}}}).
\end{aligned}
\right.$$
For our second observation, let’s rewrite more symmetrically as follows: $$\label{eq:triple-trivial}
\tilde{g}_0\tilde{g}_1\tilde{g}_2 = {\mathbbm{1}}\quad \Longrightarrow \quad
|\mathrm{rot}(\tilde{g}_0) + \mathrm{rot}(\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2)| \leq 1.$$
\[th:sharpened-inequality\] (i) In the situation of , suppose that the image $g_k$ of some $\tilde{g}_k$ is elliptic. Then the inequality is strict.
\(ii) In the same situation, suppose that $g_k$ is a nonnegative parabolic. Then $$\mathrm{rot}(\tilde{g}_0) + \mathrm{rot}(\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2) \leq 0.$$
\(iii) Correspondingly, if $g_k$ is a nonpositive parabolic, one has a lower bound of $0$ for the sum of rotation numbers.
\(i) Without loss of generality, one can assume that $k = 0$ and $\mathrm{rot}(\tilde{g}_0) \in (0,1)$. Then there is a nonnegative path from the identity to $g_0$, which lifts to a path with endpoint $\tilde{g}_0$. Therefore, $$\mathrm{rot}(\tilde{g}_0) + \mathrm{rot}(\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2) \leq
\mathrm{rot}(\tilde{g}_0) +
\mathrm{rot}(\tilde{g}_0\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2) = \mathrm{rot}(\tilde{g}_0) < 1.$$ Passing to inverses yields the corresponing lower bound.
\(ii) has the same proof, except that this time, a nonnegative path from the identity to $\tilde{g}_0$ exists for $\mathrm{rot}(\tilde{g}_0) = 0$. Again, one passes to inverses to obtain (iii).
\[th:example-path\] Let $g = \exp(\gamma) \in G$, with $\gamma \in {\mathfrak{g}}$ sufficiently small. Then, for any integer $r \geq 1$, $$g(t) = \big( \begin{smallmatrix} \cos(r \pi t) & -\sin(r \pi t) \\ \sin(r \pi t) & \cos(r \pi t) \end{smallmatrix} \big)\exp(t \gamma)$$ is a positive path from the identity to $g$, because of . The lift $\tilde{g} \in \tilde{G}$ of $g$ associated to that path differs from the obvious lift of $g$ (close to the identity) by $r \in {{\mathbb Z}}\subset \tilde{G}$; hence, $\mathrm{rot}(\tilde{g}) \approx r$, with equality unless $g$ is elliptic.
\[th:three-classes\] Fix hyperbolic conjugacy classes $C_i$ ($i = 0,1,2$). Up to simultaneous conjugation, there is exactly one triple $(g_0,g_1,g_2)$ with $g_i \in C_i$, $g_0 = g_1g_2$, and such that if we take lifts $\tilde{g}_i$ satisfying $\tilde{g}_0 = \tilde{g}_1\tilde{g}_2$, then $$\label{eq:-1rot}
\mathrm{rot}(\tilde{g}_0) = \mathrm{rot}(\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2) - 1.$$ To see that, let’s suppose that our conjugacy classes are determined by $\mathrm{tr}(\tilde{g}_i) = \pm( \lambda_i + \lambda_i^{-1})$ for some $\lambda_i > 1$. By , if we choose representatives in $\mathit{SL}_2({{\mathbb R}})$ such that $\mathrm{tr}(g_1) = \lambda_1 + \lambda_1^{-1}$ and $\mathrm{tr}(g_2) = \lambda_2 + \lambda_2^{-1}$, then $\mathrm{tr}(g_0) = -\lambda_0 - \lambda_0^{-1}$. With that in mind, after common conjugation, the only solutions are $$\label{eq:g1g2}
g_1 = \begin{pmatrix} \lambda_1 & 0 \\ 0 &\lambda_1^{-1} \end{pmatrix},
\quad
g_2 = \begin{pmatrix} \lambda_2 - t & \pm (\lambda_2-\lambda_2^{-1} - t) \\ \pm t & \lambda_2^{-1} + t \end{pmatrix}, \text{ where }
t = \frac{\lambda_0 + \lambda_0^{-1} + \lambda_1\lambda_2 + \lambda_1^{-1}\lambda_2^{-1}}{\lambda_1 - \lambda_1^{-1}}.$$ By looking at the case where $\lambda_1,\lambda_2 \approx 1$ and deforming the matrices $g_1$, $g_2$ to the identity, one sees that the solution of with the $-$ sign satisfies (the solution with the $+$ sign satisfies the version of with a $+1$ instead of $-1$).
\[th:2-elliptic\] Consider elements $\tilde{g}_i$ of elliptic conjugacy classes $\tilde{C}_i \subset \tilde{G}$ ($i = 1,2$), with $$\mathrm{rot}(\tilde{g}_1) \in (0,1), \;\; \mathrm{rot}(\tilde{g}_2) \in (0,1), \;\;
\mathrm{rot}(\tilde{g}_1) + \mathrm{rot}(\tilde{g}_2) > 1.$$ For the product $\tilde{g} = \tilde{g}_1\tilde{g}_2$ and its image $g = g_1g_2$ in $G$, the following is a complete list of possibilities: $$\label{eq:3-possi}
\left\{
\begin{aligned}
& \text{$g$ can be elliptic, with any value $\mathrm{rot}(\tilde{g}) \in (1,\mathrm{rot}(\tilde{g}_1)+\mathrm{rot}(\tilde{g}_2)]$;} \\
& \text{$g$ can be nonnegative parabolic, and $\mathrm{rot}(\tilde{g}) = 1$;} \\
& \text{$g$ can lie in any hyperbolic conjugacy class, and $\mathrm{rot}(\tilde{g}) = 1$.}
\end{aligned}
\right.$$ Moreover, each possibility happens in a way which is unique up to overall conjugation. To see that, we will carry out computations in the associated conjugacy classes in $\mathit{SL}_2({{\mathbb R}})$, which correspond to rotations with angles $\theta_i = \pi \mathrm{rot}(\tilde{g}_i) \in (0,\pi)$. The conjugacy class of rotations with angle $\theta_1$ can be identified with the open unit disc in ${{\mathbb C}}$, by writing $$g_1 = \begin{pmatrix} \cos(\theta_1) - \sin(\theta_1) \frac{2\, \mathrm{im}(w)}{1-|w|^2} &
-\sin(\theta_1) \frac{|1+w|^2}{1-|w|^2} \\
\sin(\theta_1) \frac{|1-w|^2}{1-|w|^2} &
\cos(\theta_1) + \sin(\theta_1) \frac{2\, \mathrm{im}(w)}{1-|w|^2} \end{pmatrix}, \;\; |w| < 1.$$ Conjugation with standard (Euclidean) rotations acts by rotating $w$. If $g_2$ is the standard rotation with angle $\theta_2$, we have $$\label{eq:g1g2-trace}
\mathrm{tr}(g_1 g_2) = 2 \cos(\theta_1)\cos(\theta_2) - 2 \sin(\theta_1)\sin(\theta_2) {\textstyle \frac{1+|w|^2}{1-|w|^2}}.$$ The values of are $(-\infty, 2\cos(\theta_1+\theta_2)]$, each being reached exactly once. The maximum is reached when $g_1$ is also a standard rotation ($w = 0$), so that $g_1g_2$ has angle $\theta_1+\theta_2>\pi$; from there, as the trace decreases, so does the angle. The case when $\mathrm{tr}(g_1g_2) = -2$ happens as a limit of rotations whose angle approaches $\pi$ from above. Since $g_1g_2$ can’t be $-{\mathbbm{1}}$, it must therefore be a nonnegative parabolic. We know (from the existence of suitable nonnegative paths) that $\mathrm{rot}(\tilde{g}_i) \leq \mathrm{rot}(\tilde{g}_1\tilde{g}_2) \leq \mathrm{rot}(\tilde{g}_i) + 1$. This, together with , implies the statements about rotation numbers made in .
Nonnegative paths and the rotation number quasimorphism generalize to linear symplectic groups (see [@lalonde-mcduff97; @slimowitz01] and [@barge-ghys92; @salamon-zehnder92; @bensimon-salamon10], respectively). The original definition of the rotation number as , due to Poincar[é]{}, applies to orientation-preserving homeomorphisms of the circle (for a basic exposition, see [@ghys01]; and for a proof of the quasimorphism property with the optimal bound , see [@zhuang08 Proposition 4.2]). Example \[th:three-classes\] will turn out to be a special case of a general result (Corollary \[th:teichmuller1b\]).
Krein theory
------------
Let’s consider the behaviour of nonnegative paths on the level of conjugacy classes. In the elliptic locus, motion can go in one direction only (Krein’s Lemma). Namely, for $\gamma \in {\mathfrak{g}}_{\geq 0}$ written as in , $$\label{eq:trace-decreases}
g = \big(\begin{smallmatrix} \cos(\theta) & -\!\sin(\theta)\! \\ \sin(\theta) & \cos(\theta) \end{smallmatrix} \big) \text{ with $\theta \in (0,\pi)$} \quad \Longrightarrow \quad
\mathrm{tr}(g \gamma) = -2\alpha \sin(\theta) \leq 0,$$ so the angle of rotation $\theta$ can’t decrease along a nonnegative path (and its derivative vanishes exactly when that of the path vanishes). There is no comparable constraint in the hyperbolic locus: $$\label{eq:trace-moves}
g = \big(\begin{smallmatrix} \lambda & 0 \\ 0 & \lambda^{-1}\end{smallmatrix}\big) \text{ with $\lambda>1$} \quad \Longrightarrow \quad
\mathrm{tr}(g \gamma) = (\lambda - \lambda^{-1}) \epsilon,$$ which can have either sign. The two parabolic conjugacy classes are each permeable in one direction (from elliptic to hyperbolic or vice versa) only. This is again elementary: $$g = \left(\begin{smallmatrix} 1 & 0 \\ \pm 1 & 1 \end{smallmatrix}\right) \quad \Longrightarrow \quad
\pm \mathrm{tr}(g \gamma) = \delta - \alpha \leq 0$$ (in this case, $\mathrm{tr}(g\gamma) = 0$ iff $\gamma = \left(\begin{smallmatrix} 0 & 0 \\ 2\alpha & 0 \end{smallmatrix} \right)$; hence, a nonnegative path intersects the parabolic locus either transversally, or tangentially to the one-parameter subgroup at the intersection point). Finally, if a nonnegative path goes through the identity in $G$, it must leave that element through either nonnegative parabolic elements, or elliptic elements which are small positive rotations (and conversely, it must arrive through nonpositive parabolic elements, or elliptic elements which are small negative rotations). These observations are summarized in Figure \[fig:traffic\].
(0,0)![Traffic flow in nonnegative direction on conjugacy classes (in reality, the space of conjugacy classes is one-dimensional and non-Hausdorff; we’ve drawn it thickened for the sake of legibility).\[fig:traffic\]](traffic.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(5602,2958)(-39,-1930) (3526,-1486)[(0,0)\[lb\]]{} (2626,164)[(0,0)\[lb\]]{} (976,-811)[(0,0)\[lb\]]{} (3601,-361)[(0,0)\[lb\]]{} (2626,-1861)[(0,0)\[lb\]]{}
The discussion above is taken from [@lalonde-mcduff97; @slimowitz01] (see more specifically [@slimowitz01 Lemma 2.2]), and we will continue to fundamentally follow those references, while making changes and adding material as required for our intended applications. Specifically, Lemma \[th:path\] below is a special case of [@lalonde-mcduff97 Proposition 2.4] (our proof fills in some technical details); and Lemmas \[th:short-path\]–\[th:short-path-2\] can be viewed as modified versions of [@slimowitz01 Theorem 3.3] (note however that [@slimowitz01 Theorem 3.6] is incorrect).
Our next topic is the converse question, of constructing nonnegative paths with prescribed motion on conjugacy classes. We will need a technical refinement of the computations from and . Writing $g = \left( \begin{smallmatrix} a & c \\ b & d \end{smallmatrix} \right)\in \mathit{SL}_2({{\mathbb R}})$, the condition $\mathrm{det}(g) = 1$ is equivalent to $$\label{eq:det-1}
(a-d)^2 + (c+b)^2 = \mathrm{tr}(g)^2 - 4 + (c-b)^2.$$ As a consequence, if we consider a subset of elements with fixed trace, then $T(g) = c-b$ is a proper function on that subset. Note also that $T(g^2) = T(g)\mathrm{tr}(g)$.
\[th:bound-vector-fields\] (i) Let $C_{\mathit{ell},\theta} \subset \mathit{SL}_2({{\mathbb R}})$ be the conjugacy class of rotations with angle $\theta \in (0,\pi)$. There is a function $$\label{eq:elliptic-gamma-function}
\left\{
\begin{aligned}
& \gamma_{\theta}: C_{\mathit{ell},\theta} \longrightarrow {\mathfrak{g}}_{>0}, \\
& |T(g\gamma_{\theta}(g))| \leq A +B|T(g)|, \text{ for some constants $A,B>0$;} \\
& \mathrm{tr}(g \,\gamma_\theta(g)) \text{ is negative and bounded away from $0$.} \\
\end{aligned}
\right.$$
\(ii) Let $C_{\mathit{hyp},\lambda} \subset \mathit{SL}_2({{\mathbb R}})$ be the conjugacy class of hyperbolic elements with eigenvalues $\lambda^{\pm 1}$, for some $\lambda>1$. There are functions $$\left\{
\begin{aligned}
& \gamma_{\lambda,\pm}: C_{\mathit{hyp},\lambda} \longrightarrow {\mathfrak{g}}_{>0}, \\
& |T(g\gamma_{\lambda,\pm}(g))| \leq A +B|T(g)|, \text{ for some constants $A,B>0$;} \\
& \pm\mathrm{tr}(g \,\gamma_{\lambda,\pm}(g)) \text{ is positive and bounded away from $0$.} \\
\end{aligned}
\right.$$
\(i) Take $$\gamma_{\theta}(g) = \frac{g - \cos(\theta) {\mathbbm{1}}}{\sin(\theta)},$$ which is conjugate to $\big(\begin{smallmatrix} 0 & - 1 \\ 1 & 0 \end{smallmatrix} \big)$, hence lies in ${\mathfrak{g}}_{>0}$. Then $$\begin{aligned}
& \mathrm{tr}(g\gamma_{\theta}(g)) = -2\sin(\theta), \\
& T(g \gamma_{\theta}(g)) = \frac{\cos(\theta)}{\sin(\theta)} T(g).\end{aligned}$$
\(ii) If $r = r(g)$ is the quantity on either side of , $$\label{eq:test-gamma}
\gamma_{\lambda,\pm}(g) = \sqrt{r} \big(\begin{smallmatrix} 0 & -1 \\ 1 & 0 \end{smallmatrix}\big) \pm
\big(\begin{smallmatrix}
a-d
&
b+c
\\
b+c
&
-(a-d)
\end{smallmatrix}\big) \in {\mathfrak{g}}_{\geq 0}$$ satisfies $$\begin{aligned}
& \mathrm{tr}(g\gamma_{\lambda, \pm}(g)) = (c-b)\sqrt{r}\, \pm\, r
= \textstyle r \Big( \mathrm{sign}(c-b) \sqrt{1 - \frac{\mathrm{tr}(g)^2 - 4}{r}} \pm 1 \Big)
\\ \notag
& \quad \quad \Longrightarrow \pm \mathrm{tr}(g\gamma_{\lambda,\pm}(g)) \geq {{\textstyle\frac{1}{2}}}(\mathrm{tr}(g)^2-4), \\
& T(g \gamma_{\lambda,\pm}(g)) = - \mathrm{tr}(g) \sqrt{r} = - \mathrm{tr}(g) \sqrt{\mathrm{tr}(g)^2 - 4 + T(g)^2}. \end{aligned}$$ Our choice always lies on the boundary of the nonnegative cone. However, one can perturb it into the interior ${\mathfrak{g}}_{>0}$, and this perturbation can be chosen sufficiently small so as not to disturb its other properties.
\[th:weak-homotopy\] When we talk about weak homotopy equivalences for spaces of paths in $G$ (and, later on, for spaces of connections as well), this is understood to refer to smooth maps from finite-dimensional compact manifolds (with boundary or corners) to the relevant space. For instance, saying that some space ${\EuScript P}$ as in is weakly contractible means that for every smooth map $P \rightarrow {\EuScript P}$ from a compact manifold $P$, there is a smooth extension $P \times [0,1] \rightarrow {\EuScript P}$ which is constant on the other endpoint of the interval. This technical clarification may not be essential, but it slightly simplifies some proofs, and is anyway the natural context for our applications.
\[th:path\] (i) Fix a nondecreasing function $\theta(t) \in (0,\pi)$ ($0 \leq t \leq 1$), and a $g_0 \in C_{\mathit{ell},\theta(0)}$. Then, the space of nonnegative paths $g(t)$, with $g(0) = g_0$ and such that $g(t) \in C_{\mathit{ell},\theta(t)}$, is weakly contractible. A corresponding result holds for (strictly) increasing functions and positive paths.
\(ii) Fix a function $\lambda(t) > 1$ ($0 \leq t \leq 1$), and a $g_0 \in C_{\mathit{hyp},\lambda(0)}$. Then, the space of nonnegative paths $g(t)$, with $g(0) = g_0$ and such that $g(t) \in C_{\mathit{hyp},\lambda(t)}$, is weakly contractible. The same holds for positive paths.
\(i) Let’s suppose that we have a smooth family of functions $$\label{eq:elliptic-gamma}
\left\{
\begin{aligned}
& \; \gamma_t: C_{\mathit{ell},\theta(t)} \longrightarrow {\mathfrak{g}}_{\geq 0}, \\
& \; |T(g\gamma_t(g))| \leq A + B|T(g)|, \text{ for some constants $A,B>0$;}\\
& \; \mathrm{tr}(g \,\gamma_t(g)) = -2\sin(\theta(t))\theta'(t).
\end{aligned}
\right.$$ The boundedness condition ensures that the ODE $$\label{eq:ode-elliptic}
g'(t) = g(t)\gamma_t(g(t))$$ can be integrated for any initial value $g(0) \in C_{\mathit{ell},\theta(0)}$. By construction, each solution is a nonnegative path with $g(t) \in C_{\mathit{ell},\theta(t)}$.
Functions can be constructed from those in Lemma \[th:bound-vector-fields\](i) by multiplying with the bounded nonnegative function $$\label{eq:bounded-fn}
g \longmapsto -2\sin(\theta(t))\theta'(t) \mathrm{tr}(g \gamma_{\theta(t)}(g))^{-1}.$$ On the other hand, the conditions in are convex, allowing us to interpolate between two such functions, and to construct them using partitions of unity. In particular, given a nonnegative path $g(t)$ with the desired behaviour, one can choose $\gamma_t$ so that this path is a solution of . More precisely, this kind of argument shows that the space of paths is weakly homotopy equivalent to the space of all , which implies the desired result. In the positive version of the same result, one takes $\gamma_t$ which have values in ${\mathfrak{g}}_{>0}$, and those exist because is strictly positive.
The proof of (ii) is parallel, using the corresponding part of Lemma \[th:bound-vector-fields\].
Spaces of nonnegative paths
---------------------------
Using Lemma \[th:path\] as our basic ingredient, we will now discuss the topology of spaces of nonnegative paths with constraints on the endpoints. The general notation will be as follows: if $U_0,U_1$ are subsets of $G$, $$\label{eq:path-spaces}
\begin{aligned}
& {\EuScript P}(G,U_0,U_1) = \{ g: [0,1] \rightarrow G \;:\; g(0) \in U_0, \, g(1) \in U_1\}, \\
& {\EuScript P}_{\geq 0}(G,U_0,U_1) = \text{subspace of nonnegative paths}, \\
& {\EuScript P}_{>0}(G,U_0,U_1) = \text{subspace of positive paths.}
\end{aligned}$$ One can introduce similar notation ${\EuScript P}_{\leq 0}(G,U_0,U_1)$, ${\EuScript P}_{<0}(G,U_0,U_1)$ for nonpositive and negative paths. We will use superscripts ${\EuScript P}(G,U_0,U_1)^r$ to restrict to paths along which the rotation number increases by exactly $r$; and similarly for inequalities, such as ${\EuScript P}(G,U_0,U_1)^{<r}$.
\[th:pre-short\] Fix hyperbolic conjugacy classes $C_0$ and $C_1$. Then, evaluation at either endpoint yields a weak homotopy equivalence $${\EuScript P}_{\geq 0}(G,C_0,C_1)^0 \stackrel{{\simeq}}{\longrightarrow} C_i.$$
By Figure \[fig:traffic\], any such path remains within the hyperbolic locus, hence falls into the class considered in Lemma \[th:path\](ii). The rest is straightforward.
\[th:pp\] Fix hyperbolic conjugacy classes $C_i$ ($i = 0,\dots,m$, for some $m \geq 2$). Consider $g_i \in C_i$ together with a nonnegative path $h(t)$ from $h(0) = g_1\cdots g_m$ to $h(1) = g_0$, such that if we take lifts of our elements to $\tilde{g}_i \in \tilde{G}$, compatible with the product and path, then $$\label{eq:lose-one}
\mathrm{rot}(\tilde{g}_0) = \mathrm{rot}(\tilde{g}_1) + \cdots + \mathrm{rot}(\tilde{g}_m) + 1-m.$$ Any $g_i$ yields a weak homotopy equivalence between the space of all $(g_0,\dots,g_m,h(t))$ and $C_i$.
Because of and the fact that rotation numbers don’t decrease along nonnegative paths, requires that, for any $i \geq 1$, $$\label{eq:productminus1}
\mathrm{rot}(\tilde{g}_1 \cdots\tilde{g}_i) = \mathrm{rot}(\tilde{g}_1\cdots \tilde{g}_{i-1}) + \mathrm{rot}(\tilde{g}_i) - 1.$$ Suppose that there is a smallest $i \geq 2$ for which the product $g_1 \cdots g_i$ is not hyperbolic. By and Lemma \[th:sharpened-inequality\], it would then have to be a nonnegative parabolic. If $i<m$, one would then apply Lemma \[th:sharpened-inequality\] again, to get $$\mathrm{rot}(\tilde{g}_1 \cdots \tilde{g}_{i+1}) \geq \mathrm{rot}(\tilde{g}_1 \cdots \tilde{g}_i) + \mathrm{rot}(\tilde{g}_{i+1}),$$ which is a contradiction to for $i+1$. If $i = m$, the path $h$ would have to go from a nonnegative parabolic to a hyperbolic element without raising the rotation number, which is impossible. We have therefore shown that all products $g_1\cdots g_i$ are hyperbolic.
Consider the space of all $(g_1,\dots,g_m)$ such that $g_i \in C_i$, the product $g = g_1\cdots g_m$ is a fixed hyperbolic element, and $\mathrm{rot}(\tilde{g}_1\cdots \tilde{g}_m) = \mathrm{rot}(\tilde{g}_1) + \cdots + \mathrm{rot}(\tilde{g}_m) + 1-m$. Then, that space is contractible. One can derive this from Example \[th:three-classes\] by induction on $m$, and the hyperbolicity of the intermediate products. We omit the details, since there is a more conceptual alternative, namely to think of it as a special case of Corollary \[th:teichmuller1b\] below. On the other hand, the space of nonnegative paths $h(t)$ from our fixed hyperbolic $g$ to $C_0$, along which the rotation number remains constant, is weakly contractible by Lemma \[th:pre-short\]. By taking the product of those spaces, and then letting $g$ vary, one sees that evaluation at $h(0) = g = g_1 \cdots g_m$ gives a weak homotopy equivalence between the space of all $(g_0,g_1,\dots,g_m,h(t))$ and $G_{\mathit{hyp}}$. Since evaluation at any $h(t)$ lands in the hyperbolic locus, we can evaluate at $h(1) = g_0$ instead. By applying simultaneous conjugation, one extends the result to the other evaluation maps.
One can interpret Lemma \[th:pre-short\] as follows. Let ${\EuScript P}(G_{\mathit{hyp}}, C_0, C_1)$ be the space of all paths from $C_0$ to $C_1$, but which remain inside the hyperbolic locus. ${\EuScript P}_{\geq 0}(G, C_0,C_1)^0$ is a subspace of ${\EuScript P}(G_{\mathit{hyp}},C_0,C_1)$, and we have shown that the inclusion yields a weak homotopy equivalence $${\EuScript P}_{\geq 0}(G,C_0,C_1)^0 \stackrel{{\simeq}}{\longrightarrow} {\EuScript P}(G_{\mathit{hyp}},C_0,C_1).$$ Alternatively, let’s think of the case $C_0 = C_1 = C$. In that case, inclusion of the constant paths yields a weak homotopy equivalence $$C \stackrel{{\simeq}}{\longrightarrow} {\EuScript P}_{\geq 0}(G,C,C)^0.$$ This second kind of interpretation also applies to Lemma \[th:pp\], which turned out to be weakly homotopy equivalent to the subspace for which the path $h(t)$ is constant. From now on, when deriving results about path spaces, we will often only give the principal steps of the proof, omitting “$\epsilon$–$\delta$ level” details (readers interested in seeing such kind of arguments in full may want to look at [@lalonde-mcduff97]).
\[th:short-path\] Fix an elliptic conjugacy class $C_0$, and a hyperbolic conjugacy class $C_1$. Then, evaluation at the hyperbolic endpoint yields a weak homotopy equivalence $${\EuScript P}_{\geq 0}(G,C_0,C_1)^{<1} \stackrel{{\simeq}}{\longrightarrow} C_1.$$ The same applies to ${\EuScript P}_{\geq 0}(G,C_1,C_0)^{<1}$, still taking the evaluation map at the hyperbolic endpoint.
We begin with a slightly different problem. Let’s consider positive paths which start in $C_0$, end in $G_{\mathit{hyp}}$, and along which the rotation number increases by less than $1$. Such a path necessarily has the following structure: $$\text{for some $T \in (0,1)$,} \; \left\{
\begin{aligned}
& \text{$g(t)$ is elliptic for $t < T$;} \\
& \text{$g(T)$ is a nonpositive parabolic;} \\
& \text{$g(t)$ is hyperbolic for $T < t$.}
\end{aligned}
\right.$$ If one fixes $T$ and $g(T)$, then by analyzing $g|[0,T-\epsilon]$ and $g|[T+\epsilon,1]$ for small $\epsilon>0$, using Lemma \[th:path\], one sees that the resulting space of paths is weakly contractible. Allowing $T$ and $g(T)$ to vary yields a path space that is weakly homotopy equivalent to the nonpositive parabolic stratum, hence also to $G$. Instead of using evaluation at $t = T$, one can use the (homotopic) evaluation at $t = 1$.
We can weaken strict positivity to nonnegativity, without changing the weak homotopy type of the space, since any nonnegative path can be perturbed to a positive one (see and the subsequent discussion); for this to work, it’s important that the endpoint condition $g(1) \in G_{\mathit{hyp}}$ is an open one. Next, again without changing the weak homotopy type, we can require that the hyperbolic endpoint should lie in a fixed conjugacy class, thanks to Lemma \[th:path\](ii). This brings us to the desired situation. The other direction (from hyperbolic to elliptic) is proved in the same way.
\[th:short-path-3\] Fix a hyperbolic conjugacy class $C$. Evaluation at the endpoint yields a weak homotopy equivalence $${\EuScript P}_{\geq 0}(G,\mathit{id},C)^1 \stackrel{{\simeq}}{\longrightarrow} C.$$
Consider first positive paths which start at the identity, end in $G_{\mathit{hyp}}$, and along which the rotation number increases by $1$. Such a path has the property that $g(t)$ is elliptic for all sufficiently small $t>0$. By looking at $g|[\epsilon,1]$ for small $\epsilon>0$, one reduces their study to Lemma \[th:short-path\]. The translation back to the original situation also follows that model.
\[th:short-path-2\] Fix two hyperbolic conjugacy classes $C_0$ and $C_1$. Then, evaluation at both endpoints yields a weak homotopy equivalence $${\EuScript P}_{\geq 0}(G,C_0,C_1)^1 \stackrel{{\simeq}}{\longrightarrow} C_0 \times C_1.$$
Once more, we start with a slightly different problem. Namely, consider positive paths, such that both endpoints lie $G_{\mathit{hyp}}$, and along which the rotation number increases by $1$. There is exactly one point $T$ where such a path intersects a fixed elliptic conjugacy class. By breaking up the path into pieces $g|[0,T]$ and $g|[T,1]$, one can think of the whole path space as a fibre product of two spaces of the kind considered in Lemma \[th:short-path\]. Hence, what we are looking at is the fibre product of two hyperbolic conjugacy classes, formed over a contractible space (the elliptic conjugacy class). As in Lemma \[th:short-path\], positivity can be weakened to nonnegativity, and one can use Lemma \[th:path\](ii) to replace the condition on the endpoints by one of lying in specified hyperbolic conjugacy classes.
One can interpret these results as follows: the inclusion of nonnegative paths into all paths yields weak homotopy equivalences $$\label{eq:into-top}
\begin{aligned}
& \left.\begin{aligned}
& {\EuScript P}_{\geq 0}(G,C_0,C_1)^{<1} \stackrel{\simeq}\longrightarrow {\EuScript P}(G,C_0,C_1)^{<1} \\
& {\EuScript P}_{\geq 0}(G,C_1,C_0)^{<1} \stackrel{\simeq}\longrightarrow {\EuScript P}(G,C_1,C_0)^{<1}
\end{aligned}\right\}
&& \text{for Lemma \ref{th:short-path},}
\\
& \left.{\EuScript P}_{\geq 0}(G,\mathit{id},C)^1 \stackrel{\simeq}\longrightarrow {\EuScript P}(G,\mathit{id},C)^1\right. && \text{for Lemma \ref{th:short-path-3},} \\
& \left.{\EuScript P}_{\geq 0}(G,C_0,C_1)^1 \stackrel{{\simeq}}{\longrightarrow} {\EuScript P}(G,C_0,C_1)^1\right. && \text{for Lemma \ref{th:short-path-2}.}
\end{aligned}$$ We will see a few more instances of such behaviour in Section \[sec:elliptic\]. Note that this contrasts with Lemmas \[th:pre-short\] and \[th:pp\], where including nonnegative paths into all paths (with the same rotation behaviour) would not have yielded a homotopy equivalence.
One can conjecture that statements similar to should hold more generally for spaces of nonnegative paths which are “sufficiently long” (meaning that the rotation number increases by a sufficient amount to force transitions between elliptic and hyperbolic elements). This idea is in tune with the results of [@lalonde-mcduff97; @slimowitz01].
The next result can be viewed as a weak version of Lemma \[th:short-path-3\] under extra constraints. Instead of determining the topology of the path space under consideration, we will argue by an explicit construction.
\[th:unit-path\] Fix two hyperbolic conjugacy classes $C_1$ and $C_2$. Then there is a nonnegative path $g_1(t)$, starting at the identity and ending in $C_1$, along which the rotation number increases by $1$; together with another path $g_2(t) \in C_2$, such that at all times, $g_1(t)g_2(t) \in C_2$. Moreover, one can write $$\label{eq:conjugating-path}
g_1(t)g_2(t) = k(t) g_2(t) k(t)^{-1}$$ for some path $k(t)$ starting at the identity, whose endpoint $k(1)$ is hyperbolic and (if one lifts the path to $\tilde{G}$ in the obvious way) has rotation number $0$.
Let’s set $g_2 = \mathrm{diag}(\lambda,\lambda^{-1})$ for some $\lambda>1$. Consider $$\label{eq:trace-h}
H = \{g \in \mathit{SL}_2({{\mathbb R}})\;:\; \mathrm{tr} (g g_2) = \lambda+\lambda^{-1} \} \stackrel{\mathrm{tr}}{\longrightarrow} {{\mathbb R}}.$$ $H$ is a connected hyperboloid in ${{\mathbb R}}^3$, and the trace is a Morse function on it, with index $1$ critical points at the identity and $g_2^{-2}$, and corresponding critical levels $2$ and $\lambda^2 +\lambda^{-2} > 2$. The open sublevel set $U = \{\mathrm{tr}(g)<2\} \subset S$ consists of two connected components $U_{\pm}$. If we write $g = \left(\begin{smallmatrix} a & c \\ b & d \end{smallmatrix} \right)$, those components are $$U_{\pm} = \{ g \in H \;:\; {\pm } b > 0, \text{ or equivalently } {\pm } c < 0 \}.$$ We construct a path in $H$ of the following kind. Starting at the identity, the path first moves inside the level set $\{\mathrm{tr}(g) = 2\}$ to $\left( \begin{smallmatrix} 1 & 0 \\ 1 & 1 \end{smallmatrix} \right)$. From there, we strictly decrease the trace, moving into $U_+$, and then further inside that subset until $\{\mathrm{tr}(g) = -2\}$. At that point, the resulting matrix is necessarily conjugate to $\left(\begin{smallmatrix} -1 & 0 \\ 1 & -1 \end{smallmatrix} \right)$, because it still belongs to $U_+$. We continue our path downwards into hyperbolic level sets $\{\mathrm{tr}(g)<-2\}$, until we reach the value of the trace (up to sign) prescribed by our choice of $C_1$. One sees easily that the rotation number increases by $1$ along this path.
Denote the path we have just constructed by $g_1(t)$, and take $g_2(t) = g_2$ to be the constant path. Since $g_1(t)g_2$ remains in the conjugacy class of $g_2$, one can write $g_1(t)g_2 = k(t) g_2 k(t)^{-1}$, for some smooth $k(t) = \left( \begin{smallmatrix} p(t) & r(t) \\ q(t) & s(t) \end{smallmatrix} \right)$ with $k(0) = {\mathbbm{1}}$. Then, $$\label{eq:g2-trace}
\begin{aligned}
\mathrm{tr}(g_1(t)) = \mathrm{tr}( k(t) g_2 k(t)^{-1} g_2^{-1}) & =
2p(t)s(t) - (\lambda + \lambda^{-2}) q(t)r(t)
\\ & =
(1-p(t)s(t)) (\lambda - \lambda^{-1})^2 + 2.
\end{aligned}$$ By construction $\mathrm{tr}(g_1(1)) < -2$, which with yields $p(1) s(1) > 1$, hence $$\mathrm{tr}(k(1))^2 = (p(1)+s(1))^2 = (p(1)-s(1))^2 + 4 p(1)s(1)> 4.$$ This shows that $k(1)$ must be hyperbolic (one could derive the same result more elegantly from Theorem \[th:teichmuller\], applied to a punctured torus). Now, bear in mind that $\mathrm{tr}(g_1(t)) \leq 2$ for all $t$. From this and , it follows that $p(t) s(t) > 0$ for all $t$. Hence, both entries $p(t)$ and $s(t)$ remain positive throughout. By tracking how points on the circle move under the action of $k(t)$, one easily concludes that the rotation number of $k(1)$ must vanish.
At this point, it seems that we have been missing an essential condition, namely, that $g_1(t)$ should be a nonnegative path. However, an inspection of each step of the construction (and the use of Lemma \[th:path\]) shows that $g_1(t)$ is conjugate to a nonnegative path, by a time-dependent conjugation. One can apply the same conjugation to $g_2(t)$ and $k(t)$, and that yields the desired result.
It would be interesting to consider similar questions for nonnegative paths in the larger group $\mathit{Diff}^+(S^1)$ (this idea was already mooted by Abouzaid and Ganatra). Enlarging the group affords one greater freedom, which may simplify the construction of homotopies. On a more fundamental level, the larger group contains many more conjugacy classes: for instance, the pullbacks of hyperbolic elements by finite covers $S^1 \rightarrow S^1$, which would be relevant to Fukaya categories with multiple fibres as “stops” [@sylvan16].
$\mathfrak{sl}_2({{\mathbb R}})$-connections on surfaces\[sec:connections\]
===========================================================================
For our TQFT, we will need to consider connections with nonnegative curvature. There is an obvious relation between such connections and nonnegative paths, and (with one exception, Proposition \[th:bochner\]) all our results will be derived from the elementary considerations in Section \[sec:sl2\] in that way. On the other hand, if we restrict attention to flat connections, there is a large body of relevant literature; [@goldman88] will be particularly important for us.
Basic notions\[subsec:connections\]
-----------------------------------
Let $S$ be a connected compact oriented surface with nonempty boundary. We consider ${\mathfrak{g}}$-connections on the trivial bundle over $S$, meaning operators $\nabla_A = d-A$ for some $A \in \Omega^1(S,{\mathfrak{g}})$ (we will often refer to $A$ as the connection). Gauge transformations are maps $\Phi: S \rightarrow G$, and act on connections (covariantly) by $$\Phi_*A = \Phi A \Phi^{-1} + (d\Phi) \Phi^{-1}.$$ With our sign convention for $A$, the curvature is $$\label{eq:curvature}
F_A = -dA + A \wedge A \in \Omega^2(S,{\mathfrak{g}}).$$ A connection is said to have nonnegative curvature if it satisfies the (gauge invariant) condition $$\label{eq:nonneg-curvature}
F_A(\xi_1,\xi_2) \in {\mathfrak{g}}_{\geq 0} \quad \text{for any oriented basis $(\xi_1,\xi_2) \in TS$.}$$ Obviously, there is a parallel notion of nonpositive curvature; one can switch from one to the other by changing the orientation of $S$, or by applying a gauge transformation in $\mathit{PGL}_2^-({{\mathbb R}})$. Similarly, one could consider connections with (strictly) positive or negative curvature; but we have no use for those notions.
Given a connection, parallel transport along some $c: [0,1] \rightarrow S$ yields a path $g:[0,1] \rightarrow G$, $$\label{eq:parallel-transport}
g(0) = {\mathbbm{1}}, \quad g'(t)g(t)^{-1} = A_{c(t)}(c'(t)).$$ (If $c$ is a closed loop, $g(1)$ is the holonomy of the connection around $c$.) That path has an obvious lift to $\tilde{G}$, with $\tilde{g}(0)$ the identity; from that, we get a rotation number $$\mathrm{rot}_c(A) \stackrel{\mathrm{def}}{=} \mathrm{rot}(\tilde{g}(1)) \in {{\mathbb R}}.$$ Since $G$ is homotopy equivalent to a circle, the homotopy class of a gauge transformation can be written as $[\Phi] \in H^1(S)$. The effect of gauge transformations on rotation numbers along closed loops $c$ is $$\label{eq:gauge-c}
\mathrm{rot}_{c}(\Phi_*A) = \mathrm{rot}_{c}(A) + \int_{c} [\Phi].$$ Let’s denote the boundary circles by $\partial_i S$, where $i \in I = \pi_0(\partial S)$. By , the following is a gauge invariant quantity: $$\label{eq:boundary-rotation}
\mathrm{rot}_{\partial S}(A) = \sum_i \mathrm{rot}_{\partial_i S}(A).$$
\[th:bochner\] Let $A$ be a connection on $S$, such that the holonomies around the boundary circles are hyperbolic. If the connection is nonnegatively curved, $$\label{eq:bochner}
\mathrm{rot}_{\partial S}(A) \leq -\chi(S).$$
We consider a slightly different but equivalent situation. Namely, let $S$ be a Riemann surface with tubular ends, carrying a connection $A$ which, in coordinates $(s,t) \in [0,\infty) \times S^1$ on each end, is of the form $a_{i,t} \mathit{dt}$, and has hyperbolic holonomy around the $S^1$ factor. One can define an analogue of , which we denote by $\mathrm{rot}_{\mathit{ends}}(A) \in {{\mathbb Z}}$. Consider the ${{\mathbb R}}$-linear Dolbeault operator associated to the connection, $$\label{eq:dbar-e}
\bar\partial_A = \nabla^{0,1}_A: {C^\infty}_{\mathit{cpt}}(S,{{\mathbb C}}) \longrightarrow \Omega_{\mathit{cpt}}^{0,1}(S).$$ As stated, this applies to compactly supported smooth functions, but in fact one wants to pass to suitable Sobolev completions, let’s say from $W^{1,2}$ to $L^2$. The index theorem (see e.g. [@schwarz95 Section 3.3]) says that the completed operator is Fredholm, with $$\mathrm{index}(\bar\partial_A) = \chi(S) + \mathrm{rot}_{\mathit{ends}}(A).$$ The Bochner identity, which applies to any $\xi \in W^{1,2}(S,{{\mathbb C}})$, is $$\int_S |\bar\partial_A \xi|^2 = \int_S |\nabla \xi|^2 + \int_S \mathrm{det}(\xi, F_A \xi).$$ If we assume that our connection has nonnegative curvature, it follows that any $W^{1,2}$ solution of $\bar\partial_A \xi = 0$ must be zero, hence the index is necessarily $\leq 0$.
By reversing orientation, we get the following:
\[th:mw\] If $A$ is a flat connection with hyperbolic boundary holonomies, $$|\mathrm{rot}_{\partial S}(A)| \leq -\chi(S).$$
Corollary \[th:mw\], proved in [@goldman88 Theorem 3.3], is a version of the Milnor-Wood inequality [@milnor58; @wood71]. The Milnor-Wood inequality has numerous generalizations (to other Lie groups, and to homeomorphisms of the circle [@wood71]). The proof used here applies to linear symplectic groups. One can consider results such as those in [@entov01; @lalonde-teleman14], proved using pseudo-holomorphic sections of symplectic fibrations over surfaces, as nonlinear analogues.
We will be interested in spaces of connections with fixed boundary behaviour. Concretely, given $S$, choose $a_i \in \Omega^1(\partial_i S,{\mathfrak{g}})$ for each $i \in I$, with hyperbolic holonomy. We will use the following associated notation: $$\label{eq:a-space-notation}
\begin{aligned}
& g_i && \text{holonomy of $a_i$}, \\
& C_i && \text{conjugacy class of $g_i$ in $G$}, \\
& \tilde{g}_i && \text{preferred lift of $g_i$ to $\tilde{G}$}, \\
& \tilde{C}_i && \text{conjugacy class of $\tilde{g}_i$ in $\tilde{G}$.}
\end{aligned}$$ The relevant spaces of connections are $$\label{eq:moduli-spaces-2}
\begin{aligned}
{\EuScript A}(S,\{a_i\}) & = \{ A \in \Omega^1(S,{\mathfrak{g}})\;:\; A|\partial_iS = a_i\}, \\
{\EuScript A}_{\geq 0}(S,\{a_i\}) & \;\; \text{subspace of nonnegatively curved connections}, \\
{\EuScript A}_{\mathit{flat}}(S,\{a_i\}) & \;\; \text{subspace of flat connections.}
\end{aligned}$$ These spaces are acted on by the group $$\label{eq:gauge-2}
{\EuScript G}(S,\{a_i\}) = \{\Phi: S \rightarrow G \;:\; (\Phi|\partial_iS)_* a_i = a_i\} {\simeq}H^1(S,\partial S).$$ The weak homotopy equivalence in comes from the fact that the stabilizer of each $a_i$ in ${C^\infty}(\partial_i S,G)$ is a copy of ${{\mathbb R}}$, hence contractible. In the case of flat connections, the quotient $$\label{eq:moduli-1}
{\EuScript M}_{\mathit{flat}}(S,\{a_i\}) = {\EuScript A}_{\mathit{flat}}(S,\{a_i\})/{\EuScript G}(S,\{a_i\})$$ has a well-known topological interpretation: $$\label{eq:rep-spaces}
{\EuScript M}_{\mathit{flat}}(S,\{a_i\}) {\cong}\big\{ \sigma: \pi_1(S)^{\mathit{op}} \rightarrow G \;:\; \sigma(\partial_iS) \in C_i, \text{ and } e_{\mathit{rel}}(\sigma) = \textstyle\sum_i \mathrm{rot}(a_i) \big\}/G.
$$ Here, $e_{\mathit{rel}}(\sigma)$ is the relative Euler number of the associated ${{\mathbb R}}P^1$-bundle, with its section over $\partial S$ given by an eigenvector for the holonomy; and $G$ acts by overall conjugation.
In the discussion above, for instance in , we have always used $\partial_iS$ with its boundary orientation. When looking at specific examples of surfaces, we will sometimes adopt different orientation conventions on some boundary components (and will warn the reader when that happens).
Spaces of flat connections
--------------------------
For our purpose, what matters is the weak homotopy type of the spaces (in the same sense as in Convention \[th:weak-homotopy\]). Flat connections are easier to understand in that respect, because of .
Let’s begin with a case in which the topological aspect is trivial, namely that of the annulus $S = [0,1] \times S^1$. Our convention is that we identify each boundary circle $\partial_i S = \{i\} \times S^1$ with $S^1$ in the obvious way (which is orientation-reversing for $i = 0$, and orientation-preserving for $i = 1$). In order for ${\EuScript A}_{\mathit{flat}}(S,a_0,a_1)$ to be nonempty, the following condition must be satisfied: $$\label{eq:conjugate-holonomies}
\parbox{38em}{$a_0, a_1$ have conjugate holonomies ($C_0 = C_1$), and the same rotation number $r = \mathrm{rot}(a_0) = \mathrm{rot}(a_1)$. Equivalently, the conjugacy classes in $\tilde{G}$ are the same ($\tilde{C}_0 = \tilde{C}_1$).}$$
\[th:z-homotopy\] Take $S = [0,1] \times S^1$, with boundary conditions . Then there is a weak homotopy equivalence $$\label{eq:z-homotopy}
{\EuScript A}_{\mathit{flat}}(S,a_0,a_1) {\simeq}{{\mathbb Z}},$$ unique up to adding a constant, and which is compatible with the action of ${\EuScript G}(S,a_0,a_1) {\simeq}{{\mathbb Z}}$. To state this more concretely, let $C = C_0 = C_1$. By associating to a connection $A$ its family of holonomies around the loops $\{s\} \times S^1$, one defines a canonical weak homotopy equivalence $$\label{eq:z-homotopy-2}
{\EuScript A}_{\mathit{flat}}(S,a_0,a_1) \longrightarrow {\EuScript P}(C,g_0,g_1).$$
Clearly, ${\EuScript G}(S,a_0,a_1)$ acts transitively on ${\EuScript A}_{\mathit{flat}}(S,a_0,a_1)$, with stabilizer ${{\mathbb R}}$ at every point. This implies that each orbit gives a weak homotopy equivalence $$\label{eq:gauge-orbit}
{\EuScript G}(S,a_0,a_1) \stackrel{{\simeq}}{\longrightarrow} {\EuScript A}_{\mathit{flat}}(S,a_0,a_1).$$ We know from that ${\EuScript G}(S,a_0,a_1) {\simeq}{{\mathbb Z}}$. The map is equivariant with respect to the action of $\Omega G$ (thought of as gauge transformations in the $[0,1]$ variable on the domain, and as acting by pointwise conjugation on the target space). From that and , it follows that is a weak homotopy equivalence.
\[th:z-homotopy-2\] Suppose that $g_0 = g_1 = g$. Take a path $$\label{eq:stupid-path}
c: [0,1] \longrightarrow S, \quad c(0) = (0,0), \;\; c(1) = (1,0).$$ For any $A \in {\EuScript A}_{\mathit{flat}}(S,a_0,a_1)$, parallel transport along $c$ yields an element of $G$ which commutes with $g$, hence is either trivial or hyperbolic. The rotation number $\mathrm{rot}_c(A) \in {{\mathbb Z}}$ then defines an isomorphism $\pi_0({\EuScript A}_{\mathit{flat}}(S,a_0,a_1)) {\cong}{{\mathbb Z}}$. One can take the path to have winding number zero around the $S^1$ factor, and that gives a preferred choice of , which is the same one gets from identifying $\pi_0({\EuScript P}(C,g,g)) {\cong}{{\mathbb Z}}$ in .
\[th:z-homotopy-3\] In the situation of Proposition \[th:z-homotopy\], let $\tau \in \mathit{Diff}(S,\partial S)$ be the (positive, or right-handed) Dehn twist along $\{{{\textstyle\frac{1}{2}}}\} \times S^1$. The action of $\tau^*$ on ${\EuScript A}_{\mathit{flat}}(S,a_0,a_1)$ corresponds to subtracting $r$ on the right hand side of .
We omit the proofs, which are straightforward. The interesting point about Addendum \[th:z-homotopy-3\] is that, even though the topology of the space of flat connections is always the same, the action of the diffeomorphism group depends on the rotation number.
Let’s consider more general surfaces $S$. A classical result concerns the case of extremal rotation numbers, by which we mean $$\label{eq:rot-chi}
\sum_i \mathrm{rot}(a_i) = \pm \chi(S).$$
\[th:teichmuller\] Suppose that $\chi(S)<0$, and that the $a_i$ satisfy . Then, the representation of $\pi_1(S)$ associated to any $A \in {\EuScript M}_{\mathit{flat}}(S,\{a_i\})$ is (conjugate to) the holonomy representation of a hyperbolic metric on $S$ with geodesic boundaries of length $l_i = \cosh^{-1}\big( {{\textstyle\frac{1}{2}}}|\mathrm{tr}(g_i)|\big)$.
The holonomy representation assumes that we have represented our hyperbolic surface $S$ as a quotient of the hyperbolic disc. The sign in depends on whether this representation agrees with the given orientation of $S$ or not. Here are two useful implications:
\[th:all-hyperbolic\] For a flat connection as in Theorem \[th:teichmuller\], the holonomies along non-contractible closed curves is hyperbolic.
\[th:teichmuller1b\] In the situation of Theorem \[th:teichmuller\], ${\EuScript M}_{\mathit{flat}}(S,\{a_i\}) {\cong}{{\mathbb R}}^{-3\chi(S)-|\pi_0(\partial S)|}$.
Theorem \[th:teichmuller\] is a version of [@goldman88 Theorem 3.4]. The results from Teichmüller theory needed for Corollary \[th:teichmuller1b\] can be proved using Fenchel-Nielsen coordinates, see e.g. [@abikoff80] (or for another proof, [@luo07]). Even though we will not need that, it is worth while pointing out that Higgs bundle methods can be used to determine the topology of other components of the representation space (as long as the Euler number is nonzero). That idea was pioneered by Hitchin [@hitchin87] for closed surfaces, and has been generalized to surfaces with boundary by Mondello [@mondello16].
For us, what’s important is the following immediate consequence of Corollary \[th:teichmuller1b\] and :
\[th:connections-rel-boundary\] In the situation of Theorem \[th:teichmuller\], there is a weak homotopy equivalence $$\label{eq:htp}
{\EuScript A}_{\mathit{flat}}(S,\{a_i\}) {\simeq}H^1(S,\partial S),$$ unique up to addition of a constant, and which is compatible with the action of .
The statement is along the same lines as our previous Proposition \[th:z-homotopy\] for the cylinder. In parallel with Addendum \[th:z-homotopy-2\], one can sometimes use rotation numbers along paths which connect different boundary components to reduce the ambiguity in the choice of the map . This relies on the following:
\[th:two-agree\] In the situation of Theorem \[th:teichmuller\], choose two distinct boundary components, and assume that, when we identify them in an orientation-reversing way, the boundary conditions agree. More precisely, denoting our boundary components by “left” and “right”, what we want is: $$\label{eq:left-right}
\begin{aligned}
&
\left\{
\begin{aligned}
& \epsilon_{\mathit{left}}: S^1 \stackrel{\cong}\longrightarrow \partial_{\mathit{left}} S && \text{orientation-reversing}, \\
& \epsilon_{\mathit{right}}: S^1 \stackrel{\cong}\longrightarrow \partial_{\mathit{right}} S && \text{orientation-preserving.}
\end{aligned}
\right.
\\
& \quad \text{such that} \quad \epsilon_{\mathit{left}}^* a_{\mathit{left}} = \epsilon_{\mathit{right}}^* a_{\mathit{right}} \in \Omega^1(S^1,{\mathfrak{g}}).
\end{aligned}$$ Then, parallel transport along a path $$\label{eq:connecting-path}
c: [0,1] \longrightarrow S, \quad c(0) = \epsilon_{\mathit{left}}(0), \;\; c(1) = \epsilon_{\mathit{right}}(0),$$ associates to $A \in {\EuScript A}_{\mathit{flat}}(S,\{a_i\})$ a hyperbolic element of $G$.
Glue together $\partial_{\mathit{left}} S$ and $\partial_{\mathit{right}} S$ to obtain a surface of genus one higher than $S$. Given any $A \in {\EuScript A}_{\mathit{flat}}(S,\{a_i\})$, that surface inherits a flat connection, to which Theorem \[th:teichmuller\] applies. The holonomy of the glued connection around any non-contractible loop must be hyperbolic, by Corollary \[th:all-hyperbolic\]; in particular, this holds for the loop obtained by identifying the endpoints of $c$.
There is a variation of that idea, which involves the following terminology. For any nonzero integer $\mu$, the $\mu$-fold cover $a^\mu$ of some $a \in \Omega^1(S^1,{\mathfrak{g}})$ is defined to be the pullback by the standard $\mu$-fold self-map of the circle: $$\label{eq:a01-cover}
\text{if } a = a_t \mathit{dt}, \quad \text{then } a^\mu = \mu a_{\mu t}\,\mathit{dt}.$$
\[th:two-are-powers\] Suppose that $S$ has at least three boundary components, and that holds. Choose parametrizations of two boundary components, as before, but now assume that $$\left\{
\begin{aligned}
& \epsilon_{\mathit{left}}^*a_{\mathit{left}} = a^{\mu_{\mathit{left}}}, \\
& \epsilon_{\mathit{right}}^*a_{\mathit{right}} = a^{\mu_{\mathit{right}}},
\end{aligned}
\right.
\quad \text{for some $a \in \Omega^1(S^1,{\mathfrak{g}})$ and $\mu_{\mathit{left}},\mu_{\mathit{right}} > 0$.}$$ Then, parallel transport along any path is again hyperbolic.
Let $\mu = \mu_{\mathit{left}}\mu_{\mathit{right}}$. There is a ${{\mathbb Z}}/\mu$-cover $\tilde{S} \rightarrow S$, such that the preimage of $\partial_{\mathit{left}} S$ has $\mu_{\mathit{left}}$ connected components, and similarly for $\partial_{\mathit{right}} S$ (here, we are using the existence of a third boundary component). Choose one component in each preimage, denoting them by $\partial_{\mathit{left}} \tilde{S}$ and $\partial_{\mathit{right}} \tilde{S}$. The pullback of our boundary conditions to those components can be identified with $a^\mu$, in both cases. Hence, if we have $A \in {\EuScript A}_{\mathit{flat}}(S,\{a_i\})$, we can pull it back to $\tilde{S}$, and it will then induce a connection on surface obtained by gluing together $\partial_{\mathit{left}}\tilde{S}$ and $\partial_{\mathit{right}}\tilde{S}$. One can choose the components and gluing map so that a lift of our path $c$ becomes a loop on the glued surface. With this in mind, the same argument as in Lemma \[th:two-agree\] goes through.
Returning to our discussion of , these Lemmas see the following use:
\[th:partial-fix\] In the situation of Proposition \[th:connections-rel-boundary\], suppose that additionally, Lemma \[th:two-agree\] or Lemma \[th:two-are-powers\] applies. Fix a path . Then, one can choose the map so that its composition with $\int_c: H^1(S,\partial S) \rightarrow {{\mathbb Z}}$ equals $\mathrm{rot}_c: {\EuScript A}_{\mathit{flat}}(S,\{a_i\}) \rightarrow {{\mathbb Z}}$.
We now turn to the action of diffeomorphisms $\phi \in \mathit{Diff}(S,\partial S)$. Under , the pullback action $\phi^*$ on ${\EuScript A}_{\mathit{flat}}(S,\{a_i\})$ corresponds to an affine automorphism, whose linear part is the usual action of $\phi$ on $H^1(S,\partial S)$. In particular, if $\phi$ acts trivially on $H^1(S,\partial S)$, its pullback action on ${\EuScript A}_{\mathit{flat}}(S,\{a_i\})$ can be described up to homotopy as a translation by a certain element $$\label{eq:translation-class}
T(\phi) \in H^1(S,\partial S).$$ We will need the following computation, which can be reduced to Addendum \[th:z-homotopy-3\] by looking at a tubular neighbourhood of the twisting curve:
\[th:t-class\] In the situation of Proposition \[th:connections-rel-boundary\], let $\tau_d$ be the Dehn twist along a nontrivial simple closed curve $d$ which separates the surface into two parts $S_{\mathit{left}}$ and $S_{\mathit{right}}$; more precisely, we choose an orientation of $d$, and take $S_{\mathit{right}} \subset S$ to be the part for which that orientation agrees with the boundary orientation. Then, the associated class is $$\label{eq:t-class}
\begin{aligned}
T(\tau_d)
& = \Big( \pm \chi(S_{\mathit{right}}) - \!\! \sum_{\partial_i S \subset \partial S_{\mathit{right}}} \mathrm{rot}(a_i) \Big)[d]
\\
& = \Big(\mp \chi(S_{\mathit{left}}) + \!\! \sum_{\partial_i S \subset \partial S_{\mathit{left}}} \mathrm{rot}(a_i) \Big)[d].
\end{aligned}$$
The coefficient in front of $[d]$ in is the rotation number of our connections along $d$, which is prescribed by (the sign there determines those in our formula) and the Milnor-Wood inequality for $S_{\mathit{left}}$, $S_{\mathit{right}}$.
Spaces of nonnegatively curved connections
------------------------------------------
We begin by going back to the basic relationship between curvature and the variation of parallel transport.
\[th:connections-on-the-square\] Given a nonnegatively curved connection on $[0,1]^2$, let $g(s) \in G$ be obtained by parallel transport along the family of paths $c_s$ from Figure \[fig:path\]. Then $g(s)$ is a nonpositive path. Conversely, fix a nonpositive path $g(s)$, and a connection $a_0 \in \Omega^1([0,1],{\mathfrak{g}})$ whose parallel transport along $[0,1]$ yields $g(0)$. Then there is a nonnegatively curved connection $A$ on $[0,1]^2$, with $A|(\{0\} \times [0,1]) = a_0$, such that parallel transport along each $c_s$ recovers $g(s)$. Moreover, the space of all such $A$ is weakly contractible.
(0,0)![\[fig:path\]The paths $c_s$ on the square and the cylinder.](path.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(3328,1594)(1036,-1184) (1051,239)[(0,0)\[lb\]]{} (1801,239)[(0,0)\[lb\]]{} (1801,-1111)[(0,0)\[lb\]]{} (1051,-1111)[(0,0)\[lb\]]{}
We can work up to gauge transformations which are trivial on $\{0\} \times [0,1]$ (the group of such gauge transformations is contractible). Hence, we may restrict our attention to connections which are trivial in $s$-direction, meaning of the form $A(s,t) \mathit{dt}$; the curvature is then just given by $-\partial_s A$. Let’s write $\Pi(s,t) \in G$ for the parallel transport along the straight line from $(s,0)$ to $(s,t)$. As in , this is given by $$\label{eq:a-from-phi}
\Pi(s,0) = {\mathbbm{1}}, \quad (\partial_t \Pi) \Pi^{-1} = A.$$ Set $$\label{eq:big-gamma}
\Gamma(s,t) = \partial_t (\Pi^{-1}\, \partial_s \Pi) = \Pi^{-1} (\partial_s A) \Pi \in {\mathfrak{g}}.$$ Integrating out yields $$\label{eq:phi-equation}
\Pi^{-1} (\partial_s \Pi) = \int_0^t \Gamma(s,\tau) \mathit{d\tau}.$$ If the curvature is nonnegative, $\Gamma$ is nonpositive, which by implies that $g(s) = \Pi(s,1)$ is a nonpositive path.
In the other direction, given a nonpositive path, set $\gamma = g^{-1} \,\partial_s g \in {\mathfrak{g}}_{\leq 0}$. For $A$ to have nonnegative curvature and the desired parallel transport maps, we need $$\label{eq:gamma-constraints}
\Gamma(s,t) \in {\mathfrak{g}}_{\leq 0} \quad
\text{and} \quad
\int_0^1 \Gamma(s,t) \mathit{dt} = \gamma(s).$$ The space of $\Gamma$ satisfying is clearly contractible. Given $\Gamma$ and the parallel transport maps $\Pi(0,t)$ coming from the given $a_0$, one gets maps $\Pi(s,t)$ by thinking of in reverse. This means that we set $\Xi(s,t) = \int_0^t \Gamma(s,\tau) d\tau$, and then define $\Pi(s,t)$ as the solution of the ODE $$\partial_s \Pi = \Pi\, \Xi.$$ Once one has the family $\Pi(s,t)$, there is a unique connection satisfying , which then automatically restricts to $a_0$ on $\{0\} \times S^1$.
\[th:connections-on-the-cylinder\] Fix a nonpositive path $g(s)$, and a connection $a_0 \in \Omega^1(S^1,{\mathfrak{g}})$ whose holonomy is $g(0)$. Then there is a nonnegatively curved connection $A$ on $S = [0,1] \times S^1$, with $A|(\{0\} \times S^1) = a_0$, such that the holonomy along each $c_s$ (see again Figure \[fig:path\]) is $g(s)$. Moreover, the space of all such $A$ is weakly contractible.
We follow the strategy from Lemma \[th:connections-on-the-square\], but pull everything back to the universal cover $[0,1] \times {{\mathbb R}}\rightarrow S$, and work ${{\mathbb Z}}$-periodically in $t$-direction. Given $\gamma = g^{-1} \partial_s g$, consider $\Gamma : [0,1] \times {{\mathbb R}}\rightarrow {\mathfrak{g}}$ which satisfy as well as the periodicity condition $$\Gamma(s,t+1) = g(s)^{-1} \Gamma(s,t) g(s).$$ This means that for a solution of , $$\label{eq:s-t-1}
\begin{aligned}
\Pi(s,t+1)^{-1} \partial_s \Pi(s,t+1) & = \gamma(s) + g(s)^{-1} (\Pi(s,t)^{-1} \partial_s \Pi(s,t)) g(s)
\\ & \quad = (\Pi(s,t) g(s))^{-1} \partial_s (\Pi(s,t) g(s)).
\end{aligned}$$ We start with $a_0 \in \Omega^1({{\mathbb R}},{\mathfrak{g}})$ which is ${{\mathbb Z}}$-periodic, and whose parallel transport along $[0,1]$ is $g(0)$. This means that $\Pi(0,t+1) = \Pi(0,t) g(0)$. Because of , we then have $\Pi(s,t+1) = \Pi(s,t) g(s)$ everywhere. As a consequence, the connection determined by $\Pi(s,t)$ through is again ${{\mathbb Z}}$-periodic.
With that in hand, let us return to the spaces of nonnegatively curved connections from . We remind the reader of the standing assumption for such spaces, which is that the boundary conditions $a_i$ are connections with hyperbolic holonomy.
\[th:nonnegative-connection\] Take $S = S^1 \times [0,1]$. As boundary conditions, fix $a_0,a_1 \in \Omega^1(S^1,{\mathfrak{g}})$ with the same rotation number $\mathrm{rot}(a_0) = \mathrm{rot}(a_1) = r$. Then there is a weak homotopy equivalence $$\label{eq:first-h}
{\EuScript A}_{\geq 0}(S,a_0,a_1) {\simeq}{{\mathbb Z}},$$ unique up to adding a constant on the right, and which is compatible with the action of ${\EuScript G}(a_0,a_1) {\simeq}{{\mathbb Z}}$. More explicitly, by associating to a connection $A$ its holonomies around the loops $d_s: S^1 \rightarrow S$, $d_s(t) = (s,t)$, one gets a weak homotopy equivalence $$\label{eq:connections-to-paths}
{\EuScript A}_{\geq 0}(S,a_0,a_1) \longrightarrow {\EuScript P}(G_{\mathit{hyp}},g_0,g_1).$$
Recall our general notation : $C_1$ is the conjugacy class of the holonomy $g_1$ of $A_1$, and $\tilde{C}_1$ its lift to $\tilde{G}$. Consider the space ${\EuScript A}_{\geq 0}(S,a_0,\tilde{C}_1)$ of nonnegatively curved connections which restrict to $a_0$ on $\{0\} \times S^1$, and whose lifted holonomy around $\{1\} \times S^1$ lies in $\tilde{C}_1$ (or equivalently, the holonomy lies in $C_1$ and $\mathrm{rot}_{\partial_1S}(A) = r$). By Lemma \[th:connections-on-the-cylinder\], there is a weak homotopy equivalence between that and a suitable space of nonpositive paths: $$\label{eq:a-p}
{\EuScript A}_{\geq 0}(S,a_0,\tilde{C}_1) {\simeq}{\EuScript P}_{\leq 0}(G,g_0,C_1)^0.$$ Lemma \[th:pre-short\] (applied to the reversed path) shows that the path space in is contractible. On the other hand, we have a weak fibration $$\label{eq:fi}
{\EuScript A}_{\geq 0}(S,a_0,a_1) \longrightarrow {\EuScript A}_{\geq 0}(S,a_0,\tilde{C}_1) \longrightarrow C_1.$$ The last space in should really be the space of connections on $S^1$ with (lifted) holonomy in $\tilde{C}_1$. By looking at parallel transport, one sees that this is weakly homotopy equivalent to the path space ${\EuScript P}(G,\mathit{id},C_1)^r$. In turn, evaluation at the endpoint yields a weak homotopy equivalence between that path space and $C_1$, which is the form in which has been stated. It follows that ${\EuScript A}_{\geq 0}(S,a_0,a_1) {\simeq}\Omega C_1 {\simeq}{{\mathbb Z}}$. For any $A \in {\EuScript A}_{\geq 0}(S,a_0,a_1)$, the holonomy around $\{s\} \times S^1$ is hyperbolic. This follows from Lemma \[th:connections-on-the-cylinder\] and the behaviour of nonnegative paths, as described in Lemma \[th:pre-short\]. Hence, makes sense. The proof that it is a homotopy equivalence involves looking at gauge transformations in the $[0,1]$-variable, as in Proposition \[th:z-homotopy\]). More precisely, if $PG$ is the space of all paths in $G$ starting ${\mathbbm{1}}$, the gauge action on a fixed connection yields a commutative diagram extending , $$\label{eq:fi2}
\xymatrix{
{\EuScript A}_{\geq 0}(S,a_0,a_1) \ar[r] & {\EuScript A}_{\geq 0}(S,a_0,\tilde{C}_1) \ar[r] & C_1 \\
\Omega G \ar[u] \ar[r] & PG \ar[u] \ar[r] & G \ar[u]
}$$ Since the middle and right hand $\uparrow$ are weak homotopy equivalences, so is the left hand one, and that implies the desired statement concerning .
\[th:nonnegative-connection-addendum\] In the situation from Proposition \[th:nonnegative-connection\], assume additionally that the boundary holonomies belong to the same one-parameter semigroup, meaning that they are of the form $$\label{eq:semigroup}
g_i = e^{t_i \gamma} \quad \text{with } t_i >0 \text{ and a common } \gamma.$$ Then there is a preferred choice of isomorphism $\pi_0({\EuScript P}(G_{\mathit{hyp}},g_0,g_1)) {\cong}{{\mathbb Z}}$, by looking at how the eigenvectors rotate; hence, the same holds for .
\[th:dehn-twist-pullback\] In the situation of Proposition \[th:nonnegative-connection\], let $\tau \in \mathit{Diff}(S,\partial S)$ be the Dehn twist along $\{{{\textstyle\frac{1}{2}}}\} \times S^1$. The action of $\tau^*$ on ${\EuScript A}_{\geq 0}(S,a_0,a_1)$ corresponds to subtracting $r$ on the right hand side of .
The first of these Addenda is obvious. The second one follows from the description (alternatively, one can deform $a_1$ so that holds, and reduce the computation to Addendum \[th:z-homotopy-3\]; this works because the inclusion of the flat connections into the nonnegatively curved ones is a weak homotopy equivalence, by Propositions \[th:z-homotopy\] and \[th:nonnegative-connection\]).
\[th:nonnegative-connection-2\] Consider the situation from Proposition \[th:nonnegative-connection\], except that now, $\mathrm{rot}(a_0) = \mathrm{rot}(a_1) + 1$. Then ${\EuScript A}_{\geq 0}(S,a_0,a_1)$ is weakly contractible.
The proof is similar to that of Proposition \[th:nonnegative-connection\]. By Lemmas \[th:short-path-2\] and \[th:connections-on-the-cylinder\], $${\EuScript A}_{\geq 0}(S,a_0,\tilde{C}_1) {\simeq}{\EuScript P}_{\leq 0}(G,g_0,C_1)^{-1} {\simeq}C_1.$$ One has the same fibration , and this time it follows that the fibre is weakly contractible.
\[th:nonnegative-connection-3\] Take $S$ to be a disc, with the boundary parametrized clockwise. As boundary condition, fix $a \in \Omega^1(S^1,{\mathfrak{g}})$ with $\mathrm{rot}(a) = 1$. Then ${\EuScript A}_{\geq 0}(S,a)$ is weakly contractible.
Once more, consider first the space ${\EuScript A}_{\geq 0}(S, \tilde{C})$ where we only prescribe the (lifted) conjugacy class of the boundary holonomy. Fix $\ast \in S \setminus \partial S$, and a family of maps $\phi_t: S \rightarrow S$ with $\phi_0 = \mathit{id}$, $\phi_t(\ast) = \ast$, $\phi_t|\partial S = \mathit{id}$, $\mathrm{det}(D\phi_t) \geq 0$ everywhere; and such that $\phi_1$ contracts a neighbourhood of $\ast$ to that point. By pulling back via those maps, one sees that ${\EuScript A}_{\geq 0}(S,\tilde{C})$ is weakly homotopy equivalent to the subspace of connections which are trivial near the chosen point. Using this and Lemma \[th:connections-on-the-cylinder\], one shows that the holonomy around concentric loops yields a weak homotopy equivalence $${\EuScript A}_{\geq 0}(S,\tilde{C}) {\simeq}{\EuScript P}_{\leq 0}(G,\mathit{id},C^{-1})^{-1} = {\EuScript P}_{\geq 0}(G,\mathit{id},C)^1.$$ By combining this with Lemma \[th:short-path-3\], one finds that the holonomy around $\partial S$ yields a weak homotopy equivalence ${\EuScript A}_{\geq 0}(S,\tilde{C}) {\simeq}C$. Our space ${\EuScript A}_{\geq 0}(S,a)$ is homotopically the fibre of that map.
\[th:nonnegative-connection-4\] Let $S$ be a genus zero surface with $m+1 \geq 3$ boundary components, labeled by $I = \{0,\dots,m\}$. We choose identifications $\partial_iS {\cong}S^1$ which are orientation-reversing for $i = 0$, and orientation-preserving for $i > 0$. As boundary conditions, take $a_i \in \Omega^1(S^1,{\mathfrak{g}})$ which satisfy $$\label{eq:m-minus-1}
\mathrm{rot}(a_0) = \mathrm{rot}(a_1) + \cdots + \mathrm{rot}(a_m) + 1-m.$$ Then there is a weak homotopy equivalence, unique up to adding a constant, and compatible with the action of , $$\label{eq:pants-connections}
{\EuScript A}_{\geq 0}(S,\{a_i\}) {\simeq}H^1(S,\partial S).$$
(0,0)![\[fig:holes\]The loops $c_i$ from the proof of Proposition \[th:nonnegative-connection-4\], for $m = 2$.](holes.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(2730,2972)(2086,-2025) (3276,-1336)[(0,0)\[lb\]]{} (3826,-1186)[(0,0)\[lb\]]{} (2926,-1186)[(0,0)\[lb\]]{} (3751,764)[(0,0)\[lb\]]{} (3276,-1861)[(0,0)\[lb\]]{}
Let’s start with the space ${\EuScript A}_{\geq 0}(S,\{\tilde{C}_i\})$ where only the conjugacy classes of the (lifted) boundary holonomies are fixed. As in , this fits into a weak fibration $$\label{eq:fi3}
{\EuScript A}_{\geq 0}(S,\{a_i\}) \longrightarrow {\EuScript A}_{\geq 0}(S,\{\tilde{C}_i\}) \longrightarrow \prod_{i=0}^m C_i.$$ Let’s think of $S$ as being glued together, rather trivially, from two parts: $S_{\mathit{left}}$ is a collar neighborhood of $\partial_0 S$, so that the other part $S_{\mathit{right}} \subset S$ is a deformation retract, containing $\partial_i S$ for $i>0$ (see Figure \[fig:holes\]). One can find a family of self-maps $(\phi_t)$ of $S$, starting at the identity, which are all the identity on $\partial S$, such that $\mathrm{det}(D\phi_t) \geq 0$ everywhere, and where $\phi_1$ retracts $S_{\mathit{right}}$ to a one-dimensional skeleton. From that, it follows (as in the proof of Proposition \[th:nonnegative-connection-3\]) that ${\EuScript A}_{\geq 0}(S,\{\tilde{C}_i\})$ is weakly homotopy equivalent to the subspace of those connections which are flat on $S_{\mathit{right}}$.
The space of flat connections on $S_{\mathit{right}}$, with holonomy around $\partial_i S$ ($i>0$) lying in $\tilde{C}_i$, is weakly homotopy equivalent to $\prod_{i=1}^m C_i$, where we can think of the homotopy equivalence as being given by holonomies around loops $c_i$ with a common base point (see again Figure \[fig:holes\]). In a slight break with our usual notation, we will use $g_i$ to denote those holonomies, and $\tilde{g}_i$ for their lifts to $\tilde{C}_i$. Extending such a connection on $S_{\mathit{right}}$ to a nonnegatively curved one on the whole of $S$, with the desired behaviour over $\partial_0 S$, amounts (up to homotopy, thanks to Lemma \[th:connections-on-the-cylinder\]) to choosing a nonnegative path from $\tilde{g}_1\cdots \tilde{g}_m$ to some point $\tilde{g}_0 \in \tilde{C}_0$ (the holonomy around $c_0$). To conclude, we have now shown that ${\EuScript A}_{\geq 0}(S,\{\tilde{C}_i\})$ is weakly homotopy equivalent to the space considered in Lemma \[th:pp\], hence to $C_0$. By looking at , one sees that ${\EuScript A}_{\geq 0}(S,\{a_i\}) {\simeq}\prod_{i>0} \Omega C_i = {{\mathbb Z}}^m$. Inspection of the argument shows that each gauge orbit yields a weak homotopy equivalence ${\EuScript G}(S,\{a_i\}) \rightarrow {\EuScript A}_{\geq 0}(S,\{a_i\})$, and from that we get the statement in its desired form.
We have arranged the results above roughly in order of complexity of their proofs, which was convenient since they all follow the same strategy. However, this might not be the best way of understanding their meaning. As the reader may have noticed, Proposition \[th:nonnegative-connection\] was a close cousin of the previous discussion concerning flat connections, in Proposition \[th:z-homotopy\]; similarly, one can compare Propositions \[th:connections-rel-boundary\] (specialized to genus zero) and \[th:nonnegative-connection-4\]. The outcome is that we have weak homotopy equivalences $$\label{eq:compare-flat}
\begin{aligned}
& {\EuScript A}_{\mathit{flat}}(S,a_0,a_1) \stackrel{{\simeq}}{\longrightarrow} {\EuScript A}_{\geq 0}(S,a_0,a_1) && \text{ for Proposition \ref{th:z-homotopy} (as mentioned before)},
\\ &
{\EuScript A}_{\mathit{flat}}(S,\{a_i\}) \stackrel{{\simeq}}{\longrightarrow} {\EuScript A}_{\geq 0}(S,\{a_i\}) && \text{ for Proposition \ref{th:nonnegative-connection-4}}.
\end{aligned}$$ The other two spaces of nonnegatively curved connections that we have analyzed show a different behaviour: they are contractible, which one could interpret as a comparison with the spaces of all connections, $$\label{eq:compare-all}
\begin{aligned}
& {\EuScript A}_{\geq 0}(S,a_0,a_1) \stackrel{{\simeq}}{\longrightarrow} {\EuScript A}(S,a_0,a_1) && \text{ for Proposition \ref{th:nonnegative-connection-2},}
\\ &
{\EuScript A}_{\geq 0}(S,a) \stackrel{{\simeq}}{\longrightarrow} {\EuScript A}(S,a) && \text{ for Proposition \ref{th:nonnegative-connection-3}.}
\end{aligned}$$ This reflects our previous distinction between spaces of “short paths” (whose topology could be related to constant paths) and “long paths” (related to spaces of all paths).
It is tempting to think that a comparison result such as should hold for higher genus surfaces, as long as the rotation numbers are chosen extremal, meaning that they saturate the bound in Proposition \[th:bochner\]; and similarly for closed surfaces, with rotation numbers replaced by the Euler class.
TQFT considerations\[sec:tqft\]
===============================
We now introduce our toy model formalism, and explore its implications. In fact, we will go over similar terrain three times: first in a way that only covers TQFT operations in a standard sense; then (in Section \[sec:diff-axiom\]) with additional features that lead to the construction of connections; and finally, in a mild generalization which allows elliptic holonomies as well (Section \[sec:elliptic\]). Overall, the discussion roughly follows [@seidel16 Sections 5–6].
Disc configurations
-------------------
Our formalism is built as a version of the classical framed little disc operad [@getzler94b; @salvatore-wahl03], enriched with connections (but not in the gauge theory sense: we keep a fixed trivialization of the bundle).
- Write $D \subset {{\mathbb C}}$ for the closed unit disc. An $m$-disc configuration, for $m \geq 0$, is an ordered collection of (closed round) discs $D_1,\dots,D_m$, which are contained in the interior of $D$, and are pairwise disjoint. To such a configuration, we can associate the (genus zero oriented) surface $$\label{eq:surface}
S = D \setminus \bigcup_{i=1}^m (D_i \setminus \partial D_i),$$ whose boundary circles we label by $\partial_0 S = \partial D$, and $\partial_i S = \partial D_i$ for $i=1,\dots,m$. We also want to allow one exceptional case, called the identity configuration, which consists of a single disc $D_1 = D$. For that configuration, becomes a circle, thought of as an annulus of thickness zero.
- A framing of an $m$-disc configuration consists of boundary parametrizations $$\label{eq:boundary-parametrizations}
\epsilon_i: S^1 \longrightarrow \partial_i S, \quad
\epsilon_i(t) = \rho_i e^{-2\pi i t} + \zeta_i, \quad i = 1,\dots,m.$$ Here, $\zeta_i$ is the center of $D_i$, and $|\rho_i| > 0$ is its radius, so only $\mathrm{arg}(\rho_i)$ is a free parameter; for the identity configuration, we specify that the parameter must be $\rho_1 = 1$. To the we add a fixed parametrization of the boundary of $D$, $$\label{eq:boundary-parametrizations-0}
\epsilon_0: S^1 \longrightarrow \partial_0 S, \quad
\epsilon_0(t) = e^{-2\pi it}.$$ The clockwise parametrization convention means that $\epsilon_0$ is inverse to the boundary orientation, while the other $\epsilon_i$ are compatible with it.
- Suppose that we have a framed disc configuration, together with $a_i \in \Omega^1(S^1,{\mathfrak{g}})$ for $i = 0,\dots,m$, which have hyperbolic holonomy (as before, we will write $g_i$ for the holonomies, and $C_i$ for their conjugacy classes). A compatible connection is given by an $A \in \Omega^1(S,{\mathfrak{g}})$ with nonnegative curvature, and whose pullback by , equals $a_i$. We also impose the following technical condition: for every boundary circle $\partial_iS$, let $r_i$ be the radial retraction of a neighbourhood onto that circle; then, $A = r_i^*(A|\partial_i S)$ near $\partial_iS$. This means that the behaviour of $A$ close to $\partial S$ is entirely determined by the $a_i$. Again, we have to mention the exceptional case of the identity configuration, where necessarily $a_0 = a_1$, and there is no additional freedom in choosing $A$.
A decorated disc configuration is a disc configuration together with a choice of framing and compatible connection. We write $Y = (\epsilon_1,\dots,\epsilon_m,A)$ for the decoration, and usually represent the decorated disc configuration by the pair $(S,Y)$.
There is a gluing process for decorated disc configurations. Start with an $m_1$-configuration and an $m_2$-configuration. Choose some $i \in \{1,\dots,m_1\}$, and insert the second configuration into the first one in place of the $i$-th disc, rescaling and rotating it as prescribed by $i$-th boundary parametrization. In terms of the associated surfaces, this means that we glue them together along one of their boundary circles, using the given parametrizations. Assuming that the boundary conditions for the connections agree, those will give rise to a connection on the glued surface. The outcome is a decorated configuration consisting of $m = m_1+m_2-1$ discs. The gluing construction is strictly associative. One can formally include the identity configuration in the gluing process, for which (as the name suggests) it is the neutral element.
We will use families of disc configurations, parametrized by smooth manifolds (possibly with boundary or corners) $P$, and equipped with the same kind of decorations. Aside from one exception, this notion is straightforward:
- The disc configurations vary smoothly depending on the parameter, giving rise to a fibration with fibres , which we denote by $\underline{S} \rightarrow P$.
- Similarly, the framings vary smoothly with $P$. The connection is given by a one-form in fibre direction on $\underline{S}$. The boundary conditions $(a_0,\dots,a_m)$ are always constant (independent of where we are on $P$). We usually denote such a decoration by $\underline{Y} = (\underline{\epsilon}_1,\dots,\underline{\epsilon}_m,\underline{A})$.
The exceptional case occurs when we have a family of annuli with the same boundary conditions ($m = 1$ and $a_0 = a_1$). That could include instances of the identity configuration, which means that $\underline{S}$ will be singular in general. This is potentially a problem for the definition of connection on the fibres: to avoid that, we assume that in a neighbourhood of the subset of $P$ where the identity configuration occurs, the framing should be trivial (no rotation), and the connection should be pulled back from radial projection to $S^1$. The gluing construction extends to families (one starts with a family over $P_1$ and another over $P_2$, and ends up with one over $P_1 \times P_2$).
Spaces of decorations
---------------------
Fix an $m$-disc configuration, which is not the identity configuration, and boundary values $(a_0,\dots,a_m)$. Write $$\label{eq:y-space}
\begin{aligned}
{\EuScript Y}_{\geq 0}(S,\{a_i\}) & \;\; \text{space of all decorations $Y = (\epsilon_1,\dots,\epsilon_m,A)$ of $S$,} \\
{\EuScript Y}_{\mathit{flat}}(S,\{a_i\}) &\;\; \text{subspace where the connection is flat.}
\end{aligned}$$ With respect to the previously considered spaces ${\EuScript A}_{\geq 0}(S,\{a_i\})$ and ${\EuScript A}_{\mathit{flat}}(S,\{a_i\})$, the only difference (apart from certain conventions concerning boundary orientations) is the additional datum given by the framings. Nevertheless, we find it convenient to explicitly translate the outcome of the discussion from Section \[sec:connections\] into the language of , since that will be relevant for our applications. The most obvious case is that of the empty disc configuration ($m = 0$), where there is no framing, and Proposition \[th:nonnegative-connection-3\] says the following:
\[th:the-disc\] Let $S$ be a disc, with $\mathrm{rot}(a_0) = 1$. Then ${\EuScript Y}_{\geq 0}(S,a_0)$ is contractible.
Next, consider the case of the annulus ($m = 1$). Let’s introduce the following abelian topological group (depending on $r \in {{\mathbb Z}}$): $$\Gamma_r = \begin{cases} {{\mathbb Z}}/r & r \neq 0, \\ {{\mathbb Z}}\times S^1 & r = 0. \end{cases}$$
\[th:x-annulus\] (i) Let $S$ be an annulus, with boundary conditions which satisfy . Then there is a weak homotopy equivalence $$\label{eq:x-annulus}
{\EuScript Y}_{\mathit{flat}}(S,a_0,a_1) {\simeq}\Gamma_r$$ which is unique up to (homotopy and) adding an element of ${{\mathbb Z}}/r$. To state this more precisely, choose an identification $[0,1] \times S^1 {\cong}S$ which restricts to $\epsilon_i$ along $\{i\} \times S^1$. Then, the holonomy around the circles $\{s\} \times S^1$ gives a map $$\label{eq:explicit-mod-r}
{\EuScript Y}_{\mathit{flat}}(S,a_0,a_1) \longrightarrow \pi_0({\EuScript P}(C,g_0,g_1)) \times_{{{\mathbb Z}}} {{\mathbb Z}}/r,$$ where $C$ is the common conjugacy class of the $g_i$, and ${{\mathbb Z}}= \pi_1(C)$ acts by composition of paths with loops. The map is canonical (independent of the choice of identification), and yields the discrete part of ; for $r = 0$, $\mathrm{arg}(\rho_1)$ yields the continuous part.
\(ii) Take boundary conditions $a_0,a_1$ which have the same rotation number $r$. Then there is a homotopy equivalence $$\label{eq:x-annulus-2}
{\EuScript Y}_{\geq 0}(S,a_0,a_1) {\simeq}\Gamma_r.$$ This has the same uniqueness property as , and can be described as in , except that the map now takes values in $\pi_0({\EuScript P}(G_{\mathit{hyp}},g_0,g_1)) \times_{{{\mathbb Z}}} {{\mathbb Z}}/r$. In the more restricted case , the inclusion ${\EuScript Y}_{\mathit{flat}}(S,a_0,a_1) \subset {\EuScript Y}_{\geq 0}(S,a_0,a_1)$ is a weak homotopy equivalence, and our construction reduces to that in (i).
\(iii) Take boundary conditions $a_0,a_1$ such that $\mathrm{rot}(a_0) = \mathrm{rot}(a_1) + 1$. Then $\mathrm{arg}(\rho_1)$ yields a weak homotopy equivalence $${\EuScript Y}_{\geq 0}(S,a_0,a_1) {\simeq}S^1.$$
The proofs are straightforward, given our previous discussion. Each of them is based on considering the fibration $$\label{eq:rho-fibration-1}
{\EuScript A}_{\mathit{flat}}(S,a_0,a_1) \longrightarrow {\EuScript Y}_{\mathit{flat}}(S,a_0,a_1) \xrightarrow{\mathrm{arg}(\rho_1)} S^1,$$ or its counterpart for ${\EuScript Y}_{\geq 0}(S,a_0,a_1)$. For (i), Proposition \[th:z-homotopy\] tells us that the fibre is homotopy equivalent to ${{\mathbb Z}}$, and Addendum \[th:z-homotopy-3\] shows that going around the $S^1$ base corresponds to subtracting $r$. The same goes for (ii) and (iii), using Proposition \[th:nonnegative-connection\] with Addendum \[th:dehn-twist-pullback\], respectively Proposition \[th:nonnegative-connection-2\].
\[th:x-annulus-a\] (i) In the situation of Proposition \[th:x-annulus\](i), suppose that $g_0 = g_1 = g$. Then, following Addendum \[th:z-homotopy-2\], there is a preferred choice of . One can see that from ; or equivalently, it can be described as the map $$\label{eq:rot-mod-r}
(\epsilon_1,A) \longmapsto \mathrm{rot}_c(A) \in {{\mathbb Z}}/r,$$ where $c$ is any path from $\epsilon_0(0)$ to $\epsilon_1(0)$ (the rotation number mod $r$ is independent of the choice of path).
\(ii) In the situation of Proposition \[th:x-annulus\](ii), suppose that $g_0$ and $g_1$ lie in the same one-parameter semigroup. Then, there is again a preferred choice of , see Addendum \[th:nonnegative-connection-addendum\] (but it is unclear whether the alternative description would still work).
\[th:fractional-rotation\] Let $r \neq 0$. Consider the configuration consisting of a disc centered at zero, with parametrization $\epsilon_1(t) = \rho_1 e^{-2\pi i (t+k/r)}$, for some $k \in {{\mathbb Z}}/r$. Suppose that $a \in \Omega^1(S^1,{\mathfrak{g}})$ is an $r$-fold cover, in the sense of . Radial projection from the disc to the circle equips $S$ with a flat connection $A$, whose boundary values are $a_0 = a_1 = a$. By looking at , one sees that this belongs to the $k$-th component of ${\EuScript Y}_{\mathit{flat}}(S,a,a)$.
\[th:decorations\] (i) Consider a configuration of $m \geq 2$ discs, with boundary conditions such that holds, and write $r_i = \mathrm{rot}(a_i)$. Then, there is a weak homotopy equivalence $$\label{eq:x-space}
{\EuScript Y}_{\mathit{flat}}(S,\{a_i\}) {\simeq}\prod_{i=1}^m \Gamma_{r_i},$$ unique up to adding an element of $\prod_{i=1}^m {{\mathbb Z}}/r_i$.
\(ii) In the same situation, the inclusion ${\EuScript Y}_{\mathit{flat}}(S,\{a_i\}) \subset {\EuScript Y}_{\geq 0}(S,\{a_i\})$ is a weak homotopy equivalence. Hence, ${\EuScript Y}_{\geq 0}(S,\{a_i\})$ admits the same description .
To prove this (let’s say, for flat connections), one considers the fibration $$\label{eq:multiple-rho}
{\EuScript A}_{\mathit{flat}}(S,\{a_i\}) \longrightarrow {\EuScript Y}_{\mathit{flat}}(S,\{a_i\}) \xrightarrow{(\mathrm{arg}(\rho_1),\dots,\mathrm{arg}(\rho_m))} (S^1)^m.$$ By Proposition \[th:connections-rel-boundary\], the fibre is homotopy equivalent to $H^1(S,\partial S)$, which we identify with ${{\mathbb Z}}^m$ by taking generators dual to paths $c_i$ ($1 \leq i \leq m$) from $\partial_0 S$ to $\partial_i S$. The monodromy of around the $i$-th $S^1$ factor agrees with the pullback action $(\tau_{d_i}^{-1})^*$ of an inverse (or left-handed) Dehn twist along a loop $d_i$ which is a parallel copy of $\partial_i S$. Let’s orient this loop opposite to the boundary orientation of $\partial_iS$; it then satisfies $d_i \cdot c_i = -1$, hence its Poincar[é]{} dual is minus the $i$-th unit element in our basis. Addendum \[th:t-class\] tells us that the action of $\tau_{d_i}^{-1}$ corresponds to adding the element $$\label{eq:rot-ai}
-T(\tau_{d_i}) = -r_i [d_i] = (0,\dots,r_i,\dots,0) \in H^1(S,\partial S) {\cong}{{\mathbb Z}}^m,$$ which implies (i). For (ii), one compares Propositions \[th:connections-rel-boundary\] and \[th:nonnegative-connection-4\], as already mentioned in .
\[th:two-agree-3\] In the situation of Proposition \[th:decorations\](i), suppose that $a_0 = a_m = a$ has rotation number $r$, and that $\mathrm{rot}(a_i) = 1$ for $i = 1,\dots,m-1$. In analogy with Addendum \[th:x-annulus-a\](i), an explicit choice of (the discrete part of) is then given by $$\label{eq:rot-mod-2}
(\epsilon_1,\dots,\epsilon_m,A) \longmapsto \mathrm{rot}_c(A) \in {{\mathbb Z}}/r,$$ where $c$ is any path from $\epsilon_0(0)$ to $\epsilon_m(0)$.
Most of the necessary argument is carried over from Addendum \[th:partial-fix\]. What remains to be shown is that is independent of the choice of path, or equivalently, is invariant under $\mathit{Diff}(S,\partial S)$. To see that, note that $\pi_0(\mathit{Diff}(S,\partial S))$ is generated by Dehn twists along (necessarily separating) curves $d$. If $d$ does not separate $\partial_0 S$ and $\partial_1 S$, its intersection number with $c$ must be zero, which in terms of Addendum \[th:t-class\] means that $T(\tau_d) \cdot [c] = 0$. In the remaining case, $d \cdot c = \pm 1$ and $T(\tau_d)$ is a multiple of $r[d]$, so using Addendum \[th:t-class\] again, we get $T(\tau_d) \cdot [c] \in r{{\mathbb Z}}$. In either case, is unchanged under $\tau_d^*$.
One can deepen this discussion by allowing the disc configuration to vary. Let $\mathit{Conf}_m$ (for some $m \geq 2$) be the space of all $m$-disc configurations (this is homotopy equivalent to the standard ordered configuration space). Over it, we have a universal family of surfaces $\underline{S}$; and on the fibres of that family, we have a space of all possible choices of decorations (let’s say, flat ones), denoted by ${\EuScript Y}_{\mathit{flat}}(\underline{S},\{a_i\})$. This all fits into a diagram, of which the columns and rows are fibrations: $$\label{eq:monodromy-diagram}
\xymatrix{
{{{\mathbb Z}}^m} \ar[d]_-{\prod_i r_i} \ar[r]
&
{\mathit{point}} \ar[d] \ar[r]
&
(S^1)^m \ar[d]^-{\text{inclusion}}
\\
{{{\mathbb Z}}^m} \ar[r] \ar[d]
&
{{\EuScript Y}_{\mathit{flat}}(\underline{S},\{a_i\})}
\ar[r] \ar@{=}[d]
&
(S^1)^m \times \mathit{Conf}_m \ar[d]^-{\text{projection}}
\\
\prod_{i=1}^m \Gamma_{r_i} \ar[r]
&
{\EuScript Y}_{\mathit{flat}}(\underline{S},\{a_i\}) \ar[r]
& \mathit{Conf}_m
}$$ Here, $(S^1)^m \times \mathit{Conf}_m$ is the space of framed configurations. The middle row of restricts to on each configuration, and the description of the map ${{\mathbb Z}}^m \rightarrow {{\mathbb Z}}^m$ reflects our previous discussion of . Here is one important additional piece of information concerning the middle row:
\[th:twist-loop\] In the situation of Proposition \[th:decorations\], consider a loop in $\mathit{Conf}_m$ formed by moving the $i$-th and $j$-th disc anticlockwise around each other, for some $1 \leq i < j \leq m$. Embed that into $(S^1)^m \times \mathit{Conf}_m$ by taking the $(S^1)^m$ component to be constant. Then, the monodromy of ${\EuScript Y}_{\mathit{flat}}(\underline{S},\{a_i\}) \rightarrow (S^1)^m \times \mathit{Conf}_m$ along that loop is given by adding the element $$\label{eq:ij-vector}
(0,\dots,\overbrace{1-\mathrm{rot}(a_j)}^{\text{$i$-th position}},\dots,\overbrace{1-\mathrm{rot}(a_i)}^{\text{$j$-th position}},\dots,0) \in {{\mathbb Z}}^m.$$
To see this, take the loop $d_{ij}$ which encloses exactly the $i$-th and $j$-th discs. In other words, it divides the surface into pieces $S_{\mathit{left}}$ and $S_{\mathit{right}}$, of which $S_{\mathit{right}}$ has boundary consisting exactly of $d_{ij}$, $\partial_i S$ and $\partial_j S$. We orient $d_{ij}$ compatibly with $\partial S_{\mathit{right}}$, which means that its Poincar[é]{} dual in $H^1(S,\partial S)$ is minus the $i$-th and $j$-th unit vectors, analogously to . By Addendum \[th:t-class\], the pullback action of $\tau_{d_ij}^{-1}$ is given by adding $$\label{eq:rot-aij}
-T(\tau_{d_{ij}}) = (1 - \mathrm{rot}(a_i) - \mathrm{rot}(a_j))[d_{ij}].$$ The monodromy we are looking for (corresponding to moving the discs, but not changing the angles $\rho_i$ or $\rho_j$) is actually the pullback action of $\tau_{d_{ij}}^{-1}\tau_{d_i}\tau_{d_j}$, where $d_i$ and $d_j$ are as in . Combining that equation with yields the desired result. Of course, by virtue of Proposition \[th:decorations\](ii), all this carries over to spaces ${\EuScript Y}_{\geq 0}(\underline{S},\{a_i\})$.
The TQFT axioms
---------------
Fix a commutative coefficient ring $R$. We suppose that the following data are given. For every $a \in \Omega^1(S^1,{\mathfrak{g}})$ with hyperbolic holonomy (our standing assumption), one has a chain complex of $R$-modules, denoted by $C^*(a)$; its cohomology will be written as $H^*(a)$. For any family of decorated disc configurations whose base is an oriented compact manifold with corners $P$, and whose boundary values are $a_0,\dots,a_m$, one has a map $$\label{eq:operation}
C^*(a_1) \otimes \cdots \otimes C^*(a_m) \longrightarrow C^{*-\mathrm{dim}(P)}(a_0).$$ We will denote by $\phi_P$, which is concise if somewhat incomplete. The behaviour of these maps is governed by three (groups of) axioms. These are very similar to those in [@seidel16], hence will be only stated briefly:
- Exchanging the labeling of the $m$ discs permutes the inputs of the associated operation , with the expected Koszul signs. Reversing the orientation of $P$ flips the sign of . Taking the disjoint union of two parameter spaces yields the sum of the associated operations. Finally, pulling back a family by a finite covering $\tilde{P} \rightarrow P$ of degree $\mu$ yields $$\label{eq:covering-axiom}
\phi_{\tilde{P}} = \mu \phi_P$$ (the last part, concerning coverings, will be needed only once, in the construction underlying Proposition \[th:bigraded-lie\]).
- The coboundary (defined using the differentials on the chain complexes) of equals the sum of the operations associated to the codimension one boundary faces of $P$. If those boundary faces are written as $\partial_j P$, then the sign convention is that $$\label{eq:boundary-sign}
\begin{aligned}
& (-1)^{\mathrm{dim}(P)} d\phi_P(x_1,\dots,x_m) - \phi_P(dx_1,\dots,x_m) -
\cdots
\\ & \qquad \qquad \cdots - (-1)^{|x_1|+\cdots+|x_{m-1}|} \phi_P(x_1,\dots,dx_m) + \sum_j \phi_{\partial_j P}(x_1,\dots,x_m) = 0.
\end{aligned}$$
- Gluing of families corresponds to composition of operations. More precisely, suppose that we have families of $m_k$-disc configurations, with parameter spaces $P_k$ ($k = 1,2$), and glue the second one into the $i$-th place of the first one. The composition axiom says that $$\label{eq:tqft}
\begin{aligned}
\phi_{P_1 \times P_2}(x_1,\dots, x_{m_1+m_2-1}) & =
(-1)^{(\mathrm{dim}(P_1)+|x_1|+\cdots+|x_{i-1}|)\mathrm{dim}(P_2)}
\\ & \phi_{P_1}(x_1, \dots, x_{i-1},
\phi_{P_2}(x_i,\dots,x_{i+m_2-1}),\dots,x_{m_1+m_2-1}).
\end{aligned}$$ Moreover, the identity configuration induces the identity map $C^*(a) \rightarrow C^*(a)$.
We will now explore the consequences of these axioms, remaining on the level of the cohomology groups $H^*(a)$. Take Proposition \[th:x-annulus\](i) in the case $a_0 = a_1 = a$. This implies that $H^*(a)$ carries an action of $H_{-*}(\Gamma_r)$, for $r = \mathrm{rot}(a)$, or in a more concrete formulation:
\[th:auto\] $H^*(a)$ carries the following endomorphisms: $$\label{eq:endomorphisms}
\left\{
\begin{aligned}
& \text{if $r = \mathrm{rot}(a) \neq 0$, an action of ${{\mathbb Z}}/r$;} \\
& \text{if $r = 0$, a ${{\mathbb Z}}$-action; and additionally, an endomorphism of degree $-1$.}
\end{aligned}
\right.$$
We denote the generator of the ${{\mathbb Z}}/r$-action by $\Sigma$. More generally, Proposition \[th:x-annulus\](ii), together with Addendum \[th:x-annulus-a\], implies the following:
\[th:well-defined\] (i) Given $a_0,a_1$ with the same rotation number $r$, there is an isomorphism $$\label{eq:01-iso}
H^*(a_1) {\cong}H^*(a_0),$$ which commutes with the action of , and is unique up to composition with powers of $\Sigma$.
\(ii) If the holonomies of $a_0$ and $a_1$ lie in the same one-parameter semigroup, meaning that holds, there is a distinguished choice of , and these choices are compatible with composition.
The situation for unequal rotation numbers is quite different. Applying Proposition \[th:x-annulus\](iii) yields the following:
\[th:canonical-continuation\] Suppose that $\mathrm{rot}(a_0) = \mathrm{rot}(a_1) + 1$. Then, there are canonical maps $$\begin{aligned}
\label{eq:a-continuation}
& C: H^*(a_1) \longrightarrow H^*(a_0), \\
\label{eq:pseudo-bv}
& D: H^*(a_1) \longrightarrow H^{*-1}(a_0).\end{aligned}$$ The “continuation map” is invariant under composition with $\Sigma$ on the left or right. It is also compatible with (any choice of) the isomorphisms relating different choices of $a_0$ or $a_1$, with the same rotation number. The map has the same properties, and additionally, it anticommutes with .
In a slightly flamboyant formulation, one could say that the continuation maps are more robustly independent of the choice of $a$ than the groups $H^*(a)$ themselves. From Proposition \[th:the-disc\], we get:
\[th:unit\] Suppose that $\mathrm{rot}(a) = 1$. Then there is a distinguished element $$\label{eq:e}
e \in H^*(a).$$ For different choices of $a$, these elements are related by (which is canonical for this particular rotation number).
We now turn to multilinear operations:
\[th:gerstenhaber-algebra\] For $\mathrm{rot}(a) = 1$, $H^*(a)$ carries a canonical structure of a Gerstenhaber algebra, for which is the unit.
Take the universal family with boundary conditions $\{a,\dots,a\}$. Consider the ${{\mathbb Z}}^m$-covering which is the middle row of . Addendum \[th:twist-loop\] and determine this covering, which turns out to be the universal cover of the $(S^1)^m$ factor. This means that the map $$\label{eq:f-map}
{\EuScript Y}_{\mathit{flat}}(\underline{S},a,\dots,a) \longrightarrow (S^1)^m \times \mathit{Conf}_m$$ has the following properties: its first component is homotopic to a constant, and its second component is a weak homotopy equivalence. Given any map $P \rightarrow \mathit{Conf}_m$, one therefore has a homotopically unique lift to ${\EuScript Y}_{\mathit{flat}}(\underline{S},a,\dots,a)$. More concretely, a single configuration of two discs ($P = \mathit{point}$) induces the product; two-disc configurations moving anticlockwise around each other ($P = S^1$) induce the bracket; and all the necessary relations between these operations can be implemented by explicit cobordisms, which means that they come from relations in the groups $\mathit{MSO}_d(\mathit{Conf}_m)$ (for $d \leq 2$). We omit the details, since the relation between the little disc operad and Gerstenhaber algebras is classical.
In the construction of the Gerstenhaber structure, we used flat connections, but of course, the same argument goes through with a more general choice of nonnegatively curved ones, thanks to Proposition \[th:decorations\](ii). Nonnegatively curved connections appear naturally when discussing unitality, where they arise from the definition of . The proof of unitality goes as follows: take two copies of the disc which defines , and glue them into the pair-of-pants which defines the product. After appealing to Proposition \[th:the-disc\], we conclude that the outcome again defines . In other words, $e$ is an idempotent. Now take the same pair-of-pants as before, and glue in just one copy of the disc. The outcome is an annulus, whose associated operation, multiplication with $e$, is necessarily some power of $\Sigma$. Because $e$ is idempotent and multiplication is associative, it follows that $e$ is a unit.
\[th:all-a\] Another application of nonnegatively curved connections is a mild generalization of Proposition \[th:gerstenhaber-algebra\], where one considers “all boundary conditions with rotation number $1$ simultaneously”. What this means is that for any such $a_k$, there are maps $$\begin{aligned}
\label{eq:012-product}
& \cdot: H^*(a_1) \otimes H^*(a_2) \longrightarrow H^*(a_0), \\
\label{eq:012-bracket}
& [\cdot,\cdot]: H^*(a_1) \otimes H^*(a_2) \longrightarrow H^{*-1}(a_0).\end{aligned}$$ These satisfy the Gerstenhaber algebra axioms, for all situations where the compositions make sense. For instance, the associativity equation $$(x_1x_2)x_3 = x_1(x_2x_3) \in H^*(a_0), \quad \text{where } x_k \in H^*(a_k)$$ always holds, where the intermediate terms can be taken to lie in arbitrary groups, $x_1x_2 \in H^*(a_4)$ and $x_2x_3 \in H^*(a_5)$. Similarly, inserting a unit element on either the left or right into always results in the isomorphism . From this, it follows that is an isomorphism of Gerstenhaber algebras.
Now fix $a_1,a_2$, with $\mathrm{rot}(a_1) = 1$, and $\mathrm{rot}(a_2) = r$ arbitrary. Consider the universal family of $m$-disc configurations, and its decorations, where the connections are flat and have boundary values $(a_2,a_1,\dots,a_1,a_2)$. We can use Addendum \[th:two-agree-3\] to pick a preferred subspace $$\label{eq:map-to-gamma}
{\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,\dots,a_1,a_2) \subset {\EuScript Y}_{\mathit{flat}}(\underline{S},a_2,a_1,\dots,a_1,a_2),$$ namely that on which the map vanishes (and in the case where $r = 0$, we require additionally that the angle $\arg(\rho_m)$ should be trivial). In other words, we use to provide a trivialization of the fibration which forms the bottom row of . The resulting homotopy equivalence ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,\dots,a_1,a_2) \rightarrow \mathit{Conf}_m$ gives a distinguished choice of decorations (up to homotopy) for each disc configuration. This choice is compatible with the two gluing processes that will be relevant for us:
- .5em Suppose that we take two configurations, with boundary values $(a_2,a_1,\dots,a_1,a_2)$ and $(a_1,\dots,a_1)$, respectively, and glue the second one into the first one (in one of the possible ways). The rotation number of the first configuration remains unchanged under gluing. Hence, this gluing process preserves the subspace ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,\dots,a_1,a_2)$.
- Suppose that we take two decorated disc configurations, both having boundary values $(a_2,a_1,\dots,a_1,a_2)$, and glue them (in the unique possible way). This gluing process adds up the rotation numbers . Hence, again, if the original configurations lie in ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,\dots,a_1,a_2)$, then so does the glued one.
Let’s pass to concrete implications. Using a two-disc configuration, respectively an $S^1$-family of such configurations, and their lifts to ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,a_2)$, one defines a product and bracket $$\begin{aligned}
\label{eq:r-product}
& \cdot: H^*(a_1) \otimes H^*(a_2) \longrightarrow H^*(a_2), \\
\label{eq:r-bracket}
& [\cdot,\cdot]: H^*(a_1) \otimes H^*(a_2) \longrightarrow H^{*-1}(a_2).\end{aligned}$$
\[th:rot-bracket\] The family of decorated disc configurations underlying has the following geometry: two discs move around each other; during that rotation, the framing of the first disc rotates $(1-r)$ times, while that of the second one remains constant (up to homotopy).
Consider the middle row of , specialized to our case: $$\label{eq:m-2}
{{\mathbb Z}}^2 \longrightarrow {\EuScript Y}_{\mathit{flat}}(\underline{S},a_2,a_1,a_2) \longrightarrow (S^1)^2 \times \mathit{Conf}_2.$$ By construction, the holonomy around the two $S^1$ factors is $(-\mathrm{rot}(a_1),0)$ and $(0,-\mathrm{rot}(a_2))$, respectively. The operation is defined using a loop in ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,a_2)$ which consists of the two discs moving around each other once, and where the framings rotate by some amounts $d_1,d_2 \in {{\mathbb Z}}$. The image of that loop in $(S^1)^2 \times \mathit{Conf}_2$ must be a loop along which has trivial monodromy. On the other hand, by Addendum \[th:twist-loop\], that monodromy is $$(1-\mathrm{rot}(a_2), 1-\mathrm{rot}(a_1)) + d_1 (-\mathrm{rot}(a_1),0) + d_2 (0,-\mathrm{rot}(a_2)) \\
= (1-r-d_1,-rd_2).$$ If $r \neq 0$, it follows that $$\label{eq:rot12}
d_1 = 1-r, \;\; d_2 = 0.$$ If $r = 0$, we know a priori that $d_2 = 0$, by definition of ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,a_2)$, which leads to the same conclusion .
\[th:gerstenhaber-module\] The operations and make $H^*(a_2)$ into a (unital) Gerstenhaber module over the Gerstenhaber algebra $H^*(a_1)$. For $a_2 = a_1$, this is the diagonal module.
Let’s consider the desired relations (the defining properties of a Gerstenhaber module; see e.g. [@kowalzig-kraehmer14 Definition 1.1], which however uses different sign conventions), for $x_1,x_2 \in H^*(a_1)$ and $x_3 \in H^*(a_2)$: $$\begin{aligned}
\label{eq:r-relation-1}
& (x_1 x_2) x_3 = x_1 (x_2 x_3), \\
\label{eq:r-relation-2}
&
[x_1,x_2x_3] = [x_1,x_2] x_3 + (-1)^{(|x_1|-1)|x_2|} x_2[x_1,x_3], \\
\label{eq:r-relation-4}
& [[x_1,x_2],x_3] + (-1)^{|x_1|} [x_1,[x_2,x_3]] + (-1)^{(|x_1|-1)|x_2|} [x_2,[x_1,x_3]] = 0.\end{aligned}$$ Because the choices of decorations, taken to lie in ${\EuScript Y}_{\mathit{flat}}^0(\underline{S},a_2,a_1,a_1,a_2)$, are homotopically unique, all we need to show is that the underlying relations hold up to bordism in $\mathit{Conf}_3$.
The only nontrivial case is , since there the framings, as analyzed in Lemma \[th:rot-bracket\], affect the families obtained from the gluing process. To spell this out in the simplest possible terms, let’s temporarily ignore the sizes of discs, and just think of $\mathit{Conf}_3$ as the ordered three-point configuration space in ${{\mathbb C}}$. The only thing that depends on $r$ is the first term in , which (up to homotopy) comes from the family $$S^1 \times S^1 \longrightarrow \mathit{Conf}_3, \quad
(p_1,p_2) \longmapsto (0, {{\textstyle\frac{1}{2}}}e^{2\pi i (1-r) p_1 + 2\pi i p_2}, e^{2\pi i p_1}).$$ Even there, the $r$-dependence can be removed by a change of parameters $p_2 \mapsto p_2 + (r-1)p_1$. The remaining argument is therefore the same as in the case $r = 1$, which is standard (and indeed, already entered into Proposition \[th:gerstenhaber-algebra\]).
Even though our definition of was restricted to flat connections, there is an alternative formulation which works more generally for nonnegatively curved ones. Namely, given a decorated $m$-disc configuration with boundary values $(a_2,a_1,\dots,a_1,a_2)$, glue in $(m-1)$ discs with boundary value $a_1$. The outcome is an annulus, with boundary values $(a_2,a_2)$, to which one can associate an element in ${{\mathbb Z}}/r$ as in Proposition \[th:x-annulus-a\](ii). The subspace ${\EuScript Y}_{\geq 0}^0(\underline{S},a_2,a_1,\dots,a_1,a_2)$ is defined by asking that to be zero. One can use this approach to prove unitality (as in Proposition \[th:gerstenhaber-algebra\]), and also to prove that the Gerstenhaber module structure is independent of the choice of $a_1$ and $a_2$ (as in Addendum \[th:all-a\]). More precisely, it is compatible with the canonical isomorphisms between different choices of $a_1$; the isomorphisms between different choices of $a_2$ are not canonical, but any such choice is a map of Gerstenhaber modules.
\[th:cop-out\] The indirect algebraic argument for unitality (based on first showing $e^2 = e$) used in Propositions \[th:gerstenhaber-algebra\] and \[th:gerstenhaber-module\] may strike the reader as a cop-out. We will therefore outline a geometric alternative, which uses the paths constructed in Lemma \[th:unit-path\]. Take the pair-of-pants surface $S_1$, and equip it with a flat connection whose holonomies around the boundary components, parametrized as in , are $(g_2(1),g_1(1),g_2(1))$. More precisely, this connection should have boundary values $(a_2,a_1,a_2)$, where $\mathrm{rot}(a_1) = 1$; and additionally, parallel transport along a path from $\epsilon_0(0)$ to $\epsilon_1(0)$ is $k(1)$, whereas parallel transport along a path from $\epsilon_1(0)$ to $\epsilon_2(0)$ is the identity ${\mathbbm{1}}$ (we refer to Figure \[fig:unit\], with $t$ set to $1$, for the precise picture). The TQFT operation associated to $S_2$ defines the product $H^*(a_1) \otimes H^*(a_2) \rightarrow H^*(a_2)$. Using the nonnegative path $g_1(t)$, one can equip the disc $S_2$ with a nonnegatively curved connection with boundary value $a_1$. Gluing together $S_1$ and $S_2$ yields the geometric description of the action of on $H^*(a_2)$ (to check this, one uses the fact that $k(1)$ has rotation number $0$, which is part of Lemma \[th:unit-path\]).
There is a one-parameter family of deformations of this construction, where the boundary holonomies on $S_1$ are $(g_2(t),g_1(t),g_2(t))$, and the parallel transport maps are $k(t)$ resp. the identity (see again Figure \[fig:unit\]). Correspondingly, on $S_2$ one has a connection with boundary holonomy $g_1(t)$. After gluing, one gets a one-parameter family of connections on the cylinder, whose boundary values can still be assumed to be $(a_2,a_2)$ (since the conjugacy class of the holonomies remains the same). For $t = 0$, the outcome is a flat connection on the cylinder, with the additional property that parallel transport along a path joining the two boundary circles is $k(0) = {\mathbbm{1}}$. Hence, the associated operation is the identity map on $H^*(a_2)$.
(0,0)![\[fig:unit\]The geometric version of the proof that $e$ is a unit (Remark \[th:cop-out\]). The flat connection on the pair-of-pants with the given parallel transport maps exists thanks to .](unit.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(3827,2274)(1186,-2084) (2551,-1186)[(0,0)\[lb\]]{} (3506,-2011)[(0,0)\[lb\]]{} (3511, 19)[(0,0)\[lb\]]{} (1201,-1186)[(0,0)\[lb\]]{} (4351,-836)[(0,0)\[lb\]]{} (3401,-1011)[(0,0)\[lb\]]{}
There are more restrictive notions of Gerstenhaber module than that used above, which are obtained by imposing a relation on $[x_1 x_2,x_3]$. The relation that holds in our context depends on the rotation number:
\[th:strange-relation\] The Gerstenhaber module structure from Proposition \[th:gerstenhaber-module\] satisfies $$\label{eq:strange-relation}
[x_1x_2,x_3] = (1-r)[x_1,x_2]x_3 + (-1)^{(|x_1|-1)|x_2|} x_2[x_1,x_3] + (-1)^{|x_1|} x_1[x_2,x_3].$$
The idea is the same as for , with simpler underlying geometry. For $1 \leq i < j \leq 3$, let $\lambda_{ij} \in H_1(\mathit{Conf}_3)$ be the class of a loop where the $i$-th and $j$-th point rotate anticlockwise once around each other (while the other point is far away and remains constant). Because of Lemma \[th:rot-bracket\], the geometric object underlying $[x_1x_2,x_3]$ is a loop of configurations whose homology class is $(1-r) \lambda_{12} + \lambda_{13} + \lambda_{23}$. The equality then amounts to giving each of those three summands a separate geometric representation.
As an analogue of Proposition \[th:strange-relation\] for $r = 0$, consider the Gerstenhaber algebra $A^* = \Lambda^*(TX)$ of polyvector fields on a manifold $X$, and the Gerstenhaber module structure of the space $M^* = \Omega^{-*}(X)$ of differential forms. The first nontrivial instances of this structure (omitting signs) are as follows: $$\label{eq:schouten}
\begin{array}{l|ll}
& \text{algebra action on $\theta \in M^*$} & \text{Lie action on $\theta \in M^*$} \\
\hline
\text{functions $f \in A^0$} & \theta \mapsto f\theta & \theta \mapsto df \wedge \theta \\
\text{vector fields $Z \in A^1$} & \theta \mapsto i_Z\theta & \theta \mapsto L_Z\theta
\end{array}$$ Then, for $x_1 = f$, $x_2 = Z$, $x_3 = \theta$ (and $r = 0$) becomes $$L_{fZ} \theta = (Z.f) \theta - i_Z (df \wedge \theta) + f (L_Z \theta),$$ which is indeed an easy consequence of the standard Cartan formalism. More generally, suppose that we look at a version of the Gerstenhaber module which consists of differential forms twisted by the $(-r)$-th power of canonical bundle (the bundle of top degree forms). With respect to , the difference is that the Lie action of $Z \in A^1$ is now written as (with respect to a local volume form $\eta$) $$\theta \otimes \eta^{-r} \longmapsto \Big( L_Z \theta - r \theta \frac{L_Z\eta}{\eta} \Big) \otimes \eta^{-r}.
$$ This satisfies an instance of for the given value of $r$. If, instead of differential topology, one considers the same formulae in algebraic geometry, this would be (roughly speaking) the $B$-model mirror of our symplectic geometry constructions.
For the next construction of operations, we switch to a more constrained setup. Fix some $a$ with $\mathrm{rot}(a) = 1$. All our boundary conditions will be covers $a^{r_i}$, where the degrees $r_i>0$ satisfy $r_1 + \dots + r_m - r_0 = m-1$. Moreover, the disc configurations will be such that all discs have centers on the imaginary axis, ordered (say) from bottom to top. We also restrict to the trivial choice of framings, meaning that $\rho_i > 0$. For each $i = 1,\dots,m$, fix a path $c_i$, going from $\epsilon_0(0) = 1 \in \partial_0 S$ to $\epsilon_i(1)$, and which lies in $S \cap \{\mathrm{re}(z) \geq 0\}$; this assumption fixes the homotopy class of the path (rel endpoints). As explained in Addendum \[th:partial-fix\], we can use such paths to define a map $$\label{eq:m-rotations}
(\mathrm{rot}_{c_1}, \dots, \mathrm{rot}_{c_m}): {\EuScript A}_{\mathit{flat}}(S,\{a_i\}) \longrightarrow {{\mathbb Z}}^m$$ which, by Proposition \[th:connections-rel-boundary\], is a weak homotopy equivalence. For any framed disc configuration in our class, there is a choice of $A$ for which vanishes. Such choices are unique up to homotopy, and moreover, they are compatible with gluing discs. This leads to the following:
\[th:bigraded-algebra\] Fix some $a$ with $\mathrm{rot}(a) = 1$. Then, $\bigoplus_{s \geq 0} H^*(a^{s+1})$ has the structure of a bigraded associative algebra, for which is the unit.
Because of our restrictions, the spaces of choices involved are weakly homotopy equivalent to ordered configuration spaces of points on the real line, which are of course contractible. A single configuration of two discs, with $(r_0 = s_1+s_2+1,\, r_1 = s_1+1,\, r_2 = s_2+1)$, defines the desired product structure. Associativity follows by gluing and considering the resulting three-disc configurations. Since the product $H^*(a) \otimes H^*(a^{s+1}) \rightarrow H^*(a^{s+1})$ agrees with the one previously defined in Proposition \[th:gerstenhaber-module\], we already know that $e$ is a left unit; the proof for right unitality is parallel.
Even though we will not discuss that explicitly, a refined version of our arguments coud be used to construct algebraic structures on the cochain level: for instance, by looking at families parametrized by associahedra, one can get a bigraded $A_\infty$-algebra structure underlying that from Proposition \[th:bigraded-algebra\].
Multivalued framings\[subsec:multivalued\]
------------------------------------------
For the final piece of algebraic structure we want to extract from our TQFT, it is convenient to think of disc configurations in a slightly different way.
- When choosing boundary values $a \in \Omega^1(S^1,{\mathfrak{g}})$, we allow only ones with $r = \mathrm{rot}(a) \geq 1$. Moreover, each such $a$ must be invariant under $1/r$-rotation of the circle. Equivalently, it must be an $r$-fold cover, in the sense of (of some underlying connection, which necessarily has rotation number $1$).
- Our disc configurations will come with multivalued framings, which are maps like , but for which the $\rho_i$ are determined only up to multiplication with $r_i$-th roots of unity; here, $r_i$ is the rotation number of the associated boundary value $a_i$.
Given a disc configuration, let ${\EuScript Y}_{\mathit{multi}, \geq 0}(S,\{a_i\})$ be the space of multivalued decorations, and ${\EuScript Y}_{\mathit{multi, flat}}(S,\{a_i\})$ the subspace where the connection is flat. The analogue of Proposition \[th:decorations\] (with a similar proof, which we omit) is the following:
\[th:multi-contractible\] If holds, both ${\EuScript Y}_{\mathit{multi,flat}}(S,\{a_i\})$ and ${\EuScript Y}_{\mathit{multi},\geq 0}(S,\{a_i\})$ are weakly contractible.
The relevant version of the gluing process goes as follows. Start with two families of configurations with multivalued decorations, parametrized by $P_1$ and $P_2$, respectively, and whose associated rotation numbers are $(r_{1,0},\dots,r_{1,m_1})$ and $(r_{2,0},\dots,r_{2,m_2})$. Assume additionally that $P_2$ comes with an (orientation-preserving) action of ${{\mathbb Z}}/r_{2,0}$, which is such that acting by the generator corresponds to rotating the disc configuration (including multivalued decorations) by $1/r_{2,0}$. Suppose that we have an $1 \leq i \leq m_1$ such that $r_{1,i} = r_{2,0}$, and such that the relevant boundary values agree. Gluing produces a family over a twisted product, which is a fibration $$\label{eq:twisted-product}
P_2 \longrightarrow P_1 \tilde\times_i P_2 \longrightarrow P_1.$$ More precisely, there is a ${{\mathbb Z}}/r_{1,i}$-covering $\tilde{P}_{1,i} \rightarrow P_1$, which corresponds to choosing lifts of the multivalued framing of the $i$-th disc to an actual framing. One then takes the quotient $$\label{eq:twisted-product-2}
\tilde{P}_{1,i} \times P_2 \longrightarrow P_1 \tilde\times_i P_2 \stackrel{\text{def}}{=} \tilde{P}_{1,i} \times_{{{\mathbb Z}}/r_{2,0}} P_2.$$ It is straightforward to see that the gluing process, defined a priori on $\tilde{P}_{1,i} \times P_2$, indeed descends to the quotient. Finally, note that if $P_1$ also carries an action of ${{\mathbb Z}}/r_{1,0}$ which rotates disc configurations, then the glued family inherits that action.
\[th:fractional-loop\] Consider two discs moving around each other, parametrized by $p \in S^1$. Using Proposition \[th:multi-contractible\], we can equip these discs with multivalued decorations, which vary smoothly along our loop, for any rotation numbers $(r_0, r_1, r_2)$ such that $r_0 = r_1+r_2-1$. The associated multivalued framings are, up to homotopy, $$\label{eq:exact-framings}
\left\{
\begin{aligned}
& \epsilon_1(t) = \rho_1 e^{2\pi i (1-r_2)r_1^{-1} p - 2\pi i t} =
\rho_1 e^{2\pi i (1-r_0r_1^{-1}) p - 2\pi i t}, \\
& \epsilon_2(t) = e^{2\pi i p} \zeta_2 + \rho_2 e^{2\pi i (1-r_1)r_2^{-1} p - 2\pi i t} \rho_2 z =
e^{2\pi i p} \zeta_2 + \rho_2 e^{2\pi i (1-r_0r_2^{-1}) p - 2\pi i t}.
\end{aligned}
\right.$$ (For the formula to make sense, $|\rho_1|, |\rho_2|$ need to be sufficiently small, and $\zeta_2 \neq 0$). In words, as the discs move, the first one rotates by a fractional amount $1-r_0/r_1$, and the second one by $1-r_0/r_2$.
This is the analogue of Lemma \[th:rot-bracket\], and is proved in the same way, by appealing to Lemma \[th:twist-loop\]; we omit the argument.
\[th:fractional-lie-family\] We may in fact choose the framings to be exactly given by . Then, rotating such a configuration by $1/r_0$ yields $$\label{eq:rotated-conf}
\left\{
\begin{aligned}
& e^{2\pi i r_0^{-1}} \epsilon_1(t) = \rho_1 e^{2\pi i r_1^{-1}} e^{2\pi i (1-r_0r_1^{-1}) (p+r_0^{-1}) - 2\pi i t}, \\
& e^{2\pi i r_0^{-1}} \epsilon_2(t) = e^{2\pi t}(p+r_0^{-1}) \zeta_2 + \rho_2 e^{2\pi i r_2^{-1}} e^{2\pi i (1-r_0r_2^{-1}) (p+ r_0^{-1}) - 2\pi i t}.
\end{aligned}
\right.$$ By definition, the $e^{2\pi i r_k^{-1}}$ factors on the right hand side of are irrelevant for multivalued framings. It follows that, after a suitable choice of connections, our family admits a ${{\mathbb Z}}/r_0$-action of the kind needed in the gluing process, namely $p \mapsto p+r_0^{-1}$.
Let’s assume that the coefficient ring of our TQFT satisfies ${{\mathbb Q}}\subset R$. Consider families of disc configurations with multivalued decorations. Given such a family, with parameter space $P$, there is a cover $\tilde{P} \rightarrow P$ with structure group ${{\mathbb Z}}/r_1 \times \cdots \times {{\mathbb Z}}/r_m$, $r_i = \mathrm{rot}(a_i)$, on which the framings become single-valued. We use those covers and our usual TQFT operation to define $$\label{eq:divide-operation}
\phi_{\mathit{multi},P} = r_1^{-1}\cdots r_m^{-1} \phi_{\tilde{P}}.$$ The induced cohomology level map $$\label{eq:multivalued-operation}
H^*(\phi_{\mathit{multi},P}): H^*(a_1) \otimes \cdots \otimes H^*(a_m) \longrightarrow H^{*-\mathrm{dim}(P)}(a_0)$$ is invariant under the action of ${{\mathbb Z}}/r_1 \times \cdots \times {{\mathbb Z}}/r_m$ on its source. If we assume in addition that our family carries an action of ${{\mathbb Z}}/r_0$ rotating the discs, then invariant under the action of that group on its target. To understand the behaviour of these invariants under gluing, we consider the operation associated to . If we take the families of decorated disc configurations over $\tilde{P}_1$ and $\tilde{P}_2$ and glue them together, then the outcome descends to $\tilde{P}_1 \times_{{{\mathbb Z}}/r_{2,0}} \tilde{P}_2$. As a consequence of the axiom concerning coverings, and the composition axiom, we get, in abbreviated notation, $$\phi_{\tilde{P}_1 \times_{{{\mathbb Z}}/r_{2,0}} \tilde{P}_2} = r_{2,0}^{-1} \phi_{\tilde{P}_1 \times \tilde{P}_2} = r_{2,0}^{-1} \phi_{\tilde{P}_1} \phi_{\tilde{P}_2},$$ and hence $$\phi_{\mathit{multi},P_1 \tilde{\times}_i P_2} = \phi_{\mathit{multi},P_1} \phi_{\mathit{multi},P_2}.$$ Here, “abbreviated notation” means that these composition formulae are shorthand for appropriate analogues of (with the same signs). We have been heading for the following application:
\[th:bigraded-lie\] Suppose that the coefficient ring $R$ is a ${{\mathbb Q}}$-algebra. For each $r > 0$, fix some $a^r \in \Omega^1(S^1,{\mathfrak{g}})$ with rotation number $r$, and which is an $r$-fold cover (not necessarily of the same underlying connection). Then, $\bigoplus_{s \geq 0} H^{*+1}(a^{s+1})^{{{\mathbb Z}}/(s+1)}$ has the structure of a bigraded Lie algebra.
Consider a loop of two-disc configurations as in Lemma \[th:fractional-loop\]. One gets a map $$H^*(\phi_{\mathit{multi},P}): H^*(a^{r_1}) \otimes H^*(a^{r_2}) \longrightarrow H^{*-1}(a^{r_0}),$$ where $r_0 = r_1+r_2-1$. By definition, this map is ${{\mathbb Z}}/r_1$-invariant, which means that it remains the same if we compose it with projection (defined by averaging over orbits) to the invariant part of $H^*(a^{r_1})$. In other words, no information is lost if we restrict it to that invariant part. The same applies to ${{\mathbb Z}}/r_2$ acting on $H^*(a^{r_2})$. Finally, the map takes value in the ${{\mathbb Z}}/r_0$-invariant part of $H^*(a^{r_0})$. We use it to define our Lie bracket.
Each term in the Jacobi identity corresponds to a family of disc configurations with multivalued decorations, parametrized by $T^2$. Thanks to Proposition \[th:multi-contractible\], we only have to check that these families satisfy the desired relation in the configuration space $\mathit{Conf}_3$. Let’s look at the expression $[x_1,[x_2,x_3]]$. Geometrically, this is given by gluing together two copies of the family from Lemma \[th:fractional-loop\], with rotation numbers $(r_{1,0} = r_0, r_{1,1} = r_1, r_{1,2} = r_2+r_3-1)$ and $(r_{2,0} = r_2+r_3-1, r_{2,1} = r_2, r_{2,2} = r_3)$ and parameters $p_1,p_2$. Again using Lemma \[th:fractional-loop\], the relevant covering is $$\tilde{P}_{1,2} = \{ (\tilde{p}_1, p_1) \in S^1 \times S^1 \;:\; (r_2+r_3-1)\tilde{p}_1 = (1-r_1) p_1 \}.$$ With this at hand, the families over $\tilde{P}_{1,2}$ and $P_2$ are: $$\label{eq:p12}
\begin{aligned}
&
\left\{
\begin{aligned}
& \epsilon_{1,1}(t) = \rho_{1,1} e^{2\pi i (2-r_2+r_3) r_1^{-1} p_1 - 2\pi i t}, \\
& \epsilon_{1,2}(t) = e^{2\pi i p_1} \zeta_{1,2} + \rho_{1,2} e^{2\pi i \tilde{p}_1 - 2\pi i t}.
\end{aligned}
\right.
\\ &
\left\{
\begin{aligned}
& \epsilon_{2,1}(t) = \rho_{2,1} e^{2\pi i (1-r_3)r_2^{-1} p_2 - 2\pi i t} , \\
& \epsilon_{2,2}(t) = e^{2\pi i p_2} \zeta_{2,2} + \rho_{2,2} e^{2\pi i (1-r_2)r_3^{-1} p_2 - 2\pi i t}.
\end{aligned}
\right.
\end{aligned}$$ The common action of ${{\mathbb Z}}/(r_2+r_3-1)$ by which we have to divide in is generated by $(\tilde{p}_1,p_1, p_2) \mapsto (\tilde{p}_1 + (r_2+r_3-1)^{-1}, p_1, p_2 - (r_2+r_3-1)^{-1})$. Hence, the quotient is parametrized by $(p_1, \tilde{p}_2 = \tilde{p}_1+p_2)$. The glued family is then of the form $$\label{eq:glue-3}
\left\{
\begin{aligned}
& \epsilon_1(t) = \rho_1 e^{2\pi i (2-r_2+r_3) r_1^{-1} p_1 - 2\pi i t}, \\
& \epsilon_2(t) = e^{2\pi i p_1} \zeta_2 + \rho_2 e^{2\pi i ((1-r_3) \tilde{p}_2 + (1-r_1) p_1)r_2^{-1} - 2\pi i t}, \\
& \epsilon_3(t) = e^{2\pi i p_1} \zeta_2 + e^{2\pi i \tilde{p}_2} \zeta_3 + \rho_3 e^{2\pi i ((1-r_1) p_2 + (1-r_2) \tilde{p}_2)r_3^{-1} - 2\pi i t}.
\end{aligned}
\right.$$ We have not spelled out the relation between the constants appearing here and those in ; the only relevant point is that $|\zeta_3| < |\zeta_2|$, so that the three discs in our configuration (in the given order) form a sun-earth-moon system. In particular, if we ignore framings, and just focus on the position of the celestial bodies, then those are independent of the rotation number. The same holds for the other terms of the Jacobi identity, which are just permuted versions of ; hence, the proof of that identity reduces to a standard argument.
The differentiation axiom\[sec:diff-axiom\]
===========================================
Following the model of [@seidel16], we now build an additional differentiation operator into the TQFT setup. Because of the intended application, there is also a shift in emphasis, from the cohomology level to the chain level (still remaining fairly ad hoc, rather than giving a description of all the resulting chain level operations).
The formalism
-------------
Assume that our coefficient ring $R$ comes with a derivation $\partial: R \rightarrow R$; and that all $C^*(a)$ come with additive graded endomorphisms $$\label{eq:qabla}
\begin{aligned}
& {\nabla\mkern-12mu/\mkern2mu}: C^*(a) \longrightarrow C^*(a), \\
& {\nabla\mkern-12mu/\mkern2mu}(fx) = f ({\nabla\mkern-12mu/\mkern2mu}x) + (\partial f) x \quad \text{for $f \in R$, $x \in C^*(a)$.}
\end{aligned}$$ Note that we do not ask that ${\nabla\mkern-12mu/\mkern2mu}$ should commute with the differential. These endomorphisms interact with the geometry as follows:
- The surface associated to a decorated disc configuration is allowed to optionally carry a marked point $Z \in S$ (in the exceptional case of the identity configuration, this is a point on the circle.) This additional datum behaves in the obvious way with respect to gluing (where of course only one of the two surfaces involved may carry a marked point). For a family of configurations over $P$, we similarly allow an optional smooth section $\underline{Z}: P \rightarrow \underline{S}$. If our family includes instances of the identity configuration, we ask that in some neighbourhood of the subset of $P$ where that configuration occurs, the section $\underline{Z}$ should take values in the boundary component $\partial_0\underline{S}$. For the associated TQFT operations, we keep the same axioms as before, treating the additional marked point as formally having degree $2$. This means that a family over a compact oriented $P$ with a $\underline{Z}$ section gives rise to an operation $$\label{eq:phi-map-with-z-point}
\phi_P: C^*(a_1) \otimes \cdots \otimes C^*(a_m) \longrightarrow C^{*-\mathrm{dim}(P)+2}(a_0).$$
- Consider the identity configuration, for some $a$, and turn this into a family over $P = S^1$ by adding the point $Z_p = e^{2\pi i p}$. We denote the associated operation by $$\label{eq:r-map}
R: C^*(a) \longrightarrow C^{*+1}(a),$$ and require that it should measure the failure of to be a chain map: $$\label{eq:differentiation-1}
R= {\nabla\mkern-12mu/\mkern2mu}d - d {\nabla\mkern-12mu/\mkern2mu}.$$
- Consider any family $\underline{S} \rightarrow P$, where none of the fibres is the identity configuration, and which does not come with a marked point (section). We can associate to this a new family, denoted by $\underline{S}_\bullet \rightarrow P_\bullet$, as follows. The new parameter space is the total space of the original family, $P_\bullet = \underline{S}$, and $\underline{S}_\bullet$ is the pullback of $\underline{S}$ to that space. This comes with a tautological section $\underline{Z}$. We call this process “adding a marked point in all possible places”, and require that the associated operation should be related to the original one by $$\label{eq:differentiation-2}
\phi_{P_\bullet}(x_1,\dots,x_m) = {\nabla\mkern-12mu/\mkern2mu}\phi_P(x_1,\dots,x_m) - \phi_P({\nabla\mkern-12mu/\mkern2mu}x_1,\dots,x_m) - \cdots - \phi_P(x_1,\dots,{\nabla\mkern-12mu/\mkern2mu}x_m).$$
The second and third part of the requirements above form the “differentiation axiom”. We will now explore some immediate consequences. Take the empty configuration ($m = 0$, which means that $S = D$), with the point $Z = 0 \in S$. After equipping this with a nonnegatively curved connection whose boundary value $a$ has rotation number $1$ (Proposition \[th:the-disc\]), we get a cocycle $$\label{eq:ks}
k \in C^2(a).$$ We call this the Kodaira-Spencer cocycle, and its class in $H^2(a)$ the Kodaira-Spencer class.
\[th:bracket-with-k\] Fix $a_1,a_2$ as in Proposition \[th:gerstenhaber-module\]. Consider a chain level version of the bracket constructed there, $$\label{eq:chain-bracket}
[\cdot,\cdot]: C^*(a_1) \otimes C^*(a_2) \longrightarrow C^{*-1}(a_2).$$ Let $k$ be the Kodaira-Spencer cocycle in $C^2(a_1)$. Then there is a chain homotopy $$\label{eq:rho-homotopy}
\begin{aligned}
& \rho: C^*(a_2) \longrightarrow C^*(a_2), \\
& d\rho(x) - \rho(dx) - R(x) + [k,x] = 0.
\end{aligned}$$
Take for $a = a_1$, and the family underlying ; glue them together; and then (temporarily) forget the marked point. The outcome is a family, still parametrized by $p \in P = S^1$, in which the surfaces $S_p$ are all annuli, carrying boundary parametrizations and nonnegatively curved connections, with boundary values $(a_2,a_2)$. Proposition \[th:x-annulus\](ii) and Addendum \[th:x-annulus-a\](ii) say that up to homotopy, our family is classified by a map $$\label{eq:nullhomotopic-s1}
P \longrightarrow \Gamma_r.$$ Assume first that $r \neq 0$, so that $\Gamma_r$ is discrete. Let’s look at a single value of the parameter. Then, the gluing construction precisely describes the module action of from Proposition \[th:gerstenhaber-module\]. In Remark \[th:cop-out\], we have given a direct geometric proof of the unitality of that action; that argument implies that gluing yields the trivial element of $\Gamma_r$, hence that is trivial. In the remaining case $r = 0$, what we have just said shows that takes values in the trivial connected component of $\Gamma_0 = {{\mathbb Z}}\times S^1$; one can then use Lemma \[th:rot-bracket\] to show that the map $\pi_1(P) \rightarrow \pi_1(\Gamma_0)$, which describes the $p$-dependence of the boundary parametrization of our annuli, is also trivial.
We now know that the glued family can be deformed to a one that is constant in $p$ except for the marked point, which moves once around the annulus. For concreteness, let’s give a specific formula for what we can deform our family to (as decorated disc configurations): $$\label{eq:124}
\left\{
\begin{aligned}
& D_1 = {\textstyle\frac{1}{4}}D, \\
& \epsilon_1(t) = {\textstyle\frac{1}{4}}e^{-2\pi i t}, \\
& A = \text{the pullback of $a_2$ by radial projection}, \\
& Z_p = {{\textstyle\frac{1}{2}}}e^{2\pi i p}.
\end{aligned}
\right.
$$ By deforming the constants ${\textstyle\frac{1}{4}}$ and ${{\textstyle\frac{1}{2}}}$ to $1$, can be connected to the family underlying $R$. Combining this with the previous deformation yields a family parametrized by an annulus, whose associated TQFT operation $\rho$ precisely satisfies , as an instance of .
In principle, these constructions should be supplemented by suitable uniqueness statements (since we are on the cochain level, this always means uniqueness up to chain homotopy). As a particularly simple instance, is homotopically unique, by which we mean the following: if $k$ and $\tilde{k}$ are the cochains arising from two choices of decorations, then $$\label{eq:secondary-k}
\tilde{k}-k = d\kappa;$$ and $\kappa \in C^1(a)$, which comes from a family of decorations parametrized by $[0,1]$, is itself unique up to adding coboundaries of elements in $C^0(a)$ (one could continue this with higher homotopies). This follows immediately from the contractibility statement in Proposition \[th:the-disc\]. Similarly (but with a slightly more complicated argument), $\rho$ is homotopically unique, and there is a relation between the maps $\rho$ obtained from different choices of cochain representative $k$. Finally, $k$ and $\rho$ are compatible, in a suitable sense, with the quasi-isomorphisms relating the complexes $C^*(a)$ for different choices of $a$. We omit the details.
The connection
--------------
From now on, our discussion no longer applies to the TQFT formalism in general, but requires a special additional assumption:
\[th:ks-vanishes\] Fix some $a \in \Omega^1(S^1,{\mathfrak{g}})$ with rotation number $1$. Assume that the class of $k$ in $H^2(a)$ is nullhomologous. Given that, we choose a bounding cochain $\theta \in C^1(a)$, $d\theta = k$. Two such cochains will be considered equivalent if their difference is a nullhomologous cocycle.
By adding arbitrary cocycles $t \in C^1(a)$ to $\theta$, the set of equivalence classes becomes an affine space over $H^1(a)$. Note that equivalence classes of $\theta$ can be transferred from one representative of $k$ to another: given and a choice of $\theta$ such that $d\theta = k$, one sets $\tilde\theta = \theta + \kappa$. Similarly, one can pass between different $a$.
\[th:hyperbolic-connection\] Suppose that Assumption \[th:ks-vanishes\] holds. Then, any group $H^*(a_2)$ carries a differentiation operator $$\label{eq:nabla-h}
\begin{aligned}
& \nabla: H^*(a_2) \longrightarrow H^*(a_2), \\
& \nabla(fx) = f \nabla x + (\partial f) x.
\end{aligned}$$ This operator depends on the equivalence class of the bounding cochain $\theta$; changing that by some $t \in H^1(a)$ has the effect of adding the Lie action $x \mapsto -[t,x]$ to .
The differentiation axiom and Lemma \[th:bracket-with-k\] show that the Kodaira-Spencer class encodes the essential failure of ${\nabla\mkern-12mu/\mkern2mu}$ to be a chain map. Correspondingly, under Assumption \[th:ks-vanishes\], one can remedy that failure: $$\label{eq:nabla}
\begin{aligned}
& \nabla: C^*(a_2) \longrightarrow C^*(a_2), \\
& \nabla x = {\nabla\mkern-12mu/\mkern2mu}x + \rho(x) - [\theta,x]
\end{aligned}$$ is a chain map. We define to be the induced operation on cohomology.
There is a uniqueness statement for , proving the independence of $\nabla$ from all choices made its construction, and its compatibility with the isomorphisms between the groups $H^*(a_2)$ for different choices of $a_2$. The argument is relatively straighforward, and we omit it. Going further along the lines of [@seidel16], one could consider the compatibility of $\nabla$ with each of the previously defined algebraic structures; since we have no immediate need or applications for such compatibility results, we prefer not to pursue that line of inquiry here.
Elliptic holonomy\[sec:elliptic\]
=================================
We now enlarge our TQFT by additionally allowing, as boundary values, connections with elliptic (which by definition still means nontrivial) holonomy. This allows us to incorporate the parallel theory from [@seidel16]. There is little new in the strategies of proof; we will therefore focus on a few basic technical statements, and then on the main result, which is the comparison theorem for connections (Proposition \[th:comparison-of-connections\]).
Spaces of nonnegative paths, revisited\[subsec:additional\]
-----------------------------------------------------------
We temporarily return to the subject of Section \[sec:sl2\]. Our first observation concerns the issue of perturbing nonnegative paths to positive ones, while keeping the conjugacy class of the endpoints the same. We have encountered this implicitly before, for instance, in the proof of Lemma \[th:short-path\]: there, our path ended in the hyperbolic locus, and we used the fact that positive motion on hyperbolic conjugacy classes is unconstrained. Here is an alternative approach, which uses the elliptic locus instead:
\[th:make-positive\] Let $g(t)$ be a nonconstant nonnegative path, such that $g(1)$ is elliptic. Then there is a deformation $g_s(t)$, with $g_0(t) = g(t)$, such that: $$\label{eq:per-pos}
\left\{
\begin{aligned}
& g_s(0) = g(0), \\
& \text{$g_s(1)$ lies in the same conjugacy class as $g(1)$,} \\
& \text{$g_s(t)$ is a positive path, for all $s>0$.}
\end{aligned}
\right.$$ The same statement applies to families of paths.
This is an adaptation of the argument from Lemma \[th:path\]. Let’s first assume, for simplicity, that our path remains entirely inside the elliptic locus, so that $g(t)$ is a rotation with angle $\theta(t) \in (0,\pi)$. Because the path is nonnegative and not constant, we know from that $\theta(1) > \theta(0)$. Consider time-dependent functions . There is such a $\gamma_t$ which additionally satisfies $g'(t) = g(t)\gamma_t(g(t))$. There is also one more choice $\tilde\gamma_t$, obtained from Lemma \[th:bound-vector-fields\](i) by rescaling as in , which has the additional property that $\tilde\gamma_t$ takes values in ${\mathfrak{g}}_{> 0}$ whenever $\theta'(t) > 0$. Take the paths $\tilde{g}_s(t)$ obtained by solving the ODE $$\tilde{g}_s(0) = g(0), \quad \tilde{g}_s'(t) = \tilde{g}_s(t)\big((1-s)\gamma_t(g_s(t)) + s\tilde{\gamma}_t(g_s(t))\big).$$ By construction, these move through the same conjugacy classes as the original path, to which they specialize for $s = 0$. For $s>0$, the path $\tilde{g}_s$ is positive at all times $t$ where $\theta'(t)>0$. Now we “redistribute the positivity”: $$\begin{aligned} \label{eq:redistribute}
& g_s(t) =\begin{pmatrix} \cos(\alpha_s(t)) & -\sin(\alpha_s(t)) \\ \sin(\alpha_s(t)) & \cos(\alpha_s(t))
\end{pmatrix} \tilde{g}_s(t), \\
&
\text{where} \quad
\left\{
\begin{aligned}
& \alpha_0(t) = 0, \\
& \alpha_s(0) = \alpha_s(1) = 0, \\
& \text{for $s>0$, $\alpha_s'(t)>0$ at all $t$ where $\tilde{g}_s(t)$ is not (strictly) positive.}
\end{aligned}
\right.
\end{aligned}$$ If we choose $\alpha_s$ sufficiently small, will be a positive path for all $s>0$, satisfying .
In the general case, fix some $T$ such that $g|[T,1]$ is elliptic; one can also arrange that there is a point in $[T,1]$ where the angle of rotation has positive derivative. Now, apply the first part of the argument to $g|[T,1]$, but multiplying $\tilde{\gamma}_t$ by a cutoff function which vanishes on $[0,T]$. The outcome is a family of paths $\tilde{g}_s$ which agree with the original $g$ on $[0,T]$, and which are positive somewhere in $[T,1]$. We then apply the same “redistribution” argument to those paths. The parametrized situation is similar.
\[th:elliptic-to-elliptic\] Fix elliptic conjugacy classes $C_0,C_1 \subset G$. Let $r_0,r_1 \in {{\mathbb R}}\setminus {{\mathbb Z}}$ be rotation numbers of some lifts of those classes to $\tilde{G}$, chosen so that $(r_0,r_1)$ contains exactly one integer, and set $\rho = r_1-r_0$. Then, ${\EuScript P}_{\geq 0}(G,C_0,C_1)^\rho$ is weakly contractible.
Lemma \[th:make-positive\] shows that ${\EuScript P}_{\geq 0}(G,C_0,C_1)^\rho$ is weakly homotopy equivalent to its subspace ${\EuScript P}_{>0}(G,C_0,C_1)^\rho$ of positive paths, so we’ll consider that instead. From Figure \[fig:traffic\], one sees that paths in ${\EuScript P}_{>0}(G,C_0,C_1)^\rho$ can meet the trivial element ${\mathbbm{1}}$ at most once. Let’s write $$\label{eq:pcc}
{\EuScript P}_{>0}(G,C_0,C_1)^\rho = {\EuScript S}_{\mathit{open}} \cup {\EuScript S}_{\mathit{id}},$$ where ${\EuScript S}_{\mathit{id}}$, the subspace of paths which go through ${\mathbbm{1}}$, is a codimension two submanifold; and ${\EuScript S}_{\mathit{open}}$ is its complement.
[**Claim (i):**]{} [*${\EuScript S}_{\mathit{id}}$ is weakly contractible.*]{}
Except for the time $T \in (0,1)$ where $g(T) = {\mathbbm{1}}$, a path $g(t)$ in ${\EuScript S}_{\mathit{id}}$ must remain in the elliptic locus. To first order near $T$, the path is determined by $g'(T) \in {\mathfrak{g}}_{>0}$, which is a contractible choice. On the other hand, Lemma \[th:path\](i) applies to $g|[0,T-\epsilon]$ and $g|[T+\epsilon,1]$ for any $\epsilon>0$, which are therefore determined uniquely up to homotopy by their endpoints $g(T \pm \epsilon)$. Combining these two ideas yields the claimed fact.
[**Claim (ii):**]{} [*${\EuScript S}_{\mathit{open}}$ is weakly homotopy equivalent to a circle.*]{}
Any path $g(t)$ in ${\EuScript S}_{\mathit{open}}$ has the following structure: $$\text{for some $0<T_0<T_1< 1$,}\;
\left\{
\begin{aligned}
& \text{$g(t)$ is elliptic for $t<T_0$;} \\
& \text{$g(T_0)$ is a nonpositive parabolic;} \\
& \text{$g(t)$ is hyperbolic for $T_0<t<T_1$;} \\
& \text{$g(T_1)$ is a nonnegative parabolic;} \\
& \text{$g(t)$ is elliptic for $T_1<t$.}
\end{aligned}
\right.$$ A more precise form of our claim says that, by evaluating such paths at a point in $(T_0,T_1)$ (smoothly depending on the path), we get a weak homotopy equivalence between ${\EuScript S}_{\mathit{open}}$ and the hyperbolic locus $G_{\mathit{hyp}}$. Indeed, by breaking up the path at such a point, we can reduce our claim to (a fibre product of) two copies of Lemma \[th:short-path\], in a version for positive rather than nonnegative paths.
With that settled, we will now show that ${\EuScript P}_{>0}(G,C_0,C_1)^\rho$ is weakly contractible. Take a path in ${\EuScript S}_{\mathit{id}}$, and a two-parameter family of perturbations of that path, which moves it in all directions transverse to ${\EuScript S}_{\mathit{id}}$. As a local model near the identity element, if the original path $g(t)$ is a standard rotation by $t$, the family could be $$g_{\delta,\epsilon}(t) = \exp \left( \begin{smallmatrix} \epsilon & \delta-t \\ \delta+t & -\epsilon \end{smallmatrix} \right).$$ For $(\delta,\epsilon)$ nontrivial, $g_{\delta,\epsilon}(0)$ is a hyperbolic element, whose eigenvector for the larger eigenvalue $\exp(\sqrt{\delta^2+\epsilon^2})$ is $$\begin{pmatrix} (\sqrt{\delta^2+\epsilon^2} + \epsilon)^{1/2} \\ \mathrm{sign}(\delta) (\sqrt{\delta^2+\epsilon^2} - \epsilon)^{1/2}\end{pmatrix}.$$ As $(\delta,\epsilon)$ moves in a small circle around $0$, that eigenvector goes once (up to sign, since there’s no preferred orientation of the $(\delta,\epsilon)$ plane) around ${{\mathbb R}}P^1$. The same holds for general paths and perturbations (by a deformation argument). This shows that the link of the stratum ${\EuScript S}_{\mathit{id}}$ yields a weak equivalence from the circle to ${\EuScript S}_{\mathit{open}}$. Weak contractibility of ${\EuScript P}_{>0}(G,C_0,C_1)^\rho$ follows from that and the previous claims.
\[th:ehe\] Fix conjugacy classes $C_0,C_1,C_2$ which are, respectively, elliptic, hyperbolic, and elliptic. For the elliptic ones, we require that they correspond to rotations with angles $\theta_0 > \theta_2$ in $(0,\pi)$. Consider $g_i \in C_i$ together with a nonnegative path $h(t)$ from $h(0) = g_1g_2$ to $h(1) = g_0$, such that if we take lifts of our elements to $\tilde{g}_i \in \tilde{G}$, compatible with the product and path, the numbers $\mathrm{rot}(\tilde{g}_0)$ and $\mathrm{rot}(\tilde{g}_1)+\mathrm{rot}(\tilde{g}_2)$ lie in the same component of ${{\mathbb R}}\setminus {{\mathbb Z}}$. Then, the space of all such $(g_0,g_1,g_2,h(t))$ is weakly homotopy equivalent to $C_1$, by projecting to $g_1$.
We will consider positive paths instead of nonnegative ones. This does not affect the weak homotopy type, by the same argument as in Lemma \[th:short-path\], applied to the path $h(t)g_2^{-1}$ (which goes from the hyperbolic element $g_1$ to $g_0g_2^{-1}$). Suppose (without loss of generality) that $\mathrm{rot}(\tilde{g}_0) \in (0,1)$, $\mathrm{rot}(\tilde{g}_2) \in (-1,0)$, which means that $\mathrm{rot}(\tilde{g}_1) = 1$. One can apply the elementary considerations from Example \[th:2-elliptic\] to the pair $(\tilde{g}_0,\tilde{g}_2^{-1})$, and that describes all the possible options for $\tilde{g}_0\tilde{g}_2^{-1}$. In particular, we see that the function $(\tilde{g}_0,\tilde{g}_2) \mapsto \mathrm{tr}(g_0g_2^{-1})$ is a submersion everywhere except at its maximum value (which is achieved when $g_0,g_2$ commute). Let’s decompose the space ${\EuScript S}$ of all $(g_0,g_1,g_2,h(t))$ into three parts ${\EuScript S}_{\mathit{ell}}$, ${\EuScript S}_{\mathit{para}}$, ${\EuScript S}_{\mathit{hyp}}$, depending on the conjugacy type of $g = g_0g_2^{-1}$. The two and last subsets are open, and ${\EuScript S}_{\mathit{para}}$ is a hypersurface separating them.
[**Claim (i):**]{} [*The projection ${\EuScript S}_{\mathit{hyp}} \rightarrow C_1$ is a weak homotopy equivalence.*]{}
Given any hyperbolic element $g$, there is a contractible space of positive paths from $C_1$ to $g$ along which the rotation number is constant, by (the positive path version of) Lemma \[th:pre-short\]. On the other hand, by Example \[th:2-elliptic\], there is an ${{\mathbb R}}$ worth of choices of $(g_0,g_2)$ such that $g_0g_2^{-1} = g$. Hence, the map $g_0g_2^{-1}: {\EuScript S}_{\mathit{hyp}} \rightarrow G_{\mathit{hyp}}$ is a weak homotopy equivalence; but it is also homotopic to the map appearing in our claim.
[**Claim (ii):**]{} [*The projection ${\EuScript S}_{\mathit{para}} \rightarrow C_1$ is a weak homotopy equivalence.*]{}
This is similar to the previous case. It requires a statement about positive paths from hyperbolic to nonnegative parabolic elements, but that can be obtained from Lemma \[th:pre-short\] and a separate consideration of positive paths near the parabolic endpoint, as in Lemma \[th:short-path-3\].
[**Claim (iii):**]{} [*The projection ${\EuScript S}_{\mathit{ell}} \rightarrow C_1$ is a weak homotopy equivalence.*]{}
Let’s work with a fixed hyperbolic $g_1$. Given an elliptic $g$, the space of positive paths connecting $g_1$ to $g$, along which the rotation number increases by less than $1$, is weakly contractible, by (the positive path version of) Lemma \[th:short-path\]. Hence, the homotopy fibre of ${\EuScript S}_{\mathit{ell}} \rightarrow C_1$ is weakly homotopy equivalent to the space of $(g_0,g_2)$ such that $g_0g_2^{-1}$ is elliptic. By the computation in Example \[th:2-elliptic\], this space is an open disc.
Putting together the three claims above yields the desired result.
Spaces of connections, revisited
--------------------------------
As in Section \[sec:connections\], statements about nonnegative paths translate into ones about nonnegatively curved connections. We use the same notation ${\EuScript A}_{\geq 0}(S,\{a_i\})$ as in , but now allowing boundary values $a_i$ with elliptic as well as hyperbolic holonomy.
\[th:grow-elliptic\] Let $S = [0,1] \times S^1$. As boundary conditions, fix $a_0,a_1 \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic holonomy, such that $\mathrm{rot}(a_0) \geq \mathrm{rot}(a_1)$, and where the interval $(\mathrm{rot}(a_1), \mathrm{rot}(a_0))$ contains at most one integer. Then ${\EuScript A}_{\geq 0}(S,a_0,a_1)$ is contractible.
\[th:mixed-continuation\] (i) Let $S = [0,1] \times S^1$. As boundary conditions, fix $a_0 \in \Omega^1(S^1,{\mathfrak{g}})$ with hyperbolic holonomy, $a_1 \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic holonomy, and such that $\mathrm{rot}(a_1) \in (\mathrm{rot}(a_0)-1,\mathrm{rot}(a_0))$. Then ${\EuScript A}_{\geq 0}(S,a_0,a_1)$ is weakly contractible.
\(ii) The same is true if $a_0$ has elliptic holonomy, $a_1$ has hyperbolic holonomy, and $\mathrm{rot}(a_0) \in (\mathrm{rot}(a_1),\mathrm{rot}(a_1)+1)$.
Both statements are proved by the same strategy as Proposition \[th:nonnegative-connection\]. The relevant results about path spaces are: for Proposition \[th:grow-elliptic\], Lemma \[th:path\](i) or Lemma \[th:elliptic-to-elliptic\] (depending on whether the relevant interval contains an integer or not); and for Proposition \[th:mixed-continuation\], Lemma \[th:short-path\].
\[th:elliptic-disc\] Let $S$ be a disc, with the boundary parametrized clockwise. As boundary condition, fix $a \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic holonomy and $\mathrm{rot}(a) \in (0,1) \cup (1,2)$. Then ${\EuScript A}_{\geq 0}(S,a)$ is weakly contractible.
One can reduce Proposition \[th:elliptic-disc\] to Lemma \[th:path\](i) or Lemma \[th:elliptic-to-elliptic\] (depending on which interval $\mathrm{rot}(a)$ lies in), just like Proposition \[th:the-disc\] was derived (via Lemma \[th:short-path-3\]) from Lemma \[th:short-path\]. Our final result follows from Lemma \[th:ehe\], in the same way as in Proposition \[th:nonnegative-connection-4\]:
\[th:ehe-pants\] Let $S$ be a pair-of-pants, with boundary parametrizations as in Proposition \[th:nonnegative-connection-4\] (specialized to $m = 2$). As boundary conditions, take $a_0, a_2 \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic holonomy, and $a_1 \in \Omega^1(S^1, {\mathfrak{g}})$ with hyperbolic holonomy, such that for some $r \in {{\mathbb Z}}$, $$\mathrm{rot}(a_0) \in (\mathrm{rot}(a_2)+1,r+2), \;\; \mathrm{rot}(a_1) = 1, \;\; \mathrm{rot}(a_2) \in (r,r+1).$$ Then ${\EuScript A}_{\geq 0}(S, a_0,a_1,a_2)$ is weakly contractible.
The formal setup
----------------
We continue in the same axiomatic framework as in Section \[sec:tqft\], except that now, the boundary conditions are allowed to have holonomies that are elliptic or hyperbolic. This gives rise to additional cohomology groups $H^*(a)$, whose basic properties can be summarized as follows:
\[th:only-elliptic\] (i) Consider connections $a \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic holonomy. Up to canonical isomorphism, $H^*(a)$ depends only on $\mathrm{rot}(a) \in {{\mathbb R}}\setminus {{\mathbb Z}}$.
\(ii) Each such group carries a degree $-1$ endomorphism, the BV operator $\Delta$.
\(iii) For $a_0, a_1$ as in Proposition \[th:grow-elliptic\], there is a canonical continuation map $C: H^*(a_1) \rightarrow H^*(a_0)$.
\(iv) If $\mathrm{rot}(a) \in (0,1) \cup (1,2)$, there is a canonical element $e \in H^0(a)$.
Parts (i) and (ii) can be constructed using flat connections on an annulus. The third part, obviously, uses Proposition \[th:grow-elliptic\]; and (iv) follows from Proposition \[th:elliptic-disc\]. Our next topic is the comparison between the elliptic and hyperbolic groups $H^*(a)$.
\[th:mixed-continuation-2\] (i) Take $a_0,a_1$ as in Proposition \[th:mixed-continuation\](i). Then there is a canonical continuation map $C: H^*(a_1) \rightarrow H^*(a_0)$. In the case where $\mathrm{rot}(a_0) = 1$ and $\mathrm{rot}(a_1) \in (0,1)$, this takes the element $e \in H^0(a_1)$ from Proposition \[th:only-elliptic\](iv) to the previously constructed element of the same name in $H^0(a_0)$.
\(ii) For $a_0,a_1$ as in Proposition \[th:mixed-continuation\](ii), there is a canonical continuation map as before. If $\mathrm{rot}(a_0) \in (1,2)$ and $\mathrm{rot}(a_1) = 1$, this takes the element in $H^0(a_1)$ to its counterpart from Proposition \[th:only-elliptic\](iv).
\(iii) Composition of the maps from (i) and (ii), in either order, yields the continuation maps from , respectively from Proposition \[th:only-elliptic\](iii). Hence, in the limit we get an isomorphism $$\label{eq:lim-lim}
\underrightarrow{\lim}_{r \in {{\mathbb Z}}}\, H^*(a^r) {\cong}\underrightarrow{\lim}_{r \in {{\mathbb R}}\setminus {{\mathbb Z}}}\, H^*(a^r).$$ Here, on the left hand side of , we choose for each $r \in {{\mathbb Z}}$ a connection $a^r$ with hyperbolic holonomy and rotation number $r$; the direct limit is independent of all choices up to canonical isomorphism, thanks to the properties of continuation maps (Proposition \[th:canonical-continuation\]). On the right hand side, we do the same for elliptic holonomies.
\(iv) The map factors through the BV operator. More precisely, given $a_0,a_1,a_2$ which, respectively, have hyperbolic, elliptic, and hyperbolic holonomy, and such that $\mathrm{rot}(a_2) < \mathrm{rot}(a_1) < \mathrm{rot}(a_0) = \mathrm{rot}(a_2)+1$, we have a commutative diagram $$\xymatrix{
\ar@/_1pc/[rrr]_-{D}
H^*(a_2) \ar[r]^-{C} & H^*(a_1)
\ar[r]^-{\Delta}
& \ar[r]^-{C} H^{*-1}(a_1) &
H^{*-1}(a_0).
}$$
Existence and uniqueneness of these “mixed” continuation maps follows from Proposition \[th:mixed-continuation\]. The statements about the unit elements are consequences of Propositions \[th:nonnegative-connection-3\] and \[th:elliptic-disc\]. In part (iii), the composition of continuation maps is given geometrically by connections on the annulus to which Propositions \[th:nonnegative-connection-2\] or \[th:grow-elliptic\] apply. The same applies to (iv), except that in that case, we are dealing with families of connections in which the parametrization of one of the boundary components rotates.
Let’s return for a moment to considering just connections with elliptic holonomy. The simplest way to approach the construction of further operations on the associated groups $H^*(a)$ is to significantly constrain the class of connections involved. Concretely, let’s consider only those $a = a^r \in \Omega^1(S^1,{\mathfrak{g}})$ of the form $$\label{eq:standard-connection}
a^r = r \alpha\, \mathit{dt}, \quad r \in {{\mathbb R}}\setminus {{\mathbb Z}}, \;\; \alpha = \left( \begin{smallmatrix} 0 & -\pi \\ \pi & 0 \end{smallmatrix} \right) \in {\mathfrak{g}}_{>0}.$$ These obviously satisfy $\mathrm{rot}(a^r) = r$. Similarly, on any surface , we consider $$\label{eq:d-beta}
A = \beta \otimes \alpha \in \Omega^1(S,{\mathfrak{g}}), \quad \text{for some } \beta \in \Omega^1(S) \text{ with } d\beta \leq 0.$$ Additionally, the restriction of $\beta$ to each boundary circle must lie in the class , which means that it has constant coefficients (with respect to the given parametrization) and integral in ${{\mathbb R}}\setminus {{\mathbb Z}}$. The inequality in expresses the nonnegative curvature condition (see for the sign). The advantage of this framework is that is a convex condition, hence the choice of $\beta$ always lies in a contractible space (this is what one usually encounters in discussions of the operations on symplectic cohomology, see e.g. [@seidel07 Section 8] or [@ritter10]). Moreover, the choice of boundary parametrizations now becomes independent of that of the connection. By applying this idea to the disc, annulus, and pair-of-pants, one gets operations $$\begin{aligned}
\label{eq:standard-unit}
& e \in H^0(a^r), && r>0, \\
\label{eq:standard-continuation}
& C: H^*(a^{r_1}) \longrightarrow H^*(a^{r_0}), && r_0 \geq r_1, \\
\label{eq:standard-product}
& \cdot: H^*(a^{r_1}) \otimes H^*(a^{r_2}) \longrightarrow H^*(a^{r_0}), && r_0 \geq r_1+r_2.\end{aligned}$$ The product is associative and commutative in an appropriate sense, and inserting into one of its inputs recovers . Of course, and agree with the previously defined structures of the same name, in their domains of definition (here, we are using special choices of nonnegatively curved connections, whereas before we allowed any choice and used more sophisticated methods to prove well-definedness). We will be particularly interested in a further family of operations, the twisted brackets $$\label{eq:twisted-bracket}
[\cdot,\cdot]_c : H^*(a^{r_1}) \otimes H^*(a^{r_2}) \longrightarrow H^{*-1}(a^{r_0}), \quad r_0 \geq r_1+r_2, \;\; c \in {{\mathbb Z}}.$$ These are obtained from a one-parameter family of two-disc configurations, in which: the two discs move once around each other; the boundary parametrization of the first disc rotates $-c$ times, while that of the second disc remains constant. Concretely, one could take the dependence on the parameter $p \in P = S^1$ to be $$\left\{
\begin{aligned}
& \epsilon_1(t) = \zeta_1 e^{2\pi i p} + \rho_1 e^{-2\pi i (t + cp)}, \\
& \epsilon_2(t) = \rho_2 e^{-2\pi i t}.
\end{aligned}
\right.$$ In the case $c = 0$, we write $[\cdot,\cdot]$, without the superscript. The bracket for general $c$ is related to that special case as in .
A family of connections
-----------------------
Into the enlarged framework including elliptic holonomy, we now re-introduce the differentiation axiom, as in Section \[sec:diff-axiom\]. This yields Kodaira-Spencer classes for connections with both hyperbolic and elliptic holonomy, related to each other in the same way as the unit elements in Proposition \[th:mixed-continuation-2\].
\[th:bracket-with-elliptic-k\] Fix $r_0, r_1, r_2 \in {{\mathbb R}}\setminus {{\mathbb Z}}$, such that $r_1>0$ and $r_0 \geq r_1+r_2$, and take $a^{r_i}$ as in . Consider the following structures on the associated complexes $C^*(a^{r_i})$: $$\left\{
\begin{aligned}
& \text{the Kodaira-Spencer cocycle $k \in C^2(a^{r_1})$,} \\
& \text{the continuation map $C: C^*(a^{r_2}) \rightarrow C^*(a^{r_0})$,} \\
& \text{the bracket $[\cdot,\cdot]_c: C^*(a^{r_1}) \otimes C^*(a^{r_2}) \rightarrow C^{*-1}(a^{r_0})$, for some $c \in {{\mathbb Z}}$,} \\
& \text{the map $R: C^*(a^{r_2}) \rightarrow C^{*+1}(a^{r_2})$.}
\end{aligned}
\right.$$ Then there is a chain homotopy $$\label{eq:rho-homotopy-elliptic}
\begin{aligned}
& \rho_c: C^*(a^{r_2}) \longrightarrow C^*(a^{r_0}), \\
& d\rho_c(x) - \rho_c(dx) - C(R(x)) + [k,x]_c \\ & \quad = d\rho_c(x) - \rho_c(dx) - C {\nabla\mkern-12mu/\mkern2mu}(dx) + d C{\nabla\mkern-12mu/\mkern2mu}(x) + [k,x]_c= 0.
\end{aligned}$$
This is the analogue of Lemma \[th:bracket-with-k\]. We omit the proof, which uses the same geometric construction as before; the details are actually simpler this time, because they just amount to choosing suitable one-forms .
\[th:ks-vanishes-2\] For some $r_1 \in {{\mathbb R}}^{>0} \setminus {{\mathbb N}}$, the Kodaira-Spencer cocycle $k \in C^2(a^{r_1})$ is nullhomologous. Given that, we choose a bounding cochain $\theta \in C^1(a^{r_1})$, $d\theta = k$.
\[th:elliptic-connection-2\] Suppose that Assumption \[th:ks-vanishes-2\] holds. Fix $r_0, r_2 \in {{\mathbb R}}\setminus {{\mathbb Z}}$ such that $r_0 \geq r_1+r_2$. Then, there is a family of “connections” with respect to the continuation map $C: H^*(a^{r_2}) \rightarrow H^*(a^{r_0})$, indexed by $c \in {{\mathbb Z}}$. These are maps $$\label{eq:nabla-h-elliptic}
\begin{aligned}
& \nabla_c: H^*(a^{r_2}) \longrightarrow H^*(a^{r_0}), \\
& \nabla_c(fx) = f\, \nabla_c x + (\partial f) C(x).
\end{aligned}$$
This is the counterpart of Proposition \[th:hyperbolic-connection\], using the bracket and the associated homotopy : the chain map underlying is $$\label{eq:elliptic-nabla}
\nabla_c x = C({\nabla\mkern-12mu/\mkern2mu}x) + \rho_c(x) - [\theta,x]_c.$$ We will not discuss how $\nabla_c$ depends on $c$ (but see [@seidel16 Section 6], which considers the relation between $c = -1$ and $c = 0$; the general case would work in the same way).
\[th:comparison-of-connections\] Suppose that Assumption \[th:ks-vanishes\] holds (and hence, so does Assumption \[th:ks-vanishes-2\] for any $r_1>1$). Take $r_0, r_2 \in {{\mathbb R}}\setminus {{\mathbb Z}}$ such that $r_1 = r_0 - r_2 \in (1,2)$, and where the interval $(r_0,r_2)$ contains exactly one integer $r$. Fix some $a \in \Omega^1(S^1,{\mathfrak{g}})$ whose holonomy is hyperbolic, and which has rotation number $r$. Then, for appropriately coordinated choices of bounding cochains, for $c = r-1$ factors through , as follows: $$\label{eq:comparison-of-connections}
\xymatrix{
\ar@/_1pc/[rrr]_-{\nabla_c}
H^*(a^{r_2}) \ar[r]^-{C} & H^*(a)
\ar[r]^-{\nabla}
& \ar[r]^-{C} H^*(a) &
H^*(a^{r_0}).
}$$
Each step in the proof is a fairly straightforward deformation argument, but there are a number of them. Part of the expository challenge is keeping the notation unambiguous, which we will attempt to do by adding rotation numbers as superscripts everywhere. In particular, for the given $r_i \in {{\mathbb R}}\setminus {{\mathbb Z}}$ ($i = 0,1,2$), we keep the notation $a^{r_i}$ for the connections (with elliptic holonomy) defined in ; but now also write $a^r = a$ for the connection with (hyperbolic holonomy and) rotation number $r \in {{\mathbb Z}}$, chosen arbitrarily, which appears in . Additionally, we choose another connection $a^1$ with (hyperbolic holonomy and) rotation number $1$.
(0,0)![\[fig:relate-k\]The family of surfaces underlying .](relate-k.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(5593,915)(2011,-3214) (4776,-3186)[(0,0)\[lb\]]{} (4201,-3186)[(0,0)\[lb\]]{} (2701,-3186)[(0,0)\[lb\]]{} (2046,-2661)[(0,0)\[lb\]]{} (3601,-2661)[(0,0)\[lb\]]{} (6901,-3186)[(0,0)\[lb\]]{} (5951,-2661)[(0,0)\[lb\]]{}
As a preliminary step, we need to explain what “appropriately coordinated choices of bounding cochains” means. Two version of the Kodaira-Spencer cocycle are relevant: $$\begin{aligned}
\label{eq:kkk}
& k^1 \in C^2(a^1) \quad \;\;\text{(hyperbolic case),} \\
\label{eq:kkk1}
& k^{r_1} \in C^2(a^{r_1}) \quad \text{(elliptic case).} \end{aligned}$$ These are related by $C^{r_1,1}: C^*(a^1) \rightarrow C^*(a^{r_1})$. More precisely, there is a cochain $$\label{eq:kkk2}
\begin{aligned}
& \kappa^{r_1} \in C^1(a^{r_1}), \\
& k^{r_1} = C^{r_1,1}(k^1) + d\kappa^{r_1},
\end{aligned}$$ whose geometric construction is shown schematically in Figure \[fig:relate-k\]. It relies on Proposition \[th:elliptic-disc\], which says that if we glue together the surfaces which define $C^{r_1,1}$ and $k^1$, the outcome is a surface which (together with its nonnegatively curved connection) can be deformed to that underlying $k^{r_1}$. We assume that a choice of bounding cochain $\theta^1 \in C^1(a^1)$ for is given, and then take that for to be $$\label{eq:kkk3}
\theta^{r_1} = C^{r_1,1}(\theta^1) + \kappa^{r_1} \in C^1(a^{r_1}).$$
In our enhanced notation, the definition of the “elliptic” connection reads as follows: $$\label{eq:elliptic-nabla-2}
\nabla_c^{r_0,r_2} x = C^{r_0,r_2} ({\nabla\mkern-12mu/\mkern2mu}^{r_2} x) + \rho_c^{r_0,r_2}(x) - [\theta^{r_1},x]^{r_0,r_1,r_2}_c.$$ Our first task is to break up the continuation map in into two pieces. As a consequence of Proposition \[th:grow-elliptic\], the following diagram of continuation maps is homotopy commutative: $$\label{eq:c-is-cc}
\xymatrix{
\ar@/_1pc/[rrrr]_-{C^{r_0,r_2}}
C^*(a^{r_2}) \ar[rr]^-{C^{r,r_2}} && C^*(a^r) \ar[rr]^-{C^{r_0,r}} && C^*(a^{r_0}).
}
$$ Using Proposition \[th:grow-elliptic\] in the same way, one can modify $\rho_c^{r_0,r_2}$ so that it satisfies a version of where the continuation map involved is $C^{r_0,r} C^{r,r_2}$ instead of $C^{r_0,r_2}$. Denote this new version by $\rho_{\mathit{broken},c}^{r_0,r_2}$. On the cohomology level, can be rewritten as $$\label{eq:elliptic-nabla-3}
\left[\nabla_{c}^{r_0,r_2} x\right] = \left[C^{r_0,r} C^{r,r_2} {\nabla\mkern-12mu/\mkern2mu}^{r_2} x + \rho_{\mathit{broken},c}^{r_0,r_2}(x) - [\theta^{r_1},x]_{c}^{r_0,r_1,r_2}\right] \in H^*(a^{r_0}).$$ More precisely, let $\gamma^{r_0,r_2}$ be the chain homotopy which appears in . The relation between $\rho_c^{r_0,r_2}$ and $\rho_{\mathit{broken},c}^{r_0,r_2}$ can be encoded schematically as follows: $$\label{eq:magic-triangle}
\xymatrix{
C^{r_0,r_2} R^{r_2} \ar@{-}[dd]_-{\gamma^{r_0,r_2} R^{r_2}} \ar@{-}[drr]^-{\rho_c^{r_0,r_2}} \\
&& [k^{r_1},\cdot]_c^{r_0,r_1,r_2} \\
C^{r_0,r} C^{r,r_2} R^{r_2} \ar@{-}[urr]_-{\rho_{\mathit{broken},c}^{r_0,r_2}}
}$$ Each side of is a homotopy, induced by a family of decorated disc configuration whose parameter space is an annulus $[0,1] \times S^1$ (and over $\{0,1\} \times S^1$, the families can be identified in pairs). There is a cobordism between the union of these three annuli and a family over $T^2$ (essentially, just rounding off the corners where we have identified the boundary circles of the annuli). One then uses Proposition \[th:grow-elliptic\] to show that that family can be extended over a solid torus, which provides the (degree $-1$) operation that “fills in the interior of the triangle” in . The cochain relating the two sides of is given by that operation together with $\gamma^{r_0,r_2} {\nabla\mkern-12mu/\mkern2mu}^{r_2}$. We denote the (chain level) right hand side of by $\nabla_{\mathit{broken},c}^{r_0,r_2}x$.
Take the decorated disc configuration underlying $C^{r,r_2}$, and “add a marked point in all possible places”. The result defines an operation $C^{r,r_2}_\bullet$, and using , we can rewrite as $$\label{eq:elliptic-nabla-4}
\nabla_{\mathit{broken},c}^{r_0,r_2}x = C^{r_0,r} {\nabla\mkern-12mu/\mkern2mu}^r C^{r,r_2}(x) + C^{r_0,r} C^{r,r_2}_\bullet(x) + \rho_{\mathit{broken},c}^{r_0,r_2}(x) - [\theta^{r_1},x]_{c}^{r_0,r_1,r_2}.$$ Additionally using , we get $$\label{eq:elliptic-nabla-5}
\begin{aligned}
\nabla_{\mathit{broken},c}^{r_0,r_2}x = &\; C^{r_0,r} {\nabla\mkern-12mu/\mkern2mu}^r C^{r,r_2}(x) + C^{r_0,r} C^{r,r_2}_\bullet(x) + \rho_{\mathit{broken},c}^{r_0,r_2}(x)\\ & \quad - [\kappa^{r_1},x]_c^{r_0,r_1,r_2} - [C^{r_1,1}(\theta^1),x]_{c}^{r_0,r_1,r_2}.
\end{aligned}$$ We now want to apply a further chain homotopy, which relates two different versions of the Lie bracket: $$\begin{aligned}
& \beta^{r_0,1,r_2}_c: C^*(a^1) \otimes C^*(a^{r_2}) \longrightarrow C^{*-2}(a^{r_0}), \\
& [C^{r_1,1}(x_1),x_2]_c^{r_0,r_1,r_2} - C^{r_0,r} [x_1,C^{r_2,r}(x_2)]^{r,1,r}
\\ & \quad =
d\beta_c^{r_0,1,r_2}(x_1,x_2) + \beta_c^{r_0,1,r_2}(dx_1,x_2) + (-1)^{|x_1|} \beta_c^{r_0,1,r_2}(x_1,dx_2).
\end{aligned}$$ This follows by comparing Lemma \[th:rot-bracket\] with the definition of , where the equality $c = r-1$ is key, and then using Proposition \[th:ehe-pants\] to construct the family of connections that interpolates between the two sides. From we can therefore pass to (the chain homotopic expression) $$\label{eq:elliptic-nabla-6}
\begin{aligned}
\nabla^{\mathit{broken},c}_{r_0,r_2}x {\simeq}& \; C^{r_0,r} {\nabla\mkern-12mu/\mkern2mu}^r C^{r,r_2}(x) + C^{r_0,r} C^{r,r_2}_\bullet(x) + \rho_{\mathit{broken},c}^{r_0,r_2}(x) \\ & \quad - [\kappa^{r_1},x]_c^{r_0,r_1,r_2} - C^{r_0,r} [\theta^1,C^{r,r_2}(x)]^{r,1,r} - \beta_c^{r_0,1,r_2}(k^1,x).
\end{aligned}$$
Comparing with shows that the desired statement concerning reduces to the following nullhomotopy of maps $C^*(a^{r_2}) \rightarrow C^*(a^{r_0})$: $$\label{eq:null-last}
C^{r_0,r} C^{r,r_2}_\bullet(x) + \rho_{\mathit{broken},c}^{r_0,r_2}(x) - C^{r_0,r} \rho^{r,r} C^{r,r_2}(x) - [\kappa^{r_1},x]_c^{r_0,r_1,r_2} - \beta^{r_0,1,r_2}_c(k^1,x) {\simeq}0.
$$ Importantly, this no longer contains ${\nabla\mkern-12mu/\mkern2mu}$ or the bounding cochain $\theta^1$: it is purely a relation between TQFT operations. Each term in comes from a family parametrized by an annulus, and for the nullhomotopy, the same reasoning as in applies. The analogue of is this: $$\label{eq:magic-pentagon}
\xymatrix{
& \ar@{-}[dl]_-{C^{r_0,r} C^{r,r_2}_\bullet}
C^{r_0,r} C^{r,r_2} R^{r_2}
\ar@{-}[dr]^-{\rho^{r_0,r_2}_{\mathit{broken},c}} & \\
\ar@{-}[d]_-{C^{r_0,r}\rho^{r,r}C^{r,r_2}}
C^{r_0,r} R^r C^{r,r_2} &&
[k^{r_1},\cdot]^{r_0,r_1,r_2}_c
\ar@{-}[d]^-{[\kappa^{r_1},\cdot]^{r_0,r_1,r_2}_c} \\
C^{r_0,r} [k^1,C^{r,r_2}(\cdot)]^{r,1,r} \ar@{-}@/_2pc/[rr]_-{\beta^{r_0,1,r_2}_c(k^1,\cdot)}
&& [C^{r_1,1}(k^1), \cdot]_c^{r_0,r_1,r_2}
}$$ In order to better visualize the situation, we also include a version of where the operations are described geometrically, in Figure \[fig:pentagon\].
(0,0)![\[fig:pentagon\]The families of surfaces appearing in . The darker shaded pieces are those one which change along the sides of the diagram.](pentagon.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(10322,7211)(-1964,-5760) (1251,1064)[(0,0)\[lb\]]{} (4051,1064)[(0,0)\[lb\]]{} (2701,1064)[(0,0)\[lb\]]{} (1726,-2536)[(0,0)\[lb\]]{} (1726,-3736)[(0,0)\[lb\]]{} (-599,-3136)[(0,0)\[lb\]]{} (-1969,-3136)[(0,0)\[lb\]]{} (6151,-3736)[(0,0)\[lb\]]{} (3001,-3736)[(0,0)\[lb\]]{} (6901,-1486)[(0,0)\[lb\]]{} (751,-886)[(0,0)\[lb\]]{} (1426,-886)[(0,0)\[lb\]]{} (2776,-886)[(0,0)\[lb\]]{} (4801,1064)[(0,0)\[lb\]]{} (4476,-886)[(0,0)\[lb\]]{} (3676,-3121)[(0,0)\[lb\]]{} (6901,-261)[(0,0)\[lb\]]{} (6096,-2517)[(0,0)\[lb\]]{} (7497,-2540)[(0,0)\[lb\]]{} (-699,-886)[(0,0)\[lb\]]{}
Maps to the disc\[sec:maps-to-the-disc\]
========================================
As a toy model for the subsequent construction of Hamiltonian Floer cohomology, we look at maps from Riemann surfaces to the hyperbolic disc, which satisfy an inhomogeneous Cauchy-Riemann equation. The inhomogeneous term comes from a family of infinitesimal automorphisms, hence is not compactly supported on the disc. We are mostly interested in what happens when the maps approach the boundary circle (at infinity).
The hyperbolic disc
-------------------
Let $B \subset {{\mathbb C}}$ be the open unit disc. The holomorphic automorphism group of $B$, and its Lie algebra of holomorphic vector fields, are given by: $$\label{eq:x-gamma}
\begin{aligned}
& G = \mathit{PU}(1,1) = \left\{ \textstyle g = \left(\begin{smallmatrix} a & b \\ \bar{b} & \bar{a} \end{smallmatrix}\right)\; : \;
|a|^2 - |b|^2 = 1 \right\}/\pm\!{\mathbbm{1}},
&&
g \mapsto \rho_g(w) = \frac{aw+b}{\bar{b}w + \bar{a}},\\
& {\mathfrak{g}}= \mathfrak{u}(1,1) = \left\{ \gamma = \left(\begin{smallmatrix} i\alpha & \beta \\ \bar{\beta} & -i\alpha \end{smallmatrix}\right) \;:\;
\alpha \in {{\mathbb R}}, \;\beta \in {{\mathbb C}}\right\}, &&
\gamma \mapsto X_\gamma = (\bar{\beta} + 2i\alpha w - \beta w^2) \, \partial_w.\end{aligned}$$ The action preserves the hyperbolic metric (in our convention, normalized to have curvature $-4$) $$g_{\mathit{hyp}} = \frac{1}{(1-|w|^2)^2} \mathit{dw}\, d\bar{w},$$ hence also its symplectic form $\omega_{\mathit{hyp}}$, and distance function $\mathrm{dist}_{\mathit{hyp}}$. There is a unique choice of Hamiltonian function for each vector field $X_\gamma$ which is linear in $\gamma$ and compatible with Poisson brackets: $$\label{eq:hamiltonian}
H_\gamma = \frac{1}{1-|w|^2} \left( {{\textstyle\frac{1}{2}}}(1+|w|^2)\alpha - \mathrm{im}(\beta w) \right).
$$
The group action extends smoothly to the closed disc $\bar{B}$. Call $\gamma$ nonnegative if $X_\gamma$ points in nonnegative direction along $\partial \bar{B}$. In terms of , the nonnegative cone defined in this way is $$\label{eq:h-pos}
{\mathfrak{g}}_{\geq 0} = \{ \alpha \geq |\beta| \}.$$ In terms of , $\gamma$ is nonnegative iff $H_\gamma$ is nonnegative everywhere (moreover, if $\gamma \notin {\mathfrak{g}}_{\geq 0}$, $H_\gamma$ is not even bounded below).
As the reader may have noticed, we are switching notation: for the rest of the paper, $G$ and ${\mathfrak{g}}$ will be as in . This is not a significant issue, since the two groups that appear under that name are actually isomorphic: $$\label{eq:change-g}
\mathit{PSL}_2({{\mathbb R}}) {\cong}\mathit{PU}(1,1),$$ and such an isomorphism is canonical up to inner automorphisms (geometrically, it is induced by a holomorphic automorphism of ${{\mathbb C}}\cup \{\infty\} = {{\mathbb C}}P^1$ which carries $B$ to the open upper half plane). preserves traces; and the associated isomorphism of Lie algebras relates to our previous notion of nonnegativity in $\mathfrak{sl}_2({{\mathbb R}})$. Hence, the entire discussion from Sections \[sec:sl2\]–\[sec:elliptic\] carries over immediately to .
Let $S$ be an oriented surface, equipped with a connection $A \in \Omega^1(S,{\mathfrak{g}})$. Via the homomorphisms from ${\mathfrak{g}}$ to the Lie algebras of vector fields and of functions (with the Poisson bracket) on $B$, this yields one-forms valued in those Lie algebras. We write them as $$\begin{aligned} \label{eq:ha}
& X_A \in {C^\infty}(S \times B, T^*S \otimes_{{{\mathbb R}}} {{\mathbb C}}), \\
& H_A \in {C^\infty}(S \times B, T^*S).
\end{aligned}$$ One can apply the same process to the curvature $F_A \in \Omega^2(S,{\mathfrak{g}})$, which yields $$\label{eq:2-curvatures}
\begin{aligned}
& X_{F_A} \in {C^\infty}(S \times B, \Lambda^2_{{{\mathbb R}}} T^*S \otimes_{{{\mathbb R}}} {{\mathbb C}}), \\
& H_{F_A} \in {C^\infty}(S \times B, \Lambda^2_{{{\mathbb R}}} T^*S).
\end{aligned}$$ What $X_A$ describes is the connection induced by $A$ on the trivial fibre bundle $S \times B \rightarrow S$. Like any connection on a fibre bundle, one can also think of this as given by a complementary subbundle to the fibrewise tangent bundle. The choice of such a subbundle determines a lift of $\omega_{\mathit{hyp}}$ to a fibrewise symplectic form $\omega_{A,\mathit{geom}} \in \Omega^2(S \times B)$, and this satisfies $$\label{eq:domega-a}
d\omega_{A,\mathit{geom}} = -dH_{F_A}.$$ On the right hand side of , we think of $H_{F_A}$ as a two-form on $S \times B$ via the projection $T(S \times B) \rightarrow TS$, and then take its exterior derivative. While $\omega_{A,\mathit{geom}}$ is not itself closed, there is an obvious closed replacement (this is an instance of the general theory of symplectic fibrations, see e.g. [@mcduff-salamon99 Chapter 6]): $$\label{eq:top-omega}
\omega_{A,\mathit{top}} = \omega_{A,\mathit{geom}} + H_{F_A}.$$ Similarly, the connection determines an almost complex structure $J_A$ on $S \times B$, such that projection to $S$ is $J_A$-holomorphic, and whose restriction to each fibre is the standard complex structure on $B$. This satisfies $$\label{eq:semi-omega}
\omega_{A,\mathit{geom}}(\eta ,J_A \eta) \geq 0 \quad \text{for all $\eta \in T(S \times B)$.}
$$ By construction, the induced connection on $S \times B \rightarrow S$ extends smoothly to the compactification $S \times \bar{B}$, and so does $J_A$.
One can be more explicit by working in a local complex coordinate $z = s+it$ on $S$. If $A = A_1(s,t)\mathit{ds} + A_2(s,t)\mathit{dt}$ with $A_1(s,t),\, A_2(s,t) \in {\mathfrak{g}}$, then by definition $$\begin{aligned}
& X_A = X_{A_1} \mathit{ds} + X_{A_2} \mathit{dt}, \\
& H_A = H_{A_1} \mathit{ds} + H_{A_2} \mathit{dt}, \\
& X_{F_A} = \big( X_{\partial_t A_1} - X_{\partial_s A_2} + X_{[A_1,A_2]} \big) \, \mathit{ds} \wedge \mathit{dt}, \\
& H_{F_A} = \big( H_{\partial_t A_1} - H_{\partial_s A_2} + \omega_{\mathit{hyp}}(X_{A_1},X_{A_2}) \big) \,\mathit{ds} \wedge \mathit{dt}.\end{aligned}$$ The subbundle of $T(S \times B)$ which defines the connection on $S \times B \rightarrow S$ is $$\label{eq:horizontal-subspace}
{{\mathbb R}}(\partial_s + X_{A_1} \partial_w) \oplus {{\mathbb R}}(\partial_t + X_{A_2} \partial_w).$$ The two-form $\omega_{A,\mathit{geom}}$ is characterized by having as its nullspace. This means that $$\begin{aligned}
& \omega_{A,\mathit{geom}} = \omega_{\mathit{hyp}} - dH_{A_1} \wedge \mathit{ds} - dH_{A_2} \wedge \mathit{dt} - \omega_{\mathit{hyp}}(X_{A_1},X_{A_2}) \mathit{ds} \wedge \mathit{dt}, \\
& \omega_{A,\mathit{top}} = \omega_{\mathit{hyp}} - d(H_{A_1} \mathit{ds}) - d(H_{A_2} \mathit{dt}).\end{aligned}$$ The associated almost complex structure is similarly given by $$\label{eq:j-a}
\left\{
\begin{aligned}
& J_A (\partial_w) = i\partial_w, \\
& J_A (\partial_s + X_{A_1} \partial_w) = \partial_t + X_{A_2} \partial_w.
\end{aligned}
\right.$$
Suppose that $A$ is nonnegatively curved. This means that, for any oriented basis $(\xi_1,\xi_2)$ in $TS$, $H_{F_A}(\xi_1,\xi_2)$ is a nonnegative function; and that along $\partial\bar{B}$, $X_{F_A}(\xi_1,\xi_2)$ points in clockwise direction. From and , it then follows that $$\label{eq:o-o}
0 \leq \omega_{A,\mathit{geom}}(\eta,J_A \eta) \leq \omega_{A,\mathit{top}}(\eta, J_A \eta).$$
\[th:levi\] If $A$ has nonnegative curvature, the hypersurface $S \times \partial \bar{B} \subset S \times \bar{B}$ is Levi convex (meaning that its Levi form with respect to $J_A$ is nonnegative).
Take $\psi(z,w) = {{\textstyle\frac{1}{2}}}|w|^2: S \times \bar{B} \rightarrow {{\mathbb R}}$. In local coordinates on $S$ as before, one gets $$\label{eq:ddc}
\begin{aligned}
-d(d\psi \circ J_A)| (S \times \partial \bar{B}) = &\;
- \mathrm{im}(X_{A_1} d\bar{w}) \wedge \mathit{ds}
-\mathrm{im}(X_{A_2} d\bar{w}) \wedge \mathit{dt}
\\ &
+ \mathrm{im}(\bar{w} X_{\partial_t A_1}) \mathit{ds} \wedge \mathit{dt}
- \mathrm{im}(\bar{w} X_{\partial_s A_2}) \mathit{ds} \wedge \mathit{dt}
\\ &
- \mathrm{im}(\bar{w}\, dX_{A_1}) \wedge \mathit{ds}
- \mathrm{im}(\bar{w}\, dX_{A_2}) \wedge \mathit{dt}.
\end{aligned}$$ The Levi form $L_\psi$ is obtained by restricting to the space $T(S \times \partial \bar{B}) \cap J_A T(S \times \partial \bar{B})$, which is . From , one gets $$\label{eq:levi}
L_\psi = \mathrm{im}( \bar{w} X_{F_A} ).$$ Let’s spell out the meaning of this. $X_{F_A}$, defined in , is a complex-valued skew bilinear form on $TS$. In , we identify that space with by pullback. The desired property follows immediately from .
The Cauchy-Riemann equation
---------------------------
Now assume that $S$ is a (connected) Riemann surface. We will consider the inhomogeneous Cauchy-Riemann equation $$\label{eq:inhomogeneous}
\left\{
\begin{aligned}
& u: S \longrightarrow B, \\
& \big( Du - X_A(u) \big)^{0,1} = 0.
\end{aligned}
\right.$$ More precisely, $X_A$ is being evaluated at points $\sigma(z) = (z,u(z))$, and the $(0,1)$ part is taken with respect to the given complex structures on the source and target. Equivalently, this is the equation for $\sigma$ to be a pseudo-holomorphic section: $$\label{eq:homogeneous}
\left\{
\begin{aligned}
& \sigma = (\mathit{id},u): S \longrightarrow S \times B, \\
& D\sigma \circ i = J_A(\sigma) \circ d\sigma.
\end{aligned}
\right.$$ In local coordinates on $S$, becomes $$\label{eq:concrete-inho}
\partial_s u + i \partial_t u = X_{A_1(s,t)}(u) + i X_{A_2(s,t)}(u).$$ As one would expect from the appearance of connections, there is an action of the group of gauge transformations. Namely, if $A^\dag = \Phi_*A$ and $u^\dag(z) = \rho_{\Phi(z)}(u(z))$ for some $\Phi \in {C^\infty}(S,G)$, then $$D u^\dag - X_{A^\dag}(u^\dag) = D\rho_{\Phi} (Du - X_A(u)),$$ where on the right hand side, $D\rho_{\Phi}$ is the derivative of $\rho_{\Phi(z)} \in \mathit{Aut}(B)$ at the point $u(z)$. Hence, passing from $(u,A)$ to $(u^\dag,A^\dag)$ transforms solutions of one equation to the other. Equivalently, the gauge transformation determines a diffeomorphism of $S \times B$ which is fibrewise a hyperbolic isometry, and that diffeomorphism relates the almost complex structures $J_A$ and $J_{A^\dag}$.
We also want to quickly review the notions of energy for solutions of . The geometric and topological energies are, respectively, $$\begin{aligned}
\label{eq:geometric-energy}
& E_{\mathit{geom}}(u) = {{\textstyle\frac{1}{2}}}\int_S |du - X_A(u)|^2_{\mathit{hyp}}, \\
& \label{eq:topological-energy}
E_{\mathit{top}}(u) = \int_S u^*\omega_{\mathit{hyp}} - d (H_A(u))\end{aligned}$$ (for general maps $u$ and noncompact $S$, the integral may not converge, but we’ll ignore that issue as it will irrelevant for our applications). The difference between the two notions is given by a curvature term: $$\label{eq:e-e}
E_{\mathit{top}}(u) = E_{\mathit{geom}}(u) + \int_S H_{F_A}(u).$$ In particular, if $A$ is nonnegatively curved, $$E_{\mathit{geom}}(u) \leq E_{\mathit{top}}(u).$$ In other words, the situation is as follows: the two notions of energy are obtained by integrating $\omega_{A,\mathit{geom}}$ and $\omega_{A,\mathit{top}}$ over $\sigma$. Then, explains , and the behaviour for nonnegatively curved connections follows from .
Convergence and compactness\[subsec:schwarz\]
---------------------------------------------
If the connection $A$ is flat, one can use a local gauge transformation to reduce to $\bar\partial u^\dag = 0$, hence to elementary complex analysis.
\[th:schwarz\] Suppose that $S = B_r \subset {{\mathbb C}}$ is the open disc of radius $r$, and that the connection $A$ is flat. Then for any solution of , we have $$|du(z) - X_A(z,u(z))|_{\mathit{hyp}} \leq \frac{r}{r^2-|z|^2}.$$
After a gauge transformation, this is just the Schwarz Lemma (holomorphic maps between hyperbolic Riemann surfaces can’t increase distances).
\[th:hyp-cylinder\] Suppose that $S = (-l/2,l/2) \times S^1$ is the open cylinder of length $l$, and that the connection $A$ is flat, with hyperbolic holonomy $g \in G$ around $\{0\} \times S^1$. Write $\mathrm{tr}(g) = \pm(\lambda+\lambda^{-1})$, for some $\lambda>1$. Then, a solution of can exist only if $$\label{eq:l-ineq}
l \leq \frac{\pi}{\log(\lambda^2)}.$$
We again apply a gauge transformation to trivialize the connection, but this time working on the universal cover $(-l/2,l/2) \times {{\mathbb R}}$. After doing that and then again quotienting out by deck transformations, a solution turns into a holomorphic map $$S \longrightarrow B/\rho_g {\cong}(0,\pi/\log(\lambda^2)) \times S^1,$$ which induces an isomorphism of fundamental groups. Such a map can’t exist if the source cylinder is longer than the target one; this follows from the Schwarz Lemma, given that the hyperbolic metric on a cylinder of length $l$ has a “waist” of girth $\pi/2l$; see e.g. [@kobayashi05 p. 13] (taking into account our scaling convention for the metric).
\[th:elliptic-cylinder\] Suppose that $S = (-l/2,l/2) \times S^1$ as before, and that the connection $A$ is flat, with elliptic holonomy $g \in G$ around $\{0\} \times S^1$. Write $\mathrm{tr}(g) = \pm 2\cos(\theta)$, for some $\theta \in (0,\pi)$. Let $w \in B$ be the unique fixed point of $\rho_g$. Then, any solution of satisfies $$\sinh(2\,\mathrm{dist}_{\mathit{hyp}}(u(0,0),w)) \leq \frac{\sinh(\pi/2l)}{\sin \theta}.
$$
After the usual gauge transformation, we get a holomorphic map $u^\dag: (-l/2,l/2) \times {{\mathbb R}}\rightarrow B$, which satisfies $u^\dag(0,0) = u(0,0)$ and $u^\dag(s,t+1) = \rho_g(u^\dag(s,t))$. In the hyperbolic metric on the domain, the distance between the points $(0,0)$ and $(0,1)$ is $\pi/2l$, as in Lemma \[th:hyp-cylinder\]. By hyperbolic trigonometry, the distance between $p = u^\dag(0,0)$ and $q = u^\dag(0,1)$ satisfies $$\sin \theta = \frac{\sinh( \mathrm{dist}_{\mathit{hyp}}(p,q))}{\sinh(2\,\mathrm{dist}_{\mathit{hyp}}(p,w))}.$$ The rest is once again Schwarz’ Lemma.
\[th:flat-convergence\] Let $S$ be any Riemann surface, equipped with a flat connection $A$, and $(u_k)$ a sequence of solutions of . Suppose that there is a sequence of points $z_k$, contained in a compact subset of $S$, such that $u_k(z_k)$ goes to $\partial \bar{B}$. Then $u_k$ goes to $\partial \bar{B}$, in the sense of uniform convergence on compact subsets.
Since we can always replace $S$ by its universal cover, let’s assume that it is simply-connected. After a gauge transformation, we get a sequence of bounded holomorphic functions $u_k^\dag$. This sequence has a convergent (on compact subsets) subsequence, and hence, so does $u_k$. Let $u_\infty^\dag, u_\infty: S \rightarrow \bar{B}$ be the limits. $u_\infty^\dag$ is a holomorphic function which meets $\partial \bar{B}$, hence must be constant at a point of $\partial \bar{B}$, by the open mapping theorem. It follows that $u_\infty$ also take values in $\partial \bar{B}$. By same argument, every subsequence of $(u_k)$ has a subsequence with the desired behaviour. From this, the result follows.
To go beyond the flat case, we need to replace holomorphic function theory by more flexible analytic tools. We work with a fixed $S$ and connection $A$. Bounds for derivatives will be with respect to some metric on $S$ (the choice is irrelevant, since we will always work on compact subsets), and the Euclidean metric on $B$.
\[th:elliptic\] Given a compact subset $K \subset S$ and an $r \geq 0$, there is a constant $C$ such that $\|u|K\|_{C^r} \leq C$ for all solutions $u$ of .
Let’s work in local coordinates on $S$. There is a uniform bound on the right hand side of , and of course on $u$ itself. By elliptic regularity, we get a $W^{1,p}$ bound on $u$ (for any $1<p<\infty$). From there, one proceeds by bootstrapping.
One can also view Lemma \[th:elliptic\] as a consequence of the regularity theory for pseudo-holomorphic maps (see e.g. [@mcduff-salamon-big Appendix B]) to ; the noncompactness of $S \times B$ does not matter here, since $J_A$ extends to $S \times \bar{B}$.
\[th:levi-apply\] Suppose that $A$ has nonnegative curvature. Let $(u_k)$ be a sequence of solutions of . Suppose that $z_k$ is a sequence of points, contained in a compact subset of $S$, such that $u_k(z_k)$ goes to $\partial \bar{B}$. Then $u_k$ goes to $\partial \bar{B}$, in the sense of uniform convergence on compact subsets.
Lemma \[th:elliptic\] shows that, after passing to a subsequence, the $u_k$ converge to some $u_\infty: S \rightarrow \bar{B}$. Consider the $J_A$-holomorphic map $\sigma_\infty(z) = (z,u_\infty(z))$. We know from Lemma \[th:levi\] that $S \times \partial \bar{B}$ is Levi convex, which by [@diederich-sukhov08 Corollary 4.7] implies that $\sigma_\infty^{-1}(S \times \partial \bar{B})$ is open and closed. But by assumption, that subset is also nonempty, which shows that $u_\infty$ takes values in $\partial \bar{B}$. The rest is as in Lemma \[th:flat-convergence\].
\[th:sukhov\] The expository paper [@gaussier-sukhov11] provides helpful additional discussion of the result from [@diederich-sukhov08] which we have used. The strategy of proof goes as follows. Let $\bar{N}$ be a domain with smooth Levi convex boundary in an almost complex manifold, and $N$ its interior. Suppose that there is a pseudo-holomorphic map from a connected Riemann surface to $\bar{N}$, which meets the boundary but is not entirely contained in it. After restriction to a suitably chosen part of $S$, we get a pseudo-holomorphic map defined on the unit disc, $$\left\{
\begin{aligned}
& \sigma: \bar{B} \longrightarrow \bar{N}, \\
& \sigma^{-1}(\partial \bar{N}) = \{1\}, \text{ and $\sigma$ is tangent to $\partial \bar{N}$ at that point.}
\end{aligned}
\right.$$ Near $\sigma(1)$, there is a smooth local defining function $r$ for $\partial \bar{N}$, such that $-(-r)^\theta$ is (strictly) plurisubharmonic in $N$; the constant $0 < \theta < 1$ can be chosen arbitrarily [@gaussier-sukhov11 Theorem 2.4], and we will just need $\theta > 1/2$. After shrinking the domain further if necessary, we may assume that $r \circ \sigma$ is well-defined. Since it has vanishing derivative at $z = 1$, we have $r(\sigma(z)) \gtrsim -|z-1|^2$. Hence, $h = -(-r)^\theta \circ \sigma$ is a negative subharmonic function on $B$, which extends continuously to $\bar{B}$ and vanishes exactly at $1 \in \partial \bar{B}$, such that $$h(z) \gtrsim -|z-1|^{2\theta}.$$ But that contradicts the Hopf Lemma for such functions.
Floer cohomology\[sec:floer\]
=============================
This section sets up Hamiltonian Floer cohomology theory for Lefschetz fibrations, and the operations on it. The basic approach is classical [@hofer-salamon95]. The issues that are specific to our context arise from the noncompactness of the symplectic manifold. Compared to previous versions of the same formalism (see e.g. [@seidel14b]), the main difference is that at infinity, our Hamiltonian automorphisms are modelled on isometries of the hyperbolic disc (following Section \[sec:maps-to-the-disc\]), rather than area-preserving diffeomorphisms of the Euclidean plane. Even though the distinction may fundamentally be a technical one, the hyperbolic model turns out to be very helpful for our constructions.
Geometry of the target space
----------------------------
Let $(E^{2n},\omega_E)$ be a symplectic manifold, together with a trivialization of its anticanonical bundle (for some compatible almost complex structure), and a proper map to the hyperbolic disc. Suppose that $\bar{N} \subset \bar{B}$ of $\partial \bar{B}$ is a closed collar neighbourhood. We then call $N = \bar{N} \cap B$ a neighbourhood of infinity (for instance, $\{ |w| \geq r\}$ for some $r<1$ is such a neighbourhood). Our main condition is:
- For some neighbourhood of infinity $N = N_E$, there is a (necessarily unique) $\omega_E$-orthogonal splitting of the tangent space at any point $x \in \pi^{-1}(N)$ into horizontal and vertical parts, $$\label{eq:te-h-v}
\left\{
\begin{aligned}
& TE_x = TE_x^h \oplus TE_x^v, \\
& \mathit{TE}_x^v = \mathit{ker}(D\pi_x), \\
& D\pi_x: \mathit{TE}_x^h \longrightarrow TB_{\pi(x)} \text{ is a symplectic isomorphism.}
\end{aligned}
\right.$$
In other words: $x$ is a regular point of $\pi$; $\mathit{TE}_x^v$ is a symplectic subspace; and the symplectic form on its $\omega_E$-orthogonal complement agrees with the pullback of $\omega_{\mathit{hyp}}$. As a consequence, restricts to a symplectic fibre bundle over $N$ with a preferred connection, given by $\mathit{TE}^h$, which is flat. Using that structure, one can construct a canonical compactification $$\label{eq:compactification}
\pi: \bar{E} \longrightarrow \bar{B}.$$ The total space $\bar{E}$ is a compact manifold with boundary, whose interior is $E$, and with $\partial \bar{E} = \pi^{-1}(\partial \bar{B})$. The compactification again has the property that, over $\bar{N}$, it is a symplectic fibre bundle with a preferred flat connection. In other words, the splitting in extends to points of $\partial\bar{E}$, and so does the symplectic structure on the vertical part. Of course, $\omega_E$ itself does not extend to $\bar{E}$, because of the $\omega_{\mathit{hyp}}$ component, but there are slightly different symplectic forms that do: $$\label{eq:omega-bar-e}
\omega_{\bar{E}} = \omega_E - \pi^* (\psi \omega_{\mathit{hyp}}) + \pi^* (\psi \omega_{\bar{B}}),$$ where $\omega_{\bar{B}}$ is any (positive) symplectic form on $\bar{B}$, and $\psi: \bar{B} \rightarrow [0,1]$ is a function which vanishes outside $N$, and equals $1$ near $\partial \bar{B}$.
- We denote by ${\EuScript J}(E)$ the space of compatible almost complex structures $J$ on $E$ with the following two additional properties. First, there is some neighbourhood of infinity $N = N_J \subset N_E$, such that $\pi$ is $J$-holomorphic at all points of $\pi^{-1}(N)$. Secondly, $J$ extends to $\bar{E}$, and that extension is compatible with .
The last-mentioned condition (compatibility with $\omega_{\bar{E}})$ is independent of the choices made in . Even before imposing that condition, we already knew that $J$ preserves the extension of the splitting to points $x \in \partial \bar{E}$, and is compatible with the orientation of $T\bar{E}^h_x {\cong}T\bar{B}_{\pi(x)}$. Hence, the additional requirement is that the extension should be compatible with the symplectic structure on $T\bar{E}_x^v$. We want to make one more observation concerning such almost complex structures. Because the hyperbolic metric scales up to infinity as one approaches $\partial\bar{B}$, there is a constant $C = C_J >0$ such that $$\label{eq:omega-bar-e-2}
\omega_{\bar{E}}(X,JX) \leq C\, \omega_E(X,JX) \quad \text{for all $X \in \mathit{TE}$.}$$ In particular, the $\omega_E$-energy of non-constant $J$-holomorphic spheres is bounded below by a positive constant (since the same holds for $\omega_{\bar{E}}$, which lives on a compact space).
- For any $\gamma \in {\mathfrak{g}}$, let ${\EuScript H}(E,\gamma)$ be the space of functions $H \in {C^\infty}(E,{{\mathbb R}})$ whose restriction to $\pi^{-1}(N)$, for some neighbourhood of infinity $N = N_H \subset N_E$, equals the pullback of the function $H_\gamma$ from . Similarly, given $a = a_t \mathit{dt} \in \Omega^1(S^1,{\mathfrak{g}})$, we write ${\EuScript H}(E,a)$ for the space of $H \in {C^\infty}(S^1 \times E, {{\mathbb R}})$ such that $H_t \in {\EuScript H}(E,a_t)$ for all $t \in S^1$.
On $\pi^{-1}(N)$, the Hamiltonian vector field $X$ associated to $H \in {\EuScript H}(E,\gamma)$ is the unique horizontal lift of $X_\gamma$. Hence, $X$ extends smoothly to $\bar{E}$, and is tangent to $\partial\bar{E}$.
\[th:families-of-neighbourhoods\] In our application, we will often encounter almost complex structures or functions which depend on additional parameters (i.e. families of such objects). In that case, it is always understood that the relevant neighbourhoods of infinity can be chosen locally constant with respect to the parameter. Equivalently, if $P$ is the parameter space, we are considering a closed collar neighbourhood of $P \times \partial \bar{B}$ inside $P \times \bar{B}$, and then taking its intersection with $P \times B$.
Given $H = (H_t) \in {\EuScript H}(E,a)$ and its time-dependent vector field $X = (X_t)$, we will consider one-periodic orbits $$\label{eq:periodic-orbit}
\left\{
\begin{aligned}
& x: S^1 \longrightarrow E, \\
& dx/dt = X_t.
\end{aligned}
\right.$$
If $a$ has nontrivial holonomy, all solutions of are contained in a compact subset of $E$.
Otherwise, there would have to be a sequence of one-periodic orbits in $E$ which converge to $\partial \bar{E}$. By projection, the same would have to be true for the corresponding ODE on the hyperbolic disc, $$\label{eq:downstairs-periodic-orbit}
\left\{
\begin{aligned}
& y: S^1 \longrightarrow B, \\
& dy/dt = X_{a_t}.
\end{aligned}
\right.$$ But that is impossible, since such $y$ correspond to fixed points of the holonomy of $a$ acting on $B$ (of which there is at most one).
When describing our results in Section \[sec:results\], we included a Lefschetz (nondegeneracy of critical points) condition, since that puts them into a familiar context. In fact, that condition is irrelevant for the constructions in this paper, and therefore doesn’t appear in the definition we have just given.
Basic transversality and compactness
------------------------------------
In spite of the constraints imposed on the almost complex structures and Hamiltonian functions, the transversality results underlying the traditional construction of Floer theory still hold. Here is a sample of the kind of arguments that are required:
\[th:j-spheres\] For generic $J \in {\EuScript J}(E)$, all simple (non-multiply-covered) $J$-holomorphic spheres are regular.
A given $J$ comes with some $N = N_J$ over which $D\pi$ is $J$-holomorphic. Genericity is understood to be with respect to perturbations which keep that subset fixed. This means that on $\pi^{-1}(N)$, we may change only the $TE^v$ component of $J$. Let $u: {{\mathbb C}}P^1 \rightarrow E$ be a simple $J$-holomorphic sphere, and $v = \pi(u)$. At any point $z$ such that $v(z) \in N$, we have $\bar\partial v = 0$. Hence, one of the following applies:
\(i) There is a point where $v(z) \notin N$. Since the choice of almost complex structure is free (except for the condition of $\omega_E$-compatibility) near $u(z)$, transversality for such $u$ follows from the standard argument.
\(ii) There is a point where $v(z) \in N \setminus \partial N$. Note that the degree of $v$ over any point $w \in B$ is the same, hence zero. It follows that for each $w \in N \setminus \partial N$, $v^{-1}(w)$ is open and closed. Under our assumption, this means that $u$ must lie in a single fibre $F = E_w$. A parametrized transversality argument, using the fact that the almost complex structure in vertical direction can be varied freely, shows generic regularity for such maps. To explain the relevance of parametrized transversality, consider the commutative diagram (with exact columns) $$\label{eq:snake}
\xymatrix{
0 \ar[d] && 0 \ar[d] \\
\ar[d]
\{ \xi \in {C^\infty}({{\mathbb C}}P^1, u^*TE) \,:\, D\pi(\xi) \text{ is constant} \}
\ar[rr] && {C^\infty}({{\mathbb C}}P^1, \Omega^{0,1} \otimes_{{{\mathbb C}}} u^*TF)\ar[d] \\
\ar[d]^-{D\pi}
{C^\infty}({{\mathbb C}}P^1, u^*TE) \ar[rr]^-{D_{E,u}} && {C^\infty}({{\mathbb C}}P^1, \Omega^{0,1} \otimes_{{{\mathbb C}}} u^*TE) \ar[d]^-{D\pi} \\
\ar[d] {C^\infty}({{\mathbb C}}P^1, {{\mathbb C}})/\mathit{constants} \ar[rr]^-{\bar\partial} && {C^\infty}({{\mathbb C}}P^1,\Omega^{0,1}) \ar[d]
\\ 0 && 0.
}$$ Here, $D_{E,u}$ is the linearization of $u$ as a pseudo-holomorphic map into $E$. The top $\rightarrow$ is the linearization of $u$ as a pseudo-holomorphic map into a variable fibre (its domain is an extension of ${C^\infty}({{\mathbb C}}P^1,u^*TF)$ by ${{\mathbb C}}= TB_w$). The parametrized transversality theory shows that the top $\rightarrow$ is onto for generic $J$. But since the bottom $\rightarrow$ is invertible, we get the same surjectivity result for $D_{E,u}$.
\(iii) $v(z) \in \partial N$ for all $z$. In that case, $v$ is holomorphic, hence must again be constant. The same strategy as in (ii) applies.
A similar result holds for one-periodic orbits:
\[th:1-periodic\] Suppose that $a \in \Omega^1(S^1,{\mathfrak{g}})$ has nontrivial holonomy. Then, for generic choice of $H \in {\EuScript H}(E,a)$, all one-periodic orbits are nondegenerate.
As before, we need to be a bit more precise concerning the statement itself. Each $H_t \in {\EuScript H}(E,a_t)$ comes with some neighbourhood of infinity $N_{H_t}$, but we can fix an $N$ which is smaller than all of them (see Remark \[th:families-of-neighbourhoods\], and take into account the compactness of the parameter space $P = S^1$). In the case where the holonomy is elliptic, we assume that has no solutions satisfying $y(t) \in N$ for all $t$ (for parabolic or hyperbolic holonomy, this condition would be vacuous, since their holonomy acts fixed point freely). That being given, it is clear that perturbing $H$ only outside $\pi^{-1}(N)$ is sufficient to achieve transversality.
Let’s keep the assumption that $a$ has nontrivial holonomy, and choose $H \in {\EuScript H}(E,a)$, satisfying the nondegeneracy condition from Lemma \[th:1-periodic\]. Suppose also that we have a family $J = (J_t)$ of almost complex structures in ${\EuScript J}(E)$. Given one-periodic orbits $x_{\pm}$ as in , we consider Floer trajectories $$\label{eq:e-floer}
\left\{
\begin{aligned}
& u: {{\mathbb R}}\times S^1 \longrightarrow E, \\
& \partial_s u + J_t(u) (\partial_t u - X_t(u)) = 0, \\
& \textstyle \lim_{s \rightarrow \pm\infty} u(s,t) = x_{\pm}(t).
\end{aligned}
\right.$$ The pair $\Phi = (J,H)$ will be called a “Floer datum”, since this is the structure required to write down Floer’s equation (nondegeneracy of one-periodic orbits is part of the definition of Floer datum).
\[th:floer-transversality\] For generic choice of $J$ (keeping $H$ fixed), all solutions of are regular. Moreover, if $u$ is a solution whose linearized operator $D_u$ has index $\leq 2$, then $u(s,t)$ never lies on a nontrivial $J_t$-holomorphic sphere.
This is essentially the transversality result from [@floer-hofer-salamon94; @hofer-salamon95]. Applied to solutions $u(s,t) = x(t)$ ($D_u$ is invertible in that case, because of the nondegeneracy of $x$), the result includes the statement that for any one-periodic orbit, $x(t)$ should not lie on a nontrivial $J_t$-holomorphic sphere; this is a version of Lemma \[th:j-spheres\] including transversality of evaluations. For the remaining part of Lemma \[th:floer-transversality\], one starts with a fixed $\Phi = (J,H)$, and takes some neighbourhood of infinity $N$ which is smaller than $N_{J_t}$ and $N_{H_t}$ for any $t$, and which has the same property concerning as in the proof of Lemma \[th:1-periodic\]. Then, since the orbits $x_{\pm}$ aren’t contained in $\pi^{-1}(N)$, the maps $u$ also can’t be contained in that subset, which means that perturbing $J$ outside $\pi^{-1}(N)$ gives one enough freedom.
To show that Floer trajectories can’t escape to infinity, we will combine elementary complex analysis (as in Section \[subsec:schwarz\]) and pseudo-holomorphic curve theory. Some basic ingredients are:
- The notion of energy of a Floer trajectory, $$E(u) = \int_{{{\mathbb R}}\times S^1} \|\partial_s u\|^2_{J_t} = \int_{{{\mathbb R}}\times S^1} u^*\omega_E - \int_{S^1} H_t(x_+(t))\, \mathit{dt} + \int_{S^1} H_t(x_-(t))\, \mathit{dt}.$$
- The “gauge transformation” which makes Floer’s equation homogeneous. Namely, let $(\phi_t)$ be the Hamiltonian isotopy generated by $(X_t)$. The change of variables $u(s,t) = \phi_t(u^\dag(s,t))$ turns into $$\label{eq:u-star}
\left\{
\begin{aligned}
& u^\dag: {{\mathbb R}}^2 \longrightarrow E, \\
& u^\dag(s,t) = \phi_1(u^\dag(s,t+1)), \\ & \partial_s u^\dag + J_t^\dag(u^\dag) \partial_t u^\dag = 0, \\
& \textstyle \lim_{s \rightarrow \pm \infty} u^\dag(s,t) = x^\dag_{\pm},
\end{aligned}
\right.$$ where $J_t = (\phi_t)_* J_t^\dag$, and $x^\dag_{\pm} = x_{\pm}(0)$. Outside a compact subset, $\phi_t$ covers an isotopy in $\mathit{Aut}(B)$. Hence, $J_t^\dag$ again belongs to ${\EuScript J}(E)$.
- Projection to the base. Wherever $u$ is suffficiently close to $\partial\bar{E}$, the projection $v = \pi(u)$ satisfies a special case of , where $A$ is the pullback of $a$ to the cylinder (hence flat): $$\label{eq:downstairs-floer}
\partial_s v + i(\partial_t v - X_{a_t}(v)) = 0.$$
\[th:e-bound-0\] Suppose that $(u_k)$ is a sequence of solutions of , with an upper bound on the energy. Assume that there are points $(s_k,t_k) \in {{\mathbb R}}\times S^1$ such that $|s_k|$ is bounded, and $u_k(s_k,t_k) \rightarrow \partial \bar{E}$. Then, the $u_k$ converge to $\partial \bar{E}$ uniformly on compact subsets.
Gromov compactness (on compact subsets of the domain) can be applied to the $u_k$, seen as maps taking values in $\bar{E}$. This may seem suspicious, but is justified by the following argument. Consider the pseudo-holomorphic maps $u_k^\dag$ as in . On each compact subset of ${{\mathbb R}}^2$, we have an upper bound on their energies with respect to $\omega_E$, hence by also for $\omega_{\bar{E}}$. When applying Gromov compactness, one thinks of the latter symplectic form on $\bar{E}$.
With this in mind, the Gromov limit (of a subsequence of the $u_k$) consists of a principal component $u_\infty: {{\mathbb R}}\times S^1 \rightarrow \bar{E}$, which solves the Cauchy-Riemann equation from , together with bubble components which are pseudo-holomorphic spheres (unlike the usual compactification of Floer trajectory spaces, the notion of Gromov limit we are working with here ignores any components which might split off over the ends $s \rightarrow \pm\infty$). By assumption, one of our components must intersect $\partial \bar{E}$. If the principal component does that, then by looking at its projection to $\bar{B}$ (as in Lemma \[th:flat-convergence\]), one concludes that $u_\infty({{\mathbb R}}\times S^1) \subset \partial \bar{E}$. The same argument applies to bubble components. Because the Gromov limit is connected, a component-by-component induction shows that all of it lies in $\partial \bar{E}$. This implies that a subsequence of the $u_k$ converges to $\partial \bar{E}$ on compact subsets. Going from that to the entire sequence works as in Lemma \[th:flat-convergence\].
\[th:e-bound-1\] Assume that the holonomy of $a$ is elliptic or hyperbolic. Suppose that $(u_k)$ is a sequence of solutions of , with an upper bound on the energy. Then, there is a compact subset of $E$ which contains the images of all the $u_k$.
Let’s suppose that on the contrary, there are (possibly after passing to a subsequence) points $(s_k,t_k)$ such that $u_k(s_k,t_k) \rightarrow \partial \bar{E}$. After translation in $s$-direction, we may assume that $|s_k|$ is bounded. Lemma \[th:e-bound-0\] implies that, on each relatively compact subset of ${{\mathbb R}}\times S^1$, $v_k = \pi(u_k)$, for $k \gg 0$, is a solution of . In the hyperbolic case, one takes that subset to be a sufficiently long finite cylinder $(-l/2,l/2) \times S^1$, and gets a contradiction to Lemma \[th:hyp-cylinder\]. In the slightly more complicated elliptic case, Lemma \[th:elliptic-cylinder\] shows that $v_k(0,0)$ must lie in a specific compact subset of $B$, which leads to a contradiction to Lemma \[th:e-bound-0\].
General surfaces\[subsec:worldsheets\]
--------------------------------------
We now extend to other Riemann surfaces, building a familiar kind of TQFT framework (compare e.g. [@schwarz95]). Let $S$ be a punctured (connected, but not closed) Riemann surface. We arbitrarily divide the set $I$ of punctures into subsets $I_{\mathit{in}}$ and $I_{\mathit{out}}$. The surface should be decorated with the following additional data:
- Tubular ends surrounding the punctures, with pairwise disjoint images, $$\left\{
\begin{aligned}
& \epsilon_i: [0,\infty) \times S^1 \hookrightarrow S, && i \in I_{\mathit{in}}, \\
& \epsilon_i: (-\infty,0] \times S^1 \hookrightarrow S, && i \in I_{\mathit{out}}.
\end{aligned}
\right.$$
- A connection $A \in \Omega^1(S,{\mathfrak{g}})$ which is nonnegatively curved and compatible with the tubular ends, meaning that $$\label{eq:end-a}
\epsilon_i^*A = a_i = a_{i,t} \mathit{dt}$$ for some $a_i \in \Omega^1(S^1,{\mathfrak{g}})$. Moreover, each $a_i$ should have nontrivial holonomy.
We write $Y = (\{\epsilon_i\},A)$, and refer to $(S,Y)$ as a decorated surface (this is closely related to, but not exactly the same as, the notion from Section \[sec:tqft\]; see Section \[sec:end\] for further discussion). The connection $A$ gives rise to an equation for maps $S \rightarrow B$, which has the form on each end. The corresponding construction for maps to $E$ requires the following additional choices:
- For each end, we fix a Floer datum $\Phi_i = (J_i,H_i)$ in the class determined by $a_i$.
- Let $J = (J_z)_{z \in S}$ be a family of almost complex structures, $J_z \in {\EuScript J}(E)$, which on the tubular ends satisfies the following asymptotic condition: $$\label{eq:exponential-asymptotics}
J_{\epsilon_i(s,t)} \longrightarrow J_{i,t} \quad \text{exponentially fast as $s \rightarrow \pm\infty$.}$$
We need to make the last-mentioned condition more precise, in several respects. The difference between $J_{\epsilon_i(s,t)}$ and $J_{i,t}$ is measured after extension to the compactification $\bar{E}$. “Exponentially fast” means that in the $C^r$ sense for any $r$, the difference between the structures is bounded by a constant times $e^{-\lambda |s|}$, for some $\lambda>0$. Finally, for each end there should be a neighbourhood of infinity $N = N_{J,i} \subset N_E$, such that all $J_{\epsilon_i(s,t)}$ make $\pi$ pseudo-holomorphic on $\pi^{-1}(N)$.
- Take $K \in {C^\infty}(S \times E, T^*S)$, thought of as a one-form on $S$ with values in ${C^\infty}(E,{{\mathbb R}})$. This must be such that $K(\xi) \in {\EuScript H}(E, A(\xi))$ for all $\xi \in TS$. Over the ends, we require the following (this time, a strict equality, rather than an asymptotic condition): $$\epsilon_i^*K = H_{i,t} \mathit{dt}.$$
We will refer to the pair $\Pi = (J,K)$ as a “perturbation datum”. Let $X_K \in {C^\infty}(S \times E, T^*S \otimes_{{{\mathbb R}}} TE)$ be the Hamiltonian-vector-field-valued one-form associated to $K$. We will consider maps $$\label{eq:e-equation}
\left\{
\begin{aligned}
& u: S \longrightarrow E, \\
& (Du - X_K(u))^{0,1} = 0, \\
&\textstyle \lim_{s \rightarrow \pm\infty} u(\epsilon_i(s,t)) = x_i(t).
\end{aligned}
\right.$$ The Cauchy-Riemann equation in says that $Du - X_K(u): TS_z \rightarrow TE_{u(z)}$ should be complex-linear with respect to $J_{z}$. On each tubular end, setting $u_i(s,t) = u(\epsilon_i(s,t))$, this equation reduces to $$\label{eq:semi-floer}
\partial_s u_i + J_{\epsilon_i(s,t)} (\partial_t u_i - X_{i,t}(u_i)) = 0.$$ The limits $x_i$ are one-periodic orbits of $H_i$. In local coordinates $z=s+it$ on $S$ in which $K = K_1 \mathit{ds} + K_2 \mathit{dt}$, the equation is $$\partial_s u + J_{s,t} \partial_t u = X_{K_1(s,t)} + J_{s,t} X_{K_2(s,t)}.$$
\[th:transversality-on-s\] For generic choice of perturbation datum $\Pi$, keeping $\Phi_i$ fixed, the spaces of solutions of are regular. Moreover, if $u$ is a solution whose linearized operator $D_u$ has index $\leq 1$, then $u(z)$ never lies on a nontrivial $J_z$-holomorphc sphere.
Let’s fix a suitably small neighbourhood of infinity $N$, so that none of the $x_i$ is contained in $\pi^{-1}(N)$. Over each end, one of the following applies (this is a version of [@floer-hofer-salamon94 Lemma 4.1] with additional $s$-dependence of the almost complex structure, but has the same proof):
\(i) The set of points where $\partial_s u_i$ vanishes is discrete. If there is at least one end with this property, there are points $(s,t)$ where $\partial_s u_i \neq 0$ and $u_i(s,t) \notin \pi^{-1}(N)$. Transversality can then be achieved by perturbing the family $J_z$ near $z = \epsilon_i(s,t)$.
\(ii) $\partial_s u_i$ vanishes identically. This means that $u_i(s,t) = x_i(t)$. In particular, there are $t$ such that $z =\epsilon_i(0,t)$ satisfies $u(z) \notin \pi^{-1}(N)$. Since this is an open condition, there are also points $z \in S$ which are disjoint from the tubular ends and such that $u(z) \notin \pi^{-1}(N)$. We can then achieve transversality by varying $(J,K)$ near such a point.
We will now tackle the issue of compactness. This is based on similar ideas as in the case of Floer trajectories:
- There are two notions of energy for solutions of , paralleling and from our toy model discussion: $$\begin{aligned}
& E_{\mathit{geom}}(u) = {{\textstyle\frac{1}{2}}}\int_S |Du - X_K(u)|^2, \\
& E_{\mathit{top}}(u) = \int_S u^*\omega_E - d(K(u)).\end{aligned}$$ These are related by a curvature term, $$\label{eq:e-e-e}
E_{\mathit{top}}(u) = E_{\mathit{geom}}(u) + \int_S R_K(u).$$ We do not want to spell out the definition of $R_K$, but on a suitable neighbourhood of infinity, it agrees with the pullback of the corresponding term in . Since we are assuming that $A$ is nonnegatively curved, it follows that the integrand in is bounded below. Moreover, it vanishes identically on the tubular ends. It follows that there is a constant $C = C_K$ such that for all $u$, $$\label{eq:e-e-bounds}
E_{\mathit{geom}}(u) \leq E_{\mathit{top}}(u) + C.$$
- To any perturbation datum $\Pi = (J,K)$ one can associate an almost complex structure $J_\Pi$ on the trivial fibre bundle $S \times E \rightarrow S$, such that turns into the equation for $J_\Pi$-holomorphic sections. In local coordinates on $S$, this is given by the same kind of formula as in , $$\left\{
\begin{aligned}
& J_\Pi(Y) = J_{s,t} Y \quad \text{for $Y \in TE$,} \\
& J_\Pi(\partial_s + X_{K_1}) = \partial_t + X_{K_2}.
\end{aligned}
\right.$$ Near $\partial \bar{E}$, the vector fields $X_{K_1}, X_{K_2} \in {C^\infty}(E,TE)$ are the horizontal lifts of the corresponding vector fields $X_{A_1}, X_{A_2} \in {C^\infty}(B,TB)$. Hence, they extend smoothly to $\bar{E}$, and so does $J_\Pi$.
- At any point $z \in S$ such that $u(z)$ is sufficiently close to $\partial \bar{E}$, the projection $v = \pi(u)$ satisfies the equation associated to $A$.
\[th:u-bound-0\] Suppose that $(u_k)$ is a sequence of solutions of , with an upper bound on the topological energy. Suppose that there are points $z_k$, contained in a compact subset of $S$, such that $u_k(z_k) \rightarrow \partial \bar{E}$. Then, the $u_k$ converge to $\partial \bar{E}$ uniformly on compact subsets.
By , we have an upper bound on the geometric energy. From there on, the proof follows the same strategy as its Floer-theoretic counterpart, Lemma \[th:e-bound-0\]. This time, the applicability of Gromov compactness has to be explained in a slightly different way. Think of our maps as $J_\Pi$-holomorphic sections $S \rightarrow S \times \bar{E}$. From the bound on the geometric energy (and the fact that $X_K$ extends smoothly to $\bar{E}$), we get a $W^{1,2}$-bound on those sections, on each fixed compact subset of $S$. This already suffices for Gromov compactness, without having to introduce a symplectic structure on $S \times \bar{E}$ (see e.g. [@zinger17] for a suitable exposition). Alternatively, one can argue by the following familiar trick: since $J_\Pi | \{z\} \times \bar{E}$ is compatible with for each $z$, one can add a large positive two-form on $S$ to create a symplectic form on $S \times \bar{E}$ which at least tames $J_\Pi$.
The Gromov limit (of a subsequence, in the sense of convergence on compact subsets) consists of a solution $u_\infty: S \rightarrow \bar{E}$ of the same $\bar\partial$-equation as in , together with bubble components which are $J_z$-holomorphic spheres. By projecting to $\bar{B}$ and using the same convexity argument as in Lemma \[th:levi-apply\], one sees that $u_\infty^{-1}(\partial \bar{E})$ is open and closed. The same applies to the bubble components, as in Lemma \[th:e-bound-0\], and we again conclude that under our assumptions on $u_k$, the entire Gromov limit must lie in $\partial \bar{E}$.
\[th:u-bound-1\] Assume that the $a_i$ have elliptic or hyperbolic holonomies. Suppose that $(u_k)$ is a sequence of solutions of , with an upper bound on the topological energy. Then, there is a compact subset of $E$ which contains the images of all the $u_k$.
This is modelled on Lemma \[th:e-bound-1\]. Let’s suppose that, on the contrary (after passing to a subsequence), there are $z_k$ such that $u_k(z_k) \rightarrow \partial \bar{E}$.
\(i) If (again after passing to a subsequence) the $z_k$ are contained in a compact subset, we can apply Lemma \[th:u-bound-0\], which implies that $u_k \rightarrow \partial \bar{E}$ on compact subsets. In particular, this convergence holds for the restriction of $u_k$ to any fixed finite cylinder $\epsilon_i((s-l/2,s+l/2) \times S^1) \subset S$, which yields a sequence of solutions of that approach $\partial\bar{E}$. By taking $l$ to be sufficiently large, and looking at $v_k = \pi(u_k)$, one obtains a contradiction with Lemma \[th:hyp-cylinder\] or Lemma \[th:elliptic-cylinder\].
\(ii) The other case one has to deal with is where $z_k = \epsilon_i(s_k,t_k)$ for some $|s_k| \rightarrow \infty$. Consider the translated versions $$\label{eq:translated}
u_{k,i}(\cdot+s_k,\cdot): (-|s_k|,|s_k|) \times S^1 \longrightarrow E.$$ By definition, the image of $(0,t_k)$ under approaches $\partial \bar{E}$. Moreover, these maps satisfy a sequence of Cauchy-Riemann equations which, as $k \rightarrow \infty$, converge to an equation . A version of the argument from Lemma \[th:u-bound-0\] can be applied to this sequence; after that, one proceeds as in case (i).
(0,0)![\[fig:stretch\]Notation from the neck-stretching process.](stretch.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(6699,3336)(-2261,-3439) (2851,-3361)[(0,0)\[lb\]]{} (1876,-1561)[(0,0)\[lb\]]{} (976,-316)[(0,0)\[lb\]]{} (751,-1561)[(0,0)\[lb\]]{} (-1049,-3361)[(0,0)\[lb\]]{} (-149,-1561)[(0,0)\[lb\]]{} (-1649,-2161)[(0,0)\[lb\]]{} (1876,-2161)[(0,0)\[lb\]]{}
A frequently encountered class of generalizations is that where, instead of maps defined on a single Riemann surface, one has a family of surfaces. If the parameter space for the family is compact, the previous arguments carry over straightforwardly. A slightly more interesting situation is “neck-stretching”, where the Riemann surfaces split into pieces. Let’s consider the simplest example of such behaviour, which is one infinite cylinder splitting into two (the general case differs mainly in being even more cumbersome to formulate). The basic situation is:
- Fix $a_{\mathit{left}}, a_{\mathit{neck}}, a_{\mathit{right}} \in \Omega^1(S^1,{\mathfrak{g}})$, all of which have nontrivial holonomy.
- Fix constants $l_{\mathit{left}}, l_{\mathit{right}} > 0$, and a sequence $l_{\mathit{neck},k} > 0$, $l_{\mathit{neck},k} \rightarrow \infty$. These govern the length of various pieces of our cylinder, see Figure \[fig:stretch\].
- Let’s write $S_k$ ($k = 1,2,\dots$), $S_{\mathit{left}}$ and $S_{\mathit{right}}$ for copies of ${{\mathbb R}}\times S^1$. We will think of these as mapped to each other as follows: $$\label{eq:sigma-embeddings}
\left\{
\begin{aligned}
& \sigma_{\mathit{left}}: S_{\mathit{left}} \longrightarrow S_k, && \sigma_{\mathit{left}}(s,t) = (s-l_{\mathit{left}}/2-l_{\mathit{neck},k}/2,t), \\
& \sigma_{\mathit{right}}: S_{\mathit{right}} \longrightarrow S_k, && \sigma_{\mathit{right}}(s,t) = (s+l_{\mathit{right}}/2+l_{\mathit{neck},k}/2,t).
\end{aligned}
\right.$$ We give $S_{\mathit{left}}$ tubular ends, whose images are $(-\infty,-l_{\mathit{left}}/2] \times S^1$ and $[l_{\mathit{left}}/2,\infty) \times S^1$; and correspondingly for $S_{\mathit{right}}$. The ends of $S_k$ have image $(-\infty,-l_{\mathit{left}}-l_{\mathit{neck},k}/2] \times S^1$ and $[l_{\mathit{right}}+l_{\mathit{neck},k}/2,\infty) \times S^1$.
- Each $S_k$ carries a nonnegatively curved connection $A_k$ which restricts to $a_{\mathit{left}}$, $a_{\mathit{right}}$ on its tubular ends, and to $a_{\mathit{neck}}$ on $[-l_{\mathit{neck},k}/2, l_{\mathit{neck},k}/2] \times S^1$. Similarly, $S_{\mathit{left}}$ carries a nonnegatively curved connection $A_{\mathit{left}}$ which restricts to $a_{\mathit{left}}$, $a_{\mathit{neck}}$ on the tubular ends. Moreover, the pullbacks of $A_k$ by converge to $A_{\mathit{left}}$ on compact subsets. On $S_{\mathit{right}}$, we have a connection $A_{\mathit{right}}$ with corresponding properties.
When setting up the pseudo-holomorphic curve theory, the perturbation data have to be similarly correlated:
- We suppose that Floer data $\Phi_{\mathit{left}} = (J_{\mathit{left},t},H_{\mathit{left},t})$, $\Phi_{\mathit{neck}} = (J_{\mathit{neck},t},H_{\mathit{neck},t})$, and $\Phi_{\mathit{right}} = (J_{\mathit{right},t},H_{\mathit{right},t})$ have been chosen. The Hamiltonians should lie in the classes specified by the previous $a$’s. We also choose perturbation data $\Pi_k = (J_k,K_k)$, $\Pi_{\mathit{left}} = (J_{\mathit{left}}, K_{\mathit{left}})$, and $\Phi_{\mathit{right}} = (J_{\mathit{right}}, K_{\mathit{right}})$, whose behaviour over the ends is governed by the appropriate Floer data.
- For the inhomogeneous terms $K_k$ which are part of $\Pi_k$, we additionally prescribe the behaviour on the neck: $$K_{k,s,t} =
H_{\mathit{neck},t} \mathit{dt} \quad \text{for $-l_{\mathit{neck},k}/2 \leq s \leq l_{\mathit{neck},k}/2$.}$$ The pullback of $K_k$ by should converge on compact subsets to $K_{\mathit{left}}$, $K_{\mathit{right}}$.
- The requirements for the families of almost complex $J_k$ are the most complicated part of the definition, because of the asymptotic condition . Pulling back $J_k$ by $\sigma_{\mathit{left}}$, and then restricting to $(-\infty, -l_{\mathit{left}}/2] \times S^1$, should yield a family of almost complex structures which converges to $J_{\mathit{left}}$; and that convergence should be in any $C^r$ topology with sufficiently small exponential weights over the end $s \rightarrow -\infty$. Take the same pullback, and restrict to it $[-l_{\mathit{left}}/2, l_{\mathit{left}}/2+l_{\mathit{neck,k}}/2 + C] \times S^1$ for some constant $C$, which yields a family of almost complex structures parametrized by increasingly large parts of $S_{\mathit{left}}$. As $k \rightarrow \infty$, that family should converge to $J_{\mathit{left}}$; and that convergence should be in any $C^r$ topology with sufficiently small exponential weights (where the weight concerns the direction $s \rightarrow +\infty$). Parallel requirements apply to the pullback by $\sigma_{\mathit{right}}$.
This notion of neck-stretching is designed to allow suitable gluing arguments to go through, while preserving our previous transversality and compactness statements. More precisely, the appropriate version of Lemma \[th:transversality-on-s\] says that, keeping the $\Phi$’s fixed, a generic perturbation of the entire sequence $\Pi_k = (J_k,K_k)$ (within the allowed class) makes the associated moduli spaces regular for all $k$. Similarly, one gets a version of Lemma \[th:u-bound-1\], assuming that the holonomies involved (of $a_{\mathit{left}}, a_{\mathit{neck}}, a_{\mathit{right}}$) are elliptic or hyperbolic. The strategy of proof remains the same, with the breakdown into cases understood in an appropriate way. Namely, take maps $u_k: S_k \rightarrow E$ and points $z_k \in S_k$ such that $u_k(z_k) \rightarrow \partial \bar{E}$. (i) If (possibly after passing to a subsequence) the $z_k$ have preimages under which lie in a compact subset of $S_{\mathit{left}}$, then one considers a Gromov limit which is a solution of an equation involving $\Pi_{\mathit{left}}$; and similarly for $S_{\mathit{right}}$. (ii) In all other cases, one gets a limit which is a Floer trajectory for one of our chosen Floer data.
Formal aspects of Floer cohomology
----------------------------------
The rest of the construction of Floer cohomology is standard, and we will keep the discussion brief. From now on, we will only allow $a \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic or hyperbolic holonomy. We also restrict our notion of decorated surface, by requiring the same property for all $a_i$.
Fix an $a$. Choose an associated Floer datum $\Phi = (H,J)$, which satisfies the transversality statement from Lemma \[th:floer-transversality\]. We define a ${{\mathbb Z}}$-graded cochain complex over the Novikov field : $$\mathit{CF}^*(E,H) = \bigoplus_x {{\mathbb K}}_x,$$ where the sum is over the (finitely many) one-periodic orbits . Each ${{\mathbb K}}_x$ is an orientation space, which means that it is identified with ${{\mathbb K}}$ in a way which is canonical up to sign. To simplify the notation, we make an arbitrary choice of such an identification, and denote the resulting basis element in $\mathit{CF}^*(E,H)$ simply by $x$. The grading is given by Conley-Zehnder indices (which are well-defined since we have chosen a trivialization of the anticanonical bundle of $E$). The Floer differential is $$\label{eq:floer-diff}
d(x_+) = \sum_{u} \pm q^{E(u)} x_-,$$ where the sum is over all solutions of $u$ of which are isolated up to translation. Lemma \[th:e-bound-1\], together with more standard compactness arguments, shows that there are only finitely many such $u$ whose energy is lower than any given bound, hence that makes sense. The sign with which each $u$ contributes is determined by gluing of determinant lines. One defines $\mathit{HF}^*(E,H)$ to be the cohomology of $(\mathit{CF}^*(E,H),d)$.
Suppose that we have a decorated surface $(S,Y)$, and for each end, a Floer datum $\Phi_i$ whose Floer complex is well-defined. Choose a perturbation datum $\Pi$ satisfying Lemma \[th:transversality-on-s\]. Then, counting solutions of yields a chain map of even degree, $$\label{eq:floer-gamma}
\begin{aligned}
& \phi: \bigotimes_{i \in I_{\mathit{in}}} \mathit{CF}^*(E,H_i) \longrightarrow
\Big( \bigotimes_{i \in I_{\mathit{out}}} \mathit{CF}^*(E,H_i) \Big) [n(\chi(S)+|I_{\mathit{in}}|-|I_{\mathit{out}}|)], \\
& \phi(\otimes_{i \in I_{\mathit{in}}} x_i) = \sum_u \pm q^{E_{\mathit{top}}(u)} (\otimes_{i \in I_\mathit{out}} x_i).
\end{aligned}$$ Here, the sum is over solutions with limits $\{x_i\}$ which are isolated; the topological energy is used, since that is locally constant in families of solutions; and there is a sign, exactly as in . The same holds more generally if we have a family of such surfaces parametrized by a closed oriented manifold $P$ (the Floer data are parameter-independent, but the perturbation data obviously change). The only difference is that the degree of the associated map $\phi_P$ is $\mathrm{dim}(P)$ less than in .
The next simplest case is where the parameter space is a compact oriented manifold with corners. In that case, the associated operation is no longer a chain map, but instead a nullhomotopy for the operations associated to the boundary faces. Let’s suppose for simplicity that exactly one end is an output, say $I_{\mathit{out}} = \{0\}$ and $I_{\mathit{in}} = \{1,\dots,m\}$; this will be the case in all our applications. Then, the precise expression is exactly as in .
It remains to to take a brief look at gluing. A simple case would be where we have a parameter space $P$ with $\partial P = P_1 \times P_2$. The associated family of surfaces is defined only over $P \setminus \partial P$, and as we approach the boundary, it stretches along a neck, with the limit being two families over $P_1$ and $P_2$, respectively. The behaviour of the perturbation data must be as discussed at the end of Section \[subsec:worldsheets\]. One can then still define an operation associated to $P$, and that satisfies (under the same one-output assumption as before) $$\label{eq:boundary-stretching}
\begin{aligned}
& (-1)^{\mathrm{dim}(P)} d\phi_P(x_1,\dots,x_m) - \phi_P(dx_1,\dots,x_m) - \cdots
- (-1)^{|x_1|+\cdots+|x_{m-1}|} \phi_P(x_1,\dots,dx_m) \\ & \qquad + (-1)^{(\mathrm{dim}(P_1) + |x_1| + \cdots + |x_i|) \mathrm{dim}(P_2)} \phi_{P_1}(x_1,\dots,\phi_{P_2}(x_{i+1},\dots,x_{i+m_2}), \dots, x_m) = 0,
\end{aligned}$$ where $m = m_1+m_2-1$; note that $(-1)^{\mathrm{\dim}(P)} = (-1)^{|\phi_P|}$. As a final example, consider a mixture of the two previously encountered situations, where $\partial P = P_0 \sqcup -(P_1 \times P_2)$ (the minus sign indicates orientation-reversal). The associated family of surfaces should extend over $P_0$, but on the other boundary component, we have neck-stretching as before. Then, one gets $$\label{eq:boundary-stretching-2}
\begin{aligned}
& (-1)^{\mathrm{dim}(P)} d\phi_P(x_1,\dots,x_m) - \phi_P(dx_1,\dots,x_m) - \cdots \\ & \qquad \cdots
- (-1)^{|x_1|+\cdots+|x_{m-1}|} \phi_P(x_1,\dots,dx_m) + \phi_{P_0}(x_1,\dots,x_m) \\ & \qquad - (-1)^{(\mathrm{dim}(P_1) + |x_1| + \cdots + |x_i|) \mathrm{dim}(P_2)} \phi_{P_1}(x_1,\dots,\phi_{P_2}(x_{i+1},\dots,x_{i+m_2}), \dots, x_m) = 0.
\end{aligned}$$ In the simplest case where all the $P_k$ are closed and $m _k = 1$, this means that $(-1)^{\mathrm{dim}(P)}\phi_P$ yields a chain homotopy $$\phi_{P_0} {\simeq}(-1)^{|\phi_1| \cdot |\phi_2|} \phi_{P_1}\phi_{P_2}.$$ Comparing this with our previous formal TQFT discussion, and are a combination of and (the difference reflects the fact that “gluing” is an asymptotic process in the present context). Of course, much more complicated degenerations, where several ends are being stretched, are routinely used in Floer theory (and we will do the same). The discussion above was intended merely as a reminder, including sign conventions.
Conclusion\[sec:end\]
=====================
While Floer cohomology is not strictly speaking an instance of the TQFT formalism from Sections \[sec:tqft\]–\[sec:elliptic\], translating the formal arguments into Floer-theoretic ones is mostly a straightforward task (compare [@seidel16], which was structured in the same way). We will give a brief sketch of how this leads to the operations from Propositions \[th:operations-1\]–\[th:operations-3\]. After that, we will discuss connections, and how the material from Section \[sec:diff-axiom\] gives rise to the proofs of Propositions \[th:diff-q\] and \[th:two-connections\].
Floer cohomology groups associated to Lefschetz fibrations
----------------------------------------------------------
In Section \[sec:floer\], we explained how, given a connection $a \in \Omega^1(S^1,{\mathfrak{g}})$ with elliptic or hyperbolic holonomy, one chooses a compatible Floer datum $\Phi = (H,J)$ and obtains a Floer cochain complex $\mathit{CF}^*(E,H)$. We want to make a few additional comments concerning that construction:
- Suppose that $a$ has constant coefficients, which are sufficiently small. Then, one can choose $H$ with the same properties, and a version of the argument from [@hofer-salamon95] yields an isomorphism between the Morse and Floer cochain complexes, $$\label{eq:pss}
\mathit{CM}^*(E,H) {\cong}\mathit{CF}^*(E,H).$$
- The time variable $t \in S^1$ in the definition of Floer cohomology can be shifted. Given $a = a_t \mathit{dt}$, consider ${}^\tau a = a_{t+\tau} \mathit{dt}$, where $\tau \in S^1$ is an arbitrary constant. If $\Phi = (H_t,J_t)$ is a Floer datum for $a$, ${}^\tau \Phi = (H_{t+\tau}, J_{t+\tau})$ is a Floer datum for ${}^\tau a$. One-periodic orbits and Floer trajectories can be identified accordingly, leading to an isomorphism of chain complexes $$\label{eq:e1}
\mathit{CF}^*(E,{}^\tau H) {\cong}\mathit{CF}^*(E,H).$$
- In a similar spirit, let’s consider reversing the orientation of $S^1$, which in the notation of means passing from $a$ to $a^{-1}$. Given a Floer datum $\Phi = (H_t,J_t)$ for $a$, consider $\Phi^{-1} = (-H_{-t}, J_{-t})$. The flow of the new Hamiltonian is inverse to the old one, and one can similarly relate Floer trajectories by setting $u^{-1}(s,t) = u(-s,-t)$. Taking signs and grading into account, the outcome is an isomorphism of chain complexes $$\label{eq:e2}
\mathit{CF}^*(E,H^{-1}) {\cong}\mathit{CF}^{2n-*}(E,H)^\vee.$$
Let’s move on to the maps between Floer complexes. Following a standard strategy, one uses them first of all to construct continuation maps which show that, up to canonical isomorphism, the choice of Floer datum doesn’t affect Floer cohomology. That involves taking a cylinder $S = {{\mathbb R}}\times S^1$, with $A$ the pullback of $a$, and a perturbation datum which agrees with one choice of Floer datum on each end. We refer to [@salamon-zehnder92] for the original exposition of this process. Once it has been carried out, one can write the Floer cohomology groups as $\mathit{HF}^*(E,a)$. It is elementary to show that and are compatible with continuation maps. Therefore, one has canonical isomorphisms $$\begin{aligned}
\label{eq:rotate-time}
& \mathit{HF}^*(E,{}^\tau a) {\cong}\mathit{HF}^*(E,a) \quad \text{for any $\tau \in S^1$,} \\
\label{eq:flip-time}
& \mathit{HF}^*(E,a^{-1}) {\cong}\mathit{HF}^{2n-*}(E,a)^\vee.\end{aligned}$$
\[th:fractional-rotation-2\] The isomorphism is particularly interesting for a connection which is a $\mu$-fold pullback, in the sense of , for $\mu>1$. In that case, taking $\tau \in (\mu^{-1}{{\mathbb Z}})/{{\mathbb Z}}$ yields ${}^\tau a^\mu = a^\mu$, resulting in a ${{\mathbb Z}}/\mu$-action on $\mathit{HF}^*(E,a^\mu)$. Note that even though we are working with a connection which has a ${{\mathbb Z}}/\mu$-symmetry, the associated Floer datum does not have to obey the same restriction (thereby avoiding equivariant transversality problems). Concretely, the chain maps underlying the action are $$\xymatrix{
\mathit{CF}^*(E,H) \ar[r]^-{\eqref{eq:e1}}_-{{\cong}} & \mathit{CF}^*(E,{}^\tau H) \ar[rr]^-{\text{continuation}}_-{{\simeq}} && \mathit{CF}^*(E,H),
}$$ where ${}^\tau H \neq H$ in general; this is not a (chain level) isomorphism, but only a homotopy equivalence. Similarly, the group composition law is satisfied only up to homotopy.
Given $a_0,a_1 \in \Omega^1(S^1,{\mathfrak{g}})$ with the same rotation number, one can find a nonnegatively curved connection $A \in \Omega^1(S,{\mathfrak{g}})$ which agrees with the pullbacks of $a_0$ and $a_1$ on the ends (in the hyperbolic case, this is Proposition \[th:nonnegative-connection\]; in the elliptic case, the holonomies are conjugate, and one can choose $A$ to be flat). Via , that gives rise to a map $$\label{eq:hf-map}
\mathit{HF}^*(E,a_1) \longrightarrow \mathit{HF}^*(E,a_0).$$ To understand these maps, it is convenient to allow a little more freedom in the choice of $A$. Namely, let’s require that $$\label{eq:rotate-one-end}
A_{s,t} = \begin{cases} a_0 = a_{0,t} \mathit{dt} & s \ll 0, \\
{}^\tau a_1 = a_{1,t+\tau} \mathit{dt} & s \gg 0, \text{ for some $\tau \in S^1$,}
\end{cases}$$ and also choose the perturbation datum to have the correspondingly rotated asymptotic behaviour as $s \rightarrow +\infty$. In view of , this still gives rise to maps . An equivalent formulation would be to say that we still consider $S$ as a surface with tubular ends, but where the parametrization of one end has been modified; with respect to that, it still carries a connection that reduces to $a_0$ and $a_1$ over the ends. The implications depend on the kind of connections involved:
- For connections with hyperbolic holonomy and rotation number $r \in {{\mathbb Z}}\setminus \{0\}$, there is a ${{\mathbb Z}}/r$ worth of choices of , by Proposition \[th:x-annulus\]. A gluing argument (with geometrically nontrivial content, since the curvature condition means that one can’t just reverse orientation on $S$ to construct the connection underlying the inverse map) shows that all such maps are isomorphisms. The outcome is just as in Propositions \[th:auto\] and \[th:well-defined\]. Hence, we may write $\mathit{HF}^*(E,r)$ for the associated Floer cohomology group, which is well-defined up to isomorphism (up to canonical isomorphism if one is willing to additionally fix a choice of one-parameter subgroup, to which the holonomy of all connections will belong), and carries a canonical ${{\mathbb Z}}/r$-action. If we choose our connection to be an $r$-fold covering, the ${{\mathbb Z}}/r$-action is that from Example \[th:fractional-rotation-2\] (compare Example \[th:fractional-rotation\]).
- What we’ve said above also applies to the $r = 0$ case, but there is an additional quirk. The abstract theory (Proposition \[th:auto\]) predicts that $\mathit{HF}^*(E,0)$ has an endomorphism of degree $-1$, obtained by treating $\tau$ as a variable in , and looking at the corresponding parametrized problem (this is the BV operator in Hamiltonian Floer cohomology). However, by looking at the specific choice of connection and Floer datum giving rise to , one sees that the BV operator vanishes.
- For connections with elliptic holonomy, each group $\mathit{HF}^*(E,r)$ depends only on $r \in {{\mathbb R}}\setminus {{\mathbb Z}}$, up to canonical isomorphism, and carries a BV operator; compare Proposition \[th:only-elliptic\](i) and (ii).
\[th:geometric2\] The vanishing of the BV operator on $\mathit{HF}^*(E,0)$, and the duality , can fail in closely related geometric situations. Namely, suppose that we considered a map $\pi : E \rightarrow B$ with noncompact fibres, and which itself has a noncompact critical locus (say, of complex Morse-Bott type). We could use a version of Floer cohomology which “wraps” in fibre direction. In this context, neither property would hold any more, even though the formal TQFT framework remains relevant. Therefore, the failure of the abstract theory to predict these behaviours was to be expected.
Along the same lines, one can consider continuation maps $\mathit{HF}^*(E,r_1) \rightarrow \mathit{HF}^*(E,r_0)$ which increase the rotation number, $r_0 > r_1$. After adapting the corresponding TQFT results, we find that such maps are canonically defined in the following instances:
- $r_1 \in {{\mathbb Z}}$, and $r_0 = r_1+1$ (hyperbolic to hyperbolic, meaning the holonomies of the connections concerned); compare Proposition \[th:canonical-continuation\].
- $r_0,r_1 \in {{\mathbb R}}\setminus {{\mathbb Z}}$, and $(r_1,r_0)$ contains no integer (elliptic to elliptic); see Proposition \[th:only-elliptic\](iii). In this case, the continuation map turns out to be an isomorphism (as in Remark \[th:geometric2\], this is not a formal TQFT property, but standard from a geometric viewpoint).
- $r_0,r_1 \in {{\mathbb R}}\setminus {{\mathbb Z}}$, and $(r_1,r_0)$ contains exactly one integer (elliptic to elliptic). The corresponding TQFT result is again Proposition \[th:only-elliptic\](iii)
- $r_0 \in {{\mathbb Z}}$, and $r_1 \in (r_0-1,r_0)$ (elliptic to hyperbolic), see Proposition \[th:mixed-continuation-2\](i).
- $r_1 \in {{\mathbb Z}}$, and $r_0 \in (r_1,r_1+1)$ (hyperbolic to elliptic), see Proposition \[th:mixed-continuation-2\](ii).
Moreover, as long as one stays within this list of cases, the composition of continuation maps is again a continuation map. In particular, since $\mathit{SH}^*(E)$ is defined as a direct limit over $\mathit{HF}^*(E,r)$ for $r \notin {{\mathbb Z}}$, Proposition \[th:limit\] follows immediately, as in Proposition \[th:mixed-continuation-2\](iii). Along the same lines, we get maps , which factor through the BV operator as in Proposition \[th:mixed-continuation-2\](iv).
The algebraic structure of Floer cohomology
-------------------------------------------
Consider operations induced by surfaces other than the cylinder. The translation from the TQFT framework happens in the same way as before: we replace compact surfaces with boundary by their analogues with tubular ends (extending the connections in the obvious way), and choose perturbation data. The naive TQFT gluing operation is replaced by neck-stretching. Then,
- Proposition \[th:gerstenhaber-algebra\] carries over to give $\mathit{HF}^*(E,1)$ the structure of a Gerstenhaber algebra structure, proving Proposition \[th:operations-1\]. One can show as in Addendum \[th:all-a\] that this is well-defined (compatible with the canonical isomorphisms between different choices made in constructing the Floer cohomology group).
- Proposition \[th:gerstenhaber-module\] similarly yields a Gerstenhaber module structure on any $\mathit{HF}^*(E,r)$, which is Proposition \[th:operations-1b\] (and moreover, this module satisfies the additional relation from Proposition \[th:strange-relation\]). As before, this structure is compatible with the canonical isomorphisms that relate different constructions of $\mathit{HF}^*(E,1)$, and also with the corresponding non-canonical isomorphisms for $\mathit{HF}^*(E,r)$.
- We can define all the $\mathit{HF}^*(E,r)$, $r \in \{1,2,\dots\}$, using covers $a^r$ of the same underlying connection $a^1 = a$ with rotation number $1$. Then, by proceeding as in Proposition \[th:bigraded-algebra\], we get the algebra structure from Proposition \[th:operations-2\]. The same observation as in Example \[th:fractional-rotation-2\] applies: using $a^r$ is useful because it makes it easier to characterize which connections on surfaces to use (via Addendum \[th:partial-fix\]), but the Floer data do not have to be ${{\mathbb Z}}/r$-invariant (or related to each other for different $r$).
The remaining item is Proposition \[th:operations-3\], which deserves a little more discussion. The main point can be summarized as follows: the original TQFT construction, in Proposition \[th:bigraded-lie\], used , which was formulated as a strict chain level identity. This is potentially suspicious in a Floer-theoretic context, where everything depends on additional choices; however, we will ultimately only need the cohomology level relation, which is unproblematic. Fix positive integers $r_i$ ($i = 0,1,2$), with $r_0 = r_1+r_2-1$. Choose connections $a^{r_i}$ with rotation number $r_i$ and which are $r_i$-fold covers (not necessarily of the same connection). Also, choose corresponding Floer data $\Phi_i$. To define $$\label{eq:bi-bracket}
[\cdot,\cdot]: \mathit{HF}^*(E,r_1)^{{{\mathbb Z}}/r_1} \otimes \mathit{HF}^*(E,r_2)^{{{\mathbb Z}}/r_2} \longrightarrow \mathit{HF}^{*-1}(E,r_0)^{{{\mathbb Z}}/r_0},$$ we start with a family of surfaces parametrized by $p \in P = S^1$. Each such surface $S_p$ is a pair-of-pants, equipped with tubular ends $$\left\{
\begin{aligned}
&
\epsilon_{0,p}: (-\infty,0] \times S^1 \longrightarrow S_p, \\
&
\epsilon_{1,p},\epsilon_{2,p}: [0,\infty) \times S^1 \longrightarrow S_p.
\end{aligned}
\right.$$ of which those for $i = 1,2$ are determined only up to rotations $\epsilon_{i,p}(s,t) \mapsto \epsilon_{i,p}(s,t+k_i r_i^{-1})$, $k_i \in {{\mathbb Z}}/r_i$. The $p$-dependence follows the model from . One can arrange that there are isomorphisms $$\label{eq:p-shift-map}
\left\{
\begin{aligned}
& \phi_{0,p}: S_p \longrightarrow S_{p+r_0^{-1}}, \\
& \phi_{0,p} \epsilon_{0,p}(s,t) = \epsilon_{0,p+r_0^{-1}}(s,t-r_0^{-1}), \\
& \phi_{0,p} \epsilon_{i,p} = \epsilon_{i,p+r_0^{-1}} \quad \text{for $i = 1,2$,}
\end{aligned}
\right.$$ which generate a ${{\mathbb Z}}/r_0$-action on the family; compare . We equip our surfaces with connections $A_p$ satisfying $\epsilon_{i,p}^*A_p = a_i$, and which are compatible with . Again following , we consider the ${{\mathbb Z}}/r_1 \times {{\mathbb Z}}/r_2$-covering $\tilde{P} \rightarrow P$ associated to the element $$(1-r_2, 1-r_1) = (-r_0,-r_0) \in H^1(P;{{\mathbb Z}}/r_1 \times {{\mathbb Z}}/r_2) = {{\mathbb Z}}/r_1 \times {{\mathbb Z}}/r_2.$$ The pullback family of surfaces $S_{\tilde{p}}$ comes with well-defined tubular ends $\epsilon_{i,\tilde{p}}$ for $i = 0,1,2$, removing the previous ambiguity. The covering group acts on that family, fractionally rotating the ends. Moreover, the ${{\mathbb Z}}/r_0$-action from also lifts to $\tilde{P}$, resulting in a total action of ${{\mathbb Z}}/r_0 \times {{\mathbb Z}}/r_1 \times {{\mathbb Z}}/r_2$. Because that action rotates the tubular ends, it is not compatible with our given Floer data, which means that there is no meaningful notion of equivariant perturbation datum. Instead, we make an arbitrary choice of perturbation datum, and then divide the resulting operation by $r_1r_2$; the outcome is invariant under the action of ${{\mathbb Z}}/r_0 \times {{\mathbb Z}}/r_1 \times {{\mathbb Z}}/r_2$ on the source and target Floer cohomology groups, and we restrict to the invariant parts to get . This follows , except that we remain on the cohomology level (since the Floer cochain complexes do not carry group actions; see Example \[th:fractional-rotation-2\]).
Now consider the composition of two such brackets, say $$\label{eq:composition-bracket}
\begin{aligned}
& \mathit{HF}^*(E,a^{r_1})^{{{\mathbb Z}}/r_1} \otimes \mathit{HF}^*(E,a^{r_2})^{{{\mathbb Z}}/r_2} \otimes \mathit{HF}^*(E,a^{r_3})^{{{\mathbb Z}}/r_3} \\ & \qquad
\xrightarrow{[\cdot,\cdot] \otimes \mathit{id}} \mathit{HF}^{*-1}(E,a^{r_1+r_2-1})^{{{\mathbb Z}}/r_1+r_2-1}
\otimes \mathit{HF}^*(E,a^{r_3})^{{{\mathbb Z}}/r_3} \\\ & \qquad \xrightarrow{[\cdot,\cdot]} \mathit{HF}^{*-2}(E,a^{r_1+r_2+r_3-2})^{{{\mathbb Z}}/r_1+r_2+r_3-2}.
\\
\end{aligned}$$ Let’s write $\tilde{P}_1$ and $\tilde{P}_2$ for the parameter spaces that occur in the two steps of . Both spaces carry actions of ${{\mathbb Z}}/r_0$, where $r_0 = r_1+r_2-1$. To prove the Jacobi identity following the strategy from Proposition \[th:bigraded-lie\], one has to relate to the operation derived from the quotient space $\tilde{P}_1 \times_{{{\mathbb Z}}/r_0} \tilde{P}_2$. For that, one argues as follows: gluing writes as an operation arising from a family of surfaces over $\tilde{P}_1 \times \tilde{P}_2$ (for an arbitrary choice of perturbation data). On that family, the ${{\mathbb Z}}/r_0$-action no longer affects the tubular ends, and hence, we can choose the perturbation data to be equivariant. In other words, our family descends to $\tilde{P}_1 \times_{{{\mathbb Z}}/r_0} \tilde{P}_2$, including tubular ends; we can make a generic choice of perturbation data on that quotient space, and then pull that choice back to $\tilde{P}_1 \times \tilde{P}_2$. For those specific choices, we have an analogue of on the Floer cochain level. Hence, the corresponding Floer cohomology level identity holds.
The differentiation axiom revisited\[subsec:diff-axiom-2\]
----------------------------------------------------------
We will now implement differentiation with respect to the Novikov parameter $q$ in Floer theory. This closely follows [@seidel16 Section 9]. As a preliminary step, it is convenient to modify slightly the definition of the Floer differential and of the operations ; the new version is isomorphic to the original one (admittedly, the isomorphism is canonical only up to powers of $q$, but that is hopefully a venial sin; in any case, all previous constructions also go through with the new version, without any other modifications). Let’s fix an ${{\mathbb R}}$-cycle of codimension two, which means a formal finite linear combination $$\Omega = \sum_j m_j\Omega_j,$$ where $\Omega_j \subset E$ is a properly embedded codimension two submanifold, and the multiplicities are $m_j \in {{\mathbb R}}$. This should be such that the Poincaré dual to $\Omega$ is the symplectic class $[\omega_E]$. When carrying out any Floer-theoretic construction, we assume that the one-periodic orbits of the relevant Hamiltonians are disjoint from $\Omega$; and when counting maps $u: S \rightarrow E$ satisfying a Cauchy-Riemann equation with suitable asymptotics, we replace the expression $q^{\int_u \omega_E}$ with the intersection number $$\label{eq:usual-power}
q^{u \cdot \Omega} = q^{\sum_j m_j (u \cdot \Omega_j)}.$$ The naive differentiation operator on any Floer cochain group is $$\begin{aligned}
& {\nabla\mkern-12mu/\mkern2mu}: \mathit{CF}^*(E,H) \longrightarrow \mathit{CF}^*(E,H), \\
& {\nabla\mkern-12mu/\mkern2mu}(\textstyle \sum_x c_x x) = \sum_x (\partial_q c_x) x,
\end{aligned}$$ where the coefficients are $c_x \in {{\mathbb K}}$. Let’s assume that our Floer datum $\Phi = (H,J)$ has been chosen in such a way that the isolated solutions $u$ of Floer’s equation intersect $\Omega$ transversally, and no $J_t$-holomorphic sphere intersects $\Omega$ (these conditions are satisfied generically, thanks to transversality of evaluations). Then, one can write $$\label{eq:define-r}
\begin{aligned}
& R = {\nabla\mkern-12mu/\mkern2mu}d - d {\nabla\mkern-12mu/\mkern2mu}: \mathit{CF}^*(E,H) \longrightarrow \mathit{CF}^{*+1}(E,H), \\
& R (x_+) = \sum_{x_-} \Big( \sum_u \pm (u \cdot \Omega) q^{u \cdot \Omega-1} \Big) x_-
\end{aligned}$$ as an operation associated to a certain parametrized moduli space. Namely, for each $\Omega_j$, we consider pairs $(p,u)$ consisting of $p \in P = S^1$ and a Floer trajectory $u$, such that at the point $\zeta_p = (0,-p)$, we have $u(\zeta_p) \in \Omega_j$. We count those those solutions with multiplicity $q^{-1}m_j$ times as the usual power of $q$, and add up over $j$; see [@seidel16 Section 9a]. In future, whenever such a weighted count is used, we will refer to it by the shorthand notation “counting those $u$ such that $u(\zeta_p) \in q^{-1}\Omega$”.
For expository reasons, let’s briefly consider the Floer-theoretic analogue of , namely the unit class in $\mathit{HF}^0(E,1)$. This is constructed as follows. Choose some $a_1 \in \Omega^1(S^1,{\mathfrak{g}})$ with $\mathrm{rot}(a_1) = 1$. Take $S = {{\mathbb C}}$, considered as a Riemann surface with an output end. We equip that with a nonnegatively curved connection which reduces to $a_1$ on the end. Choose a Floer datum $\Phi_1 = (H_1,J_1)$ associated to $a_1$, and a corresponding perturbation datum. Counting solutions of the resulting equation then yields a cocycle representative for the unit in $\mathit{CF}^0(E,H_1)$. Now let’s additionally fix a point $\zeta \in S$, and impose the condition $$u(\zeta) \in q^{-1}\Omega$$ on the solutions of our equation . The outcome is the analogue of the Kodaira-Spencer class in $\mathit{HF}^2(E,1)$, represented by a Floer cocycle $$\label{eq:floer-k-s}
k \in \mathit{CF}^2(E,H_1).$$ In parallel with Lemma \[th:bracket-with-k\], one has a chain homotopy $$\label{eq:geometric-rho}
\begin{aligned}
& \rho: \mathit{CF}^*(E,H_2) \longrightarrow \mathit{CF}^*(E,H_2), \\
& d\rho(x) - \rho(dx) - R(x) + [k,x] = 0.
\end{aligned}$$ The context for this is that one takes a connection $a_2$ with hyperbolic holonomy and arbitrary rotation number $r$, as well as an associated Floer datum $(H_2,J_2)$, chosen so that the operator $R$ on $\mathit{CF}^*(E,H_2)$ admits the enumerative interpretation described above. The definition of involves a family of Riemann surfaces with marked point $(S_p,\zeta_p)$, parametrized by $p = (p_1,p_2) \in P =[0,\infty) \times S^1$, which for $p_1 = 0$ reduce to those defining . At the other end $p_1 \rightarrow \infty$, we “pull out” the marked point so that the surface degenerates into two pieces, one defining the Lie bracket and the other . The process is shown schematically in Figure \[fig:pull-out\]. Of course, our surfaces need to carry connections, and those are constructed exactly as in Lemma \[th:bracket-with-elliptic-k\]. In particular, this construction constrains how the ends of the limiting pair-of-pants ($p_1 = \infty$) change depending on the remaining parameter $p_2 \in S^1$ (see Lemma \[th:rot-bracket\]).
(0,0)![\[fig:pull-out\]The geometry behind , with the rotation numbers of the connections associated to the ends. Note that there is another parameter $p_2 \in S^1$, not shown here.](pullout.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(7910,2283)(-239,-1723) (1876,-886)[(0,0)\[lb\]]{} (3451,-661)[(0,0)\[lb\]]{} (2251,-661)[(0,0)\[lb\]]{} (7126,389)[(0,0)\[lb\]]{} (-154,-886)[(0,0)\[lb\]]{} (826,-136)[(0,0)\[lb\]]{} (6751,-1486)[(0,0)\[lb\]]{} (4396,-886)[(0,0)\[lb\]]{} (6771,-291)[(0,0)\[lb\]]{}
The corresponding construction for connections with elliptic holonomy, which is the subject of [@seidel16], is slightly more complicated. In that case, one can’t express $R$ itself as a Lie bracket, but has to combine it with a continuation map, so as to obtain the analogue of : $$\label{eq:geometric-rho-c}
\begin{aligned}
& \rho_c: \mathit{CF}^*(E,H_2) \longrightarrow \mathit{CF}^*(E,H_0), \\
& d\rho_c(x)- \rho_c(dx) - C(R(x)) + [k,x]_c = 0.
\end{aligned}$$ The context for this equation is as follows. Throughout, we work with connections of the form and . We fix numbers $r_i$ as in Lemma \[th:bracket-with-elliptic-k\], and Floer data $(H_i,J_i)$ that are compatible with the connections $r_i \alpha\, \mathit{dt}$. This means that on a neighbourhood of infinity, the Hamiltonian vector field of $H_i$ is (time-independent, and) the horizontal lift of the infinitesimal rotation of $B$ with speed $r_i$ (this is a version of the standard setup for symplectic cohomology). The family of surfaces which defines is parametrized by $p = (p_1,p_2) \in P = {{\mathbb R}}\times S^1$, and splits as we approach either end $p_1 \rightarrow \pm\infty$ (see Figure \[fig:pullout-2\]). The parameter $c \in {{\mathbb Z}}$ affects only the geometry of the limit $p_1 \rightarrow +\infty$. In that limit, the cylinder splits into a pair-of-pants and a copy of ${{\mathbb C}}$. The copy of ${{\mathbb C}}$ always carries the data defining a version of the Kodaira-Spencer cocycle, lying in $\mathit{CF}^2(E,H_1)$, which is independent of the remaining parameter $p_2$. In contrast, the pair-of-pants does depend on $p_2$ and on $c$ (taking the tubular ends into account). It is convenient to write it as $$S_{p_2} = ({{\mathbb R}}\times S^1) \setminus \{(0,-p_2)\},$$ with coordinates $(s,t)$. Then, the two ends asymptotic to $s = \pm \infty$ can be taken to be standard, hence independent of $p_2$; while the parametrization of the end asymptotic to $(0,p_2)$ rotates $(c+1)$ times as $p_2$ goes around the circle.
(0,0)![\[fig:pullout-2\]The geometric construction of .](pullout2.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(8181,2058)(-1064,-1498) (3451,-661)[(0,0)\[lb\]]{} (4318,-863)[(0,0)\[lb\]]{} (2446,-661)[(0,0)\[lb\]]{} (6301,-1336)[(0,0)\[lb\]]{} (6301,-361)[(0,0)\[lb\]]{} (6526,389)[(0,0)\[lb\]]{} (1201,-136)[(0,0)\[lb\]]{} (2101,-886)[(0,0)\[lb\]]{} (501,-886)[(0,0)\[lb\]]{} (-1049,-886)[(0,0)\[lb\]]{}
In the proof of Proposition \[th:two-connections\] (modelled on that of Proposition \[th:comparison-of-connections\]), a variant $\rho_{\mathit{broken},c}$ of occurs, in which the continuation map is replaced by a composition of two such maps. As far as the Floer-theoretic construction is concerned, the only difference is that in the limit $p_1 \rightarrow -\infty$, we have a “double stretching” of two necks as the same time (the relative speed at which this happens is irrelevant, as long as both necks stretch unboundedly). The comparison between the two constructions is given by a higher homotopy, which is the counterpart of . The three-dimensional parameter space involved is drawn schematically in Figure \[fig:triangle2\]. Topologically, this space can be thought of as a triangle, with a vertex and the opposite side removed, times $S^1$. There are two codimension $1$ boundary faces (along which the surfaces do not degenerate), plus one more such face at infinity, which would appear when compactifying it (and along which a standard stretching process happens). We want to pay particular attention to what happens at the vertices of our triangle:
- Near the vertex on the right (as drawn in Figure \[fig:triangle2\]), the parameter space has local coordinates $(p_1,p_2,p_3) \in {{\mathbb R}}\times[0,1] \times S^1$, where $p_1 \gg 0$. We can choose the Riemann surfaces and all their auxiliary data to be independent of $p_2$, while preserving regularity of the parametrized moduli space. In the region where that holds, no isolated points of that moduli space can occur; and any one-dimensional components that occur are of the form $\{p_1\} \times [0,1] \times \{p_3\}$.
- Near the top left vertex, we have local coordinates $(p_1,p_2,p_3) \in {{\mathbb R}}\times (-\infty,0] \times S^1$, where $p_1 \ll 0$ and $p_2$ is small. Here, $p_2 = 0$ is the adjacent boundary face, and $p_1 = -\infty$ the “face at infinity”. We again choose all data to be $p_2$-independent, with similar consequences as before.
- Near the bottom left vertex, we have local coordinates $(p_1,p_2,p_3) \in {{\mathbb R}}\times [0,\infty) \times S^1$, where $p_1 \ll 0$ and $p_2$ is small. Here, we have two “gluing parameters” or stretching lengths. Along the “face at infinity” $p_1 = -\infty$, the $p_2$ parameter governs the gluing of the left two of the three cylinders (the last cylinder remains the same). Along the boundary face $p_2 = 0$, we carry out both gluing processes simultaneously (with some relation between the gluing parameters). This is, roughly speaking, half of the more standard process near a codimension two corner (since we do not consider a region where only the right two cylinders are glued together).
(0,0)![\[fig:triangle2\]Relating $\rho_c$ and $\rho_{\mathit{broken},c}$. As in Figure \[fig:pull-out\], there is an additional $S^1$ degree of freedom, which we have not represented.](triangle2.pdf "fig:")
\#1\#2\#3\#4\#5[ @font ]{}
(7607,3777)(-9311,-3451) (-6974,-1636)[(0,0)\[lb\]]{}
For the other parameter spaces appearing in the proof of Proposition \[th:two-connections\] (in particular that from Figure \[fig:pentagon\]), similar considerations apply when translating them into Floer-theoretic moduli spaces.
[10]{} W. Abikoff. , volume 820 of [*Lecture Notes in Math.*]{} Springer, 1980.
J. Barge and E. Ghys. Cocycles d’[E]{}uler et de [M]{}aslov. , 294:235–265, 1992.
G. Ben Simon and D. Salamon. Homogeneous quasimorphisms on the symplectic linear group. , 175:221–224, 2010.
K. Diederich and A. Sukhov. Plurisubharmonic exhaustion functions and almost complex [S]{}tein structures. , 56:331–355, 2008.
M. Entov. Commutator length of symplectomorphisms. , 79:58–104, 2004.
A. Floer, H. Hofer, and D. Salamon. Transversality in elliptic [M]{}orse theory for the symplectic action. , 80:251–292, 1995.
H. Gaussier and A. Suhkov. Levi-flat filling of real two-spheres in symplectic manifolds [(I)]{}. , XX:515–539, 2011.
E. Getzler. atalin-[V]{}ilkovisky algebras and 2d topological field theories. , 159:265–285, 1994.
E. Ghys. Groups acting on the circle. , 47:329–407, 2001.
W. Goldman. Topological components of spaces of representations. , 93:557–607, 1988.
N. Hitchin. The self-duality equations on a [R]{}iemann surface. , (3) 55:58–126, 1987.
H. Hofer and D. Salamon. Floer homology and [N]{}ovikov rings. In [*The [F]{}loer memorial volume*]{}, pages 483–524. Birkh[ä]{}user, 1995.
S. Kobayashi. . World Scientific, 2nd edition, 2005.
N. Kowalzig and U. Kraehmer. Batalin-[V]{}ilkovisky structures on $\mathit{{E}xt}$ and $\mathit{{T}or}$. , 697:159–219, 2014.
F. Lalonde and D. McDuff. Positive paths in the linear symplectic group. In [*The [A]{}rnold-[G]{}elfand mathematical seminars*]{}, pages 361–387. Birkhäuser, 1997.
F. Lalonde and A. Teleman. The [$g$]{}-areas and the commutator length. vol. 24, 2013.
F. Luo. On [T]{}eichmüller spaces of surfaces with boundary. , 139:463–482, 2007.
D. McDuff and D. Salamon. . Oxford University Press, 2nd edition, 1999.
D. McDuff and D. Salamon. . Amer. Math. Soc., 2004.
M. McLean. Symplectic homology of [L]{}efschetz fibrations and [F]{}loer homology of the monodromy map. , 18:473–512, 2012.
J. Milnor. On the existence of a connection with curvature zero. , 32:215–223, 1958.
G. Mondello. Topology of representation spaces of surface groups in [$\mathit{PSL}_2(\mathbb{R})$]{} with assigned boundary monodromy and nonzero Euler number. , 12:399–462, 2016.
A. Ritter. Topological quantum field theory structure on symplectic cohomology. , 6:391–489, 2013.
D. Salamon and E. Zehnder. Morse theory for periodic solutions of [H]{}amiltonian systems and the [M]{}aslov index. , 45:1303–1360, 1992.
P. Salvatore and N. Wahl. Framed discs operads and [B]{}atalin-[V]{}ilkovisky algebras. , 54:213–231, 2003.
M. Schwarz. . PhD thesis, [ETH]{} [Z]{}[ü]{}rich, 1995.
P. Seidel. A biased survey of symplectic cohomology. In [*Current Developments in [M]{}athematics ([H]{}arvard, 2006)*]{}, pages 211–253. Intl. Press, 2008.
P. Seidel. Fukaya [$A_\infty$]{}-structures associated to [L]{}efschetz fibrations. [I]{}. , 10:325–388, 2012.
P. Seidel. ukaya [$A_\infty$]{}-categories associated to [L]{}efschetz fibrations. [II]{}. In [*Algebra, [G]{}eometry and [P]{}hysics in the 21st [C]{}entury ([K]{}ontsevich [F]{}estschrift)*]{}, pages 295–364. Birkh[ä]{}user, 2017.
P. Seidel. Fukaya [$A_\infty$]{}-structures associated to [L]{}efschetz fibrations. [II]{} 1/2. , 20:883–944, 2016.
P. Seidel. Fukaya [$A_\infty$]{}-structures associated to [L]{}efschetz fibrations. [III]{}. , to appear.
J. Slimowitz. The positive fundamental group of [${\rm Sp}(2)$]{} and [${\rm
Sp}(4)$]{}. , 109:211–235, 2001.
Z. Sylvan. On partially wrapped [F]{}ukaya categories. Preprint arXiv:1604.02540, 2016.
J. Wood. Bundles with totally disconnected structure group. , 46:257–273, 1971.
D. Zhuang. Irrational stable commutator length in finitely presented groups. , 2:499–507, 2008.
A. Zinger. Notes on [$J$]{}-holomorphic maps. Preprint ArXiv:1706.00331, 2017.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We show that the combination of a Dresselhaus interaction and a spatially periodic Rashba interaction leads to the formation of a helical liquid in a quantum wire when the electron-electron interaction is weakly screened. The effect is sustained by a helicity-dependent effective band gap which depends on the size of the Dresselhaus and Rashba spin-orbit couplings. We propose a design for a semiconductor device in which the helical liquid can be realized and probed experimentally.'
author:
- 'George I. Japaridze$^{1,2}$, Henrik Johannesson$^3$, and Mariana Malard$^4$'
title: Synthetic helical liquid in a quantum wire
---
The concept of a [*helical liquid*]{} $-$ a phase of matter where spin and momentum directions of electrons are locked to each other $-$ underpins many of the fascinating features of the recently discovered topological insulators [@Review]. In the case of an ideal two-dimensional (2D) topological insulator, electron states at its edges propagate in opposite directions with opposite spins, forming a one-dimensional (1D) helical liquid (HL) [@KaneMele; @BHZ]. Given the right conditions [@Wu; @Xu], the spin-filtered modes of the HL may serve as ballistic conduction channels [@Konig], holding promise for novel electronics/spintronics applications.
The HL is expected to exhibit several unusual properties, such as charge fractionalization near a ferromagnetic domain wall [@Qi], interaction-dependent response to pinching the sample into a point contact [@SJ], and enhanced superconducting correlations when two HLs are coupled together [@Tanaka]. A particularly tantalizing scenario is the appearance of Majorana zero modes in an HL in proximity to a superconductor and a ferromagnet [@FuKane]. However, testing these various predictions in experiments on the HgTe/CdTe quantum well structures in which the HL phase has been observed is a formidable challenge: The softness and reactivity of HgTe/CdTe makes it difficult to handle [@Molenkamp], and moreover, charge puddles formed due to fluctuations in the donor density may introduce a helical edge resistance [@Vayrynen]. Alternative realizations of the HL are therefore in high demand. The prospect of using the dissipationless current of an HL in future chip designs adds to the importance of this endeavor [@IBM].
![Device supporting a 1D synthetic helical liquid: Electrons in a single-channel quantum wire (blue) formed in a heterostructure supporting a Dresselhaus interaction are subject to a modulated Rashba field from a periodic sequence of charged top gates (dark grey).[]{data-label="fig:example"}](Device.pdf){width="2.4in"}
One suggestion is to use a nanowire made of a “strong topological insulator” material [@Review]. When pierced with a magnetic flux quantum, the electrons in the wire are predicted to form an interacting HL [@Egger]. In another scheme $-$ appearing in attempts to engineer Kitaev’s toy model [@Kitaev] for p-wave pairing [@LutchynOreg] $-$ electrons in a quantum wire form an HL when subject to a Rashba spin-orbit coupling combined with a transverse magnetic field [@StredaSeba]. These, like most other proposals for HLs in quantum wires [@Braunecker], specifically rely on the presence of a magnetic field.
In this article we show that an HL can be produced and controlled in a quantum wire using electric fields only. The advantages of employing electric fields rather than magnetic fields are manifold. Most importantly, an electric field does not corrupt the feature that counterpropagating helical modes carry antiparallel spins. Also, an electric field can easily be generated and applied locally, and eliminates many of the design complexities that come with the use of magnetic fields [@AwschalomSamarth]. Our proposed device (see FIG. 1) exploits an unexpected effect that appears when interacting electrons are subject to a Dresselhaus spin-orbit interaction combined with a [*spatially periodic*]{} Rashba interaction: When the electron density is tuned to a certain value, determined by the wavelength of the Rashba modulation, a band gap tied to the helicity of the electrons opens. This gives rise to an HL. Notably, the required setup for realizing this HL is built around standard nanoscale semiconductor technology, and is very different from the recently proposed all-electric setup in Ref. using carbon nanotubes. In what follows we derive an effective model that captures the surprising effect from the interplay between the Dresselhaus and the modulated Rashba interaction. We analyze the model and explain how the HL materializes, and also discuss the practicality and robustness of this novel type of a synthetic HL.
We consider a setup with a single-channel quantum wire formed in a gated 2D quantum well supported by a semiconductor heterostructure. The electrons in the well are subject to two types of spin-orbit interactions, the [*Dresselhaus*]{} and [*Rashba*]{} interactions [@Winkler]. For a heterostructure grown along $[001]$, with the electrons confined to the $xy$-plane, the leading term in the Dresselhaus interaction takes the form $H_\text{D} = \beta (k_x \sigma_x - k_y \sigma_y)$ with $\beta$ a material-specific parameter. The Rashba interaction is given by $H_\text{R}
= \alpha (k_x \sigma_y - k_y \sigma_x),$ where $\alpha$ depends on several distinct features of the heterostructure [@Sherman; @GolubIvchenko], including the applied gate electric field. The latter feature allows for a gate control of the Rashba coupling $\alpha$ [@LiangGao]. It is important to mention that large fluctuations of $\alpha$ [@Sherman] may drive the HL to an insulating state through an Anderson-type transition [@SJJ]. We shall return to this issue below.
Taking the $x$-axis along the wire, adding to $H_{\text{D}}$ and $H_{\text{R}}$ the kinetic energy of the electrons as well as the chemical potential, one obtains $-$ using a tight-binding formulation $-$ the Hamiltonian $H_0 + H_{\text{DR}}$, where $$\begin{aligned}
\label{uni}
H_0 &\! = \!&\! -t \sum_{n,\alpha}
c^{\dagger}_{n,\alpha}c^{\phantom{\dagger}}_{n+1,\alpha} +
\frac{\mu}{2}\sum_{n,
\alpha}c^{\dagger}_{n,\alpha}c^{\phantom{\dagger}}_{n,\alpha} + \mbox{h.c.}, \\
H_{\text{DR}} &\! = \!& -\!i\!\sum_{n,\alpha,\beta}
c^{\dag}_{n,\alpha}\!\left[\gamma_{\text{D}}\, \sigma^{x}_{\alpha\beta}
\!+ \!\gamma_{\text{R}}\, \sigma^{y}_{\alpha\beta}\right]
\!c^{\phantom{\dag}}_{n\!+\!1,\beta} \!+\! \mbox{h.c.},\end{aligned}$$ with $H_{\text{DR}}$ the second-quantized projection of $H_{\text{D}} + H_{\text{R}}$ along the wire. Here $c^{\dagger}_{n,\alpha}$ ($c^{\phantom{\dag}}_{n,\alpha}$) is the creation (annihilation) operator for an electron with spin ${\alpha}=\uparrow,\downarrow$ on site $n$ (with spin along the growth direction $\hat{z}$), $t$ is the electron hopping amplitude, and $\mu$ a chemical potential controllable by a back gate. The signs and magnitudes of $\gamma_{\text{D}} \!\equiv\! \beta a^{-1}$ and $\gamma_{\text{R}}
\!\equiv \!\alpha a^{-1}$ ($a$ being the lattice spacing) depend on the material as well as on the particular design of the heterostructure.
We now envision that we place a sequence of equally charged nanoscale electrodes on top of the heterostructure (cf. FIG. 1). As a result, the Rashba coupling will pick up a modulated contribution due to the modulation of the electric field from the electrodes. Taking their separation to be the same as their extension along the wire (cf. FIG. 1), we model the Rashba modulation by a simple harmonic, $$\label{mod}
H_\text{R}^\text{mod}=-i {\gamma_{\text{R}}^{\prime}} \sum_{n,\alpha,\beta}
\cos(Qna) c^{\dag}_{n,\alpha}\sigma^{y}_{\alpha\beta}c_{n+1,\beta}
\,+\, \mbox{h.c.},$$ with $\gamma_{\text{R}}^{\prime}$ the amplitude and $Q$ its wave number. Besides the modulation of the Rashba interaction, also the chemical potential gets modulated by the external gates: $$\label{modchempot}
H_\text{cp}^\text{mod}=\frac{\mu^{\prime}}{2} \sum_{n,\alpha}
\cos(Qna) c^{\dag}_{n,\alpha}c_{n,\alpha}
\,+\, \mbox{h.c.}$$ As follows from the analysis in Ref. , this term has no effect at low energies unless the electron density is tuned to satisfy the commensurability condition $|Q - 2k_{F}|<<O(1/a)\,\mbox{mod}\, 2\pi$, with $k_{F}$ the Fermi wave number: At all other densities, including those for which an HL emerges, $H_\text{cp}^\text{mod}$ in Eq. (\[modchempot\]) is rapidly oscillating and gives no contribution in the low-energy continuum limit. Hence, we shall neglect it here.
Given the full Hamiltonian $H \!=\! H_0 \!+ \!H_{\text{DR}} \!+ \!H_\text{R}^\text{mod}$, we pass to a basis which diagonalizes $H_0 \!+ \!H_{\text{DR}}$ in spin space, $$\label{spinor}
\left( \begin{array}{c}
d_{n,+} \\
d_{n,-} \end{array} \right)
\equiv \frac{1}{\sqrt{2}} \big( \begin{array}{c}
-i e^{-i\theta}c_{n,\uparrow}\, + \, e^{i\theta}c_{n,\downarrow} \\
\ \ e^{-i\theta}c_{n,\uparrow} -i e^{i\theta}c_{n,\downarrow} \end{array} \big),$$ with $2\theta=\arctan{\gamma_{\text{D}}/\gamma_{\text{R}}}$. The index $\tau\! = \!\pm$ of the operator $d_{n, \tau}$ label the new quantized spin projections along the direction of the combined Dresselhaus ($\propto \gamma_{\text{D}} \hat{x}$) and uniform Rashba ($\propto \gamma_{\text{R}} \hat{y}$) fields. Putting $\gamma_{\text{R}}^{\prime}=0$ in Eq. (\[mod\]) and using (\[spinor\]), the system is found to exhibit four Fermi points $\pm k_F + \tau q_0$, where $q_{0}a=\arctan\sqrt{(\tilde{t}/t)^{2}-1}$ with $\tilde{t}=\sqrt{t^{2}+\gamma_{\text{R}}^{2}+\gamma^{2}_{\text{D}}}$, and where $k_{F}= \pi \nu/a$ with $\nu=N_{e}/2N$, $N_e \, [N]$ being the number of electrons \[lattice sites\]. The corresponding Fermi energy $\epsilon_F$ is given by $\epsilon_F =
-2\tilde{t}\cos(k_{F}a)+\mu$.
To analyze what happens when $\gamma_{\text{R}}^{\prime}$ is switched on, we focus on the physically relevant limit of low energies, linearize the spectrum around the Fermi points and take the continuum limit $na \rightarrow x$. By decomposing $d^{\phantom{\dagger}}_{n,\tau}$ into right- and left-moving fields $R^{\phantom{\dagger}}_{\tau}(x)$ and $L^{\phantom{\dagger}}_{\tau}(x)$, $$\label{Decomposition}
d^{\phantom{\dagger}}_{n,\tau} \rightarrow \sqrt{a}
\big(\mbox{e}^{i(k_F+\tau q_0)x}
R^{\phantom{\dagger}}_{\tau}(x) + \mbox{e}^{i(-k_F +\tau q_0)x} L^{\phantom{\dagger}}_{\tau}(x)\big),$$ and choosing $|Q - 2(k_F + \tau q_0)| << O(1/a)\,\mbox{mod}\, 2\pi$ one thus obtains an effective theory with two independent branches, $H\!\rightarrow \sum_{i=1,2}\int dx \ {\cal H}_{i}$, where ${\cal
H}_{1}$ applies to the Fermi points $\pm k_F \mp q_0$, and ${\cal
H}_{2}$ to $\pm k_F \pm q_0$. We here choose $Q=2(k_F + q_0)$, and come back to the general case below. Omitting all rapidly oscillating terms that vanish upon integration, one finds $$\begin{aligned}
\label{linear-Hamiltonian}
{\cal H}_{1}&\!=\!& -iv_F
(:\!R^{\dag}_{-}\partial_{x}R^{\phantom{\dagger}}_{-}\!: -
:\!L^{\dag}_{+}\partial_{x}L^{\phantom{\dagger}}_{+}\!:) \\
{\cal H}_{2}&\!=\!& -iv_F (
:\!R^{\dag}_{+}\partial_{x}R^{\phantom{\dagger}}_{+}\!:-
:\!L^{\dag}_{-}\partial_{x}L^{\phantom{\dagger}}_{-}\!:) \nonumber \\
\label{2linear}
&+&i\lambda\, (\!R^{\dag}_{+}\partial_{x}L^{\phantom{\dagger}}_{-}\!+
\!L^{\dag}_{-}\partial_{x}R^{\phantom{\dagger}}_{+}\! ),\end{aligned}$$ where $v_{F}=2a\tilde{t}\sin(\pi\nu)$, $\lambda =
a \gamma^{\prime}_{R}\gamma_{D}(\gamma_{R}^{2}+\gamma^{2}_{D})^{-1/2}$, $:...:$ denotes normal ordering, and where we have absorbed the constant phase $\mbox{e}^{i(k_F+q_0)a}$ into $R_+(x)$.
While the nondiagonal term in Eq. (\[2linear\]) is renormalization-group (RG) irrelevant in the absence of [*e-e*]{} interactions it may turn relevant and open a gap at the Fermi points $\pm k_F \pm q_0$ when the [*e-e*]{} interaction $$\label{ee}
H_{e{\text -}e} = \sum_{n,n'; \alpha, \beta} V(n-n')c^{\dagger}_{n,\alpha} c^{\dagger}_{n',\beta}
c^{\phantom{\dagger}}_{n', \beta} c^{\phantom{\dagger}}_{n, \alpha},$$ is included. Its low-energy limit can be extracted by following the procedure from above, and we obtain $H_{e{\text -}e} \!\rightarrow \int dx \ {\cal H}_{e{\text -}e}$, where $$\begin{aligned}
{\cal H}_{e{\text -}e} & = & g_1
\!:\!R^{\dag}_{\tau}L^{\phantom{\dagger}}_{\tau}L^{\dag}_{\tau'}
R^{\phantom{\dagger}}_{\tau'}\!: + \, {g}_{2}
\!:\!R^{\dag}_{\tau}R^{\phantom{\dagger}}_{\tau}
L^{\dag}_{\tau'}L^{\phantom{\dagger}}_{\tau'}\!: \nonumber \\
&+& \frac{g_{2}}{2}(
:\!L^{\dag}_{\tau}L^{\phantom{\dagger}}_{\tau}
L^{\dag}_{\tau'}L^{\phantom{\dagger}}_{\tau'}\!: + \, L
\rightarrow R), \label{LCHee}
\end{aligned}$$ with $\tau, \tau' = \pm$ summed over, and where $g_{1} \!\sim\! \tilde{V}(k\!\sim\!2k_F)$ and $g_{2}\!\sim\! \tilde{V}(k\!\sim \!0)$, $\tilde{V}(k)$ being the Fourier transform of the screened Coulomb potential $V(n-n')$ in Eq. (\[ee\]). The backscattering $\sim g_1$ is weak in a semiconductor structure and renormalizes to zero at low energies also in the presence of spin-orbit interactions [@Schulz]. In effect we are thus left with only the dispersive and forward scattering channels $\sim g_2$ in Eq. (\[LCHee\]), to be added to ${\cal H}_1$ and ${\cal H}_2$ from Eqs. (\[linear-Hamiltonian\]) and (\[2linear\]). Passing to a bosonized formalism [@Giamarchi_book_04], the resulting full Hamiltonian density can be written as ${\cal H} = {\cal H}^{(1)} + {\cal H}^{(2)} + {\cal H}^{(12)}$ with $$\begin{aligned}
{\cal H}^{(i)} \!&\!=\!& {\cal H}^{(i)}_0 \!+\!\!\frac{\lambda \,\delta_{i2}}{\sqrt{\pi K}a}\!\cos(\!\sqrt{4\pi K}\phi_2) \partial_x\theta_2, i\!=\!1,2 \label{bosoni} \\
{\cal H}^{(12)} \!&\!=\!& \frac{g_2K}{\pi}\partial_x \phi_1 \partial_x \phi_2 \label{boson12},\end{aligned}$$ where $K \approx (1\!+\!g_2/\pi v_F)^{-1/2}$. Here ${\cal
H}^{(i)}_0\!=\! u[(\partial_x \theta_{i})^2 \!+\! (\partial_x
\phi_{i})^2]$ is a free boson theory with $u\! \approx \!v_F/2K$, and with $\theta_i$ the dual field to $\phi_i$. The indices “1” and “2” tagged to the fields label the two branches originating from Eqs. (\[linear-Hamiltonian\]) and (\[2linear\]).
We should point out that our fields $\phi_i$ ($i\!=\!1,2$) are rotated with respect to the conventional bosonic fields $\phi^{R,L}_\tau$ ($\tau\!=\!\pm$) [@Shankar] representing the original fermion fields $R_{\tau}$ and $L_{\tau}$, $\phi_i=\phi^R_\pm+\phi^L_\mp$, with upper (lower) sign attached to $i\!=\!1\ (i\!=\!2)$. This nonstandard spin-mixing basis $\{\phi_i\}$ is suitable for revealing how the non-diagonal term in Eq. (\[2linear\]) combines with the e-e interaction in Eq. (\[LCHee\]) to gap out the states near $\pm k_F \pm q_0$: The term in Eq. (\[2linear\]) transforms into the sine-Gordon-like potential in Eq. (\[bosoni\]) [@Malard2], controlled by e-e interactions through the Luttinger liquid $K$-parameter. As we shall see, the theory brought on the form of Eqs. (\[bosoni\]) and (\[boson12\]) can be efficiently handled by using an adiabaticity argument.
To make progress we pass to a Lagrangian formalism by Legendre transforming Eqs. (\[bosoni\]) and (\[boson12\]). Using that $\Pi_i = \sqrt{K}\partial_x \theta_i$ serves as conjugate momentum to $\phi_i/\sqrt{K}$, $\Pi_i$ can be integrated out from the partition function $Z$, with the result $$Z \sim \int {\cal D}\phi_1 {\cal D}\phi_2 e^{-(S^{(1)} + S^{(2)} + S^{(12)})},$$ with Euclidean actions $$\begin{aligned}
\label{Actions}
\! \! \! \! \! \! \! \! S^{(i)} \!&\!=\!&\! S_0^{(i)} \! - \delta_{i2} \frac{m_0}{\pi a}\!\int\! d\tau dx \cos(\sqrt{16\pi K}\phi_2), i\!=\!1,2 \label{branch1and2} \\
\! \! \! \! \! \! \! \! S^{(12)} \!&\!=\!&\! \frac{g_2 K}{\pi}\! \int \!d\tau dx \partial_x \phi_1 \partial_x \phi_2. \ \ \ \label{branch12}\end{aligned}$$ Here $S_0^{(i)}= (1/2) \int d\tau dx [(1/v)(\partial_{\tau}\phi_i)^2 + v(\partial_x \phi_i)^2]$ is a free action with $v=2u$, and $m_0 ={\lambda}^2 /4Kva$.
Having brought the theory on the form of Eqs. (\[branch1and2\]) and (\[branch12\]), valid for a Rashba modulation with $Q=2(k_F+q_0)$, we first consider the auxiliary problem where the amplitude $g_2$ in Eq. (\[branch12\]) is replaced by a tunable parameter, $g_2'$ call it. Putting $g_2'=0$ and refermionizing $S^{(1)}$ we then obtain a helical Dirac action for the first branch (with Fermi points $\pm k_F \mp q_0$), with the second branch (with Fermi points $\pm k_F \pm q_0$) described by a sine-Gordon action, $S^{(2)}$. The cosine term in $S^{(2)}$ becomes RG relevant for $K<1/2$, driving this branch to a stable fixed point with massive soliton-antisoliton excitations [@Malard2]. The energy to create a soliton-antisoliton pair defines an insulating gap $\Delta$, and one finds from the exact solution of the sine-Gordon model [@Zamolodchikov] that $$\label{gap}
\Delta = c(K)\Lambda(\frac{m_0}{\Lambda})^{1/(2-4K)}, \ \ K<\frac{1}{2},$$ where $\Lambda =v/a$ is an energy cutoff, and $c(K)$ is expressible in terms of products of Gamma functions. The opening of a gap implies that the field $\phi_2$ gets pinned at one of the minima of the cosine term. Thus, in the neighborhood of the fixed point its gradient is suppressed with the effect that the action $S^{(12)}$ remains vanishingly small also after $g_2'$ has been restored to its true value, $g_2' \rightarrow
g_2$. In particular, it follows that $S^{(12)}$ cannot close the gap. Note that this “argument by adiabaticity” is perfectly well controlled as the approach to a stable fixed point rules out any non-analyticities in the spectrum. In summary, when $K<1/2$, a Rashba modulation $Q=2(k_F+q_0)$ opens a gap in the second branch which becomes insulating, [*leaving behind a conducting helical electron liquid*]{} in the first branch (see FIG 2(a)).
![(Color online) Schematic plot of the dispersion relations for the two types of helical liquid phases, with (a) $Q = 2(k_F + q_0)$ and (b) $Q = 2(k_F - q_0)$.[]{data-label="Fig_Spectrum_3"}](Bands.pdf "fig:")\
The analysis above is readily adapted to the case with $Q=2(k_F-q_0)$, and one finds that the gap now opens in the first branch. Note that our results remain valid in the presence of the weakened commensurability condition $|Q-2(k_F + \tau q_0)| \ll {\cal O}(1/a) \, \mbox{mod} \,2\pi, \tau = \pm$, as this condition still allows us to throw away the rapidly oscillating terms in the low-energy limit of $H_\text{R}^\text{mod}$.
Our interpretation of the dynamically generated gap $\Delta$ as an effective band gap $-$ as in FIG. 2 $-$ draws on a result by Schulz [@hjSchulz] where a bosonized theory similar to that defined by our Eqs. (\[bosoni\]) and (\[boson12\]) is refermionized into a non-interacting two-band model, with the bands separated by a gap corresponding to the dynamic gap of the bosonized theory. This picture $-$ while heuristic only $-$ helps to conceptualize the role of the commensurability conditions for the emergence of the synthetic HL.
The fact that [*e-e*]{} interactions can open a gap in an HL is well-known from the literature [@Wu; @Xu; @LutchynOreg]. In particular, Xu and Moore [@Xu] noted that if a dynamically generated gap opens in one of two coexisting Kramers’ pairs ([*alias*]{} ‘branches’ 1 and 2 in our model), this gives rise to a stable HL in the other pair. Their observation pertains to the case where the scattering within each branch is governed by distinct strengths of the [*e-e*]{} interaction: a gap may then open in the branch with the stronger interaction. For this reason the Xu-Moore observation does not apply to the realistic case of of a single quantum wire with the same interaction strength in the two spin-split branches. This is where our proposal injects a novel element into the picture: By properly combining a modulated Rashba spin-orbit interaction with a Dresselhaus interaction we find that the gap-opening mechanism from [*e-e*]{} interactions can indeed be triggered in such a way as to open a gap in one of the branches only, leaving behind a stable HL in the other. This HL is of a new type compared to the ones hitherto probed experimentally: It owes its existence neither to being ‘holographic’ [@JapanReview] (like the edge states of an HgTe QW [@Konig]) nor to being ‘quasi-helical’ [@Braunecker] (as is the case for magnetic-field assisted HLs [@Quay]). The time-reversal analogue of the notorious fermion-doubling problem [@Nielsen] is instead circumvented by the fact that the gapped branch breaks time-reversal symmetry [*spontaneously*]{} by developing a spin-density wave [@WuDiscussion]. As there is no need to apply a magnetic field to realize the synthetic HL, it escapes the complications from time-reversal symmetry breaking that mar a quasi-helical liquid [@Braunecker]. By this, it becomes an attractive candidate for renewed Majorana fermion searches [@Kouwenhoven].
Having established a proof of concept that a synthetic HL can be sustained in a quantum wire by application of electric fields only, is our proposal also a ‘deliverable’ in the lab? The query can be broken down into three specific questions: (i) Is it feasible to realize a regime with sufficiently strong [*e-e*]{} interactions (as required by the condition $K<1/2$)? (ii) Can the size of the gap $\Delta$ be made sufficiently large to block thermal excitations at experimentally relevant temperatures? (iii) Is the synthetic HL robust against disorder?
To answer these questions, we take as case study a quantum wire patterned in an InAs quantum well (QW) [@LiangGao; @Giglberger]. Starting with (i), a detailed analysis yields that $$\tilde{V}(k \sim 0) \approx \frac{e^2}{\pi \epsilon_0 \epsilon_r} \ln(\frac{2d}{\eta}) + {\cal O}(\frac{\eta^2}{d^2})$$ with $\eta$ the half width of the wire, and where $\epsilon_r$ is the averaged relative permittivity of the dopant and capping layers between the QW and a metallic back gate at a distance $d$ from the wire [@ByczukDietl]. The commonly used In$_{1-x}$Al$_{x}$As capping layer has $\epsilon_r \!\approx\! 12$ when $x\!=\!0.25$, with roughly the same value when doped with Si. With $\eta \approx 5$ nm and $v_F
\!\approx\! 6 \times 10^5$ m/s [@PB], taking $d>1\, \mu$m and using that $g_2 = 4 \tilde{V}(k\!\sim \!0)/\pi \hbar$ [@Giamarchi_book_04], one verifies that $K \approx
(1\!+\!g_2/\pi v_F)^{-1/2} <1/2$. Thus, the desired “strong-coupling” regime is attainable without difficulty.
Turning to (ii), we need to attach a number to the gap $\Delta$ in Eq. (\[gap\]). Reading off data from Ref. , applicable when the InAs QW is separated from the top gates by a solid PEO/LiClO$_4$ electrolyte, the Rashba coupling $\hbar \alpha$ is found to change from $0.4 \times 10^{-11}$ eVm to $2.8 \times 10^{-11}$ eVm when tuning a top gate from $0.3$ to $0.8$ V. With $a \approx 5$ Å [@PB], we may thus take $\hbar \gamma_{\text{R}} = 8$ meV and $\hbar \gamma^{\prime}_{\text{R}} = 60$ meV, assuming that \[the spacers between\] the top gates in Fig. 1 are biased at \[0.3 V\] 0.8 V. As for the Dresselhaus coupling, experimental data for InAs QWs come with large uncertainties. We here take $\hbar \gamma_{\text{D}} = 5 $ meV, guided by the prediction that $1.6 < \alpha/\beta < 2.3$ in conventionally gated structures [@Giglberger]. Inserting $\lambda = a \gamma^{\prime}_{\text{R}}\gamma_{\text{D}}(\gamma_{\text{R}}^{2}+\gamma^{2}_{\text{D}})^{-1/2}$ into Eq. (\[gap\]), and choosing, say, $K=1/4$ with $c(1/4)= 1$ [@Zamolodchikov] we obtain $\Delta \approx 0.3$ meV (with smaller values of $K$ producing a larger gap). While this value of $\Delta$ is much smaller than the bulk gap in an HgTe QW [@Konig], it is still large enough $-$ with safe margins $-$ to protect the synthetic HL at sub-Kelvin temperatures. This allows to probe it by standard quantum transport experiments. It is here interesting to note that a recent proposal for an “all-electric" topological insulator in an InAs double well arrives at an inverted band gap of roughly the same size as our interaction-assisted gap [@ErlingssonEgues].
Finally, let us address (iii). As shown in Refs. and , a 1D helical liquid may undergo a localization transition due to disorder-generated correlated two-particle backscattering. A case in point is when a Rashba interaction is present [@SJJ], with a fluctuating component $\alpha_{\mbox{\footnotesize{rand}}}(x)$ from the random ion distribution in nearby doping layers [@Sherman]. Fortuitously, the localization length $\xi_{\text{rand}}$ for an InAs wire, making the usual assumption that $\sqrt{\langle \alpha^2_{\mbox{\footnotesize{rand}}}(x) \rangle} \approx \langle \alpha(x) \rangle$ [@Sherman], turns out to be much larger than the renormalization scale $\xi = \hbar v/\Delta$ at which the helicity gap develops [@AS]. Moreover, estimates of the elastic mean free path $\ell_{\text{e}}$ for InAs quantum wires [@YYCC] show that $\xi < \ell_{\text{e}} < \xi_{\text{rand}}$ when $1/5 \!<\! K \!<\! 1/2$ and $\alpha_{\mbox{\footnotesize{rand}}}(x) \!<\! 4 \times 10^{-11}$ eVm. It follows that the synthetic HL is well protected within these parameter intervals.
In summary, we have unveiled a scheme for producing an interacting helical electron liquid in a quantum wire using electric fields only, exploiting an interplay between a Dresselhaus- and a spatially periodic Rashba spin-orbit interaction. This synthetic helical liquid is of a different type than existing varieties, being neither ‘holographic’ [@Konig] nor ‘quasi-helical’ [@Quay]. While a number of nontrivial design criteria have to be satisfied for its realization in the laboratory, none of them are beyond present-day capabilities. Indeed, considering the principal simplicity and robustness of the required setup, the synthetic helical liquid could become a workhorse for exploring many of the intriguing phenomena associated with helical electrons in one dimension.
We thank D. Grundler, K. Le Hur, and A. Ström for valuable comments and suggestions. This work was supported by SCOPES Grant IZ73Z0$_{-}128058/1$ (G.I.J.) and Swedish Research Council Grant 621-2011-3942 (H.J.).
[99]{}
For reviews, see M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. [**82**]{}, 3045 (2010); X.-L. Qi and S.-C. Zhang, Rev. Mod. Phys. [**83**]{}, 1057 (2011); J. Maciejko, T. L. Hughes, and S.-C. Zhang, Annu. Rev. Condens. Matter Phys. [**2**]{}, 31 (2011).
C. L. Kane and E. J. Mele, Phys. Rev. Lett [**95**]{}, 146802 (2005).
B. A. Bernevig, T. L. Hughes, and S.-C. Zhang, Science [**314**]{}, 1757 (2006).
C. Wu, B. A. Bernevig, and S.-C. Zhang, Phys. Rev. Lett. [**96**]{}, 106401 (2006).
C. Xu and J. E. Moore, Phys. Rev. B [**73**]{}, 045322 (2006).
M. König, S. Wiedmann, C. Brüne, A. Roth, H. Buhmann, L. W. Molenkamp, X.-L. Qi, and S.-C. Zhang, Science [**318**]{}, 766 (2007); A. Roth, C. Brune, H. Buhmann, L. W. Molenkamp, J. Maciejko, X.-L. Qi, and S.-C. Zhang, Science [**325**]{}, 294 (2009).
X.-L. Qi, T. L. Hughes, and S.-C. Zhang, Nature Physics [**4**]{}, 273 (2008).
C.-Y. Hou, E.-A. Kim, and C. Chamon, Phys. Rev. Lett. [**102**]{}, 076602 (2009).; A. Ström and H. Johannesson, Phys. Rev. Lett. [**102**]{}, 096806 (2009).
Y. Tanaka and N. Nagaosa, Phys. Rev. Lett. [**103**]{}, 166403 (2009).
L. Fu and C. L. Kane, Phys. Rev. Lett. [**100**]{}, 096407 (2008); Phys. Rev. B [**79**]{}, 161408(R) (2009).
M. Baenninger, M. König, A. G. F. Garcia, M. Mühlbauer, C. Ames, P. Leubner, C. Brüne, H. Buhmann, L. W. Molenkamp, and D. Goldhaber-Gordon, J. Appl. Phys. [**112**]{}, 103713 (2012).
J. I. Väyrynen, M. Goldstein, and L. I. Glazman, Phys. Rev. Lett. [**110**]{}, 216402 (2013).
B. A. Bernevig and S.-C. Zhang, IBM J. Res. & Dev. [**50**]{}, 1 (2006).
R. Egger, A. Zazunov, and A. L. Yeyati, Phys. Rev. Lett. [**105**]{}, 136403 (2010).
A. Y. Kitaev, Physics-Uspekhi [**44**]{}, 131 (2001).
R. M. Lutchyn, J. D. Sau, and S. Das Sarma, Phys. Rev. Lett. [**105**]{}, 077001 (2010); Y. Oreg, G. Refael, and F. von Oppen, Phys. Rev. Lett. [**105**]{}, 177002 (2010).
P. Středa and P. Šeba, Phys. Rev. Lett. [**90**]{}, 256601 (2003).
See B. Braunecker, A. Ström, and G. I. Japaridze, Phys. Rev. B [**87**]{}, 075151(2013), and references therein.
D. Awschalom and N. Samarth, Physics [**2**]{}, 50 (2009).
J. Klinovaja, S. Gangadharaiah, and D. Loss, Phys. Rev. Lett. [**108**]{}, 196804 (2012).
For a review, see R. Winkler, [*Spin-Orbit Interaction Effects in Two-Dimensional Electron and Hole Systems*]{} (Springer-Verlag, Berlin, 2003).
E. Ya. Sherman, Phys. Rev. B [**67**]{}, 161303(R) (2003).
L. E. Golub and E. L. Ivchenko, Phys. Rev. B [**69**]{}, 115333 (2004).
D. Liang and X. P. A. Gao, Nano Lett. [**6**]{}, 3263 (2012).
A. Ström, H. Johannesson, and G. I. Japaridze, Phys. Rev. Lett. [**104**]{}, 256804 (2010).
M. Malard, I. Grusha, G. I. Japaridze, and H. Johannesson, Phys. Rev. B [**84**]{}, 075466 (2011).
A. Schulz, A. De Martino, P. Ingenhoven, and R. Egger, Phys. Rev. B [**79**]{}, 205432 (2009).
T. Giamarchi, [*Quantum Physics in One Dimension*]{} (Oxford University Press, Oxford, 2004).
R. Shankar, Acta Physica Polonica B [**26**]{}, No 12 1835 (1995).
For a review, see M. Malard, Braz. J. Phys. [**43**]{}, 182 (2013).
Al. B. Zamolodchikov, Int. J. Mod. Phys. [**A10**]{}, 1125 (1995).
H. J. Schulz, Phys. Rev. B [**22**]{}, 5274 (1980).
For a review, see M. König, H. Buhmann, L. W. Molenkamp, T. Hughes, C.-X. Liu, X.-L. Qi, and S.-C. Zhang, J. Phys. Soc. Jpn. [**77**]{} 031007 (2008).
C. H. L. Quay, T. L. Hughes, J. A. Sulpizio, L. N. Pfeiffer, K. W. Baldwin, K. W. West, D. Goldhaber-Gordon, and R. de Picciotto, Nature Physics [**6**]{}, 336 (2010).
H. B. Nielsen and M. Ninomiya. Nucl. Phys. B [**185**]{}, 20 (1981); [*ibid.*]{} [**193**]{}, 173 (1981).
For a discussion of dynamical breaking of time-reversal symmetry in one-dimensional electron systems, see Refs. .
V. Mourek, K. Zuo1, S. M. Frolov, S. R. Plissard, E. P. A. M. Bakkers, L. P. Kouwenhoven, Science [**336**]{}, 1003 (2012).
S. Giglberger, L. E. Golub, V. V. Bel’kov, S. N. Danilov, D. Schuh, Ch. Gerl, F. Rohlfing, J. Stahl, W. Wegscheider, D. Weiss, W. Prettl, and S. D. Ganichev, Phys. Rev. B [**75**]{}, 035327 (2007).
K. Byczuk and T. Dietl, Phys. Rev. B [**60**]{}, 1507 (1999).
P. Bhattacharaya, [*Properties of lattice-matched and strained indium gallium arsenide*]{} (EMIS Datareview Series No. 8, Stevenage, UK).
S. I. Erlingsson and J. C. Egues, arXiv:1312.2034.
C. H. Yang, M. J. Yang, K. A. Cheng, and J. C. Culbertson, Phys. Rev. B [**66**]{}, 115306 (2002).
A. Ström, [*Interaction and Disorder in Helical Conductors*]{}, University of Gothenburg Ph.D. thesis (2012).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We introduce the notion of measurable sensitivity, a measure-theoretic version of the condition of sensitive dependence on initial conditions. It is a consequence of light mixing, implies a transformation has only finitely many eigenvalues, and does not exist in the infinite measure-preserving case. Unlike the traditional notion of sensitive dependence, measurable sensitivity carries up to measure-theoretic isomorphism, thus ignoring the behavior of the function on null sets and eliminating dependence on the choice of metric.'
address:
- ' Williams College, MA 01267, USA'
- ' University of Chicago, Chicago, IL 60637, USA '
- ' Williams College, MA 01267, USA'
- |
Department of Mathematics\
Williams College\
Williamstown, MA 01267, USA
- ' Princeton University, Princeton, NJ 08544, USA'
author:
- Jennifer James
- Thomas Koberda
- Kathryn Lindsey
- 'Cesar E. Silva'
- Peter Speh
bibliography:
- 'ErgodicTheoryBib.bib'
title: Measurable Sensitivity
---
Introduction
============
Sensitive dependence on initial conditions, which is widely understood to be one of the central ideas of chaos, is a topological, rather than measurable notion. It was introduced by Guckenheimer in [@Guckenheimer]. A transformation $T$ on a metric space $(X,d)$ is said to exhibit **sensitive dependence** with respect to $d$ if there exists a $\delta> 0$ such that for all $\varepsilon > 0$ and for all $x \in X$, there exists an $n \in \mathbb{N}$ and a $y \in B_{\varepsilon}(x)$ such that $d(T^n(x),T^n(y)) > \delta$. The notion of sensitive dependence has been studied extensively and the reader is referred to [@BBCDS], [@GlasnerWeiss] and [@AEAJBK] for recent results. The relationship between measure theoretic notions, such as weak mixing, and sensitive dependence is studied in [@GlasnerWeiss], [@AEAJBK], [@BGK], and [@HYW]. A stronger notion of sensitivity, called strong sensitivity, was introduced by [@AbrahamBC]. A transformation $T$ on a metric space $(X,d)$ is said to exhibit **strong sensitive dependence** with respect to $d$ if there exists a $\delta> 0$ such that for all $\varepsilon
> 0$ and for all $x \in X$, there exists an $N \in \mathbb{N}$ so that for all integers $n \geq N$ there exists a $y \in
B_{\varepsilon}(x)$ such that $d(T^n(x),T^n(y))
> \delta$. Both sensitive dependence and strong sensitivity are topological notions, depending on both the choice of metric and the behavior of the transformation on null sets. We introduce the notions of measurable sensitivity and weak measurable sensitivity, which are ergodic-theoretic versions of strong sensitive dependence and sensitive dependence, respectively.
In Section \[S:2\] we show that a doubly ergodic (a condition equivalent to weak mixing for finite measure-preserving transformations) nonsingular transformation is weak measurably sensitive, that a lightly mixing nonsingular transformation (for example, a mixing finite measure-preserving transformation) is measurably sensitive, and that measurable sensitivity does not imply weak mixing. In Section \[S:3\] we show that if an ergodic nonsingular transformation $T$ is measurably sensitive, then there exists an integer $n>0$ so that $T^n$ has $n$ invariant subsets, and the restriction of $T^n$ to each of these subsets is weakly mixing. Section \[S:4\] shows that if an ergodic finite measure-preserving transformation $T$ is measurably sensitive, then there is an integer $n$ so that $T^n$ has $n$ invariant sets of positive measure covering $X$ a.e. and such that the restriction of $T^n$ to each is lightly mixing. The final section shows that an ergodic infinite measure-preserving transformation cannot be measurably sensitive (though it can be weak measurably sensitive by Section \[S:2\]).
All of our spaces are Lebesgue spaces with a probability or a $\sigma$-finite measure defined on them. We assume the measures to be regular.
Throughout the paper, $(X,S(X),\mu)$ denotes a Lebesgue space $X$ with a positive, finite or $\sigma$-finite non-atomic measure $\mu$, and $S(X)$ the collection of $\mu$-measurable subsets of $X$. It is a standard fact that any two such spaces are isomorphic under a nonsingular isomorphism. We let $d$ denote a metric on $X$. We shall say that a metric is **good** if all nonempty open sets have positive measure. When $X$ has a good metric we assume the measures defined on $X$ are regular. For convenience, given two non-empty sets $A$ and $B$ in a space $X$ with metric $d$, define $d(A,B) = \inf \{d(a,b):a \in A, b \in B\}$.
\[t:msdef\] Let $(X,\mu,T)$ be a nonsingular dynamical system. $T$ is **measurably sensitive** if whenever a dynamical system $(X_1,\mu_1,T_1)$ is measure-theoretically isomorphic to $(X,\mu,T)$ and $d$ is a good metric on $X_1$, then there exists a $\delta
> 0$ such that for all $x \in X_1$ and all $\varepsilon > 0$ there exists an $N \in \mathbb{N}$ such that for all integers $n \geq N$ $$\mu_1 \{y\in B_{\varepsilon}(x): d(T_1^nx,T_1^ny) > \delta \} > 0.$$
\[t:wmsdef\] Let $(X,\mu,T)$ be a nonsingular dynamical system. $T$ exhibits **weak measurable sensitivity** if whenever a dynamical system $(X_1,\mu_1,T_1)$ is measure-theoretically isomorphic to $(X,\mu,T)$ and $d$ is a good metric on $X_1$ then there exists $\delta
> 0$ such that for all $x \in X$ and $\varepsilon > 0$ there exists an $n \in \mathbb{N}$ such that $\mu_1 \{y\in B_{\varepsilon}(x):
d(T_1^n(x),T_1^n(y)) > \delta \} > 0$.
The real number $\delta$ in the above definition is referred to as a sensitivity constant.
\[t:example\] Consider a measure space X consisting of two copies of the circle $S^1$, labeled $S_1$ and $S_2$. Define a metric $d$ on $X$ as follows: for two points $x \in S_i$ and $y\in S_j$, if $i=j$ define $d(x,y)$ to be the the minimal arclength between points $x$ and $y$, and if $i \neq j$ we let $d(x,y) = r$ for some fixed $r> \frac{\pi}{2}$. In each copy of $S^1$, pick out the orbit of a fixed point $z$ under a fixed irrational rotation $R$ on $S^1$, and denote this set by $M$. Define $T:X \rightarrow X$ as follows: for $x\in M$, let $T(x)
= x$, and for $x\in X\setminus M$, let $T(x)$ map $x$ to $R(x)$ in the *other* copy of $S^1$.
Because the points in $M$ go “far away" from those in $M^c$, it is clear that the system exhibits sensitive dependence, but without this null set sensitive dependence would fail. Thus, the system does not exhibit measurable or weak measurable sensitivity. In fact, $T^n(B_{\varepsilon}(x)) = B_{\varepsilon}(R^2(x))$ for all $n = 0
\pmod{2}$.
Let $X$ be an interval of finite length in $\mathbb{R}$ and let $d$ be the standard Euclidean metric on $X$. If a continuous transformation $T:X\rightarrow X$ is sensitive with respect to $d$, then $T$ is strongly sensitive with respect to $d$.
Suppose $T$ is sensitive with sensitivity constant $\delta$. Let $I_1$,...,$I_n$ be disjoint (except at endpoints) intervals with closed or open endpoints which cover $X$ and each have length shorter than $\frac{\delta}{2}$. Every interval of length at least $\delta$ must contain one of these intervals. Since $T$ is sensitive, for each $1 \le j \le n$ there must exist a natural number $m_j$ so that $T^{m_j}(I)$ has length at least $\delta$. Consequently, for any interval $I$ with length at least $\delta$, and any $n\in \mathbb{N}$, $T^{n}(I)$ contains one of $T^{k}(I_i)$ where $1\le i \le n$ and $1\le k \le m_i$. Let $\delta'$ be one third of the minimum of the lengths of these intervals. Let $x\in
X$ and $\varepsilon > 0$. Since $T$ is sensitive, there must be some $k_0$ so that $T^{k_0}(B_\varepsilon(x))$ contains an interval of length $\delta$ and hence for any $k > k_0$, $T^{k}(B_\varepsilon(x))$ contains an interval of the length at least $3\delta'$ and hence contains a point whose distance is at least $\delta'$ from $T^{k}(x)$. Hence, $T$ is strongly sensitive with strong sensitivity constant $\delta'$.
Acknowledgements
----------------
This paper is based on research by the Ergodic Theory group of the 2006 SMALL summer research project at Williams College. Support for the project was provided by National Science Foundation REU Grant DMS - 0353634 and the Bronfman Science Center of Williams College.
Mixing Notions and Measurable and Weak Measurable Sensitivity {#S:2}
=============================================================
We begin by proving that double ergodicity implies weak measurable sensitivity. A nonsingular transformation is said to be **doubly ergodic** if for all sets $A$ and $B$ of positive measure there exists an integer $n > 0$ such that $\mu(T^{-n}(A) \cap A) > 0$ and $\mu(T^{-n}(A) \cap B) > 0$. Double ergodicity is equivalent to weak mixing for measure-preserving transformations on finite measure spaces [@Furstenberg], but strictly stronger than weak mixing in the infinite measure-preserving case [@BFMS].
\[t:doubleergodicity\] If $(X,\mu,T)$ is a nonsingular, doubly ergodic dynamical system, then $T$ exhibits weak measurable sensitivity. In particular, weakly mixing, finite measure-preserving transformations exhibit weak measurable sensitivity.
Let $(X_1,\mu_1,T_1)$ be measure-theoretically isomorphic to $(X,\mu,T)$, and let $d$ be a good metric on $X_1$. By the definition of a good metric, there exist sets $A$ and $C$ of positive measure in $S(X_1)$ such that $d(A,C) > 0$. Pick $\delta <
d(A,C)/2$. Let $x \in A$ and $\varepsilon
> 0$. As $(X_1,\mu_1,T_1)$ is doubly ergodic, there exists an $n
\in \mathbb{N}$ such that $\mu (T_1^{-n}C \cap B_{\varepsilon}(x))
> 0$ and $\mu (T_1^{-n}A \cap B_{\varepsilon}(x))
> 0$. Therefore, $\mu_1 \{y \in B_{\varepsilon}(x): T_1^n(y) \in A\} >
0$ and $\mu \{y \in B_{\varepsilon}(x):T_1^n(y) \in C\} > 0$. Since $T_1^n(x)$ cannot be within $\delta$ of both $A$ and $C$, $\mu_1 \{y
\in B_{\varepsilon}(x) : d(T_1^n(x),T_1^n(y)) > \delta \} > 0$.
We now turn our attention to measurable sensitivity and light mixing. Recall that a system $(X,\mu,T)$ on a finite measure space is said to be **lightly mixing** if $\liminf_{n \rightarrow
\infty} \mu(T^{-n}(A) \cap B) > 0$ for all sets $A$ and $B$ of positive measure.
If $(X,\mu,T)$ is a nonsingular, lightly mixing dynamical system, then $T$ is measurably sensitive.
Let $(X_1,\mu_1,T_1)$ be measure-theoretically isomorphic to $(X,\mu,T)$, and let $d$ be a good metric on $X_1$. By the definition of a good metric, there exist sets $A, C \subseteq X_1$ of positive measure with $d(A,C) > 0$. Pick $\delta <
\frac{d(A,C)}{2}$. Let $x \in X_1$ and $\varepsilon> 0$. By light mixing, there exists an $N \in \mathbb{N}$ such that for any integer $n \geq N$, $\mu_1(T_1^{-n}(C) \cap B_{\varepsilon}(x))> 0 \
\textrm{and} \ \mu_1(T_1^{-n}(A) \cap B_{\varepsilon}(x))> 0.$ Thus, $n \geq N$ implies $\mu_1\{y \in B_{\varepsilon}(x):T_1^n(y) \in A\}
> 0$ and $\mu_1\{y \in B_{\varepsilon}(x):T_1^n(y) \in C\} > 0$. Since $d(A,C)> 2\delta$, $T_1^n(x)$ cannot be within $\delta$ of both $A$ and $C$. As $T_1^n(B_{\varepsilon}(x))$ intersects both $A$ and $C$ in sets of positive measure, $\mu_1\{y \in
B_{\varepsilon}(x) : d(T_1^n(x),T_1^n(y))
> \delta\}> 0$, so $T$ is measurably sensitive.
In Proposition \[T: weak implies light\], we prove that a finite measure-preserving, weakly mixing transformation that is not lightly mixing is not measurably sensitive (but is weak measurably sensitive).
If $T_1$ is a finite measure-preserving lightly mixing transformation on $(X,\mu)$ and $T_2$ is a rotation on two points, then $T \times S$ is measurably sensitive and ergodic but not weakly mixing.
It is well known that all powers of $T$ are lightly mixing. We claim that $T\times S$ is measurably sensitive. $(T_1 \times
T_2)^2$ acts as a lightly mixing transformation on $X\times\{1\}$ and $X\times\{2\}$. Let $(S,Y)$ be isomorphic to $(T_1\times
T_2,X\times\{1,2\})$ and let $g:X\times\{1,2\}\rightarrow Y$ be the corresponding isomorphism of measure spaces. Under a good metric on $Y$, there exist sets $A_i$ and $B_i$ in $g(X\times \{i\})$ with a positive distance between $A_i$ and $B_i$. Any ball $B_\epsilon(y)$ will intersect at least one of $g(X\times \{1\})$ and $g(X\times
\{2\})$ with positive measure. Then, as a consequence of light mixing, for $n$ sufficiently large, either both $S^{-n}(A_1)$ and $S^{-n}(B_1)$ or both $S^{-n}(A_2)$ and $S^{-n}(B_2)$ intersect $B_\epsilon(x)$ with positive measure. Consequently, $S$ exhibits strong sensitive dependence with a sensitivity constant $\delta =
\min\{d(A_1,B_1\}, d(A_2, B_2)\}$. Hence, $S$ is strongly sensitive. As $T_2$ is ergodic and $T_1$ is weakly mixing, $T_1 \times T_2$ is ergodic. Finally, $T_1\times T_2$ is not weakly mixing since $-1\in
e(T_1\times T_2)$, the eigenvalue group of $T_1\times T_2$.
Measurable Sensitivity and Eigenvalues {#S:3}
======================================
We now show that if an ergodic nonsingular transformation is measurably sensitive, then it can have only finitely many eigenvalues. (Recall that $\lambda$ is an ($L^\infty$) eigenvalue of $T$ if there is a nonzero a.e. $f\in L^\infty$ such that $f\circ T=\lambda f$ a.e. Also, if $T$ is ergodic and finite measure-preserving, its $L^2$ eigenfunctions are in $L^\infty$.) This is used to give a further characterization of measurably sensitive transformations. All ($L^\infty$) eigenvalues of ergodic transformations lie on the unit circle. We refer to an eigenvalue as rational if it is of finite order and irrational if it is not.
\[T: intervals\] Suppose an ergodic nonsingular transformation $T: X\rightarrow X$ has an eigenfunction $f$ with an eigenvalue that is of the form $\exp(2\pi
iq)$ with $q$ irrational, with $|f|=1$. Then for any measurable set $A \subset
S^1$ of positive Haar measure, the backwards orbit of the set $f^{-1}(A)$ equals $X \mod \mu$.
Define $h:S^1 \rightarrow S^1$ by $h(x)
= xe^{-2\pi iq}$. The pushed measure $\mu\circ f^{-1}$ that is invariant under $h$ must be Haar measure as $h$ is an irrational rotation. Then
$$\begin{aligned}
\bigcup_{n=0}^{\infty}T^{-n}(f^{-1}(A)) =
\bigcup_{n=0}^{\infty}f^{-1}(h^{n}(A)) = f^{-1}(S^1) = X.\end{aligned}$$
\[T: irrational eval\] Suppose an ergodic nonsingular transformation $T: X\rightarrow X$ has an eigenfunction $f$ with an eigenvalue that is of the form $\exp(2\pi
iq)$ with $q$ irrational. Then T is not measurably sensitive.
Assume that $|f|=1$. Construct a nonsingular isomorphism from $(X,\mu)$ to $[0,1)$ as follows. As a consequence of Lemma \[T: intervals\] and countable subadditivity, each of the sets $f^{-1}(\exp(i(2^{1-n}\pi,2^{2-n}\pi]))$ has positive measure. Then for $n\ge 1$, there exist nonsingular isomorphisms $g_{n}$ from $f^{-1}(\exp(i(2^{1-n}\pi,2^{2-n}\pi]))$ to $(2^{-n},2^{-n+1})$ with Lebesgue measure. Let $g$ be the point map associated with $g_n$ on $f^{-1}(\exp(i(2^{1-n}\pi,2^{2-n}\pi]))$ for each positive $n$. We now define a new metric $d$ on $X$ given by $d(x,y) = |g(x)-g(y)|$. As each isomorphism is nonsingular and open sets have positive measure under Lebesgue measure, $d$ is a good metric. Let $x \in X$ be a point so that $|g(x) - 2^{-n}| < 2^{-n-1}$ and let $\varepsilon <
2{^{-n-1}}$. Then $B^{d}_{\varepsilon}(x) \subset
f^{-1}(\exp(i(0,2^{1-n}\pi]))$. Let $h$ be a translation by $q$ on $\mathbb{R}/\mathbb{Z}$. Then there is a sequence $\{m_k\}_{k=1}^{\infty}$ for which
$$h^{m_k}((0,2^{1-n}])\subset (0,2^{2-n}].$$
Consequently, $T^{m_k}(B_{\varepsilon}(x))$ is contained in $f^{-1}(\exp(i(0,2^{2-n}\pi]))$, and so for almost every point $y\in
T^{m_k}(B_{\varepsilon}(x))$, $d(T^{m_k}(y), T^{m_k}(x)) < 2^{2-n}$. As $n$ tends to infinity, $2^{2-n}$ tends to zero, and so there is no possible sensitivity constant. Hence, $T$ is not measurably sensitive.
\[t:rationaleigenvalues\] If an ergodic nonsingular transformation $T$ on a Lebesgue space $X$ has infinitely many rational eigenvalues, then $T$ is not measurably sensitive.
Since the eigenvalues of a transformation form a multiplicative group, there exists an increasing sequence of positive integers so that $e^{\frac{2\pi i}{a_i}}$ is an eigenvalue and $a_i|a_{i+1}$. These eigenvalues indicate the existence of $a_i$ invariant sets under $T^{a_i}$. There are $\frac{a_{i+1}}{a_i}$ invariant sets for $T^{a_i+1}$ contained in each invariant set for $T^{a_i}$. The invariant sets can be enumerated so that the invariant sets for $T^{a_i}$ are represented by integers $A_{i,0}, ..., A_{i,a_i-1}$ and the $T^{a_{i+1}}$-invariant sets $A_{i+1,m}$ contained in a given $T^{a_i}$-invariant set $A_{i,n}$ have $m = 2 \pmod{a_i}$. Then any sequence of integers $\{b_i\}_{i=1}^{\infty}$ for which $0
\le b_i < \frac{a_{i+1}}{a_i}$ corresponds to a sequence of sets $C_{i}$ such that $C_{i+1}\subset C_{i}$, where $C_{i}=A_{i,\sum_{j=1}^{i}b_j a_j}$. For any such sequence, the set $\bigcap_{i=1}^{\infty}C_i$ cannot intersect any analogous set corresponding to another sequence. There are uncountably many possible sequences, so there must be a sequence of sets $\{a_i\}_{i=0}^{\infty}$ so that $\bigcap_{i=1}^{\infty}C_i$ has measure $0$ for the corresponding sequence $a_i$. Choose such a sequence. Then letting $C_{0} = X$, the space may be expressed as
$$X = \left(\bigsqcup_{i=0}^{\infty}C_{i}\backslash C_{i+1}\right) \mbox{ (mod $\mu$)}.$$
Let $g_{i}$ be a nonsingular isomorphism from $C_{i}\backslash
C_{i+1}$ to $(2^{-i-1},2^{-i})$ with Lebesgue measure. Let $N$ be the backwards orbit of $\bigcap_{i=1}^{\infty}C_i$, where none of the maps $g_{i}$ are defined. This must have measure zero. Then let $g: X\backslash N \rightarrow X\backslash N$ be the union of the maps $g_i$. Let $T'$ be the restriction of $T$ to $X\backslash N$. Let $d$ be a metric on $X\backslash N$ defined by $d(x,y) =
|g(x)-g(y)|$. This metric is good since each map $g_i$ is nonsingular. Choose a point $x\in C_{i}\backslash N$ and let $\varepsilon$ be small enough so that $B^d_{\varepsilon}(x)\subset C_{i}$. Then for each positive integer $k$, $(T')^{ka_i}(B^d_{\varepsilon}(x))\subset C_{i}\backslash N$. Any two points in $C_i\backslash N$ can have a distance of at most $2^{-i}$ between them and so any sensitivity constant $\delta$ for $T'$ would have to be at most $2^{-i}$. Consequently, there can be no positive sensitivity constant, and $T$ is not measurably sensitive.
\[t:finite evals\] Any ergodic, nonsingular, measurably sensitive transformation has finitely many eigenvalues.
\[t:invariant sets\] If an ergodic, nonsingular transformation $T:X\rightarrow X$ is measurably sensitive, then for some $n\in\mathbb{N}$, $T^{n}$ has $n$ invariant subsets and the restriction of $T^{n}$ to each of these subsets is weakly mixing.
By Corollary \[t:finite evals\], $T$ must have finitely many eigenvalues. The eigenvalues form a cyclic group of finite order $n$. Let $f$ be an eigenfunction whose eigenvalue has order $n$. Then the sets $A_k = f^{-1}\left(\exp\left(i\left[\frac{2\pi
k}{n},\frac{2\pi (k+1)}{n}\right)\right)\right)$ for $0 \le k \le
n-1$ are invariant under $T^{n}$. First, we show that the restriction of $T^{n}$ to each of these sets is ergodic. Suppose there exists a $T^n$-invariant set $C$ of positive measure. Notice that since $T^{-n}(C) = C$, then $\bigcup_{i=0}^{n-1}T^{-i}(C)$ is $T$-invariant, and thus equals $X \mod \mu$. Each of $T^{-i}(C)$ for $0\le i \le n-1$ must be contained in $A_k$ for a different $k$ and so $C\cap A_k = A_k \mod \mu$ for all $k$. Hence, the restriction of $T^{n}$ to $A_{k}$ is ergodic for $0 \le k \le n-1$.
Next, we prove that the restriction of $T^{n}$ to any of these sets admits no eigenvalues other than $1$. Suppose $T^{n}$ restricted to $A_0$ admits an eigenvalue $\lambda \ne 1$. For $0\le k \le n-1$, let $h(T^{-k}(x)) = \lambda^{\frac{-k}{n}}f(x)$. Then $h$ will be an eigenfunction of $T$ which will have order greater than $n$. Thus the restriction of $T^{n}$ to $A_{i}$ for each $0\le i \le n-1$ is ergodic and admits no eigenvalues other than $1$. Thus, the restriction must be weakly mixing.
A totally ergodic, measurably sensitive transformation is weakly mixing.
Measurable Sensitivity for Finite Measure-preserving Transformations {#S:4}
====================================================================
In this section, we consider measurable sensitivity for measure-preserving transformations on finite measure spaces. Considering only such spaces, Proposition \[t:invariant sets\] is strengthened to include a requirement of light mixing. We assume spaces have total measure one.
\[T: separate sets\] Suppose $T:X\rightarrow X$ is measure-preserving and not lightly mixing. Then there exist sets $C_{1}$ and $D_{1}$ of positive measure and an infinite sequence $\{n_k\}_{k=1}^{\infty}$ such that $T^{n_k}(C_1)\cap D_{1}= \emptyset$ for all $k \in \mathbb{N}$.
From the definition of lightly mixing, we may assume that there are sets $C$ and $D$ of positive measure so that $\liminf_{n\rightarrow\infty}\mu(C\cap T^{-n}(D)) = 0$. Choose an increasing sequence of distinct natural numbers $\{n_k\}_{k=1}^{\infty}$ so that $\mu(C\cap T^{-n_k}(D)) \le
2^{-k-1}\mu(C)$. Let $C_1 =
C\backslash\bigcup_{k=1}^{\infty}(C\cap T^{-n_k}(D))$ and $D_{1} =
D$. Then $\mu(C_1)
> \frac{1}{2}\mu(C) > 0$ and so $T^{n_k}(C_{1})\cap D_{1} = \emptyset$ for every $k$.
The proof of the following lemma is standard and is omitted.
\[T: increasing continuous\] Let $f:(0,1)\rightarrow(0,1)$ be a continuous function with $f(x) >
x$ for every $x\in (0,1)$. Then for every $x\in (0,1)$, $\lim_{n\rightarrow\infty}f^{n}(x) = 1$.
\[T: sequences of subsets\] Let $T:X\rightarrow X$ be a finite measure-preserving, weakly mixing and not lightly mixing transformation. Then there exist sequences of measurable sets $\{C_i\}_{i=1}^{\infty}$ and $\{D_i\}_{i=1}^{\infty}$ satisfying the following properties.
1. \[I: positive measure\] $\mu(C_i) > 0$ and $\mu(D_i) > 0$
2. \[I: C subset\] $C_i \subset C_{i-1}$ for $i > 1$.
3. \[I: D subset\] $D_{i-1} \subset D_i$ for $i > 1$.
4. \[I: D growth\] $\lim_{i\rightarrow\infty}\mu(D_i) = 1$
5. \[I: C growth\] $\lim_{i\rightarrow\infty}\mu(C_i) = 0$
6. \[I: subsequence\] There is a sequence $\{n_k\}_{k=1}^{\infty}$ so that $T^{n_k}(C_i)\cap D_{i}= \emptyset$.
We let $C_1$, $D_1$, and $\{n_k\}_{k=1}^{\infty}$ be as defined in Lemma \[T: separate sets\]. They clearly satisfy properties 1-3 and property 6. For the inductive step, assume that $C_i$ and $D_i$ have been chosen to satisfy these properties for all $i \leq j$. As a consequence of weak mixing, there is a zero density subset $E_1\subset \mathbb{N}$ such that
$$\lim_{n\rightarrow\infty, n\not\in E_1}\mu(T^{-n}(C_j)\cap
C_j) = \mu(C_j)^{2}.$$
As a result, the set of values $n$ with $\mu(T^{-n}(C_j)\cap C_j)
> 0$ has density 1. Similarly, there exists a zero density subset $E_2\subset \mathbb{N}$ such that
$$\lim_{n\rightarrow\infty, n\not\in E_2}\mu(T^{-n}(D_{j})\cap D_j) = \mu(D_j)^2.$$ Consequently, there is a natural number $m_{j}$ so that $\mu(T^{-m_{j}}(C_j)\cap C_j)> 0$ and $\mu(T^{-m_{j}}(D_j)\cap D_j)
< \frac{1}{2}(\mu(D_j)^2 + \mu(D_j))$. Then $\mu(T^{-m_{j}}(D_j)\backslash D_j)
> \frac{1}{2}(\mu(D_j) - \mu(D_j)^{2})$. For positive integers $j$, we let $C_{j+1} =
C_{j}\cap T^{-m_j}(C_j)$ and let $D_{j+1} = D_{j} \cup
T^{-m_j}(D_j)$. Properties \[I: C subset\] and \[I: D subset\] are clear from the definitions of $C_{j+1}$ and $D_{j+1}$. Property \[I: positive measure\] for $D_{j+1}$ follows from the fact that $D_0$ has positive measure and $D_i \subset D_{i+1}$ for all $i \le
j$. Property \[I: positive measure\] for $C_{j+1}$ follows from the fact that $m_j$ is chosen to make $C_{j+1}$ have positive measure. Property \[I: D growth\] follows from Lemma \[T: increasing continuous\] applied to the function $\frac{3}{2}x -
\frac{1}{2}x^2$ and the lower bound for the measure of $D_{i+1}$ in terms of $D_{i}$. To show Property \[I: subsequence\], it suffices to see that $$\begin{aligned}
T^{n_k}(T^{-m_j}(C_j)\cap C_j)\cap (T^{-m_j}(D_j)\cup D_j)\\
\subset T^{-m_j}(T^{n_k}(C_j)\cap D_j)\cup (T^{n_k}(C_j)\cap D_j) =
\emptyset.\end{aligned}$$ Property \[I: C growth\] follows from property \[I: D growth\] and the fact that $T^{-n_k}(D_i)$ and $C_i$ must be disjoint.
\[T: weak implies light\] If $T:X\rightarrow X$ is weakly mixing, finite measure-preserving and not lightly mixing, then $T$ is not measurably sensitive.
The space $X$ can be decomposed as $$X = \left(\bigsqcup_{i,j =
0}^{\infty} (C_{i}\backslash C_{i+1})\cap(D_{j+1}\backslash
D_{j})\right)\mod \mu$$ where $C_{0} = X$ and $D_{0} = \emptyset$ and $C_i$ and $D_i$ are as in Lemma \[T: sequences of subsets\]. Let $g_{i,j}$ be a nonsingular isomorphism from $(C_{i}\backslash
C_{i+1})\cap(D_{j+1}\backslash D_{j})$ to $(2^{-j} +
2^{-i-1-j},2^{-j} + 2^{-i-j})$ with Lebesgue measure whenever $(C_{i}\backslash C_{i+1})\cap(D_{j+1}\backslash D_{j})$ has positive measure. Let $N$ be the backwards orbit of the points where no $g_{i,j}$ is defined. This set has measure zero, so the restriction of $T$ to $X\backslash N$ is isomorphic to $T$. Let $T'$ denote this restriction. Define function $g$ on $X\backslash N$ by letting $g(x) = g_{i,j}(x)$. Then $d(x,y) = |g(x)-g(y)|$ is a metric on $X\backslash N$. A ball around a point $x\in X\backslash
N$ must have positive measure as each of the maps $g_{i,j}$ is nonsingular, so the metric is good. Note that $g(D_{j})\subset(2^{-j},1)$ and $g(D_{j}^{c})\subset(0, 2^{-j})$. Let $x$ be a point in $(C_{i}\backslash N)\cap D_j$ for some $j$ where such a point exists, and choose $\varepsilon$ so that $B^{d}_{\varepsilon}(x)\subset C_{i}\cup D_j$. From property \[I: subsequence\] of Lemma \[T: sequences of subsets\], there is a sequence $\{n_{k}\}_{k=1}^{\infty}$ so that $g((T')^{n_{k}}(B_{\varepsilon}(x)))\subset g^{-1}(0,2^{-i})$. Since any sensitivity constant $\delta$ must be smaller than $2^{-i}$ for each positive integer $i$, there is no possible sensitivity constant. Hence, $T'$ does not exhibit strong sensitive dependence for any good metric $d$ and so $T$ is not measurably sensitive.
\[T: components\] Suppose a measurably sensitive transformation $T:X\rightarrow X$ has finitely many invariant subsets of positive measure $A_1$,$A_2$,..,$A_n$. Then the restriction of $T$ to each subset is measurably sensitive.
Suppose the restriction of $S = T|A_i$ is not measurably sensitive. Let $d$ be a good metric on a set $A_i^\prime$ so that $(S^\prime,A_i^\prime)$ is measurably isomorphic to $(S,A_i)$ and assume that $S^\prime$ does not exhibit strong sensitivity. Let $g$ be a measure-preserving isomorphism from $X\backslash A_i$ to (0,1) with Lebesgue measure. Let $N$ be the set of points where $g$ and the isomorphism from $A_i^\prime$ to $A_i$ are not preserved as well as their backwards orbits. Then $N$ must be a null set. For any two points $x,y \in X\backslash (A_i\cup N)$, extend $d$ so that $d(x,y)
= |g(x) - g(y)|$. Let $X^\prime = X\backslash(A_i\cup N) \cup
A_i^\prime$ and let $T^\prime:X^\prime\rightarrow X^\prime$ be equal to the restriction of $T$ on $X\backslash(A_i\cup N)$ and equal to $S^\prime$ on $A_i^\prime$. Then $T^{\prime}$ is measurably isomorphic to $T$. Now $d$ is extended to a metric on $X^\prime$. For any two points $x,y \in X\backslash (A_i\cup N)$, extend $d$ so that $d(x,y) = |g(x) - g(y)|$. Choose a point $y_0\in A_i^\prime$. For points $y\in A_i^\prime$ and for $x\in X\backslash (A_i\cup N)$, let $d(y,x)=d(x,y) = 1 + d(y,y_0)$. It is easy to verify that the extension of $d$ is a good metric on $X^\prime$.
As $S^\prime$ does not exhibit strong sensitivity on $A_i^\prime$, for any $\varepsilon
> 0$, there is a ball $B_r(x)$ with $x\in A_i^\prime$ so that for some sequence $\{n_k\}$, almost every point $y\in B_r(x)\cap A_i^\prime$ satisfies $d((S^\prime)^{n_k}(y),(S^\prime)^{n_k}(x)) <
\varepsilon$. We may assume that $r<1$. Then $B_{r}(x)= B_r(x)\cap
A_i^\prime$. As a consequence, almost every point $y\in B_{r}(x)$ satisfies $d((T^\prime)^{n_k}(y),(T^\prime)^{n_k}(x)) <
\varepsilon$. Hence, $T'$ is not strongly sensitive and so $T$ is not measurably sensitive.
Let $T$ be an ergodic transformation on a finite measure Lebesgue space $X$. If $T$ is measurably sensitive, then there is some positive integer $n$ so that $T^n$ has $n$ invariant sets of positive measure which cover almost all of $X$, and the restriction of $T^n$ to each of the sets is lightly mixing.
Suppose $T$ is measurably sensitive. By Proposition \[t:invariant sets\], there are $n$ invariant sets for $T^{n}$, each of positive measure. Every power of a measurably sensitive transformation is clearly measurably sensitive, so $T^n$ must be measurably sensitive. By Lemma \[T: components\], the restriction of $T^{n}$ to any of the sets must be measurably sensitive and, by Proposition \[t:invariant sets\], the restriction must be weakly mixing. Consequently, Proposition \[T: weak implies light\] indicates that the restriction must be lightly mixing.
Infinite Measure Spaces {#S:5}
=======================
While the existence of finite measure-preserving, measurably sensitive transformations is implied by the existence of lightly mixing finite measure-preserving transformations, there is no corresponding notion of light mixing for the infinite measure-preserving case. In fact, we show that there are no ergodic, infinite measure-preserving, measurably sensitive transformations. There exist non-conservative ergodic nonsingular transformations that are measurable sensitive; for example let $T:X\to X$ be a finite measure-preserving mixing transformation and define $S:X\times\mathbb{N}\to X\times\mathbb{N}$ by $S(x,n)= (Tx,n-1)$ if $n>1$ and $S(x,1)=(Tx,2)$.
\[t: infinite\] There are no ergodic, infinite measure-preserving, measurably sensitive transformations.
Let $T$ be an ergodic, measure-preserving transformation on a $\sigma$-finite measure space $X$ with infinite measure. Then $X =
\bigsqcup_{i=1}^{\infty} A_i$, where each set $A_i$ has positive finite measure. Let $D_i = \bigcup_{j=1}^{i}A_j$. Then, as both $A_{i+1}$ and $D_i$ have finite measure, $\liminf_{n\rightarrow\infty}\mu(T^{-n}(D_i)\cap A_{i+1}) = 0$. Choose an increasing sequence $\{n_{i,k}\}_{k=1}^{\infty}$ so that $\mu(T^{-n_{i,k}}(D_i)\cap A_{i+1}) < 2^{-k-1}\mu(A_{i+1})$. Then let
$$C_i = A_{i+1}\backslash\left(\bigcup_{k=1}^{\infty}T^{-n_{i,k}}(D_i)\right)$$
The set $C_i$ has positive measure, is contained in $D_{i+1}$ and $T^{n_{i,k}}(C_i)\cap D_i = \emptyset$ for every natural number $k$. Let $g_{i}$ be a nonsingular isomorphism from $C_i$ to $(2^{-2i},
2^{-2i+1})$ with Lebesgue measure and let $h_i$ be a nonsingular isomorphism $A_{i+1}\backslash C_i$ to $(2^{-2i+1}, 2^{-2i+2})$ with Lebesgue measure whenever $A_{i+1}\backslash C_i$ has positive measure. Let $N$ be the set where none of the functions $h_i$ and $g_i$ are defined as well as their backwards orbits. This set must have measure zero due to the nonsingularity of $T$ and so $T'$, the restriction of $T$ to $X\backslash N$, is measurably isomorphic to $T$. Let $g: X\backslash N\rightarrow (0,1)$ be equal to whichever of $h_i$ and $g_i$ is defined. Note that $g(D_i)\subset
(0,2^{-2i})$. Then define metric $d$ on $X\backslash N$ by $d(x,y) =
|g(x)-g(y)|$. A ball in metric $d$ around any point in $X\backslash
N$ must have positive measure as each isomorphism is nonsingular and so the metric is good. Let $x\in C_i\backslash N$. For sufficiently small $\varepsilon
> 0$, $B_{\varepsilon}(x)\subset C_i$. Then $(T')^{n_{i,k}}(B_{\varepsilon}(x))\subset D^{c}\backslash N$ and, as any two points in $D_{i}^{c}\backslash N$ have a maximum distance of $2^{-2i + 1}$ between them, any sensitivity constant must be at most $2^{-2i + 1}$. Consequently, $T'$ does not exhibit strong sensitive dependence in this metric. As the metric is good, $T$ is not measurably sensitive.
Although an infinite, ergodic, measure-preserving dynamical system $(X,\mu,T)$ cannot be measurably sensitive, it can, however, exhibit the desired property with respect to a good metric.
Let $(X,\mu,T)$ be the Hajian-Kakutani Skyscraper and $d$ the standard Euclidean metric $X$. Then there exists $\delta
> 0$ such that for all $x \in X$ and all $\varepsilon
> 0$ there exists $N \in \mathbb{N}$ such that $\mu \{y\in
B_{\varepsilon}(x): d(T^n(x),T^n(y))
> \delta \} > 0$ for all integers $n \geq N$.
The Hajian-Kakutani Skyscraper is an infinite measure-preserving, invertible, ergodic, rank-one transformation constructed from a recursively defined sequence of columns consisting of left-open, right-closed intervals. $C_0$ consists of $(0,1]$, and $C_{n+1}$ is formed from $C_n$ by cutting $C_n$ into two equal pieces, placing $2h_n$ spacers, which we denote $S_{n+1}$, above the right-hand half of $C_n$, and stacking right-over-left. The number of levels in column $C_n$ is denoted $h_n$. Refer to [@HK] for a full description of the construction.
We begin with some remarks about the structure of column $C_k$, for $k \geq 1$. $C_k$ consists of 4 subcolumns of height $h_{k-1}$; the bottom two subcolumns are points in $C_{k-1}$, and the top two are points in $S_k$. $S_{k+1}$ may be thought of as consisting of 8 subcolumns of height $h_{k-1}$ and width half that of those in $C_k$ positioned above the right half of $C_k$. For convenience, we refer to these subcolumns in order from bottom to top by **$K_1$** to **$K_{12}$**.
Let $\delta < \frac{3}{8}$, $x \in X$ and $\varepsilon
> 0$. There exists an integer $\alpha > 1$ such that $B_{\varepsilon}(x)$ contains a level, which we denote $I$, of $C_{\alpha}$. Define a sequence $\{l_i\}_{i=\alpha}^{\infty}$, where $l_i$ is defined to be the first time for which the top level of column $C_i$ is contained in $T^{l_i}(I)$. Set $N = l_{\alpha}$. Then for an integer $n > N$, there exists a unique integer $k \geq
1$ such that $l_k < n \leq l_{k+1}$.
Denote the left half of $I$ by $L$ and the right half by $R$. Since $T^{l_k + 8h_{k-1}}(I)$ contains the top level of column $C_{k+1}$, there exists a unique $j \in \{1,...,8\}$ such that $(j-1)h_{k-1} <
n-l_k \leq jh_{k-1}$. Then $T^{n}(L) \subset K_j$ and $T^{n}(R)
\subset K_{j+4}$. When $j = 1$, $T^{n}(L) \subset K_1 \subset
C_{k-1}$ and $T^{n}(R) \subset K_{j+4} \subset S_{k+1}$, so $S_{k}$ lies between $T^{n}(R)$ and $T^{n}(L)$ on the real line. As $\mu(S_k) >
1$, $d(T^{n}(L), T^{n}(R)) \ge 1$. For $2 \le j \le 8$, $K_{j+3}$ lies between $T^{n}(R)$ and $T^{n}(L)$ on the real line and as $\mu
(K_{j+3}) \ge \frac{1}{4}$, $d(T^{n}(L),T^{n}(R)) \ge \frac{1}{4}$. As $\delta < \frac{1}{8}$, $T^{n}(x)$ cannot be within $\delta$ of both $T^{n}(L)$ and $T^{n}(R)$. Hence $\mu\{y\in B_{\epsilon}(x):
d(T^{n}(x), T^{n}(y)) > \delta\} > 0$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this note we give some remarks and improvements on a recent paper of us [@nos] about an optimization problem for the $p-$Laplace operator that were motivated by some discussion the authors had with Prof. Cianchi.'
address:
- 'Leandro M. Del Pezzo Departamento de Matemática, FCEyN, Universidad de Buenos Aires, Pabellón I, Ciudad Universitaria (1428), Buenos Aires, Argentina.'
- 'Julián Fernández Bonder Departamento de Matemática, FCEyN, Universidad de Buenos Aires, Pabellón I, Ciudad Universitaria (1428), Buenos Aires, Argentina.'
author:
- 'Leandro M. Del Pezzo and Julián Fernández Bonder'
title: 'Remarks on an optimization problem for the $p-$Laplacian'
---
Introduction
============
In this note, we want to give some remarks and improvements on a recent paper of us [@nos] about an optimization problem for the $p-$Laplace operator.
These remarks were motivated by some discussion the authors had with Prof. Cianchi and we are grateful to him.
Let us recall the problem analyzed in [@nos].
Given a domain ${\Omega}\subset{\mathbb{R}^N}$ (bounded, connected, with smooth boundary) and some class of admissibel loads $\mathcal{A}$, in [@nos] we studied the following problem: $$\mathcal{J}(f):=\int_{\partial\Omega}f(x)u_f\,{\textrm{d}}{\mathcal{H}}^{N-1} \to \mbox{max}$$ for $f\in \mathcal{A},$ where ${\mathcal{H}}^d$ denotes the $d-$dimensional Hausdorff measure and $u$ is the (unique) solution to the nonlinear problem with load $f$ $$\label{bati}
\begin{cases}
-\Delta_p u + |u|^{p-2} u = 0 & \textrm{in } {\Omega},\\
|\nabla u|^{p-2}\frac{\partial u}{\partial\nu}= f & \textrm{on }
\partial{\Omega}.
\end{cases}$$ Where $p\in(1,\infty)$, $\Delta_p u = \mbox{div}(|\nabla u|^{p-2}\nabla u)$ is the usual $p-$Laplacian, $\frac{\partial }{\partial\nu}$ is the outer normal derivative and $f\in L^{q}(\partial {\Omega})$ with $q>\frac{p'}{N'}$ .
In [@nos], we worked with three different classes of admissible functions $\mathcal{A}$
- The class of rearrangements of a given function $f_0.$
- The (unit) ball in some $L^q.$
- The class of characteristic functions of sets of given measure.
For each of these classes, we proved existence of a maximizing load (in the respective class) and analyzed properties of these maximizer.
When we worked in the unit ball of $L^{q},$ we explicitly found the (unique) maximizar for $\mathcal{J},$ namely, the first eigenfunction of a Steklov-like nonlinear eigenvalue problem.
Whereas when we worked with the class of characteristic functions of set of given boundary measure, besides to prove that there exists a maximizer function we could give a characterization of set where the maximizer function is supported. Moreover, in order to analyze properties of this maximizer, we computed the first variation with respect respect to perturbations on the set where the characteristic function was supported. See [@nos] (section 5).
The aim of this work is to generalize the results obtained for the class of characteristic functions of set of given boundary measure to the class of rearrangements function of a given function $f_0.$
Recall that if $f_0$ is a characteristic function of a set of ${\mathcal{H}}^{N-1}$-measure $\alpha$, then [*every*]{} characteristic function of a set of ${\mathcal{H}}^{N-1}$-measure $\alpha$ is a rearrangement of $f_0$.
Characterization of Maximizer Function
======================================
In this section we give characterization of the maximizer function relative to the class of rearrangements of a given function $f_0.$
We begin by observe that has a unique weak solution $u_{f}$, for which the following equations hold $$\label{rojas}
\int_{\partial{\Omega}}fu_{f} \, {\textrm{d}}{\mathcal{H}}^{N-1} = \sup_{u\in W^{1,p}(\Omega)} \mathcal{I}(u),$$ where $$\mathcal{I}(u):=\frac{1}{p-1}\Big\{ p\int_{\partial{\Omega}} f u \, {\textrm{d}}{\mathcal{H}}^{N-1} - \int_{\Omega}|\nabla u|^p + |u|^p \, {\textrm{d}}{\mathcal{H}}^N \Big\}.$$
Let $f_0\in{L^{q}(\partial\Omega)}$, with $q=p/(p-1),$ and let $\mathcal{R}_{f_0}$ be the class of rearrangements of $f_0$. We was interested in finding $$\label{cani}
\sup_{f\in \mathcal{R}_{f_0}} \int_{\partial{\Omega}} fu_{f} \, {\textrm{d}}{\mathcal{H}}^{N-1}.$$
In [@nos], Theorem 3.1, we could proof that there exists $\hat{f}\in\mathcal{R}_{f_0}$ such that $$\int_{\partial\Omega}\hat{f}\hat{u} \, \textrm{d}\mathcal{H}^{N-1}=\sup_{f\in\mathcal{R}_{f_0}}\int_{\partial\Omega}
fu_{f}\,\textrm{d} \mathcal{H}^{N-1}.$$ where $\hat{u}=u_{\hat{f}}.$
We begin by giving a characterization of this maximizer $\hat{f}$ in the spirit of [@cuccu].
\[teo1\] $\hat{f}$ is the unique maximizer of linear functional $L(f):=\int_{\partial\Omega}f\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1},$ relative to $f\in \mathcal{R}_{f_0}.$ Therefore, there is an increasing function $\phi$ such that $\hat{f}=\phi\circ\hat{u}$ $\mathcal{H}^{N-1}-$a.e.
We proceed in three steps.
[*Step 1.*]{} First we show that $\hat{f}$ is a maximizer of $L(f)$ relative to $f\in \mathcal{R}_{f_0}.$
In fact, let $h\in \mathcal{R}_{f_0}$, since $\int_{\partial\Omega}\hat{f}\hat{u} \, \textrm{d}\mathcal{H}^{N-1}=\sup_{f\in\mathcal{R}_{f_0}}\int_{\partial\Omega}
fu_{f}\,\textrm{d} \mathcal{H}^{N-1},$ we have that $$\begin{aligned}
\int_{\partial\Omega}\hat{f}\hat{u} \, \textrm{d}\mathcal{H}^{N-1}&\ge\int_{\partial\Omega}
hu_{h}\,\textrm{d} \mathcal{H}^{N-1}\\
&= \sup_{u\in W^{1,p}(\Omega)}\frac{1}{p-1}\left\{p\int_{\partial\Omega}hu \, {\textrm{d}}\mathcal{H}^{N-1} - \int_{\partial\Omega}|\nabla u|^p
+ |u|^p \, {\textrm{d}}{\mathcal{H}}^N\right\}\\
&\ge\frac{1}{p-1}\left\{p\int_{\partial\Omega}h\hat{u} \, {\textrm{d}}\mathcal{H}^{N-1} - \int_{\partial\Omega}|\nabla \hat{u}|^p
+ |\hat{u}|^p \, {\textrm{d}}{\mathcal{H}}^N\right\},\end{aligned}$$ and, since $$\int_{\partial\Omega}\hat{f}\hat{u}\,{\textrm{d}}\mathcal{H}^{N-1}=\frac{1}{p-1}\left\{p\int_{\partial\Omega}\hat{f}\hat{u} \, {\textrm{d}}\mathcal{H}^{N-1} -
\int_{\partial\Omega}|\nabla \hat{u}|^p + |\hat{u}|^p \,
{\textrm{d}}{\mathcal{H}}^N\right\},$$ we have $$\int_{\partial\Omega}\hat{f}\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1} \ge\int_{\partial\Omega}h\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}.$$ Therefore, $$\int_{\partial\Omega}\hat{f}\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}=\sup_{f\in \mathcal{R}_{f_0}} L(f).$$
[*Step 2.*]{} Now, we show that $\hat{f}$ is the unique maximizer of $L(f)$ relative to $f\in \mathcal{R}_{f_0}.$
We suppose that $g$ is another maximizer of $L(f)$ relative to $f \in \mathcal{R}_{f_0}$. Then $$\int_{\partial\Omega}\hat{f}\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}=\int_{\partial\Omega}g\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}.$$ Thus $$\begin{aligned}
\int_{\partial\Omega}g\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}&=\int_{\partial\Omega}\hat{f}\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}\\
&\ge\int_{\partial\Omega}gu_g \,{\textrm{d}}\mathcal{H}^{N-1}\\
&=\sup_{u\in W^{1,p}(\Omega)}\frac{1}{p-1}\left\{p\int_{\partial\Omega}gu\, {\textrm{d}}\mathcal{H}^{N-1}-\int_{\partial\Omega}|\nabla u|^p + |u|^p \, {\textrm{d}}\mathcal{H}^N\right\}.\end{aligned}$$
On the other hand, $$\begin{aligned}
\int_{\partial\Omega}g\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}&=\int_{\partial\Omega}\hat{f}\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}\\
&=\frac{1}{p-1}\left\{p\int_{\partial\Omega}\hat{f}\hat{u}\, {\textrm{d}}\mathcal{H}^{N-1}-\int_{\partial\Omega}|\nabla \hat{u}|^p + |\hat{u}|^p \, {\textrm{d}}\mathcal{H}^N\right\}\\
&=\frac{1}{p-1}\left\{p\int_{\partial\Omega}g\hat{u}\, {\textrm{d}}\mathcal{H}^{N-1}-\int_{\partial\Omega}|\nabla \hat{u}|^p + |\hat{u}|^p \, {\textrm{d}}\mathcal{H}^N\right\}.\end{aligned}$$ Then $$\int_{\partial\Omega}g\hat{u} \,{\textrm{d}}\mathcal{H}^{N-1}=\sup_{u\in W^{1,p}(\Omega)}\frac{1}{p-1}\left\{p\int_{\partial\Omega}gu\, {\textrm{d}}\mathcal{H}^{N-1}-\int_{\partial\Omega}|\nabla u|^p + |u|^p \, {\textrm{d}}\mathcal{H}^N\right\}.$$
Therefore $\hat{u}=u_g.$ Then $\hat{u}$ is the unique weak solution to $$\begin{cases}
\Delta_p \hat{u} + |\hat{u}|^{p-2}\hat{u}=0 &\textrm{in } \Omega,\\
|\nabla \hat{u}|^{p-2}\frac{\partial \hat{u}}{\partial \nu}=g &\textrm{on } \partial\Omega.
\end{cases}$$ Furthermore, we now that u is the unique weak solution to $$\begin{cases}
\Delta_p \hat{u} + |\hat{u}|^{p-2}\hat{u}=0 &\textrm{in } \Omega,\\
|\nabla \hat{u}|^{p-2}\frac{\partial \hat{u}}{\partial \nu}=\hat{f} &\textrm{on } \partial\Omega.
\end{cases}$$ Therefor $\hat{f}=g$ $\mathcal{H}^{N-1}-$a.e.
[*Step 3.*]{} Finally, we have that there is an increasing function $\phi$ such that $\hat{f}=\phi\circ\hat{u}$ $\mathcal{H}^{N-1}-$a.e.
This is a direct consequence of Steps 1, 2 and Theorem \[burton\] below.
This completes the proof of Theorem \[teo1\].
In order to state Theorem \[burton\], we need the following definition
The measure space $(X,\mathcal{M},\mu)$ is called nonatomic if for $U\in \mathcal{M}$ with $\mu(U)>0$, there exists $V\in \mathcal{M}$ with $V \subset U$ and $0<\mu(V)<\mu(U).$ The measure space $(X,\mathcal{M},\mu)$ is called separable if there is a sequence $\{U_n\}_{n=1}^{\infty}$ of measurable sets such that for every $V\in \mathcal{M}$ and $\varepsilon>0$ there exists n such that $$\mu(V\setminus U_n)+\mu(U_n\setminus V) <\varepsilon.$$
\[burton\] Let $(X,\mathcal{M},\mu)$ be a finite separable nonatomic measure space, let $1\le p \le\infty,$ let $q$ be the conjugate exponent of $p,$ let $f_0\in L^p(X,\mu)$ and $g\in L^q(X,\mu)$ and let $R_{f_0}$ be the set of rearrangements of $f_0$ on $X.$ If $L(f)=\int_{X}fg \,{\textrm{d}}\mu$ has a unique maximizer $\hat{f}$ relative to $\mathcal{R}_{f_0}$ there is an increasing function $\phi$ such that $f^*=\phi\circ g$ $\mu-$a.e.
Derivate with respect to the load
=================================
Now we compute the derivate of the functional $\mathcal{J}(\hat{f})$ with respect to perturbations in $\hat{f}.$ We will consider regular perturbations and asume that the function $\hat{f}$ has bounded variation in $\partial\Omega.$
We begin by describing the kind of variations that we are considering. Let $V$ be a regular (smooth) vector field, globally Lipschitz, with support in a neighborhood of $\partial{\Omega}$ such that $\langle V,\nu \rangle=0$ and let $\psi_t:{\mathbb{R}^N}\to{\mathbb{R}^N}$ be defined as the unique solution to $$\label{moreno}
\begin{cases}
\frac{{\textrm{d}}}{{\textrm{d}}t}\psi_t (x)=V(\psi_t(x)) & t>0,\\
\psi_0(x)= x & x\in {\mathbb{R}^N}.
\end{cases}$$
We have $$\psi_t(x)=x+tV(x)+ o(t) \quad \forall x\in{\mathbb{R}^N}.$$
Thus, if $f\in \mathcal{R}_{f_0},$ we define $f_t =f\circ\psi_t^{-1}.$ Now, let $$I(t):= {\mathcal J}(f_t) = \int_{\partial\Omega}u_t f_t \textrm{d}\mathcal{H}^{N-1}$$ where $u_t\in W^{1,p}(\Omega)$ is the unique solution to $$\label{balbo}
\begin{cases}
-\Delta_p u_t + |u_t|^{p-2} = 0 & \textrm{in }\Omega,\\
|\nabla u_t|^{p-2}\frac{\partial u_t}{\partial \nu}=f_t & \textrm{on } \partial\Omega.
\end{cases}$$
\[con1\] Given $f\in L^q(\partial\Omega)$ then $$\begin{aligned}
f_t=f\circ\psi_t^{-1} &\to& f \textrm{ in } L^{q}(\partial\Omega), \textrm{ as } t\to 0.\end{aligned}$$
Let $\varepsilon>0,$ and let $g\in C^{\infty}_{c}(\partial\Omega)$ fixed such that $\|f-g\|_{L^q(\partial\Omega)}<\varepsilon.$ By the usual change of variables formula, we have, $$\|f_t-g_t\|_{L^q(\partial\Omega)}^q=\int_{\partial\Omega}|f-g|^qJ_{\tau}\psi_t {\textrm{d}}{\mathcal{H}}^{N-1},$$ where $g_t=g\circ \psi_t^{-1}$ and $J\psi$ is the tangential Jacobian of $\psi$. We also know that $$J_{\tau}\psi:=1+t \mbox{ div}_\tau V + o(t).$$ Here $\textrm{div}_\tau V$ is the tangential divergence of $V$ over $\partial\Omega.$ Then $$\|f_t-g_t\|_{L^q(\partial\Omega)}^q=\int_{\partial\Omega}|f-g|^q(1 + t \mbox{ div}_\tau V +o(t))\,{\textrm{d}}{\mathcal{H}}^{N-1}.$$ Then, there exist $t_1>0$ and such that if $0<t<t_1$ then $$\|f_t-g_t\|_{L^q(\partial\Omega)}\le C\varepsilon.$$ where $C$ is a constant independent of $t.$ Moreover, since $\psi_t^{-1}\to Id$ in the $C^1$ topology when $t\to0$ then $g_t=g\circ\psi^{-1}_{t}\to g$ in the $C^1$ topology and therefore there exists $t_2>0$ such that if $0<t<t_2$ then $$\|g_t-g\|_{L^q(\partial\Omega)}<\varepsilon.$$ Finally, we have for all $0<t<t_0=\min\{t_1,t_2\}$ then $$\begin{aligned}
\|f_t-f\|_{L^q(\partial\Omega)}&\le \|f_t-g_t\|_{L^q(\partial\Omega)}+\|g_t-g\|_{L^q(\partial\Omega)}
+\|g-f\|_{L^q(\partial\Omega)}\\
&\le C\varepsilon
\end{aligned}$$ where $C$ is a constant independent of $t.$
\[con2\] Let $u_0$ and $u_t$ be the solution of with $t=0$ and $t>0$, respectively. Then $$u_t \to u_0 \textrm{ in } W^{1,p}(\Omega), \textrm{ as } t\to 0^+.$$
The proof follows exactly as the one in Lemma 4.2 in [@cuccu]. The only difference being that we use the trace inequality instead of the Poincaré inequality.
\[remark.conver\] It is easy to see that, as $\psi_t\to Id$ in the $C^1$ topology, then from Lemma \[con2\] it follows that $$w_t:=u_t\circ\psi_t\to u_0 \quad \textrm{strongly in } W^{1,p}(\Omega).$$
With these preliminaries, the following theorem follows exactly as Theorem 5.5 of [@nos].
\[derivada\] With the previous notation, we have that $I(t)$ is differentiable at $t=0$ and $$\begin{aligned}
\frac{\textrm{d}I(t)}{\textrm{d}t}\Big|_{t=0}
& = \frac{1}{p-1} \bigg\{p\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau V\, {\textrm{d}}{\mathcal{H}}^{N-1} + p \int_{\Omega}|\nabla u_0|^{p-2}\langle\nabla u_0, ^T V' \nabla u_0^T\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& \quad - \int_{\Omega}(|\nabla u_0|^p + |u_0|^p) \mbox{ div}\, V \,
{\textrm{d}}{\mathcal{H}}^N\bigg\}.\end{aligned}$$ where $u_0$ is the solution of with $t=0.$
For the details see the proof of Theorem 5.5 of [@nos].
Now we try to find a more explicit formula for $I'(0)$. For This, we consider $f\in L^q(\partial\Omega)\cap
BV(\partial\Omega)$, where $BV(\partial\Omega)$ is the space of functions of bounded variation. For details and properties of BV functions we refer to the book [@E-G].
If $f\in L^q(\partial \Omega)\cap BV(\partial\Omega),$ we have that $$\frac{\partial I(t)}{\partial t}\Big|_{t=0}=\frac{p}{p-1}\int_{\partial\Omega}u_0 V \,{\textrm{d}}[Df].$$ where $u_0$ is the solution of with $t=0.$
In the course of the computations, we require the solution $u_0$ to $$\begin{cases}
-\Delta u_0 + |u_0|^{p-2} u_0 = 0 & \mbox{in }\Omega,\\
|\nabla u_0|^{p-2}\frac{\partial u_0}{\partial \nu} = f & \mbox{on }
\partial\Omega,
\end{cases}$$ to be $C^2$. However, this is not true. As it is well known (see, for instance, [@16]), $u_0$ belongs to the class $C^{1,\delta}$ for some $0<\delta<1$.
In order to overcome this difficulty, we proceed as follows. We consider the regularized problems $$\label{regularized}
\begin{cases}
-\mbox{div}( (|\nabla u_0^\varepsilon|^2 + \varepsilon^2)^{(p-2)/2}\nabla u^\varepsilon_0) + |u^\varepsilon_0|^{p-2}
u^\varepsilon_0 = 0 & \mbox{in }\Omega,\\
(|\nabla u^\varepsilon_0|^2 + \varepsilon^2)^{(p-2)/2}\frac{\partial u^\varepsilon_0}{\partial \nu} =
f & \mbox{on }\partial\Omega.
\end{cases}$$ It is well known that the solution $u_0^\varepsilon$ to is of class $C^{2,\rho}$ for some $0<\rho<1$ (see [@LSU]).
Then, we can perform all of our computations with the functions $u_0^\varepsilon$ and pass to the limit as $\varepsilon\to 0+$ at the end.
We have chosen to work formally with the function $u_0$ in order to make our arguments more transparent and leave the details to the reader. For a similar approach, see [@GM].
Now, by Theorem \[derivada\] and since $$\begin{aligned}
\mbox{div}(|u_0|^pV) & = & p|u_0|^{p-2}u_0\langle\nabla u_0, V\rangle + |u_0|^p \mbox{ div}\, V,\\
\mbox{div}(|\nabla u_0|^pV)&=&p|\nabla u_0|^{p-2}\langle\nabla u_0 D^2 u_0, V\rangle + |\nabla u_0|^p \mbox{ div}\, V,\end{aligned}$$ we obtain $$\begin{aligned}
I'(0) = & \frac{1}{p-1} \bigg\{p\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau V\,
{\textrm{d}}{\mathcal{H}}^{N-1} + p \int_{\Omega}|\nabla u_0|^{p-2}\langle\nabla u_0, ^T V' \nabla u_0^T\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& - \int_{\Omega}(|\nabla u_0|^p + |u_0|^p) \mbox{ div}V \, {\textrm{d}}{\mathcal{H}}^N\\
= & \frac{1}{p-1} \bigg\{p\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau
V\,{\textrm{d}}{\mathcal{H}}^{N-1} + p\int_{\Omega}|\nabla u_0|^{p-2}\langle\nabla u_0, ^T V' \nabla u_0^T\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& -\int_{\Omega}\mbox{div}((|\nabla u_0|^p + |u_0|^p)V)\, {\textrm{d}}{\mathcal{H}}^N
+ p\int_{\Omega}|\nabla u_0|^{p-2}\langle \nabla u_0 D^2 u_0, V\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& + p\int_{\Omega}|u_0|^{p-2}u_0 \langle \nabla u_0, V\rangle \, {\textrm{d}}{\mathcal{H}}^N\bigg\}.\end{aligned}$$ Hence, using that $\langle V, \nu\rangle=0$ in the right hand side of the above equality we find $$\begin{aligned}
I'(0) & = & \frac{p}{p-1}\bigg\{\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau V\, {\textrm{d}}{\mathcal{H}}^{N-1}\\
& & + \int_{\Omega}|\nabla u_0|^{p-2} \langle \nabla u_0, ^T V' \nabla u_0^T + D^2 u_0 V^T\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& & + \int_{\Omega}|u_0|^{p-2}u_0 \langle \nabla u_0, V\rangle \, {\textrm{d}}{\mathcal{H}}^N\bigg\}\\
& = & \frac{p}{p-1} \bigg\{\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau V\,
{\textrm{d}}{\mathcal{H}}^{N-1} + \int_{\Omega}|\nabla u_0|^{p-2} \langle \nabla u_0, \nabla(\langle\nabla u_0, V\rangle)\rangle\, {\textrm{d}}{\mathcal{H}}^N\\
& & + \int_{\Omega}|u_0|^{p-2}u_0 \langle \nabla u_0, V\rangle \, {\textrm{d}}{\mathcal{H}}^N\bigg\}.\end{aligned}$$ Since $u_0$ is a week solution of with $t=0$ we have $$\begin{aligned}
I'(0) & = & \frac{p}{p-1}\bigg\{\int_{\partial{\Omega}} u_0 f \mbox{ div}_\tau
V\, {\textrm{d}}{\mathcal{H}}^{N-1} + \int_{\partial{\Omega}} \langle \nabla u_0, V\rangle f\, {\textrm{d}}{\mathcal{H}}^{N-1}\bigg\}\\
& = & \frac{p}{p-1} \int_{\partial{\Omega}} \mbox{div}_\tau(u_0 V) f\, {\textrm{d}}{\mathcal{H}}^{N-1}\\\end{aligned}$$ Finally, since $f\in BV(\partial\Omega)$ and $V\in C^1(\partial\Omega;{\mathbb{R}^N}),$ $$\begin{aligned}
I'(0)& = & \frac{p}{p-1} \int_{\partial{\Omega}} \mbox{div}_\tau(u_0 V) f\, {\textrm{d}}{\mathcal{H}}^{N-1}\\
&=&\frac{p}{p-1}\int_{\partial\Omega}u_0 V \,{\textrm{d}}[Df].\end{aligned}$$ The proof is now complete.
[99]{}
Burton, G. R. Rearrangements of functions, maximization of convex functionals, and vortex rings. Math. Ann. 276 (1987), no. 2, 225–253.
F. Cuccu, B. Emamizadeh and G. Porru. [*Nonlinear elastic membranes involving the $p$-Laplacian operator*]{}. Electron. J. Differential Equations 2006, No. 49, 10 pp.
Del Pezzo, L. M. and Fernández Bonder, J. Some optimization problems for nonlinear elastic membranes. To appear in Appl. Math. Optim.
Evans, Lawrence C.; Gariepy, Ronald F. [*Measure theory and fine properties of functions*]{}. Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1992.
J. García Meliá, J. Sabina de Lis. [*On the perturbation of eigenvalues for the $p$-Laplacian*]{}. C. R. Acad. Sci. Paris Sér. I Math. 332 (2001), no. 10, 893–898.
O.A. Ladyzhenskaya, V.A. Solonnikov, N.N. Ural’tseva, Linear and Quasilinear Equations of Parabolic Type, Transl. Math. Monographs, Vol. 23, Amer. Math. Soc., Providence, R.I., 1968.
P. Tolksdorf. [*Regularity for a more general class of quasilinear elliptic equations*]{}. J. Differential Equations, 51 (1984), 126–150.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This manuscript investigates the existence and uniqueness of solutions to the first order fractional anti-periodic boundary value problem involving Caputo-Katugampola (CK) derivative. A variety of tools for analysis this paper through the integral equivalent equation of the given problem, fixed point theorems of Leray–Schauder, Krasnoselskii’s, and Banach are used. Examples of the obtained results are also presented.'
author:
- |
Saleh S. Redhwan [^1]\
Department of Mathematics\
Dr.Babasaheb Ambedkar Marathwada\
University, Aurangabad, (M.S),431001, India\
`saleh.redhwan909@gmail.com`\
Sadikali L. Shaikh\
Department of Mathematics\
Maulana Azad College of arts,\
Science and Commerce\
Aurangabad, (M.S),431001, India\
``\
Mohammed S. Abdo$^1$$^,$$^2$\
$^1$Department of Mathematics\
Dr.Babasaheb Ambedkar Marathwada University\
Aurangabad, (M.S),431001, India\
$^2$Hodeidah University\
Al-Hodeidah, 3114, Yemen\
`msabdo1977@gmail.com`\
title: 'Theory of Nonlinear Caputo-Katugampola Fractional Differential Equations'
---
Introduction
============
In this manuscript, we investigate the sufficient conditions of the existence and uniqueness of solutions for the CK fractional differential equation$$^{c}D_{a^{+}}^{\alpha ;\rho }y(t)=f(t,y(t)),\text{ where }0<\alpha <1,\text{
}t\in J=[a,b]. \label{1}$$with the anti-period boundary condition $$y(a)+y(b)=0, \label{2}$$where $^{c}D_{a^{+}}^{\alpha ;\rho }$ is CK fractional derivatives of order $%
\alpha $, and $f:J\times
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
\longrightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ is the appropriate function.
The subject of fractional differential equations has recently evolved as an interesting and popular field of research; see the interesting books [@k] [@p]. In fact, fractional derivatives provide an excellent tool for the description of memory and hereditary properties of various materials and processes. More and more researchers have found that fractional differential equations play important roles in many research areas, such as physics, chemical technology, population dynamics, biotechnology, and economics. For examples and recent developments of the topic, we refer to the papers [AZ]{} [@BK] [@BO] and the references cited therein.
In [@K1] the author introduced a new fractional integral, which generalizes the Riemann-Liouville and the Hadamard integral into a single form. For further properties such as expansion formulas, variational calculus applications, control theoretical applications, convexity and integral inequalities and Hermite-Hadamard type inequalities of this new operator and similar operators, for example, see [@AG] [@SG] [RH]{} [@GJ]. The corresponding fractional derivatives were introduced in [@K2] [@K3] [@MO], Which so-called katugampola fractional operators. The existence and uniqueness results for the CK derivative are given in [@K4], the author used the Peano theorem to obtain the existence and uniqueness results of the Cauchy type problem$$^{c}D_{0^{+}}^{\alpha ;\rho }y(t)=f(t,y(t)),\text{ \ \ }t\in \lbrack 0,b],
\label{3}$$$$y^{(k)}(0)=y_{0}^{(k)},\text{ \ }k=0,1,...,m-1, \label{4}$$where $m=[\alpha ].$ In the same context, R. Almeida in [@RA], proved the uniqueness result of the problem (\[3\])-(\[4\]) involving $%
^{c}D_{a^{+}}^{\alpha ;\rho }$ via Gronwall inequality type.
On the other hand, Oliveira and de Oliveira in [@AM], considered the initial value problem for a nonlinear fractional differential equation including Hilfer-Katugampola derivative of the form$$^{\rho }D_{a^{+}}^{\alpha ,\beta }y(t)=f(t,y(t)),\text{ \ \ }t\in J,
\label{eq17}$$$$^{\rho }I_{a^{+}}^{1-\gamma }y(a)=c,\text{ \ }\gamma =\alpha +\beta -\alpha
\beta . \label{eq18}$$They used the generalized Banach fixed point theorem to investigate the existence and uniqueness results on problems (\[eq17\])-(\[eq18\]).
The recent development of fractional differential equations and the theoretical analysis can be seen in [@AP1; @AP2; @AP3; @AP4; @BB; @K4]. As such, the CK type fractional operators have not been studied and investigated in much detail as yet as compared to the other classical operators. In this manuscript, we introduce new results of the solution to the problem (\[1\])-(\[2\]) involving CK fractional operator. Further, we use some fixed point theorems to analyize our results. To the best of our knowledge, the fractional boundary value problems involving Caputo-Katugampola type fractional operators have not yet been investigated and developed till the present day.
The paper is organized as follows. In the next section, we begin by summarizing the forms of Katugampola fractional integral and CK type fractional derivative, and we also present the background material related to our work and prove an important lemma which plays a key role in the sequel. The third section contains the existence and uniqueness results to the problem (\[1\])-(\[2\]) by means of fixed point theorems of (Banach, Leray-schauder, Krasnoselskii’s). In the last section, we present some illustrative examples.
Preliminaries
=============
In this partition, we recall some essential basic definitions, lemmas, and preliminary facts related to our results throughout the paper. Let $J=[a,b]$ $\left( -\infty <a<b<\infty \right) \ $be a finite interval of $%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$. Denote $\mathcal{C(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ be the Banach space of all continuous functions $h:J\longrightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ endowed with the norm given by$$\left\Vert h\right\Vert _{\mathcal{C}}=\sup_{t\in J}\left\vert
h(t)\right\vert :t\in J\},$$where $h\in \mathcal{C(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$. $\mathcal{C}^{n}\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ $(n\in
%TCIMACRO{\U{2115} }%
%BeginExpansion
\mathbb{N}
%EndExpansion
_{0})$ denotes the set of mappings having $n$ times continuously differentiable on $J.$
For $a<b$, $c\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ and $1\leq p<\infty ,$ define the function space$$X_{c}^{p}(a,b)=\left\{ h:J\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
:\left\Vert h\right\Vert _{X_{c}^{p}}=\left( \int_{a}^{b}\left\vert
t^{c}h(t)\right\vert ^{p}\frac{dt}{t}\right) ^{\frac{1}{p}}<\infty \right\} .$$for $p=\infty ,$ $$\left\Vert h\right\Vert _{X_{c}^{p}}=ess\sup_{a\leq t\leq b}\left[
\left\vert t^{c}h(t)\right\vert \right] .$$
Now we state some definitions of the generalized fractional operators were introduced in [@K2; @K4; @K1].
\[de1\] Let $t>a$ be two reals, $\alpha >0$, $\rho >0,$ and $%
h:J\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ be an integrable function. The left-sided Katugampola fractional integral of order $\alpha $ and type $\rho $ is defined by $$I_{a^{+}}^{\alpha ;\rho }h(t)=\frac{1}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho
-1}\left( \frac{t^{\rho }-s^{\rho }}{\rho }\right) ^{\alpha -1}h(s)ds,
\label{a1}$$where, $\Gamma (.)$ is a gamma function.
Let $n-1<\alpha <n$ , $(n=[\alpha ]+1),$ $\rho >0,$ and $h\in \mathcal{C}^{n}%
\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
).$ The left-sided Katugampola fractional derivative of order $\alpha $ with dependence on a parameter $\rho $ is defined as$$D_{a^{+}}^{\alpha ;\rho }h(t)=\left( t^{1-\rho }\frac{d}{dt}\right)
^{n}I_{a^{+}}^{n-\alpha ;\rho }h(t)=\frac{\gamma ^{n}}{\Gamma (1-\alpha )}%
\int_{a}^{t}\left( \frac{t^{\rho }-s^{\rho }}{\rho }\right) ^{n-\alpha -1}%
\frac{h(s)}{s^{1-\rho }}ds\text{, }t>a, \label{e2}$$where $\gamma =\left( t^{1-\rho }\frac{d}{dt}\right) .$ In particular, if $%
0<\alpha <1,$ $\rho >0,$ and $h\in C^{1}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
),$ we have $$D_{a^{+}}^{\alpha ;\rho }h(t)=\left( t^{1-\rho }\frac{d}{dt}\right)
I_{a^{+}}^{1-\alpha ;\rho }h(t)=\frac{\gamma }{\Gamma (1-\alpha )}%
\int_{a}^{t}\left( \frac{t^{\rho }-s^{\rho }}{\rho }\right) ^{-\alpha }\frac{%
h(s)}{s^{1-\rho }}ds\text{, }t>a.$$
Let $\alpha \geq 0,$ $n=[\alpha ]+1.$ If $h\in \mathcal{C}^{n}\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
),$ we define the left sided CK fractional derivatives of $h$ of order $%
\alpha $ with a parameter $\rho >0$ by$$^{C}D_{a^{+}}^{\alpha ;\rho }h(t)=D_{a^{+}}^{\alpha ;\rho }\left[
h(t)-\sum_{k=0}^{n-1}\frac{h_{\rho }^{(k)}(a)}{k!}\left( \frac{t^{\rho
}-a^{\rho }}{\rho }\right) ^{k}\right] , \label{e3}$$where $h_{\rho }^{(k)}(t)=\left( t^{1-\rho }\frac{d}{dt}\right) ^{k}h(t).$ In case $0<\alpha <1$, and $h\in \mathcal{C}^{1}\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ we have $$^{C}D_{a^{+}}^{\alpha ;\rho }h(t)=D_{a^{+}}^{\alpha ;\rho }\left[ h(t)-h(a)%
\right] .$$ From last equation and (\[e2\]), one deduces $$^{C}D_{a^{+}}^{\alpha ;\rho }h(t)=\frac{\gamma }{\Gamma (1-\alpha )}%
\int_{a}^{t}\left( \frac{t^{\rho }-s^{\rho }}{\rho }\right) ^{-\alpha }\frac{%
\left[ h(s)-h(a)\right] }{s^{1-\rho }}ds\text{, }t>a.$$Notice that, if $\alpha \notin
%TCIMACRO{\U{2115} }%
%BeginExpansion
\mathbb{N}
%EndExpansion
_{0}$, $h$ is an absolutely continuous function on $J,$ then the CK fractional derivative exists a.e. Moreover, we have$$\begin{aligned}
^{C}D_{a^{+}}^{\alpha ;\rho }h(t) &=&\frac{1}{\Gamma (1-\alpha )}%
\int_{a}^{t}\left( \frac{t^{\rho }-s^{\rho }}{\rho }\right) ^{-\alpha }\frac{%
h_{\rho }^{(1)}(s)}{s^{1-\rho }}ds\text{, }t>a, \\
&=&I_{a^{+}}^{1-\alpha ;\rho }h_{\rho }^{(1)}(t).\end{aligned}$$Also, if $\alpha \in
%TCIMACRO{\U{2115} }%
%BeginExpansion
\mathbb{N}
%EndExpansion
,$ $^{C}D_{a^{+}}^{\alpha ;\rho }h(t)=h_{\rho }^{(n)}(t).$ Particularly, $%
^{C}D_{a^{+}}^{0;\rho }h(t)=h_{\rho }^{(0)}(t)=h(t).$
$I_{a^{+}}^{\alpha ;\rho }$ is bounded on the function space $%
X_{c}^{p}(a,b). $
\[L2\] Let $\alpha >0,$ $\beta >0,$ $h$ $\in X_{c}^{p}(a,b)$ $(1\leq
p<\infty ),$ $\rho ,c\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
,$ $\rho \geq c.$ Then we have$$I_{a^{+}}^{\alpha ;\rho }I_{a^{+}}^{\beta ;\rho }h(t)=I_{a^{+}}^{\alpha
+\beta ;\rho }h(t),\qquad ^{c}D_{a^{+}}^{\alpha ;\rho }\text{ }^{\rho
}I_{a^{+}}^{\alpha }h(t)=h(t).$$
\[L5\] Let $t>a,$ $\alpha ,\delta \in (0,\infty ),$ and $%
I_{a^{+}}^{\alpha ;\rho },D_{a^{+}}^{\alpha ;\rho }$ and $%
^{C}D_{a^{+}}^{\alpha ;\rho }$are according on (\[a1\]), (\[e2\]) and (\[e3\]) respectively. Then we have$$I_{a^{+}}^{\alpha ;\rho }\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{\delta -1}=\frac{\Gamma (\delta )}{\Gamma (\delta +\alpha )}\left( \frac{%
t^{\rho }-a^{\rho }}{\rho }\right) ^{\alpha +\delta -1},$$$$^{C}D_{a^{+}}^{\alpha ;\rho }\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{\delta -1}=\frac{\Gamma (\delta )}{\Gamma (\delta -\alpha )}\left( \frac{%
t^{\rho }-a^{\rho }}{\rho }\right) ^{\delta -\alpha -1},$$and$$^{C}D_{a^{+}}^{\alpha ;\rho }\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{k}=0,\text{ \ }\alpha \geq 0,\text{ }k=0,1,...,n-1.$$Particularly, $^{C}D_{a^{+}}^{\alpha ;\rho }\left( 1\right) =0.$
[@BK] (Banach fixed point theorem) Let $(X;d)$ be a nonempty complete metric space with $T:X\rightarrow X$ is a contraction mapping. Then map $T$ has a fixed point.
[@BK] (Krasnoselskii’s fixed point theorem) Let $X$ be a Banach space, let $\Omega $ be a bounded closed convex subset of $X$ and let $T_{1},T_{2}$ be mapping from $\Omega $ into $X$ such that $T_{1}x+T_{2}y$ $\in $ $\Omega
$ for every pair $x,y\in \Omega $. If $T_{1}$ is contraction and $T_{2}$ is completely continuous, then the equation $T_{1}x+T_{2}x$ $=x$ has a solution on $\Omega .$
[@BK] (Leray-Schauder Nonlinear Alternative). Let $X$ be a Banach space and $\Omega \subseteq X$ closed and convex. Assume that $K$ is a relatively open subset of $\Omega $ with $0\in K$ and $T:\overline{K}\longrightarrow
\Omega $ is a compact and continuous mapping. Then ethier
1. $T$ has a fixed point in $\overline{K},$ or
2. there exists $x\in \partial K$ such that $x=\lambda Tx$ for some $%
\lambda \in (0,1)$, where $\partial K$ is boundary of $K.$
Existence and uniqueness theorems
=================================
In this partition, we present the existence and uniqueness results of fractional boundary value problem (\[1\])-(\[2\]) involving CK fractional derivatives. To prove the existence of solutions to (\[1\])-(\[2\]), we need the following auxiliary Lemmas
\[L3\] Let $\alpha ,\rho >0\ $and $y\in \mathcal{C(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)\cap \mathcal{C}^{1}\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
).$ Then
1. the CK fractional deferential equation$$^{c}D_{a^{+}}^{\alpha ;\rho }y(t)=0$$has a solutions $$y(t)=c_{0}+c_{1}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right) +c_{2}\left(
\frac{t^{\rho }-a^{\rho }}{\rho }\right) ^{2}+....+c_{n-1}\left( \frac{%
t^{\rho }-a^{\rho }}{\rho }\right) ^{n-1},$$where $c_{i}\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$, $i=0,1,2,...,n-1$ and $n=[\alpha ]+1.$
2. If $y,^{C}D_{a^{+}}^{\alpha ;\rho }y\in \mathcal{C(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)\cap \mathcal{C}^{1}\mathcal{(}J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$. Then $$I_{a^{+}}^{\alpha ;\rho }\text{ }^{C}D_{a^{+}}^{\alpha ;\rho
}y(t)=y(t)+c_{0}+c_{1}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
+c_{2}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{2}+....+c_{n-1}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right) ^{n-1},
\label{u}$$ where $c_{i}\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$, $i=0,1,2,...,n-1$ and $n=[\alpha ]+1.$
The first part, follows immediately from the fact$$^{C}D_{a^{+}}^{\alpha ;\rho }\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{k}=0,\text{ \ }k=0,1,2,...,n-1.$$
To prove the second part, we apply the operator $^{C}D_{a^{+}}^{\alpha ;\rho
}$ to $I_{a^{+}}^{\alpha ;\rho }$ $^{C}D_{a^{+}}^{\alpha ;\rho }y(t)-y(t),$ and use Lemma \[L2\], it follows that$$^{C}D_{a^{+}}^{\alpha ;\rho }\left[ I_{a^{+}}^{\alpha ;\rho }\text{ }%
^{C}D_{a^{+}}^{\alpha ;\rho }y(t)-y(t)\right] =\text{ }^{C}D_{a^{+}}^{\alpha
;\rho }I_{a^{+}}^{\alpha ;\rho }\text{ }^{C}D_{a^{+}}^{\alpha ;\rho
}y(t)-^{C}D_{a^{+}}^{\alpha ;\rho }y(t)=0.$$
From the first part, we deduce there exist $c_{i}\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ ($i=0,1,2,...,n-1$) such that$$I_{a^{+}}^{\alpha ;\rho }\text{ }^{C}D_{a^{+}}^{\alpha ;\rho
}y(t)-y(t)=c_{0}+c_{1}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
+c_{2}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right)
^{2}+....+c_{n-1}\left( \frac{t^{\rho }-a^{\rho }}{\rho }\right) ^{n-1},$$which implies the law of composition (\[u\]). The proof is completed.
\[L1\] Let $0<\alpha <1,$ $\rho >0$ and $g\in \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
).$ Then the linear anti-preiodic boundary value problem $$^{c}D_{a^{+}}^{\alpha ;\rho }y(t)=g(t),\text{ \ }t\in J, \label{eq1}$$$$y(a)+y(b)=0,\qquad \qquad \label{eq2}$$has a unique solution defined by $$y(t)=-\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}g(s)ds+\frac{\rho
^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho }-s^{\rho
})^{\alpha -1}g(s)ds. \label{7}$$
Applying the Katugampola fractional integral of order $\alpha $ to both sides of equation in (\[eq1\]), and using Lemma \[L3\], we get$$y(t)=c_{0}+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho
-1}(t^{\rho }-s^{\rho })^{\alpha -1}g(s)ds\text{,} \label{8}$$where $c_{0}\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$. Using (\[8\]) in the boundary conditions of (\[eq2\]), we get $$c_{0}=-\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}g(s)ds.$$which, on substituting in (\[8\]), yields the solution (\[7\]). The converse follows by direct calculations. The proof is completed.
Assume that (H$_{1}$) holds. A function $y(t)$ solves the problem (\[1\])-(\[2\]) if and only if it is a fixed-point of the operator $T:\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)\rightarrow \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ defined by$$\begin{aligned}
Ty(t) &=&-\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}f(s,y(s))ds \notag
\\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}f(s,y(s))ds. \label{r}\end{aligned}$$Our first result is based on the Banach fixed point theorem to obtain the existence of a unique solution of problem (\[1\])-(\[2\]).
\[Th1\] Assume that $f:J\times
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ be a continuous function satisfying the Lipschitz condition:
(H$_{1}$)
: There exists a constant $L_{f}>0$ such that : $$\left\vert f(t,x)-f(t,y)\right\vert \leq L\left\vert x-y\right\vert ,\text{
\ }\forall t\in J,\text{ }x,y\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
.$$If $L\mathcal{N}<1,$ where $\mathcal{N}:=\frac{3}{2}\frac{\rho ^{-\alpha }}{%
\Gamma (\alpha +1)}(b^{\rho }-a^{\rho })^{\alpha },$ then the boundary value problem (\[1\])-(\[2\]) has a unique solution on $J$.
Now, we first show that the operator $T:\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)\rightarrow \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ defined by (\[r\]) is well-defined, i.e., we show that $T\mathcal{B}%
_{r}\subseteq \mathcal{B}_{r}\ $where $$\mathcal{B}_{r}=\{y\in \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
),\left\Vert y\right\Vert \leq r\}, \label{ew}$$ with choose $r\geq \frac{\mu \mathcal{N}}{1-L\mathcal{N}},$ and $\sup_{t\in
J}\left\vert f(t,0)\right\vert =\mu <\infty .$ For any $y\in \mathcal{B}_{r}$, we obtain by our hypotheses that$$\begin{aligned}
\left\vert Ty(t)\right\vert &\leq &\underset{t\in J}{\sup }\left\vert
Ty(t)\right\vert \\
&\leq &\underset{t\in J}{\sup }\left\{ \frac{1}{2}\frac{\rho ^{1-\alpha }}{%
\Gamma (\alpha )}\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha
-1}\left\vert f(s,y(s))-f(t,0)\right\vert +\left\vert f(t,0)\right\vert
ds\right. \\
&&\left. +\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho
-1}(t^{\rho }-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))-f(t,0)\right\vert
+\left\vert f(t,0)\right\vert ds\right\} \\
&\leq &\frac{3}{2}\frac{\rho ^{-\alpha }}{\Gamma (\alpha +1)}(b^{\rho
}-a^{\rho })^{\alpha }\left( Lr+\mu \right) \\
&<&\left( Lr+\mu \right) \mathcal{N} \\
&\leq &r.\end{aligned}$$
which implies that $Ty\in \mathcal{B}_{r}.$Moreover, by (\[r\]), and lammas \[L2\], \[L5\], we obtain $$\text{ }^{C}D_{a^{+}}^{\alpha ;\rho }Ty(t)=\text{ }^{C}D_{a^{+}}^{\alpha
;\rho }\text{ }I_{a^{+}}^{\alpha ;\rho }f(t,y(t))=f(t,y(t)).$$
Since $f(t,y(t))$ is continuous on $J,$ Then $^{c}D_{a^{+}}^{\alpha ;\rho
}Ty(t)$ is continuous on $J,$ that is $T\mathcal{B}_{r}\subseteq \mathcal{B}%
_{r}.$
Next, we apply the Banach fixed point theorem to prove that $T$ has a fixed point. Indeed, it enough to show that $T$ is contraction map$.$ Let $%
y_{1},y_{2}\in \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ and for $t\in J,$ we have $$\begin{aligned}
\left\vert Ty_{1}(t)-Ty_{2}(t)\right\vert &\leq &\frac{1}{2}\frac{\rho
^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho
})^{\alpha -1}\left\vert f(s,y_{1}(s))-f(s,y_{2}(s))\right\vert ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\left\vert f(s,y_{1}(s))-f(s,y_{2}(s))\right\vert ds
\\
&\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}L\left\Vert
y_{1}-y_{2}\right\Vert ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}L\left\Vert y_{1}-y_{2}\right\Vert ds \\
&\leq &\frac{3}{2}\frac{\rho ^{-\alpha }L}{\Gamma (\alpha +1)}(b^{\rho
}-s^{\rho })^{\alpha }\left\Vert y_{1}-y_{2}\right\Vert \\
&=&L\mathcal{N}\left\Vert y_{1}-y_{2}\right\Vert ,\end{aligned}$$
which gives $\left\Vert Ty_{1}-Ty_{2}\right\Vert \leq L\mathcal{N}\left\Vert
y_{1}-y_{2}\right\Vert $. The inequality $L\mathcal{N}<1$ shows that $T$ is contraction mapping. As a consequence of Banach fixed point theorem. Then the problem (\[1\])-(\[2\]) has a unique solution. This complete the proof.
Next, we prove an existence result for the problem (\[1\])-(\[2\]) by using Leray–Schauder nonlinear alternative.
\[Th2\] Assume that $f:J\times
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
$ is continuous on $J.$ In addition, we assume that:
(H$_{2}$)
: There exist two functions $\psi :%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
^{+}\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
^{+}$ be a nondecreasing continuous$,$ and $\eta :J\rightarrow
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
^{+}$ is a continuous such that $$\left\vert f(t,y)\right\vert \leq \eta (t)\psi (\left\Vert y\right\Vert ),%
\text{ \ }\forall t\in J,\text{ }y\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
.$$
(H$_{3}$)
: There exists a constant $M>0$ such that $$\frac{\mathcal{N}\left\Vert \eta \right\Vert \psi \left( M\right) }{M}<1,$$where $\mathcal{N}\ $is defined as in Theorem \[Th1\]$.$ Then the boundary value problem (\[1\])-(\[2\]) has at least one solution on $J$.
Firstly, we will prove that the operator $T$ defined by (\[r\]), maps bounded sets into bounded sets in $\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$. For a positive number $r$, let $\mathcal{B}_{r}$ be a bounded ball in $%
\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ defined by (\[ew\]). Then, for $t\in J$, we have$$\begin{aligned}
\left\vert Ty(t)\right\vert &\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{%
\Gamma (\alpha )}\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha
-1}\left\vert f(s,y(s))\right\vert ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))\right\vert ds \\
&\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}\eta (s)\psi \left(
\left\Vert y\right\Vert \right) ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\eta (s)\psi \left( \left\Vert y\right\Vert \right)
ds \\
&\leq &\frac{3}{2}\frac{\rho ^{-\alpha }}{\Gamma (\alpha +1)}(b^{\rho
}-a^{\rho })^{\alpha }\left\Vert \eta \right\Vert \psi \left( r\right) \\
&=&\mathcal{N}\left\Vert \eta \right\Vert \psi \left( r\right) .\end{aligned}$$
In view of (H$_{3}$), we obtain $\left\Vert Ty\right\Vert \leq r,$ i.e. $%
\left( T\mathcal{B}_{r}\right) $ is uniformly bounded.
Next, we prove that $T$ maps bounded sets into equicontinuous sets of $%
\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$, i.e. $\left( T\mathcal{B}_{r}\right) $ is equicontinuous. Let $%
t_{1},t_{2}\in J,$ with $t_{1}<t_{2}$ and for any $y$ $\in \mathcal{B}_{r}$, then we have$$\begin{aligned}
\left\vert Ty(t_{1})-Ty(t_{2})\right\vert &\leq &\frac{\rho ^{1-\alpha }}{%
\Gamma (\alpha )}\int_{a}^{t_{1}}s^{\rho -1}\left[ (t_{1}^{\rho }-s^{\rho
})^{\alpha -1}-(t_{2}^{\rho }-s^{\rho })^{\alpha -1}\right] \left\vert
f(s,y(s)\right\vert )ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{t_{1}}^{t_{2}}s^{\rho
-1}(t_{2}^{\rho }-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))\right\vert ds \\
&\leq &\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t_{1}}s^{\rho -1}%
\left[ (t_{1}^{\rho }-s^{\rho })^{\alpha -1}-(t_{2}^{\rho }-s^{\rho
})^{\alpha -1}\right] \eta (s)\psi \left( \left\Vert y\right\Vert \right) ds
\\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{t_{1}}^{t_{2}}s^{\rho
-1}(t_{2}^{\rho }-s^{\rho })^{\alpha -1}\eta (s)\psi \left( \left\Vert
y\right\Vert \right) ds \\
&\leq &\left\Vert \eta \right\Vert \psi \left( r\right) \frac{\rho
^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t_{1}}s^{\rho -1}\left[
(t_{1}^{\rho }-s^{\rho })^{\alpha -1}-(t_{2}^{\rho }-s^{\rho })^{\alpha -1}%
\right] ds \\
&&+\left\Vert \eta \right\Vert \psi \left( r\right) \frac{\rho ^{1-\alpha }}{%
\Gamma (\alpha )}\int_{t_{1}}^{t_{2}}s^{\rho -1}(t_{2}^{\rho }-s^{\rho
})^{\alpha -1}ds \\
&\leq &\frac{2\rho ^{-\alpha }\left\Vert \eta \right\Vert \psi \left(
r\right) }{\Gamma (\alpha +1)}(t_{2}^{\rho }-t_{1}^{\rho })^{\alpha }.\end{aligned}$$
As $t_{1}\longrightarrow t_{2}$ the right-hand side of the preceding inequality is not dependent on $y$ and goes to zero. Consequently, $T%
\mathcal{B}_{r}$ is equicontinuous i.e.$$\left\vert Ty(t_{1})-Ty(t_{2})\right\vert \rightarrow 0\text{, \ }\forall
\text{ }\left\vert t_{2}-t_{1}\right\vert \rightarrow 0,\text{ }y\in
\mathcal{B}_{r}.$$So, the compactness of $T$ follows by Ascoli Arzela’s theorem, we conclude that $T$ is completely continuous.
Finally, we show there exists an open set $\mathcal{U}\subseteq \mathcal{C}%
(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ with $y\neq \lambda Ty$ for $\lambda \in (0,1)$ and $y\in \partial
\mathcal{U}$.
Let $y\in \mathcal{B}_{r}$ be any solution of $$y=\lambda Ty,\text{ }\lambda \in (0,1). \label{E3'}$$
Then$$\begin{aligned}
\left\vert y(t)\right\vert &=&\lambda \left\vert Ty(t)\right\vert \notag
\\
&<&\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{b}s^{\rho
-1}(b^{\rho }-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))\right\vert ds
\notag \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))\right\vert ds \notag \\
&\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}\eta (s)\psi \left(
\left\Vert y\right\Vert \right) ds \notag \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\eta (s)\psi \left( \left\Vert y\right\Vert \right)
ds \notag \\
&\leq &\frac{3}{2}\left\Vert \eta \right\Vert \psi \left( \left\Vert
y\right\Vert \right) \frac{\rho ^{-\alpha }}{\Gamma (\alpha +1)}(b^{\rho
}-a^{\rho })^{\alpha }, \label{E1'}\end{aligned}$$which leads to $$\frac{\left\Vert y\right\Vert }{\mathcal{N}\left\Vert \eta \right\Vert \psi
\left( \left\Vert y\right\Vert \right) }<1.$$In view of (H$_{2}$), there exists $M$ such that $\left\Vert y\right\Vert
\neq M$. This means that, any solution $y$ of equation (\[E3’\]) satisfies $\left\Vert y\right\Vert \neq M,$ let $$\mathcal{U=\{}z\in \mathbb{K};\left\Vert y\right\Vert <M\}.$$Thus, the Leray–Schauder nonlinear alternative guarantees that (\[E3’\]) has a fixed point in $\partial \mathcal{U}$, which is a solution of the boundary value problem (\[1\])-(\[2\]), The proof is over.
We will study the next existence result by using Krasnoselskii$^{^{\prime }}$s fixed point theorem. To this end, we change hypothesis (H$_{2}$) into the following one:
(H$_{4}$)
: There exists a function $q(t)\in \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ suth that $$\left\vert f(t,y)\right\vert \leq q(t),\text{ \ }\forall t\in J,\text{ }y\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
.$$
\[Th3\] Assume that $(H_{4})$ holds. In addition, we assume that:
$($H$_{5})$
: There exists a function $\delta (t)\in \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ suth that $$\left\vert f(t,x)-f(t,y)\right\vert \leq \delta (t)\left\vert x-y\right\vert
,\text{ }\forall t\in J,\text{ }x,y\in
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
.$$If $$\Lambda :=\frac{1}{2}\frac{\rho ^{-\alpha }\left\Vert \delta \right\Vert }{%
\Gamma (\alpha +1)}(b^{\rho }-a^{\rho })^{\alpha }<1, \label{w}$$then the boundary value problem (\[1\])-(\[2\]) has at least one solution on $J$.
Consider the operator $T:\mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)\longrightarrow \mathcal{C}(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
)$ defined by (\[r\]). We define $\mathcal{B}_{r_{0}}:=\{y\in \mathcal{C}%
(J,%
%TCIMACRO{\U{211d} }%
%BeginExpansion
\mathbb{R}
%EndExpansion
):\left\Vert y\right\Vert \leq r_{0}\}$, $\sup_{t\in J}\left\vert
q(t)\right\vert =\left\Vert q\right\Vert ,$ and select a suitable constant $%
r_{0}$ such that $$r_{0}=\mathcal{N}\left\Vert q\right\Vert +1,$$where $\mathcal{N}\ $is defined as in Theorem \[Th1\]. Furthermore, we need to analyze the operator $T$ into sum two operators $T_{1}+T_{2}$, as follows $$T_{1}y(t)=-\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}f(s,y(s))ds,$$and$$T_{2}y(t)=\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho
-1}(t^{\rho }-s^{\rho })^{\alpha -1}f(s,y(s))ds.$$Taking into account that $T_{1}$ and $T_{2}$ are defined on $\mathcal{B}%
_{r_{0}}.$ The proof will be given in several steps.
**Step 1**: $T_{1}y_{1}+T_{2}y_{2}$ $\in \mathcal{B}_{r_{0}}$ for every $y_{1},y_{2}\in \mathcal{B}_{r_{0}}$.
For $y_{1},y_{2}\in \mathcal{B}_{r_{0}}$, we have$$\begin{aligned}
\left\vert T_{1}y_{1}(t)+T_{2}y_{2}(t)\right\vert &\leq &\left\vert
T_{1}y_{1}(t)\right\vert +\left\vert T_{2}y(t)\right\vert \\
&\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}\left\vert
f(s,y_{1}(s))\right\vert ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t}s^{\rho -1}(t^{\rho
}-s^{\rho })^{\alpha -1}\left\vert f(s,y_{2}(s))\right\vert ds \\
&\leq &\frac{3}{2}\frac{\rho ^{-\alpha }}{\Gamma (\alpha +1)}(b^{\rho
}-a^{\rho })^{\alpha }\left\Vert q\right\Vert ,\end{aligned}$$which gives $$\left\Vert T_{1}y_{1}+T_{2}y_{2}\right\Vert \leq \mathcal{N}\left\Vert
q\right\Vert \leq r_{0}. \label{q1}$$
This proves that $T_{1}y_{1}+T_{2}y_{2}$ $\in \mathcal{B}_{r_{0}}$ for every $y_{1},y_{2}\in \mathcal{B}_{r_{0}}.$
**Step 2**: $T_{1}$ is a contration mapping on $\mathcal{B}_{r_{0}}$.
In view of hypothesis (H$_{5}$), then for each $y_{1},y_{2}\in \mathcal{B}%
_{r_{0}}$ and $t\in $ $J$; we have$$\begin{aligned}
\left\vert T_{1}y_{1}(t)-T_{1}y_{2}(t)\right\vert &\leq &\frac{1}{2}\frac{%
\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{b}s^{\rho -1}(b^{\rho
}-s^{\rho })^{\alpha -1}\left\vert f(s,y_{1}(s))-f(s,y_{2}(s))\right\vert ds
\\
&\leq &\frac{1}{2}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}%
\int_{a}^{b}s^{\rho -1}(b^{\rho }-s^{\rho })^{\alpha -1}\left\Vert \delta
\right\Vert \left\Vert y_{1}-y_{2}\right\Vert ds \\
&\leq &\frac{1}{2}\frac{\rho ^{-\alpha }\left\Vert \delta \right\Vert }{%
\Gamma (\alpha +1)}(b^{\rho }-a^{\rho })^{\alpha }\left\Vert
y_{1}-y_{2}\right\Vert ,\end{aligned}$$which implies$$\left\Vert T_{1}y_{1}-T_{1}y_{2}\right\Vert \leq \Lambda \left\Vert
y_{1}-y_{2}\right\Vert ,$$It follows from the inequality (\[w\]) that $T_{1}$ is contraction mapping.
**Step 3**: The operator $T_{2}$ is completely continuous on $\mathcal{B%
}_{r_{0}}$.
First, from the continuity of $f$, we conclude that the operator $T_{2}$ is continuous.
Next, It is easy to verify that$$\left\Vert T_{2}y\right\Vert \leq \frac{\rho ^{-\alpha }\left\Vert
q\right\Vert }{\Gamma (\alpha +1)}(b^{\rho }-a^{\rho })^{\alpha }=\frac{2}{3}%
\mathcal{N}\left\Vert q\right\Vert =\frac{2}{3}r_{0}-1<r_{0}.$$
This proves that $T_{2}$ is uniformly bounded on $\mathcal{B}_{r_{0}}$.
Finally, we show that $T_{2}$ is equicontinuous on $\mathcal{B}_{r_{0}}$. Let us set $sup_{(t,y)\in J\times \mathcal{B}_{r_{0}}}$ $\left\vert
f(t,y\right\vert =f_{0}<\infty ,$ and let $y\in \mathcal{B}_{r_{0}}$ and $%
t_{1},t_{2}\in $ $J$ with $t_{1}<t_{2}$. Then we have$$\begin{aligned}
\left\vert T_{2}y(t_{1})-T_{2}y(t_{2})\right\vert &\leq &\frac{\rho
^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t_{1}}s^{\rho -1}\left[
(t_{1}^{\rho }-s^{\rho })^{\alpha -1}-(t_{2}^{\rho }-s^{\rho })^{\alpha -1}%
\right] \left\vert f(s,y(s)\right\vert ds \\
&&+\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{t_{1}}^{t_{2}}s^{\rho
-1}(t_{2}^{\rho }-s^{\rho })^{\alpha -1}\left\vert f(s,y(s))\right\vert ds \\
&\leq &f_{0}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{a}^{t_{1}}s^{%
\rho -1}\left[ (t_{1}^{\rho }-s^{\rho })^{\alpha -1}-(t_{2}^{\rho }-s^{\rho
})^{\alpha -1}\right] ds \\
&&+f_{0}\frac{\rho ^{1-\alpha }}{\Gamma (\alpha )}\int_{t_{1}}^{t_{2}}s^{%
\rho -1}(t_{2}^{\rho }-s^{\rho })^{\alpha -1}ds \\
&\leq &\frac{\rho ^{-\alpha }f_{0}}{\Gamma (\alpha +1)}(t_{2}^{\rho
}-t_{1}^{\rho })^{\alpha }.\end{aligned}$$
As $t_{1}\longrightarrow t_{2}$ the right-hand side of the last inequality is not dependent on $y$ and goes to zero. Consequently,$$\left\vert T_{2}y(t_{1})-T_{2}y(t_{2})\right\vert \rightarrow 0\text{, \ }%
\forall \text{ }\left\vert t_{2}-t_{1}\right\vert \rightarrow 0,\text{ }y\in
\mathcal{B}_{r_{0}}.$$This proves that $T_{2}$ is equicontinuous on $\mathcal{B}_{r_{0}}$. In view of Arzela-Ascoli Theorem, it follows that $T_{2}$ is relatively compact on $%
\mathcal{B}_{r_{0}}$. Thus, all the assumptions of Krasnosel’skii fixed point theorem are satisfied. Therefore, we conclude that the boundary value problem (\[1\])-(\[2\]) has at least one solution on $J.$
Examples
========
Will be provided in the revised submission.
[99]{}
Abdo, M.S. and Panchal, S.K. (2019), Fractional integro-differential equations involving $\psi $-Hilfer fractional derivative, *Advances in Applied Mathematics and Mechanics***11** , 338-359. https://doi.org/10.4208/aamm.OA-2018-0143.
Abdo, M.S. Panchal, S.K. and Shafei, H.H., Fractional integro-differential equations with nonlocal conditions and $\psi $–Hilfer fractional derivative.* Mathematical Modelling and Analysis*, **24**(4), (2019), 564-584. https://doi.org/10.3846/mma.2019.034.
Abdo, M.S., Panchal, S.K. and Sandeep, P.B. (2019), Existence of solution for Hilfer fractional differential equations with boundary value conditions, *preprint: arXiv:1909.13680*, in press.
M. S. Abdo and S. K. Panchal, Weighted Fractional Neutral Functional Differential Equations, J. Sib. Fed. Univ. Math. Phys. 11 (2018), pp. 535-549. https://doi.org/10.17516/1997-1397-2018-11-5-535-549.
R. Almeida, [A Gronwall inequality for a general Caputo fractional operator]{}, [ *arXiv preprint arXiv:1705.10079*]{}, (2017).
R. Almeida, N. R. Bastos, [An approximation formula for the Katugampola integral]{}. [*arXiv preprint arXiv:1512.03791*]{}, (2015).
R. Almeida, A. B. Malinowska, T. MonteiroM, [Fractional differential equationswith a Caputo derivative with respect to a kernel function and their applications]{}, [ *Meth. Appl. Sci.*]{}, [**41**]{}(2018), 336-352.
R.P. Agarwal, Y. Zhou, Y. He, [Existence of fractional neutral functional differential equations]{}, [ *Comput. Math. Appl.*]{}, [**59**]{}( 2010), 1095-1100.
M. Benchohra, S. Bouriah, [Existence and stability results for nonlinear boundary value problem for implicit differential equations of fractional order]{}, [ *Moroccan J. Pure Appl. Anal.*]{},[**1**]{}(2015), 22-37.
D. Baleanu, K. Diethelm, E. Scalas, J.J. Trujillo , [Fractional Calculus Models and Numerical Methods. Series on Complexity]{}, [ *Nonlinearity and Chaos. World Scientific*]{},[**3**]{}( 2012).
T. A. Burton, C. A Kirk [Fixed Point Theorem of Krasnoselskii Schaefer Type]{},[ *Mathematische Nachrichten*]{},[**189**]{}(1998), 23-31.
D. Baleanu, O.G. Mustafa, R. P. Agarwal, [On L]{}$^{p}$*-solutions for a class of sequential fractional differential equations*. [ *Appl.Math. Comput.*]{},[**218**]{}(2011), 2074-2081.
A. G. Butkovskii, S. S. Postnov, E. A. Postnova, [Fractional integro-differential calculus and its control-theoretical applications i - mathematical fundamentals and the problem of interpretation]{}, [*Automation and Remote Control*]{},[**74, no. 4**]{}( 2013), 543-574.
S. Gaboury, R. Tremblay, B.J. Fugere, [Some relations involving a generalized fractional derivative operator]{}, [ *Journal of Inequalities and Applications*]{},[**167, no. 1**]{}( 2013).
R. Herrmann, [Fractional Calculus: An Introduction for Physicists]{}, [ *World Scienti c, River Edge, New Jerzey*]{}, 2 edition, 2014.
G. Jumarie,[On the solution of the stochastic di erential equation of exponential growth driven by fractional brownian motion]{}. [*Applied Mathematics Letters*]{}, [**18, no. 7**]{}(2005), 817-826.
U.N. Katugampola, [A new approach to generalized fractional derivatives]{},[*Bull. Math. Anal. Appl.*]{},[**6, no. 4**]{}(2014), 1-15.
U. N. Katugampola, [Existence and uniqueness results for a class of generalized fractional differential equations]{}, [*Preprint. arXiv:1411.5229*]{}, (2014).
U. N. Katugampola, [New approach to a generalized fractional integral]{}, [*Appl. Math. Comput.*]{},[**218, no. 3**]{}(2011), 860-865.
U.N. Katugampola, [Mellin transforms of the generalized fractional integrals and derivatives]{}, [*Appl. Math. Comput.*]{},[**257**]{}(2015), 566-580.
A. A. A. Kilbas, H. M. Srivastava , J. J. Trujillo, [Theory and applications of fractional differential equations]{}, [*Elsevier Science Limited*]{}, [**204**]{}(2006).
A. B. Malinowska, T. Odzijewicz, D. F. M. Torres [ Advanced Methods in the Fractional Calculus of Variations]{}, [*Springer,Berlin*]{}, (2015).
I. Podlubny, [Fractional Differential Equations]{}, [*Academic Press, San Diego*]{}, (1999).
I. A. Rus, [Ulam stabilities of ordinary differential equations in a Banach space]{}, [*Carpath. J. Math.*]{}, [**26**]{}(2010), 103–107.
J. Wang, Y. Zhou, M. Medved, [Existence and stability of fractional differential equations with Hadamard derivative]{}, [*Topol. Methods Nonlinear Anal.*]{}, [**41**]{}(2013), 113-133.
[^1]: S.S. Redhwan, saleh.redhwan909@gmail.com.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, we consider the quadratic nonlinear Schrödinger system in three space dimensions. Our aim is to obtain sharp scattering criteria. Because of the mass-subcritical nature, it is difficult to do so in terms of conserved quantities. The corresponding single equation is studied by the second author and a sharp scattering criteria is established by introducing a distance from a trivial scattering solution, the zero solution. By the structure of the nonlinearity we are dealing with, the system admits a scattering solution which is a pair of zero solution and a linear solution. Taking this fact into account, we introduce a new optimizing quantity and give a sharp scattering criterion in terms of it.'
address:
- 'Department of Mathematics, Graduate School of Science and Engineering Saitama University, Shimo-Okubo 255, Sakura-ku, Saitama-shi, Saitama 338-8570, Japan'
- 'Department of Systems Innovation, Graduate School of Engineering Sciences, Toyonaka, Osaka, Japan'
author:
- Masaru Hamano and Satoshi Masaki
title: 'A sharp scattering threshold level for mass-subcritical nonlinear Schrödinger system'
---
Introduction
============
We consider the following quadratic Schrödinger system in three space dimensions: $$\tag{NLS}\label{NLS}
\begin{cases}
\hspace{-0.4cm}&\displaystyle{i\partial_tu+\Delta u=-2v\bar{u}\qquad(t,x)\in\mathbb{R}\times\mathbb{R}^3,}\\
\hspace{-0.4cm}&\displaystyle{i\partial_tv+\frac{1}{2}\Delta v=-u^2\qquad(t,x)\in\mathbb{R}\times\mathbb{R}^3,}
\end{cases}$$ where $i=\sqrt{-1}$, $u,v:\mathbb{R}\times\mathbb{R}^3\longrightarrow\mathbb{C}$ are unknown functions, $\Delta=\sum_{j=1}^3\frac{\partial^2}{\partial x_j^2}$, and $\overline{u}$ is the complex conjugate of $u$. In this paper, we study in a weighted space. For $s\in[0,\frac{3}{2})$, we define the homogeneous weighted space $\mathcal{F}\dot{H}^s(\mathbb{R}^3)$ by the norm $$\begin{aligned}
\|f\|_{\mathcal{F}\dot{H}^s(\mathbb{R}^3)}:=\|\mathcal{F}f\|_{\dot{H}^s(\mathbb{R}^3)}=\||x|^sf\|_{L_x^2(\mathbb{R}^3)}.\end{aligned}$$ Here, $\mathcal{F}$ denotes the Fourier transform on $\mathbb{R}^3$, that is, $$\begin{aligned}
\mathcal{F}f(\xi):=\widehat{f}(\xi):=(2\pi)^{-\frac{3}{2}}\int_{\mathbb{R}^3}e^{-ix\cdot\xi}f(x)dx.\end{aligned}$$ We consider the system with the initial condition $$\begin{aligned}
(u(0),v(0))=(u_0,v_0)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}.\label{003}\end{aligned}$$
At least formally, the system has the following two conserved quantities: One is *mass* $$\label{D:mass}
M[u,v] := \int_{\mathbb{R}^3} (|u(x)|^2 + 2|v(x)|^2 ) dx$$ and the other is *energy* $$\label{D:energy}
E[u,v] := \int_{\mathbb{R}^3} \(|\nabla u(x)|^2 + \frac12 |\nabla v(x)|^2 - 2 {\operatorname{Re}}(u(x)^2\overline{v(x)}) \) dx.$$
Let us make the notion of solution clear. We need a slight modification of the notion compared with $L^2$ or $H^1$ solutions because the Schrödinger flow is not unitary in the homogeneous weighted space $\mathcal{F}\dot{H}^\frac{1}{2}$. The corresponding single equation is studied in this space in [@Mas15] (see also [@GinOzaVel94; @NakOza02] and references therein).
\[D:sol\] Let $I\subset\mathbb{R}$ be a nonempty time interval. We say that a pair of functions $(u,v) : I\times\mathbb{R}^3\rightarrow\mathbb{C}^2$ is a solution to on $I$ if $(e^{-it\Delta}u(t),e^{-\frac{1}{2}it\Delta}v(t))\in(C(I;\mathcal{F}\dot{H}^\frac{1}{2}))^2$ and the Duhamel formula $$\begin{cases}
\hspace{-0.4cm}&\displaystyle{e^{-it\Delta}u(t)=e^{-i\tau\Delta}u(\tau)+2i\int_{\tau}^te^{-is\Delta}(v\overline{u})(s)ds},\\
\hspace{-0.4cm}&\displaystyle{e^{-\frac{1}{2}it\Delta}v(t)=e^{-\frac{1}{2}i\tau\Delta}v(\tau)+i\int_{\tau}^te^{-\frac{1}{2}is\Delta}(u^2)(s)ds}
\end{cases}$$ holds in $\mathcal{F}\dot{H}^\frac{1}{2}$ for any $t,\tau\in I$, where $e^{it\Delta}:=\mathcal{F}^{-1}e^{-it|\xi|^2}\mathcal{F}$ is the Schrödinger group. We express the maximal interval of existence of $(u,v)$ by $I_{\text{max}}=(T_{\text{min}},T_{\text{max}})$. We say $(u,v)$ is forward time-global (resp. backward time-global) if $T_{\text{max}}=\infty$ (resp. if $T_{\text{min}}=-\infty$).
This definition of solutions is not time-translation invariant. That is, if $(u,v)$ is a solution to , then $(u(\cdot+\tau),v(\cdot+\tau))$ is not necessarily a solution for $\tau\in\mathbb{R}$. On the other hand, solutions to remains invariant under the rescaling $$\begin{aligned}
(u(t,x),v(t,x))\mapsto (u_{[\lambda]}(t,x),v_{[\lambda]}(t,x)):=(\lambda^2u(\lambda^2t,\lambda x),\lambda^2v(\lambda^2t,\lambda x))\end{aligned}$$ for $\lambda>0$. Corresponding transform of initial data is as follows: $$\begin{aligned}
(\phi(x),\psi(x))\mapsto(\phi_{\{\lambda\}}(x),\psi_{\{\lambda\}}(x)):=(\lambda^2\phi(\lambda x),\lambda^2\psi(\lambda x))\label{002}\end{aligned}$$ for $\lambda>0$. The $\dot{H}_x^{-\frac{1}{2}}$-norm is invariant under the above scaling transformation. In other words, the equation with initial condition is scaling critical. In this sense our problem is mass-subcritical. Remark that the equation with initial condition is also a critical problem, in the sense that the $\mathcal{F}\dot{H}^\frac{1}{2}$-norm is invariant under the scaling . To state the local well-posedness for , we introduce the function spaces $\dot{X}_m^{s,r}(t)$, $W_1$, and $W_2$ defined by norms $$\begin{aligned}
\|f\|_{\dot{X}_m^{s,r}(t)}:=\Bigl\|\Bigl(-\frac{t^2}{m^2}\Delta\Bigr)^\frac{s}{2}e^{-\frac{im|x|^2}{2t}}f\Bigr\|_{L_x^r},\ \ \
\|f\|_{W_j}:=\|f\|_{L_t^{6,2}\dot{X}_{2^{{}^{j-2}}}^{\frac{1}{2},\frac{18}{7}}}.\end{aligned}$$ For a space $\dot{X}_m^{s,r}(t)$, we omit the second exponent when $r=2$, that is, $\dot{X}_m^s(t)=\dot{X}_m^{s,2}(t)$. We discuss these function spaces in more detail in Subsection \[subsec:Function spaces\] and Subsection \[ss:space\], below. The following is our result on the local well-psedness. More detailed version is given later as Theorem \[Th. Local well-posedness2\].
\[Th. Local well-posedness\] For any $(u_0,v_0)\in \mathcal{F}\dot{H}^\frac{1}{2}\times \mathcal{F}\dot{H}^\frac{1}{2}$ there exist an open interval $I \ni 0$ and a unique solution $(u,v)\in(C(I;\dot{X}_{1/2}^{1/2})\cap W_1(I)) \times (C(I;\dot{X}_1^{1/2})\cap W_2(I))$ to with a initial condition $(u(0),v(0))=(u_0,v_0)$.
Now, we turn to the large time behavior of solutions to . There are various types of possible behavior. In this paper, we are interested in scattering solutions defined as follows:
We say that a solution $(u,v)$ scatters forward in time (resp. backward in time) if $(u,v)$ is forward time-global (resp. backward time-global) and there exists $(u_+,v_+)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$ (resp. $(u_-,v_-)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$) such that $$\begin{aligned}
\label{E:definition of scattering}
\lim_{t\rightarrow+\infty}\|(e^{-it\Delta}u(t),e^{-\frac{1}{2}it\Delta}v(t))-(u_+,v_+)\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}}=0\hspace{0.28cm}&\\
\left(\text{resp.~} \lim_{t\rightarrow-\infty}\|(e^{-it\Delta}u(t),e^{-\frac{1}{2}it\Delta}v(t))-(u_-,v_-)\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}}=0\right)&.\notag\end{aligned}$$ We say that a solution $(u,v)$ scatters when it scatters both forward and backward in time.
One has several equivalent characterizations of the scattering. For example, the forward-in-time scattering is equivalent to the boundedness of ${\lVert u\rVert}_{W_1((\tau, T_{\max}))}$ or of ${\lVert v\rVert}_{W_2((\tau, T_{\max}))}$ for some $\tau \in I_{\max}$. See Proposition \[Scattering criterion\] for the details.
Criterion for scattering by conserved quantities
------------------------------------------------
We are interested in obtaining a sharp condition for scattering. This subject is recently extensively studied based on a concentration compactness/rigidity type argument after Kenig and Merle [@KenMer06]. As for the Schrödinger system , the first author treated five dimensions [@Ham18] and Inui, Kishimoto, and Nishimura treated four dimensions [@InuKisNis18]. In these results, a sharp condition for scattering is given in terms of conserved quantities. For example, in the four dimensions, the equation is mass-critical and we have the following simple criterion in terms of the mass: a solution scatters for both time direction if the mass of a solution is smaller than that of the ground state solution. This is a natural extension of the single equation case [@Dod15; @KilTaoVis09; @KilVisZha08]. In five dimensions, condition for blowup is also studied.
Remark that if a solution $(u,v)$ scatters then any scaled solution $(u_{[\lambda]},v_{[\lambda]})$ scatters also. Consequently, any criterion for scattering is scaling invariant.[^1] Hence, if we look for a criteria given in terms of some characteristic quantity of a solution, it is natural that the quantity is scaling invariant. Recall that mass $M$ is scaling invariant in four dimensions, and the product of two quantities $ME$ is a scaling invariant in five dimensions. In the previous results [@Ham18; @InuKisNis18], these quantities play a crucial role in criterion there.
However, in the three dimensional case, it would be difficult to give a criteria in terms of the mass and the energy. They are not scaling invariant: $$M[\phi_{\{\lambda\}},\psi_{\{\lambda\}}] = \lambda M[\phi,\psi], \quad
E[\phi_{\{\lambda\}},\psi_{\{\lambda\}}] = \lambda^3 E[\phi,\psi]$$ for $\lambda>0$. Furthermore, the both right hands sides has the positive power of $\lambda$. It means that one can make the both magnitudes of $M$ and of $E$ small or large at the same time by scaling. This is a feature of the mass-subcritical case. Thus, we may not have a criteria similar to those in four or five dimensional cases, as one may not construct a scaling invariant quantity by a combination or product of (positive powers of) these quantities.
Hence, in the sequel, we look for a criteria which is not given in terms of the conserved quantities, as in [@Mas15; @Mas17; @Mas16; @MasSeg18].
Trivial scattering set and minimization of non-scattering solutions
-------------------------------------------------------------------
It can be said that the main purpose of this paper is to investigate transition phenomena between scattering solutions near the trivial scattering set and other solutions. For comparison, let us recall the single NLS equation with the gauge invariant quadratic nonlinearity: $$\begin{aligned}
\label{E:NLSsingle}
i \partial_t w + \Delta w = |w|w, \qquad(t,x)\in\mathbb{R}\times\mathbb{R}^3\end{aligned}$$ with $w(0)=w_0 \in \mathcal{F} \dot{H}^{1/2}$. Let $\mathcal{S}_{+,\mathrm{single}}$ be the set of initial data for which the corresponding solution scatters forward in time. Since the zero solution is a scattering solution, we have $$\begin{aligned}
\{ 0 \} \subset \mathcal{S}_{+,\mathrm{single}}.\end{aligned}$$ Then, it is natural to define a “size" of a solution by the distance from the zero solution. This fact leads us to a study of a quantity like $$\begin{aligned}
\label{E:singlelc}
\inf \left\{ \| w_0 \|_{\mathcal{F} \dot{H}^\frac12} : w_0 \not\in \mathcal{S}_{+,\mathrm{single}} \right\}.\end{aligned}$$ It is known [@Mas15; @Mas17] that this infimum value is strictly smaller than the size of the ground state solution for and further that there exists a minimizer to this quantity.
Let us go back to the system case. We define $\mathcal{S}_{+}$ as the set of initial data $(u_0,v_0) \in \mathcal{F} \dot{H}^{\frac12} \times \mathcal{F} \dot{H}^{\frac12}$ for which the corresponding solution scatters forward in time. A straightforward generalization of the quantity is $$\begin{aligned}
\label{E:systemlc}
\inf \left\{ (\| u_0 \|_{\mathcal{F} \dot{H}^\frac12}^2 + \alpha \| v_0 \|_{\mathcal{F} \dot{H}^\frac12}^2)^{1/2} : (u_0,v_0) \not\in \mathcal{S}_{+} \right\}\end{aligned}$$ with some constant $\alpha>0$. However, there may not be a strong motivation to study this quantity other than the similarity to because the quantity is not relevant to conserved quantities of . Hence, we want to find a different way of sizing which is based on a system nature. To this end, we look at the fact that not only the zero solution $(0,0)$ but also all solutions of the form $(0,e^{\frac12it\Delta}v_0)$ can be also regarded as a trivial scattering solution for arbitrary $v_0 \in \mathcal{F} \dot{H}^{\frac12}$. Taking this fact into account, one natural choice of the “size" of an initial data would be the distance from the set $\{0\} \times \mathcal{F} \dot{H}^{\frac12}$. This choice leads us to consider the following optimization problem: $$\label{D:ellv0}
\ell_{v_0}:=\inf \{\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}: (u_0,v_0)\not\in \mathcal{S}_+ \} \in (0,\infty].$$ By using a stability type argument, we will show that $\ell_{v_0}>0$ for any $v_0 \in \mathcal{F} \dot{H}^{\frac12}$ (see, Proposition \[Small data scattering\]). The following criteria is obvious by the definition of $\ell_{v_0}$.
\[P:SSDS\] Let $(u_0,v_0)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$ and let $(u,v):I_{\text{max}}\times\mathbb{R}^3\longrightarrow\mathbb{C}^2$ be the solution to with the initial condition . If $\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}<\ell_{v_0}$ then $(u,v)$ scatters forward in time.
The above criteria “$\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}<\ell_{v_0}$" is sharp in such a sense that the number $\ell_{v_0}$ may not be replaced with any larger number. The questions which we address in this paper are the following two: (a) to obtain a condition which implies $\ell_{v_0}$ is finite; (b) to show the existence of a minimizer to $\ell_{v_0}$ (when $\ell_{v_0}$ is finite).
Stability of ground state
-------------------------
A characteristic property of the mass-subcritical case is that the ground state is orbitally stable in $H^1\times H^1$ [@Din20]. The ground state solution for is a solution of the form $$(e^{i\omega t} Q_{1,\omega}(x), e^{2i \omega t} Q_{2,\omega}(x)),$$ where $\omega>0$ and $(Q_{1,\omega}(x), Q_{2,\omega}(x))$ is a positive radial solution to the elliptic equation $$- \Delta Q_{1,\omega} + \omega Q_{1,\omega} = 2 Q_{1,\omega} Q_{2,\omega},\quad
- \frac12 \Delta Q_{2,\omega} + 2\omega Q_{2,\omega} = Q_{1,\omega}^2.$$ The orbital stability implies that there exists an open neighborhood $\mathcal{N} \subset \mathbb{R}^2$ of $(1,1)$ such that $$(c_1 Q_{1,\omega}(x), c_2 Q_{2,\omega}(x)) \notin S_+$$ for all $(c_1,c_2) \in \mathcal{N}$. (This also follows from Theorem \[Nonpositive energy\], below.) Hence, the ground state solutions are not optimizers to our problems. In particular, $\ell_{Q_{2,\omega}}$ is strictly smaller than ${\lVert Q_{1,\omega}\rVert}_{\mathcal{F} \dot{H}^\frac{1}{2}}$. Similarly, $(Q_{1,\omega}(x), Q_{2,\omega}(x))$ is not a solution to for any choice of $\alpha>0$.
Main results
------------
It will turn out that the following quantity $\ell_{v_0}^\dagger$ plays an important role in the analysis of $\ell_{v_0}$.
For $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$ and $0\leq \ell<\infty$, we let $$L_{v_0}(\ell):=\sup\left\{\|(u,v)\|_{W_1([0,T_{\max}))\times W_2([0,T_{\max}))}\middle|
\begin{array}{l}
(u,v)\text{ is the solution to \eqref{NLS} on }[0,T_{\max}),\\[0.1cm]
v(0)=v_0,\,\,
\|u(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell,\ u(0)\in\mathcal{F}\dot{H}^\frac{1}{2}
\end{array}
\right\},$$ where $W_j([0,T_{\max}))$ $(j=1,2)$ is a Strichartz-like function space defined in Subsection \[ss:space\], below. Further, define $$\label{D:ellv0d}
\ell_{v_0}^\dagger:=\sup\{\ell:L_{v_0}(\ell)<\infty\} \in (0,\infty].$$
We have $\ell_{v_0}^\dagger \le \ell_{v_0}$ by the definition of the each quantities (see Lemma \[Comparison of ell and ell\^dagger\] for more detail). Intuitively, this can be seen by noticing that if $\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}<\ell_{v_0}^\dagger$ then not only $(u,v)$ scatters forward in time but also we have a priori bound $\|(u,v)\|_{W_1([0,\infty))\times W_2([0,\infty))}\le L_{v_0}(\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}) <\infty$. As for the single-equation , it is known that these two kinds of quantities coincide each other (see [@Mas16]). Our first result is as follows.
\[T:l0\] $\ell_{v_0}^\dagger=\min (\ell_0,\ell_{v_0})$ is true for any $v_0 \in \mathcal{F} \dot{H}^{\frac12}$, including the case where the both sides are infinite. In particular, $\ell_{0}^\dagger=\ell_0$ holds.
It is worth noting that $\ell_{v_0}^\dagger=\infty$ guarantees $\ell_{v_0}=\infty$ but the inverse is not necessarily true. Our interest in the sequel is to see what happens when $\ell_{v_0}^\dagger<\infty$.
In the case $v_0=0$, we have $\ell_{0}^\dagger = \ell_0$, including the case both are infinite, as seen in Theorem \[T:l0\].
\[T:case0\] If $\ell_{0}^\dagger<\infty$ then there exists a minimizer $(u^{(0)}(t),v^{(0)}(t))$ to $\ell_{0}(=\ell_0^\dagger)$ such that
1. $v^{(0)}(0)=0$ and ${\lVert u^{(0)}(0)\rVert}_{\mathcal{F} \dot{H}^{\frac12}} = \ell_{0}$;
2. $(u^{(0)}(t),v^{(0)}(t))$ does not scatter forward in time.
So far, we do not know whether $\ell_0^\dagger<\infty$ or not. It will turn out that this question is important to understand the attainability of $\ell_{v_0}$ for all non-zero $v_0$. One quick conclusion of $\ell_0^\dagger=\infty$ is $\ell_{v_0}=\ell_{v_0}^\dagger$ for all $v_0$, which follows from Theorem \[T:l0\]. We will resume this subject later.
Let us move on to the case $v_0\neq 0$. Suppose $\ell_{v_0}^\dagger <\infty$. Then, we have either $$\label{E:CaseAB}
\ell_{v_0}^\dagger=\ell_{v_0}\quad \text{ or } \quad
\ell_{v_0}^\dagger< \ell_{v_0}.$$ The following Theorem \[T:case1\] is about the first case and Theorem \[T:case2\] is about the second case, respectively.
\[T:case1\] Fix $v_0 \in \mathcal{F} \dot{H}^{\frac12}\setminus \{0\}$. Suppose that $\ell_{v_0}^\dagger=\ell_{v_0}<\ell_0$. Then, there exists a minimizer $(u^{(v_0)}(t),v^{(v_0)}(t))$ to $\ell_{v_0}$ such that
1. $v^{(v_0)}(0)=v_0$ and ${\lVert u^{(v_0)}(0)\rVert}_{\mathcal{F} \dot{H}^{\frac12}} = \ell_{v_0}$;
2. $(u^{(v_0)}(t),v^{(v_0)}(t))$ does not scatter forward in time.
The case $\ell_{v_0}^\dagger=\ell_{v_0}=\ell_0<\infty$ is excluded in the above theorem. We consider this exceptional case in Remark \[R:exception\], below.
Let us consider the second case of . In this case, the following strange thing occurs: Take $u_0 \in \mathcal{F}\dot{H}^{\frac12}$ with ${\lVert u_0\rVert}_{\mathcal{F}\dot{H}^{\frac12}}= \ell_{v_0}^\dagger$ and consider the corresponding solution $(u(t),v(t))$ with the data $(u_0,v_0)$. Then, in one hand, the solution $(u(t),v(t))$ scatters forward in time for any choice of such $u_0$ since ${\lVert u_0\rVert}_{\mathcal{F}\dot{H}^{\frac12}}<\ell_{v_0}$. However, on the other hand, for arbitrarily large number $N>0$, one can choose $u_0 \in \mathcal{F}\dot{H}^{\frac12}$ so that the corresponding solution $(u(t),v(t))$ satisfies $$\begin{aligned}
{\lVert (u,v)\rVert}_{W_1([0,\infty)) \times W_2([0,\infty))} \ge N.\end{aligned}$$ The next theorem tells us how this is “attained". Notice that the second case of occurs only when $\ell_0=\ell_{v_0}^\dagger <\infty$, thanks to Theorem \[T:l0\]. Consequently, there is a minimizer to $\ell_0$ in this case, by means of Theorem \[T:case0\].
\[T:case2\] Fix $v_0 \in \mathcal{F} \dot{H}^{\frac12}\setminus\{0\}$. Suppose that $\ell_{v_0}^\dagger < \ell_{v_0}$. Pick a sequence $\{ u_{0,n} \}_{n} \subset \mathcal{F} \dot{H}^{\frac12}$ satisfying ${\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^{\frac12}}< \ell_{v_0}^\dagger$ for all $n\ge1$, $$\begin{aligned}
\lim_{n\to\infty} {\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^{\frac12}} = \ell_{v_0}^\dagger,\end{aligned}$$ and $$\begin{aligned}
\lim_{n\to\infty} {\lVert (u_{n},v_{n})\rVert}_{W_1([0,\infty)) \times W_2([0,\infty))} = \infty,\end{aligned}$$ where $(u_n(t),v_n(t))$ is a solution with the initial data $(u_n(0),v_n(0))=(u_{0,n},v_0)$. Then, there exist a subsequence of $n$, a minimizer $(u^{(0)},v^{(0)})$ to $\ell_0$, and two sequences $\{\xi_n\}_n \subset \mathbb{R}^3$ and $\{h_n\}_n \subset 2^\mathbb{Z}$ such that $$\begin{aligned}
|\log h_n| + |\xi_n| \longrightarrow \infty\end{aligned}$$ and $$\begin{aligned}
e^{-ix\cdot h_n \xi_n }(u_{0,n})_{\{h_n\}} \longrightarrow u^{(0)}(0) \quad \text{in }\mathcal{F} \dot{H}^{\frac12}\end{aligned}$$ hold along the subsequence. In particular, along the same subsequence, it holds for any $\tau \in (0, T_{\max}(u^{(0)},v^{(0)}))$ that $$\begin{aligned}
(u_n(t),v_n(t))
&=\(e^{-it|\xi_n|^2+ix\cdot\xi_n}u^{(0)}_{[h_n^{-1}]}\(t,x-2\xi_nt\),e^{-2it|\xi_n|^2+2ix\cdot\xi_n}v^{(0)}_{[h_n^{-1}]}\(t,x-2\xi_nt\)\) \\
&\hspace{8.0cm} +(0,e^{\frac{1}{2}it\Delta}v_0) + o_{\dot{X}_\frac{1}{2}^{\frac{1}{2}}(t)\times \dot{X}_1^{\frac{1}{2}}(t)}(1)\end{aligned}$$ for $0 \le t \le \tau h_n^2$.
\[R:exception\] The special case $\ell_{v_0}^\dagger = \ell_0 = \ell_{v_0}<\infty$ ($v_0\neq 0$) is not included in the above two theorems. In this exceptional case, the conclusion of Theorem \[T:case1\] and/or Theorem \[T:case2\] holds. Namely, if there does not exist a minimizer to $\ell_{v_0}$ as in Theorem \[T:case1\], then the conclusion of Theorem \[T:case2\] is true.
Let us summarize the above results. Let $v_0 \in \mathcal{F} \dot{H}^{\frac12}$ be a given function. If $\ell_{v_0}^\dagger=\infty$ then we have $\ell_{v_0}=\infty$ (Theorem \[T:l0\]) and hence any solution satisfies $v(0)=v_0$ scatters forward in time (Proposition \[P:SSDS\]). On the other hand, if $\ell_{v_0}^\dagger<\infty$ and $v_0\neq0$ then we have either Theorem \[T:case1\] or Theorem \[T:case2\] according to the dichotomy . Remark that the first case in contains an exceptional case discussed in Remark \[R:exception\]. When $v_0=0$ then we do not have the dichotomy, we have $\ell_0=\ell^\dagger_0$ (Theorem \[T:l0\]). If $\ell_0$ is finite then there exists a minimizer (Theorem \[T:case0\]).
The question whether $\ell_0=\infty$ or not would be an interesting question to the system . We do not have the answer yet. Let us formulate the problem without using our terminology:
In , does $v_0=0$ implies scattering of the corresponding solution for any $u_0$?
If it were true, that is, if $\ell_0=\ell_0^\dagger=\infty$ then Theorem \[T:l0\] tells us that $\ell_{v_0}^\dagger=\ell_{v_0}$ is true for any $v_0$, as mentioned above.
Although we do not know the exact value of $\ell_{v_0}$, we are able to have a condition which implies the finiteness of $\ell_{v_0}$ and to give an upper bound for $\ell_{v_0}$. A simple one is a condition in terms of the energy.
\[Nonpositive energy\] Fix nontrivial $u_0, v_0 \in \mathcal{F} \dot{H}^{\frac12} \cap H^1$. If $E[u_0,v_0]\le 0$ and then the corresponding solution $(u,v)$ does not scatter. In particular, $\ell_{v_0} \le {\lVert u_0\rVert}_{\mathcal{F}\dot{H}^{\frac12}}$.
In our context, we want to find a condition which is stated in terms of $v_0$ only. We give two criteria in this direction. The first one is for large data case:
\[Large data nonscattering\] For any $v_0 \in \mathcal{F} \dot{H}^{\frac12} \cap H^1$ with $v_0\neq0$, there exists $c_0>0$ such that the estimate $\ell_{cv_0} \lesssim_{v_0} c^{\frac12}$ holds for any $c \ge c_0$.
The second one is criterion for a specific $v_0$:
\[Sufficient condition of finite ell\] Pick $v_0 \in \mathcal{F} \dot{H}^{\frac12} \cap H^1$. If there exists $\theta \in \mathbb{R}$ such that a Schrödinger operator $-\Delta - 2 \text{Re}(e^{i\theta} v_0)$ has a negative eigenvalue then $\ell_{v_0}<\infty$. Moreover, if $\varphi \in \mathcal{F}\dot{H}^\frac{1}{2}\cap H^1$ is a real-valued eigenfunction associated with a negative eigenvalue $\tilde{e}<0$ of $-\Delta - 2 \text{Re}(e^{i\theta} v_0)$ then the estimate $$\ell_{v_0} \le \frac{{\lVert \varphi\rVert}_{\mathcal{F}\dot{H}^{\frac12}}}{\sqrt{2|\tilde{e}|} {\lVert \varphi\rVert}_{L^2}} {\lVert \nabla v_0\rVert}_{L^2}$$ is true.
The estimate given in Corollary \[Sufficient condition of finite ell\] is scaling invariant. Indeed, if $\varphi(x)$ is an eigenfunction of $-\Delta - 2{\operatorname{Re}}(e^{i\theta} v_0 )$ associated with a negative eigenvalue $\tilde{e}$ then $\varphi_{\{\lambda\}}(x)$ is an eigenfunction of $-\Delta - 2{\operatorname{Re}}(e^{i\theta} (v_0)_{\{\lambda\}} )$ and the corresponding eigenvalue is $\lambda^2 \tilde{e}$, where $f_{\{\lambda\}}$ denotes the scaling of $f$ defined in .
It is possible to study the optimizing problem . Let us formulate in an abstract setting. Let $f(x,y)$ be a function on $[0,\infty)\times [0,\infty)$ such that it is continuous and strictly increasing with respect to the both variables and that $f(0,0)=0$. We define $$\label{E:definition of lf}
\ell_f := \inf \{ f({\lVert u_0\rVert}_{\mathcal{F} \dot{H}^{\frac12} },{\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12} }) \ :\ (u_0,v_0) \not \in \mathcal{S}_+ \}.$$ Then, we have the relation $$\ell_f = \inf_{v_0 \in \mathcal{F} \dot{H}^{\frac12}} f(\ell_{v_0}, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}})
=\inf_{v_0 \in \mathcal{F} \dot{H}^{\frac12}} f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}).$$ Moreover, there exists a minimizer, say $(u_{0,f}, v_{0,f})$, to $\ell_f$. The minimizer satisfies $\ell_{v_{0,f}}^\dagger = \ell_{v_{0,f}}$, and $u_{0,f}$ is a minimizer to $\ell_{v_{0,f}}$, i.e., $\ell_{v_{0,f}} = {\lVert u_{0,f}\rVert}_{\mathcal{F} \dot{H}^{\frac12} }$. (See Theorem \[T:ellfminimizer\] for the details.) Intuitively, this implies the following: Let us consider the level set $\Omega_r := \{(u_0,v_0): f({\lVert u_0\rVert}_{\mathcal{F} \dot{H}^{\frac12} },{\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12} }) \le r\}$. Then $\Omega_0 = \{(0,0)\}$ and, as the “radius" $r$ increases from zero, $\Omega_{r}$ first contains a non-scattering initial data exactly at it touches the curve $v_0 \mapsto \ell_{v_0}^\dagger$. And it touches the curve $v_0 \mapsto \ell_{v_0}$ at the same point. We would emphasize that it is true for any choice of $f$. It can be said that a function $v_0$ such that $\ell_{v_0}^\dagger < \ell_{v_0}$ is, if exists, never found as a minimizer to a minimizing problem of the type . Further, for any choice of $f$, the ground state solution is not a minimizer, as mentioned above.
The rest of the paper is organized as follows. In Section \[Preliminary\], we collect some notations and inequalities. Then, we define function spaces. In Section \[Local well-posedness and Stability\], we prove local well-posedness for and give a necessary and sufficient condition for scattering. Then, we check that the solutions to with nonpositive energy does not scatter (Theorem \[Nonpositive energy\]). In Section \[Properties of L and l\], we investigate properties of $L_{v_0}^\dagger$ and $\ell_{v_0}^\dagger$. In Section \[Sec:Linear profile decomposition\], we obtain linear profile decomposition theorem (Theorem \[Linear profile decomposition\]). In Section \[Proof of main theorem\], we prove Theorem \[T:l0\], Theorem \[T:case1\], and Theorem \[T:case2\] and consider the optimizing problem $\ell_f$. In Scetion \[Proof of corollaries\], we prove corollaries of Theorem \[Nonpositive energy\].
Preliminary {#Preliminary}
===========
In this section, we prepare some notations and estimates used throughout the paper.
Notations
---------
For non-negative $X$ and $Y$, we write $X\lesssim Y$ to denote $X\leq CY$ for a constant $C>0$. If $X\lesssim Y\lesssim X$, we write $X\sim Y$. The dependence of implicit constants on parameters will be indicated by subscripts when necessary, e.g. $X\lesssim_uY$ denotes $X\leq CY$ for some $C=C(u)$. We write $a'\in[1,\infty]$ to denote the Hölder conjugate to $a\in[1,\infty]$, that is, $\frac{1}{a}+\frac{1}{a'}=1$ holds. For $s\in\mathbb{R}$, the operator $|\nabla|^s$ is defined as the Fourier multiplier operator with multiplier $|\xi|^s$, that is, $|\nabla|^s=\mathcal{F}^{-1}|\xi|^s\mathcal{F}$. For a set $A \in \mathbb{R}^d$, ${\bf 1}_A(x)$ stands for the characteristic function of $A$.
We recall the standard Littlewood-Paley projection operators. Let $\phi$ be a radial cut-off function satisfies ${\bf 1}_{\{ |\xi| \le 4/3\}} \le \phi \le {\bf 1}_{\{ |\xi| \le 5/3\}}$. For $N\in2^\mathbb{Z}$, the operators $P_N$ is defined as $$ \widehat{P_Nf}(\xi):=\widehat{f_N}(\xi):=\psi_N(\xi)\widehat{f}(\xi),
$$ where $\phi_N(x) = \phi(x/N)$ and $$\begin{aligned}
\psi_N(x)=\phi_N(x)-\phi_{N/2}(x). \label{069}\end{aligned}$$
The Galilean transform and the Galilean operator
------------------------------------------------
The identities $$\begin{split}
[e^{it\Delta}(e^{ix\cdot\xi_0}f)](x)
&=e^{-it|\xi_0|^2+ix\cdot\xi_0}(e^{it\Delta}f)(x-2t\xi_0),\\[0.1cm]
[e^{\frac{1}{2}it\Delta}(e^{2ix\cdot\xi_0}f)](x)
&=e^{-2it|\xi_0|^2+2ix\cdot\xi_0}(e^{\frac{1}{2}it\Delta}f)(x-2t\xi_0) \label{008}
\end{split}$$ imply that the class of solutions to the linear Schrödinger equation is invariant under Galilean transform: $$\begin{aligned}
(u(t,x),v(t,x))\mapsto(e^{-it|\xi_0|^2+ix\cdot\xi_0}u(t,x-2\xi_0t),e^{-2it|\xi_0|^2+2ix\cdot\xi_0}v(t,x-2\xi_0t)),\ \ \ \xi_0\in\mathbb{R}^3.\label{009}\end{aligned}$$ The invariance is inherited in the nonlinear equation . The Galilean operator $$J_m(t):=x+i\frac{t}{m}\nabla,$$ which is a multiple of the infinitesimal operator for transforms appearing in , plays an important role in the scattering theory for mass-subcritical nonlinear Schrödinger equation.
We define the multiplication operator $$\begin{aligned}
[\mathcal{M}_m(t)f](x):=e^\frac{im|x|^2}{2t}f(x)\ \ (t\neq0) \label{071}\end{aligned}$$ and the dilation operator $$\begin{aligned}
[\mathcal{D}(t)f](x):=(2it)^{-\frac{3}{2}}f\left(\frac{x}{2t}\right)\ \ (t\neq0). \label{072}\end{aligned}$$ It is well known that the Schrödinger group is factorized as $e^{it\Delta}=\mathcal{M}_\frac{1}{2}(t)\mathcal{D}(t)\mathcal{F}\mathcal{M}_\frac{1}{2}(t)$ by using these operators. This factorization deduces the identity $$\begin{aligned}
e^{it\Delta}\Phi(x)e^{-it\Delta}
=\mathcal{M}_\frac{1}{2}(t)\Phi(2it\nabla)\mathcal{M}_\frac{1}{2}(-t) \label{007}\end{aligned}$$ for suitable multiplier $\Phi$, where $\Phi(i\nabla)$ denotes the Fourier multiplier operator with multiplier $\Phi(\xi)$, that is, $\Phi(i\nabla):=\mathcal{F}^{-1}\Phi(\xi)\mathcal{F}$. The Galilean operator is written as follows: $$\begin{aligned}
J_m(t)
=e^{\frac{1}{2m}it\Delta}xe^{-\frac{1}{2m}it\Delta}
=\mathcal{M}_m(t)i\frac{t}{m}\nabla \mathcal{M}_m(-t),\end{aligned}$$ where the second equality holds for $t\neq0$. We define a fractional power of $J_m$ by $$\begin{aligned}
J_m^s(t):=e^{\frac{1}{2m}it\Delta}|x|^se^{-\frac{1}{2m}it\Delta}=\mathcal{M}_m(t)\left(-\frac{t^2}{m^2}\Delta\right)^\frac{s}{2}\mathcal{M}_m(-t)\ \ \text{ for }\ \ s\in\mathbb{R}. \label{070}\end{aligned}$$ Remark that the second formula is valid for $t\neq0$.
Function spaces {#subsec:Function spaces}
---------------
We define a time-dependent spaces $\dot{X}_m^{s,r}=\dot{X}_m^{s,r}(t)$ by using the norm $$\begin{aligned}
\|f\|_{\dot{X}_m^{s,r}}:=\|J_m^s(t)f\|_{L_x^r(\mathbb{R}^3)}\sim\||t|^s|\nabla|^s\mathcal{M}_m(-t)f\|_{L_x^r(\mathbb{R}^3)}. \label{010}\end{aligned}$$ When $r=2$, we omit it, that is, $\dot{X}_m^s=\dot{X}_m^{s,2}$. We can see immediately by the definition of $J_m^s$ that the equivalence of norms in for $t\neq0$. It is natural to write $$\begin{aligned}
f\in e^{\frac{1}{2m}it \Delta}\mathcal{F}\dot{H}^s\Longleftrightarrow e^{-\frac{1}{2m}it \Delta}f\in\mathcal{F}\dot{H}^s.\end{aligned}$$ Then, we have a change of notation:$e^{\frac{1}{2m}it\Delta}\mathcal{F}\dot{H}^s=\dot{X}_m^s(t)$ by using this description.
We use Lorentz-modified space-time norms. For an interval $I$, $1\leq q<\infty$, and $1\leq\alpha\leq\infty$, the Lorentz space $L_t^{q,\alpha}(I)$ is defined by using the quasi-norm $$\begin{aligned}
\|f\|_{L_t^{q,\alpha}(I)}
:=\|\lambda|\{t\in I:|f(t)|>\lambda\}|^\frac{1}{q}\|_{L^\alpha((0,\infty),\frac{d\lambda}{\lambda})}.\end{aligned}$$ For a Banach space $X$, $L_t^{q,\alpha}(I;X)$ is defined as the whole of functions $u:I\times\mathbb{R}^3\longrightarrow\mathbb{C}$ satisfying $$\begin{aligned}
\|u\|_{L_t^{q,\alpha}(I;X)}:=\left\|\|u(t)\|_{X}\right\|_{L_t^{q,\alpha}(I)}<\infty.\end{aligned}$$ The following equivalence is useful: $$\begin{aligned}
\|f\|_{L_t^{q,\alpha}(I)}\sim\|\|f\cdot1_{\{2^{k-1}\leq|f|\leq2^k\}}\|_{L_t^q(I)}\|_{\ell^\alpha(k\in\mathbb{Z})}.\end{aligned}$$
Strichartz estimates
--------------------
The standard Strichartz estimates for $e^{it\Delta}$ were proved in [@GinVel92; @KeeTao98; @Str77]. We also need Strichartz estimates for the spaces $L_t^{q,\alpha}\dot{X}_m^{s,r}$, which were proved in [@Mas16; @NakOza02].
If a pair $(q,r)$ satisfies $$\begin{aligned}
2<q<\infty,\ \ \ 2<r<6,\ \ \text{ and }\ \ \frac{2}{q}+\frac{3}{r}=\frac{3}{2},\end{aligned}$$ then $(q,r)$ is an admissible pair.
Remark that we do not included two end points $(\infty,2)$ and $(2,6)$ to admissible pairs. It is because they require exceptional treatments sometimes.
\[Strichartz estimates\] Let $s\geq0$ and $t_0\in I\subset\mathbb{R}$.
- For any admissible pair $(q,r)$, we have $$\begin{aligned}
\|e^{it\Delta}f\|_{L_t^\infty\dot{X}_m^s\cap L_t^{q,2}\dot{X}_m^{s,r}}\lesssim\|f\|_{\mathcal{F}\dot{H}^s}.\end{aligned}$$
- For any admissible pairs $(q,r)$ and $(\alpha,\beta)$, we have $$\begin{aligned}
\left\|\int_{t_0}^te^{i(t-s)\Delta}F(s)ds\right\|_{L_t^\infty(I;\dot{X}_m^s)\cap L_t^{q,2}(I;\dot{X}_m^{s,r})}\lesssim\|F\|_{L_t^{\alpha',2}(I;\dot{X}_m^{s,\beta'})}.\end{aligned}$$
Specific function spaces {#ss:space}
------------------------
Throughout this paper, we use the following concrete choice of function spaces. The same exponents were used in [@KilMasMurVis17; @Mas15].
We define $$\begin{aligned}
\left(\frac{1}{q_1},\frac{1}{r_1}\right)
:=\left(\frac{1}{6},\frac{7}{18}\right),\ \ \
\left(\frac{1}{\widetilde{q}},\frac{1}{\widetilde{r}}\right)
:=\left(\frac{2}{3},\frac{2}{9}\right).\end{aligned}$$ The pair $(q_1,r_1)$ is admissible. The pair $(\tilde{q},\tilde{r})$ satisfies the critical scaling relation $\frac{2}{\widetilde{q}}+\frac{3}{\widetilde{r}}=2$, and is not a admissible pair. These exponents satisfy the following relations. $$\begin{aligned}
\frac{1}{q_1'}=\frac{1}{\widetilde{q}}+\frac{1}{q_1},\ \ \ \frac{1}{r_1'}=\frac{1}{\widetilde{r}}+\frac{1}{r_1},\ \ \ \frac{1}{\widetilde{q}}-\frac{1}{q_1}=\frac{3}{r_1}-\frac{3}{\widetilde{r}}=\frac{1}{2}.\label{011}\end{aligned}$$ We define the spaces $$\begin{aligned}
S^\text{weak}
:=L_t^{\widetilde{q},\infty}L_x^{\widetilde{r}}=L_t^{\frac{3}{2},\infty}L_x^\frac{9}{2},\ \ \
S
:=L_t^{\widetilde{q},2}L_x^{\widetilde{r}}=L_t^{\frac{3}{2},2}L_x^\frac{9}{2},\ \ \
W_j
:=L_t^{q_1,2}\dot{X}_{2^{{\,}^{j-2}}}^{\frac{1}{2},r_1}=L_t^{6,2}\dot{X}_{2^{{\,}^{j-2}}}^{\frac{1}{2},\frac{18}{7}}\end{aligned}$$ for the solutions and the spaces $$\begin{aligned}
N_j
:=L_t^{q_1',2}\dot{X}_{2^{{\,}^{j-2}}}^{\frac{1}{2},r_1'}
=L_t^{\frac{6}{5},2}\dot{X}_{2^{{\,}^{j-2}}}^{\frac{1}{2},\frac{18}{11}}\end{aligned}$$ for the nonlinear terms. We use a notation $S^\text{weak}(I)$ to indicate that the norm is taken over the space-time slab $I\times\mathbb{R}^3$, and similarly for the other spaces.
Some estimates
--------------
In this section, we collect some estimates. They easily follow as in [@Mas16] (see also [@KilMasMurVis17]).
\[Embeddings\] The following inequalities hold. $$\begin{aligned}
\|u\|_{S^\text{weak}}\lesssim\|u\|_{S}\lesssim_j \|u\|_{W_j},$$ where $j=1,2$.
\[Nonlinear estimates\] The following inequalities hold. $$\begin{aligned}
\|v\overline{u}\|_{N_1}\lesssim\|u\|_{S^\text{weak}}\|v\|_{W_2}+\|u\|_{W_1}\|v\|_{S^\text{weak}}\lesssim\|u\|_{W_1}\|v\|_{W_2},\end{aligned}$$ $$\begin{aligned}
\|u_1u_2\|_{N_2}\lesssim\|u_1\|_{W_1}\|u_2\|_{S^\text{weak}}+\|u_1\|_{S^\text{weak}}\|u_2\|_{W_1}\lesssim\|u_1\|_{W_1}\|u_2\|_{W_1},\end{aligned}$$ $$\begin{gathered}
\|v\overline{u}\|_{L_t^{\frac{3}{2},2}\dot{X}_\frac{1}{2}^{\frac{1}{2},\frac{18}{13}}}\lesssim\|v\|_{L_t^\infty\dot{X}_1^\frac{1}{2}}\|u\|_{S}+\|v\|_{S}\|u\|_{L_t^\infty\dot{X}_\frac{1}{2}^\frac{1}{2}},\\
\|u_1u_2\|_{L_t^{\frac{3}{2},2}\dot{X}_1^{\frac{1}{2},\frac{18}{13}}}\lesssim\|u_1\|_{L_t^{\infty}\dot{X}_\frac{1}{2}^\frac{1}{2}}\|u_2\|_{S}+\|u_1\|_{S}\|u_2\|_{L_t^{\infty}\dot{X}_\frac{1}{2}^\frac{1}{2}},\end{gathered}$$ $$\begin{aligned}
\left\| \int_{\tau}^t e^{i(t-s)\Delta}(v\overline{u})(s) ds\right\|_{S}\lesssim
\|u\|_{S}\|v\|_{S},\end{aligned}$$ $$\begin{aligned}
\left\| \int_{\tau}^t e^{\frac{1}2i(t-s)\Delta}(u_1 u_2 )(s) ds\right\|_{S}\lesssim
\|u_1\|_{S}\|u_2\|_{S}.\end{aligned}$$
Remark that the last two are consequences of inhomogeneous Strichartz estimate for non-admissible pairs by Kato [@Kat94].
\[Interpolation in X\] Let $I\subset\mathbb{R}$. The following inequality holds. $$\begin{aligned}
\|f\|_{L^{\rho,\gamma}(I;\dot{X}_m^{s,r})}\lesssim\|f\|_{L^{\rho_1,\gamma_1}(I;\dot{X}_m^{s_1,r_1})}^{1-\theta}\|f\|_{L^{\rho_2,\gamma_2}(I;\dot{X}_m^{s_2,r_2})}^\theta\end{aligned}$$ for $1\leq\rho, \rho_1, \rho_2<\infty$, $1\leq\gamma, \gamma_1, \gamma_2\leq\infty$, $1<r, r_1, r_2<\infty$, $0<\theta<1$ with $\frac{1}{\rho}=\frac{1-\theta}{\rho_1}+\frac{\theta}{\rho_2}$, $\frac{1}{\gamma}=\frac{1-\theta}{\gamma_1}+\frac{\theta}{\gamma_2}$, $s=(1-\theta)s_1+\theta s_2$, and $\frac{1}{r}=\frac{1-\theta}{r_1}+\frac{\theta}{r_2}$.
\[Bounded multiplication operator\] Let $1\leq r<\infty$, $0<s<\frac{3}{r}$ and let $\chi\in \mathcal{S}(\mathbb{R}^3)$, where $\mathcal{S}(\mathbb{R}^3)$ denotes Schwartz space. Then, a multiplication operator $\chi\times$ is bounded on $\dot{X}_{m}^{s,r}$.
\[Square function estimate\] For $0\leq s\leq 2$ and $1<p<\infty$, we have $$\begin{aligned}
\||\nabla|^sf\|_{L_x^p}\sim\Bigl\|\Bigl(\sum_{N\in2^\mathbb{Z}}\left|P_N|\nabla|^sf\right|^2\Bigr)^\frac{1}{2}\Bigr\|_{L_x^p}.\end{aligned}$$
The following Hölder’s inequality for Lorentz spaces holds.
\[Holder in Lorentz spaces\] Let $1\leq q,q_1,q_2<\infty$ and $1\leq\alpha,\alpha_1,\alpha_2\leq\infty$ satisfy $$\begin{aligned}
\frac{1}{q}=\frac{1}{q_1}+\frac{1}{q_2}\ \ \text{ and }\ \ \frac{1}{\alpha}=\frac{1}{\alpha_1}+\frac{1}{\alpha_2}.\end{aligned}$$ Then, the following estimate holds. $$\begin{aligned}
\|fg\|_{L_t^{q,\alpha}}\lesssim\|f\|_{L_t^{q_1,\alpha_1}}\|g\|_{L_t^{q_2,\alpha_2}}.\end{aligned}$$
Local well-posedness and Stability {#Local well-posedness and Stability}
==================================
Local well-posedness
--------------------
In this subsection, we establish a local theory in $(C_t\dot{X}_{1/2}^{1/2}\cap W_1)\!\times\!(C_t\dot{X}_1^{1/2}\cap W_2)$ for . The result is given as a consequence of Strichartz estimate (Proposition \[Strichartz estimates\]) and the estimates of the previous subsection (Lemma \[Embeddings\] and Lemma \[Nonlinear estimates\]).
Let us first establish a weak version of the local well-posedness.
\[P:Slwp\] There exists a universal constant $\delta>0$ with the following property: Let $\tau \in \mathbb{R}$. If a pair of functions $(u_\tau,v_\tau) \in \mathcal{S}'(\mathbb{R}^3)^2$ satisfies $${\lVert (e^{i(t-\tau)\Delta}u_\tau,e^{\frac{1}2i(t-\tau)\Delta}v_\tau) \rVert}_{S(I) \times S(I)} \le \delta$$ for some interval $I \ni \tau$ then there exists a unique pair of functions $(u,v) \in S(I) \times S(I)$ such that $$\begin{cases}
\hspace{-0.4cm}&\displaystyle{u(t)=e^{i(t-\tau)\Delta}u_\tau+2i\int_{\tau}^te^{i(t-s)\Delta}(v\overline{u})(s)ds},\\
\hspace{-0.4cm}&\displaystyle{v(t)=e^{\frac{1}{2}i(t-\tau)\Delta}v_\tau+i\int_{\tau}^te^{\frac{1}{2}i(t-s)\Delta}(u^2)(s)ds}
\end{cases}$$ holds in $S(I)\times S(I)$ sense and satisfies $${\lVert (u,v)\rVert}_{S(I)\times S(I)} \le 2 {\lVert (e^{i(t-\tau)\Delta}u_\tau,e^{\frac{1}2i(t-\tau)\Delta}v_\tau) \rVert}_{S(I) \times S(I)}.$$
This follows by a standard contraction mapping argument with the last two estimates of Lemma \[Nonlinear estimates\]. We refer a pair of functions $(u,v)$ in this proposition to as an $S$-solution to on $I$.
Now we are able to establish the following version of the local well-posedness result. Theorem \[Th. Local well-posedness\] corresponds to the case $t_0=0$.
\[Th. Local well-posedness2\] For any $t_0\in\mathbb{R}$ and $(u_0,v_0)\in\dot{X}_{1/2}^{1/2}(t_0)\times\dot{X}_1^{1/2}(t_0)$ there exist an open interval $I \ni t_0$ and a unique solution $(u,v)\in(C_t(I;\dot{X}_{1/2}^{1/2})\cap W_1(I))\!\times\!(C_t(I;\dot{X}_1^{1/2})\cap W_2(I))$ to with a initial condition $(u(t_0),v(t_0))=(u_0,v_0)$. Moreover, there exists a universal constant $\delta>0$ such that if the data satisfies $$\begin{aligned}
\|(e^{i(t-t_0)\Delta}u_0,e^{\frac{1}{2}i(t-t_0)\Delta}v_0)\|_{W_1(I)\times W_2(I)}\leq\delta,\end{aligned}$$ then the solution satisfies $$\begin{aligned}
\|(u,v)\|_{W_1(I)\times W_2(I)}\lesssim\|(e^{i(t-t_0)\Delta}u_0,e^{\frac{1}{2}i(t-t_0)\Delta}v_0)\|_{W_1(I)\times W_2(I)}.\end{aligned}$$
The strategy of the proof is as follows: We first obtain a $S$-solution. Then, we show it is a solution in the sense of Definition \[D:sol\] by a persistence-of-regularity type argument.
By Lemma \[Embeddings\] and Proposition \[Strichartz estimates\], we have $$\begin{aligned}
{\lVert (e^{i(t-t_0)\Delta}u_0,e^{\frac{1}2i(t-t_0)\Delta}v_0)\rVert}_{S(\mathbb{R})\times S(\mathbb{R})}
&{}\lesssim {\lVert (e^{i(t-t_0)\Delta}u_0,e^{\frac{1}2i(t-t_0)\Delta}v_0)\rVert}_{W_1(\mathbb{R}) \times W_2(\mathbb{R})}\\
&{}\lesssim {\lVert (e^{-it_0\Delta}u_0,e^{-\frac{1}2it_0\Delta}v_0)\rVert}_{\mathcal{F}\dot{H}^\frac12 \times \mathcal{F}\dot{H}^\frac12}<\infty.\end{aligned}$$ Hence, we can chose an open interval $I\ni t_0$ so that $${\lVert (e^{i(t-t_0)\Delta}u_0,e^{\frac{1}2i(t-t_0)\Delta}v_0)\rVert}_{S(I)\times S(I)} \le \delta.$$ For this interval, we have a unique $S$-solution $(u,v) \in S(I)\times S(I)$ by Proposition \[P:Slwp\].
We shall show this is a solution in the sense of Definition \[D:sol\]. By Proposition \[Strichartz estimates\] and Lemma \[Nonlinear estimates\], one has $$\begin{aligned}
\|(u,v)\|_{W_1(I)\times W_2(I)}
&\leq \bigl\|\bigl(e^{i(t-t_0)\Delta}u_0,e^{\frac{1}{2}i(t-t_0)\Delta}v_0\bigr)\bigr\|_{W_1(I)\times W_2(I)}+c\|v\overline{u}\|_{N_1(I)}+c\|u^2\|_{N_2(I)}\\
&\leq \bigl\|\bigl(e^{i(t-t_0)\Delta}u_0,e^{\frac{1}{2}i(t-t_0)\Delta}v_0\bigr)\bigr\|_{W_1(I)\times W_2(I)}+c\|(u,v)\|_{W_1(I)\times W_2(I)}\|(u,v)\|_{S(I)\times S(I)}.\end{aligned}$$ By subdividing the interval $I$ into $\cup_{j=0}^J I_j$ so that we have $c\|(u,v)\|_{S(I_j)\times S(I_j)} \le \frac12$ in each interval. Suppose $t_0 \in I_0$. We have $$\|(u,v)\|_{W_1(I_0)\times W_2(I_0)} \leq 2\bigl\|\bigl(e^{i(t-t_0)\Delta}u_0,e^{\frac{1}{2}i(t-t_0)\Delta}v_0\bigr)\bigr\|_{W_1(I_0)\times W_2(I_0)}.$$ Another use of Strichartz’ estimate then shows $$\|(u,v)\|_{L_t^\infty(I_0; \dot{X}_{1/2}^{1/2})\times L_t^\infty(I_0;\dot{X}_1^{1/2})}
\lesssim \|(u_0,v_0)\|_{\dot{X}_{1/2}^{1/2}(t_0)\times \dot{X}_1^{1/2}(t_0)}.$$ Repeat the argument to obtain $(u,v) \in (L_t^\infty(I; \dot{X}^{1/2}_{1/2}) \cap W_1(I))\times (L_t^\infty(I;\dot{X}_1^{1/2})\cap W_2(I))$. The latter statement is obvious. We omit the details.
Scattering criterion
--------------------
In this subsection, we derive a necessary and sufficient condition for scattering (Proposition \[Scattering criterion\]). We also give a scattering result for small data (Proposition \[Small data scattering\]).
\[Scattering criterion\] Let $(u(t),v(t))$ be a unique solution to given in Theorem \[Th. Local well-posedness\]. Let $I_{\max}=(T_\text{min},T_{\max})$ be the maximal interval of $(u(t),v(t))$. Then, the following seven statements are equivalent.
- $(u,v)$ scatters forward in time;
- $\|(u,v)\|_{W_1([\tau,T_{\max}))\times W_2([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$;
- $\|(u,v)\|_{S([\tau,T_{\max}))\times S([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$;
- $\|u\|_{W_1([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$;
- $\|v\|_{W_2([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$;
- $\|u\|_{S([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$;
- $\|v\|_{S([\tau,T_{\max}))}<\infty$, $\exists \tau \in I_{\max}$.
Similar criterion holds for the backward-in-time scattering.
$(1)\Leftrightarrow(2)\Leftrightarrow(3)$ is standard (see, for instance, [@Mas16]). Let us prove they are also equivalent to from (4) to (7). To this end, it suffices to show that $(6)$ and $(7)$ are equivalent. It is because (3) is equivalent to “$(6)$ and $(7)$." Further, once the above equivalence is established, the rest $(4)$ and $(5)$ are handled easily: We have $(4)\Rightarrow(6)\Leftrightarrow(2)\Rightarrow(4)$ and $(5)\Rightarrow(7)\Leftrightarrow(2)\Rightarrow(5)$.
Suppose (6). Then, for any $T \in (0,T_{\max})$, one deduces from Proposition \[Strichartz estimates\] and Lemma \[Nonlinear estimates\] that $${\lVert v\rVert}_{S([0,T))} \lesssim {\lVert v_0\rVert}_{\mathcal{F}\dot{H}^\frac{1}{2}} + {\lVert u\rVert}_{S([0,T))}^2.$$ Here the implicit constant is independent of $T$. Hence, by letting $T \uparrow T_{\max}$ we obtain (7).
Next, suppose (7). Take $\tau \in (T_{\min},T_{\max})$ to be chosen later. For any $T \in (\tau,T_{\max})$, we see $${\lVert u\rVert}_{S((\tau,T))} \leq C {\lVert u(\tau)\rVert}_{\dot{X}^{1/2}_{1/2}(\tau)} + C{\lVert u\rVert}_{S((\tau,T))} {\lVert v\rVert}_{S((\tau,T))},$$ where the constant $C$ is independent of $\tau$ and $T$. We now choose $\tau$ so that $C {\lVert v\rVert}_{S((\tau,T_{\max}))}\le \frac{1}{2}$. This is possible by the property (7). Then, the above inequality implies that $${\lVert u\rVert}_{S((\tau,T))} \leq 2C {\lVert u(\tau)\rVert}_{\dot{X}^{1/2}_{1/2}(\tau)}.$$ Since $T \in (\tau,T_{\max})$ is arbitrary, we obtain the result.
We turn to a sufficient condition for scattering. One of the simplest is by the smallness of the data.
\[Small data scattering\] Let $(u_0,v_0)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$ and let $(u,v)$ be a corresponding unique solution given in Theorem \[Th. Local well-posedness\]. Then, we have the following.
- There exists $\eta_1>0$ such that if $\|(e^{it\Delta}u_0,e^{\frac{1}{2}it\Delta}v_0)\|_{S\times S}\leq\eta_1$, then $(u,v)$ scatters.
- There exists $\eta_2>0$ such that if $\|(e^{it\Delta}u_0,e^{\frac{1}{2}it\Delta}v_0)\|_{W_1\times W_2}\leq\eta_2$, then $(u,v)$ scatters.
- There exists $\eta_3>0$ such that if $\|(u_0,v_0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}}\leq\eta_3$, then $(u,v)$ scatters.
This follows from Proposition \[P:Slwp\], Proposition \[Scattering criterion\], and Proposition \[Strichartz estimates\].
Nonpositive energy implies failure of scattering {#subsec:Nonpositive energy}
------------------------------------------------
In this subsection, we give a proof of Theorem \[Nonpositive energy\]. To begin with, we will prove that if a data belongs $H^1 \times H^1$, in addition, then the corresponding solution given in Theorem \[Th. Local well-posedness\] stays in $H^1 \times H^1$ and the mass and the energy make sense and are conserved. Furthermore, as is well-known, since our equation is mass-subcritical, the conservation of mass implies the solution is global.
\[LWP H1\] For any $t_0\in\mathbb{R}$ and $(u_0,v_0)\in (\dot{X}_{1/2}^{1/2}(t_0) \cap H^1 )\times (\dot{X}_1^{1/2}(t_0)\cap H^1)$ there exists a unique time global solution $(u,v)\in(C_t(\mathbb{R};\dot{X}_{1/2}^{1/2}\cap H^1)\cap W_{1,\text{loc}}(\mathbb{R}))\!\times\!(C_t(\mathbb{R};\dot{X}_1^{1/2}\cap H^1)\cap W_{2,\text{loc}}(\mathbb{R}))$ to with the initial condition $(u(t_0),v(t_0))=(u_0,v_0)$. The solution have conserved mass and energy; $$M[u(t),v(t)] = M[u_0,v_0],\quad E[u(t),v(t)] = E[u_0,v_0].$$ Furthermore, if the solution scatters forward in time then holds also in $H^1 \times H^1$ sense.
This is done by a persistence-of-regularity argument. Now, we prove Theorem \[Nonpositive energy\].
Suppose that a solution $(u,v)$ given in Proposition \[LWP H1\] scatters forward in time. Then, the limit holds also in $H^1 \times H^1$ sense. Moreover, one sees from Proposition \[Scattering criterion\] that ${\lVert (u,v)\rVert}_{S([0,\infty))\times S([0,\infty))}<\infty$. Hence, one finds a sequence $\{t_n\}\subset [0,\infty)$, $t_n\to\infty$ as $n\to\infty$, such that $${\lVert u(t_n)\rVert}_{L^{\frac92}_x} + {\lVert v(t_n)\rVert}_{L^{\frac92}_x} \to 0$$ as $n\to\infty$. Combining with the mass conservation, the above $L^\frac{9}{2}_x$ can be replaced by any $L^p_x$ for $2<p\le \frac{9}{2}$. In particular, $p=3$ is allowed. Hence, $$\left| 2 \int_{\mathbb{R}^3} {\operatorname{Re}}(u(t_n)^2\overline{v(t_n)}) dx \right| \le 2{\lVert u(t_n)\rVert}_{L^{3}_x}^2 {\lVert v(t_n)\rVert}_{L^{3}_x} \longrightarrow 0$$ as $n\to\infty$. We deduce that $$E[u_0,v_0] = \lim_{n\to\infty} E[u(t_n),v(t_n)] = {\lVert \nabla u_+\rVert}_{L_x^2}^2 + \frac12 {\lVert \nabla v_+\rVert}_{L_x^2}^2 \ge 0.$$ Further, $E[u_0,v_0]=0$ implies $(u_+,v_+)=(0,0)$. By in $H^1$ and the mass conservation implies $(u_0,v_0)=(0,0)$
Stability
---------
In this subsection, we establish a stability result. Roughly speaking, the proposition implies that two solutions are also close each other if their initial data are close and the equations for them are close.
\[Long time perturbation\] Let $I$ be a time interval with $t_0\in I$ and $M>0$. Let $(\widetilde{u},\widetilde{v}):I\times\mathbb{R}^3\longrightarrow\mathbb{C}^2$ satisfy $$\notag
\begin{cases}
\hspace{-0.4cm}&\displaystyle{i\partial_t\widetilde{u}+\Delta\widetilde{u}=-2\widetilde{v}\overline{\widetilde{u}}+e_1},\\
\hspace{-0.4cm}&\displaystyle{i\partial_t\widetilde{v}+\frac{1}{2}\Delta\widetilde{v}=-\widetilde{u}^2+e_2}
\end{cases}$$ and $$\begin{aligned}
\|(\widetilde{u},\widetilde{v})\|_{W_1(I)\times W_2(I)}\leq M\end{aligned}$$ for some functions $e_1$, $e_2$. Let $(u_0,v_0)\in\dot{X}^{1/2}_{1/2}(t_0)\times\dot{X}^{1/2}_1(t_0)$ and let $(u,v)$ be a corresponding solution to with $(u(t_0),v(t_0))=(u_0,v_0)$ given by Theorem \[Th. Local well-posedness\]. There exist $\varepsilon_1=\varepsilon_1(M)>0$ and $c=c(M)>0$ such that if $$\begin{aligned}
\|(\widetilde{u}(t_0),\widetilde{v}(t_0))-(u_0,v_0)\|_{\dot{X}_{1/2}^{1/2}(t_0)\times\dot{X}_1^{1/2}(t_0)}+\|(e_1,e_2)\|_{N_1(I)\times N_2(I)}\le \varepsilon\end{aligned}$$ for some $0\le \varepsilon<\varepsilon_1$ then the maximal existence interval of $(u,v)$ contains $I$ and the solution satisfies $$\begin{aligned}
\|(u,v)-(\widetilde{u},\widetilde{v})\|_{(L_t^\infty(I;\dot{X}_{1/2}^{1/2})\cap W_1(I))\times (L_t^\infty(I;\dot{X}_1^{1/2})\cap W_2(I))}
\leq c\varepsilon.\end{aligned}$$
For the proof, see [@Mas15].
Properties of $L_{v_0}$ and $\ell_{v_0}^\dagger$ {#Properties of L and l}
================================================
In this section, we collect properties of $L_{v_0}$ and $\ell_{v_0}^\dagger$.
\[Control of L\] For any $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$, there exist $\varepsilon_1>0$ and $\delta>0$ such that $$\begin{aligned}
L_{v_0}(\varepsilon)\leq\|e^{\frac{1}{2}it\Delta}v_0\|_{W_2([0,\infty))}+c\delta\varepsilon\end{aligned}$$ holds for $0\le \varepsilon<\varepsilon_1$. Here, the constants $\varepsilon_1>0$ and $\delta>0$ depend only on $\|e^{\frac{1}{2}it\Delta}v_0\|_{W_2([0,\infty))}$. In particular, $\ell^{\dagger}_{v_0}>0$ for any $v_0 \in \mathcal{F} \dot{H}^{1/2}.$
Apply Proposition \[Long time perturbation\] with $(\widetilde{u},\widetilde{v})=(0,e^{\frac{1}{2}it\Delta}v_0)$, for which $e_1=e_2=0$.
In the sequel, we call a function which values in an extended real numbers $\mathbb{R} \cup \{+\infty\}$ as an extended function.
\[D:continuity\] We say that a non-decreasing extended function $f:\mathbb{R}\longrightarrow[0,\infty]$ is continuous at $x_0\in\mathbb{R}$ if $f$ satisfies the following:
- When $f(x_0)<\infty$, $f$ is continuous in the usual sense.
- When $f(x_0)=\infty$ and $0<x_0<\infty$, either “there exists $\varepsilon>0$ such that $f(x)=\infty$ for any $x\in(x_0-\varepsilon,x_0]$” or “$f(x)<\infty$ for any $x\in[0,x_0)$ and $\displaystyle\lim_{x\rightarrow x_0-0}f(x)=\infty$” is true.
Moreover, we say a non-decreasing extended function $f$ is continuous if $f(x)$ is continuous at any $x_0\in\mathbb{R}$ in the above sense. Furthermore, we say a non-decreasing extended function $f$ defined on $[0,\infty)$ is continuous if there exists a non-decreasing extended function $\tilde{f}$ (defined on $\mathbb{R}$) such that $f(x)=\tilde{f}(x)$ for all $x\ge0$.
\[Continuity of L\] For each fixed $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$, the function $L_{v_0}$ is a non-decreasing continuous extended function defined on $[0,\infty)$.
It is clear that $L_{v_0}$ is a non-decreasing extended function defined on $[0,\infty)$.
We prove the continuity in the sense of Definition \[D:continuity\]. It is obvious that $$L_{v_0}(0) = \|e^{\frac{1}{2}it\Delta}v_0\|_{W_2([0,\infty))}<\infty.$$ The continuity of $L_{v_0}(\ell)$ at $\ell=0$ holds by Proposition \[Control of L\].
Fix $\ell_0\in(0,\infty)$ such that $L_{v_0}(\ell_0)<\infty$. Let us prove right continuity of $L_{v_0}(\ell)$ at $\ell=\ell_0$. Pick $\varepsilon>0$. Take $\delta>0$ so that $\delta<\varepsilon_1$ and $c\delta<\varepsilon$, where $\varepsilon_1=\varepsilon_1(L_{v_0}(\ell_0))$ and $c=c(L_{v_0}(\ell_0))$ are the constants given in Proposition \[Long time perturbation\] with the choice $M=L_{v_0}(\ell_0)$. Fix $\ell\in(\ell_0,\ell_0+\delta)$. Then, for any $u_{0,1}\in\mathcal{F}\dot{H}^\frac{1}{2}$ satisfying $\|u_{0,1}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell$, the function $$u_{0,2}=\frac{\ell_0}{\ell_0+\delta} u_{0,1}$$ satisfies $\|u_{0,2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell_0$ and $\|u_{0,1}-u_{0,2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\delta$. Let $(u_1,v_1)$ and $(u_2,v_2)$ be two solutions to with initial data $(u_{0,1},v_0)$ and $(u_{0,2},v_0)$, respectively. Note that $$\begin{aligned}
\|(u_2,v_2)\|_{W_1([0,\infty))\times W_2([0,\infty))}\leq L_{v_0}(\ell_0)\end{aligned}$$ since $\|u_{0,2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell_0$. Hence, we have $$\begin{aligned}
\|(u_1,v_1)-(u_2,v_2)\|_{W_1([0,\infty))\times W_2([0,\infty))}\leq c\delta<\varepsilon\end{aligned}$$ by Proposition \[Long time perturbation\]. Thus, it follows that $$\begin{aligned}
\|(u_1,v_1)\|_{W_1([0,\infty))\times W_2([0,\infty))}<\|(u_2,v_2)\|_{W_1([0,\infty))\times W_2([0,\infty))}+\varepsilon\leq L_{v_0}(\ell_0)+\varepsilon.\end{aligned}$$ Taking the supremum over such $u_{0,1} \in \mathcal{F}\dot{H}^\frac{1}{2}$, we obtain $$\begin{aligned}
L_{v_0}(\ell) \le L_{v_0}(\ell_0)+\varepsilon\end{aligned}$$ for $\ell \in (\ell_0 , \ell_0 + \delta)$. This shows the right continuity of $L_{v_0}(\ell)$ at $\ell = \ell_0$ together with non-decreasing property. The left continuity is shown in a similar way. We omit the details. Let us move on to the case $L_{v_0}(\ell_0)=\infty$. We may suppose that $\ell_0:=\inf\{\ell:L_{v_0}(\ell)=\infty\}$ otherwise continuity is trivial by definition. Under this assumption, we prove that $L_{v_0}(\ell)$ goes to infinity as $\ell\uparrow\ell_0$. Assume that $$\displaystyle C_0:=\sup_{\ell<\ell_0}L_{v_0}(\ell)<\infty$$ for contradiction. Let $\varepsilon_1=\varepsilon_1(C_0)$ be the constant given in Proposition \[Long time perturbation\]. Fix $0<\varepsilon<1$ so that $\varepsilon\ell_0<\varepsilon_1$. Then, for any fixed $u_{0,1}\in\mathcal{F}\dot{H}^\frac{1}{2}$ with $\|u_{0,1}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell_0$, the function $$u_{0,2}:=(1-\varepsilon)u_{0,1}$$ satisfies $\|u_{0,2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq(1-\varepsilon)\ell_0$. Let $(u_1,v_1)$ and $(u_2,v_2)$ be two solutions to with initial data $(u_{0,1},v_0)$ and $(u_{0,2},v_0)$, respectively. One sees that $$\begin{aligned}
\|(u_2,v_2)\|_{W_1([0,\infty))\times W_2([0,\infty))}
\leq L_{v_0}\left((1-\varepsilon)\ell_0\right)
\leq C_0
<\infty.\end{aligned}$$ In addition, we have $$\begin{aligned}
\|u_{0,1}-u_{0,2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}
=\varepsilon\|u_{0,1}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}
\leq \varepsilon\ell_0.\end{aligned}$$ Applying Proposition \[Long time perturbation\], we obtain $$\begin{aligned}
\|(u_1,v_1)\|_{W_1([0,\infty))\times W_2([0,\infty))}
\leq \|(u_2,v_2)\|_{W_1([0,\infty))\times W_2([0,\infty))}+c\varepsilon \ell_0
\leq L_{v_0}\left((1-\varepsilon)\ell_0\right)+c\varepsilon \ell_0
<\infty,\end{aligned}$$ where $c=c(C_0)$ is a constant. Taking supremum over $u_{0,1}$, it follows that $$\begin{aligned}
L_{v_0}(\ell_0)\leq L_{v_0}\left((1-\varepsilon)\ell_0\right)+c\varepsilon\ell_0<\infty.\end{aligned}$$ This is a contradiction.
By using the non-decreasing property of $L_{v_0}$, we have the following:
\[Another characterization\] The following identity holds. $$\begin{aligned}
\ell_{v_0}^\dagger=\inf\{\ell:L_{v_0}(\ell)=\infty\}\end{aligned}$$ for any $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$.
When $L_{v_0}(\ell)$ is finite for any $\ell>0$, we see that the both sides are infinite. Otherwise, the two sets $\{\ell:L_{v_0}(\ell)<\infty\}$ and $\{\ell:L_{v_0}(\ell)=\infty\}$ give us a Dedekind cut of a totally ordered set $[0,\infty)$, by means of Propositions \[Control of L\] and Proposition \[Continuity of L\].
A consequence of the alternative characterization is that $$\label{E:Lvellv}
L_{v_0}(\ell_{v_0}^\dagger) = \infty$$ holds for any $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$. This follows from the continuity of $L_{v_0}$. We also have the following:
\[Comparison of ell and ell\^dagger\] $\ell_{v_0} \ge \ell_{v_0}^\dagger$ for any $v_0 \in \mathcal{F} \dot{H}^\frac{1}{2}$.
If $\ell_{v_0}=\infty$, then Lemma \[Comparison of ell and ell\^dagger\] holds. Let $\ell_{v_0}<\infty$. By the definition of $\ell_{v_0}$, for any $\varepsilon>0$, there exists $u_0\in \mathcal{F}\dot{H}^\frac{1}{2}$ such that $$\begin{aligned}
\|u_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}<\ell_{v_0}+\varepsilon\end{aligned}$$ holds and the corresponding solution $(u(t),v(t))$ with data $(u(0),v(0))=(u_0,v_0)$ does not scatter forward in time. Since Proposition \[Scattering criterion\] deduces $\|(u,v)\|_{W_1([0,T_{\max}))\times W_2([0,T_{\max}))}=\infty$ from the failure of scattering, we obtain $$\begin{aligned}
L_{v_0}(\ell_{v_0}+\varepsilon)=\infty.\end{aligned}$$ This implies the relation $\ell_{v_0}^\dagger\leq\ell_{v_0}+\varepsilon$, thanks to Proposition \[Another characterization\]. Since $\varepsilon>0$ is arbitrary, we have the desired conclusion.
The following is one of the key property to prove Theorem \[T:l0\].
\[P:l0lv0\] $\ell_0^\dagger\geq\ell_{v_0}^\dagger$ for any $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$.
Fix $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$. We assume that $\ell_0^\dagger<\ell_{v_0}^\dagger$ for contradiction. Then, we have $$L_0 \Bigl(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2}\Bigr)=\infty ,\quad L_{v_0} \Bigl(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2}\Bigr)<\infty .$$ Using the fact that $L_0 (\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2})=\infty $ and the scaling argument, one can take data $\{(U_{0,n},0)\}$ so that the corresponding solution $(U_n(t),V_n(t))$ to satisfies $$\begin{aligned}
\left\|U_{0,n}\right\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2} \label{037}\end{aligned}$$ and $$\label{E:l0pf1}
\left\|(U_n,V_n)\right\|_{W_1([0,n^{-1}])\times W_2([0,n^{-1}])}\geq n$$ for all $n\ge1$. Let $(u_{n},v_{n})$ be another solution to with the initial data $(U_{0,n},v_0)$. Since $L_{v_0} (\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2})<\infty $, one sees from that $(u_{n},v_{n})$ is global in time and $$\begin{aligned}
\left\|(u_{n},v_{n})\right\|_{W_1([0,\infty))\times W_2([0,\infty))}\leq L_{v_0}\Bigl(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2}\Bigr)<\infty.\end{aligned}$$ We now set $(\widetilde{u}_{n},\widetilde{v}_{n})=(u_{n},v_{n})-(0,e^{\frac{1}{2}it\Delta}v_0)$. Then, $(\widetilde{u}_{n},\widetilde{v}_{n})$ solves $$\left\{
\begin{aligned}
&i\partial_t\widetilde{u}_{n}+\Delta\widetilde{u}_{n}+2\widetilde{v}_{n}\overline{\widetilde{u}_{n}}=-2(e^{\frac{1}{2}it\Delta}v_0)\overline{u_{n}},\\
&i\partial_t\widetilde{v}_{n}+\frac{1}{2}\Delta\widetilde{v}_{n}+\widetilde{u}_{n}^2=0,\\
&(\widetilde{u}_{n}(0),\widetilde{v}_{n}(0))= (U_{0,n},0)
\end{aligned}
\right.$$ and so it is an approximate solution to with an error $$\begin{aligned}
e_1 = -2(e^{\frac{1}{2}it\Delta}v_0)\overline{u_{n}},\quad e_2 = 0.\end{aligned}$$ Take $\tau>0$ and set $I=[0,\tau]$. We have $$\begin{aligned}
{\lVert (e_1,e_2)\rVert}_{N_1(I)\times N_2(I)} \lesssim {\lVert e^{\frac{1}{2}it\Delta}v_0\rVert}_{W_2(I)} {\lVert u_n\rVert}_{W_1(I)} \le
{\lVert e^{\frac{1}{2}it\Delta}v_0\rVert}_{W_2(I)} L_{v_0}\Bigl(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2}\Bigr).\end{aligned}$$ The right hand side is independent of $n$, and tends to zero as $\tau \downarrow 0$.
Now we apply the Proposition \[Long time perturbation\] with $M=L_{v_0}(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2})+\|e^{\frac{1}{2}it\Delta}v_0\|_{W_2([0,\infty))}$. Choose $\tau$ sufficiently small so that the above upper bound of the error becomes smaller than the corresponding $\varepsilon_1$. Since $(U_n,V_n)$ is a solution with the same initial data as $(\widetilde{u}_{n},\widetilde{v}_{n})$, we see from Proposition \[Long time perturbation\] that $(U_n,V_n)$ extends up to time $\tau$ and obeys the bound $$\begin{aligned}
{\lVert (U_n,V_n)\rVert}_{W_1(I)\times W_2(I)}
\le {\lVert (\widetilde{u}_n,\widetilde{v}_n)\rVert}_{W_1(I)\times W_2(I)} + C\varepsilon_1
\le L_{v_0}\Bigl(\frac{\ell_0^\dagger+\ell_{v_0}^\dagger}{2}\Bigr) + \|e^{\frac{1}{2}it\Delta}v_0\|_{W_2([0,\infty))} + C\varepsilon_1.\end{aligned}$$ However, this contradicts with for large $n$.
Linear profile decomposition {#Sec:Linear profile decomposition}
============================
In this section, we obtain a linear profile decomposition (Theorem \[Linear profile decomposition\]).
Linear profile decomposition {#linear-profile-decomposition}
----------------------------
Let us first introduce several operators and give a notion of deformation, which is a specific class of bounded operator.
We define the following operators.
- A dilation $$\begin{aligned}
(D(h)(f,g))(x)=(f_{\{h\}},g_{\{h\}})=(h^2f(hx),h^2g(hx)),\ \ \ h\in2^{\mathbb{Z}},\end{aligned}$$
- A translation in Fourier space $$\begin{aligned}
(T(\xi)(f,g))(x)=(e^{ix\cdot\xi}f(x),e^{2ix\cdot\xi}g(x)),\ \ \ \xi\in\mathbb{R}^3.\end{aligned}$$
We say that a bounded operator $$\begin{aligned}
\mathcal{G}=(\mathcal{G}_{1},\mathcal{G}_{2})=T(\xi)D(h),\ \ \ (\xi,h)\in\mathbb{R}^3\!\times\!2^{\mathbb{Z}}\end{aligned}$$ on $\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ is called a deformation in $\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$. Let a set $G\subset\mathcal{L}(\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2})$ be composed of all deformations.
$G$ is a group with the functional composition as a binary operation. Id$=T(0)D(1)\in G$ is the identity element. For $\mathcal{G}=T(\xi)D(h)$, the inverse element is $\mathcal{G}^{-1}=T(-\frac{\xi}{h})D(\frac{1}{h})\in G$.
Next, we introduce a class of families of deformations.
We say that a family of deformations $\{\mathcal{G}_n=T(\xi_n)D(h_n)\}_n\subset G$ is vanishing if $
|\xi_n|+|\log h_n|\longrightarrow\infty\ \text{ as }\ n\rightarrow\infty
$ holds.
\[Symmetry vanishing\] A family $\{\mathcal{G}_n\}_n\subset G$ is vanishing if and only if a family of inverse elements $\{\mathcal{G}_n^{-1}\}_n$ is vanishing.
It is clear from $(T(\xi_n)D(h_n))^{-1}=T(-\frac{\xi_n}{h_n})D(\frac{1}{h_n})$.
The following characterization of the vanishing family is useful.
\[Equivalent vanishing 2\] For a family $\{\mathcal{G}_n\}_n\subset G$ of deformations, the following three statements are equivalent.
- $\{\mathcal{G}_n\}_n$ is vanishing.
- For any $(\phi,\psi)\in\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$, $\mathcal{G}_n(\phi,\psi)\xrightharpoonup[]{\hspace{0.4cm}}(0,0)$ weakly in $\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ as $n\rightarrow\infty$.
- For any subsequence $n_k$ of $n$, there exists a sequence $\{(f_k,g_k)\}_k\subset\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ and a subsequence $k_l$ of $k$ such that $(f_{k_l},g_{k_l})\xrightharpoonup[]{\hspace{0.4cm}}(0,0)$ and $\mathcal{G}_{n_{k_l}}^{-1}(f_{k_l},g_{k_l})\xrightharpoonup[]{\hspace{0.4cm}}(\phi,\psi)\neq(0,0)$ weakly in $\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ as $l\rightarrow\infty$.
For the proof, see [@Mas16; @MasSeg18].
We now introduce a notion of orthogonality.
We say two families of deformations $\{\mathcal{G}_n\}$, $\{\widetilde{\mathcal{G}}_n\}\subset G$ are orthogonal if $\{\mathcal{G}_n^{-1}\widetilde{\mathcal{G}}_n\}$ is vanishing.
Let $\{\mathcal{G}_n^j=T(\xi_n^j)D(h_n^j)\}$ $\subset G$ $(j=1,2)$ be two families of deformations. $\{\mathcal{G}_n^1\}$ and $\{\mathcal{G}_n^2\}$ are orthogonal if and only if $$\begin{aligned}
\frac{h_n^1}{h_n^2}+\frac{h_n^2}{h_n^1}+\frac{|\xi_n^1-\xi_n^2|}{h_n^1}\longrightarrow\infty\end{aligned}$$ as $n\rightarrow\infty$. This equivalence holds from the identity $$\begin{aligned}
(\mathcal{G}_n^1)^{-1}\mathcal{G}_n^2
=T\left(\frac{\xi_n^2-\xi_n^1}{h_n^1}\right)D\left(\frac{h_n^2}{h_n^1}\right).\end{aligned}$$
Let $\{\mathcal{G}_n\}, \{\widetilde{\mathcal{G}}_n\}\subset G$. Define the relation $\sim$ as follows: If $\{\mathcal{G}_n\}$ and $\{\widetilde{\mathcal{G}}_n\}$ are not orthogonal then $\{\mathcal{G}_n\}\sim\{\widetilde{\mathcal{G}}_n\}$. Then, $\sim$ is an equivalent relation.
For the proof, see [@Mas16; @MasSeg18].
Let us now state the linear profile decomposition result.
\[Linear profile decomposition\] Let $\{(\phi_n,\psi_n)\}$ be a bounded sequence in $\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$. Passing to a sequence if necessary, there exist profile $\{(\phi^j,\psi^j)\}\subset \mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$, $\{(\Phi_n^J,\Psi_n^J)\}\subset \mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}$, and pairwise orthogonal families of deformations $\{\mathcal{G}_n^j=T(\xi_n^j)D(h_n^j)\}_n\subset G$ $(j=1,2,\cdots)$ such that for each $J\geq 1$, $$\begin{aligned}
(\phi_n,\psi_n)=\sum_{j=1}^J\mathcal{G}_n^j(\phi^j,\psi^j)+(\Phi_n^J,\Psi_n^J)\end{aligned}$$ for any $n\geq1$. Moreover, $\{(\Phi_n^J,\Psi_n^J)\}$ satisfies $$\notag (\mathcal{G}_n^j)^{-1}(\Phi_n^J,\Psi_n^J)\xrightharpoonup[]{\hspace{0.4cm}}
\begin{cases}
&\hspace{-0.4cm}(\phi^j,\psi^j)\hspace{0.55cm}(J<j),\\
&\hspace{-0.2cm}(0,0)\hspace{0.75cm}(J\geq j)
\end{cases}$$ in $\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ as $n\rightarrow\infty$ for any $j\geq0$, where we use the convention $(\Phi_n^0,\Psi_n^0)=(\phi_n,\psi_n)$, and $$\begin{aligned}
\limsup_{n\rightarrow\infty}\|(e^{it\Delta}\Phi_n^J,e^{\frac{1}{2}it\Delta}\Psi_n^J)\|_{L_t^{q,\infty}L_x^r\times L_t^{q,\infty}L_x^r}\longrightarrow0 \label{022}\end{aligned}$$ as $J\rightarrow \infty$ for any $1<q,r<\infty$ such that $\frac{1}{q}\in(\frac{1}{2},1)$ and $\frac{2}{q}+\frac{3}{r}=2$. Furthermore, we have Pythagorean decomposition: $$\begin{aligned}
\|\phi_n\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2=\sum_{j=1}^J\|\phi^j\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+\|\Phi_n^J\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+o_n(1),\end{aligned}$$ $$\begin{aligned}
\|\psi_n\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2=\sum_{j=1}^J\|\psi^j\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+\|\Psi_n^J\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+o_n(1),\end{aligned}$$ where $o_n(1)$ goes to $0$ as $n\rightarrow\infty$.
We define $$\nu(\{(\phi_n,\psi_n)\}):=\left\{(\phi,\psi)\in\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}\left|
\begin{array}{l}
\text{There exist }\xi_n\in\mathbb{R}^3\text{ and }h_n\in2^\mathbb{Z}\text{ such that}\\
(\mathcal{G}_n^j)^{-1}(\phi_n,\psi_n)\xrightharpoonup[]{\hspace{0.4cm}}(\phi,\psi)\text{ in }\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}\\
\text{ as }n\rightarrow\infty,\text{ up to subsequence.}
\end{array}
\right.\right\}.$$ and $$\begin{aligned}
\eta(\{(\phi_n,\psi_n)\}):=\sup_{(\phi,\psi)\in\nu(\{(\phi_n,\psi_n)\})}\|(\phi,\psi)\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times\mathcal{F}\dot{H}^\frac{1}{2}}.\end{aligned}$$ Then, a standard argument shows the theorem. However, the smallness is replaced by $\eta(\{(\Phi_n^J,\Psi_n^J)\}) \to 0$ as $J\to\infty$. The following Proposition \[P:cv\] shows that this smallness is stronger.
Control of vanishing
--------------------
To complete the proof of Theorem \[Linear profile decomposition\], we show the following in this subsection.
\[P:cv\] If a sequence $\{(\Phi_n,\Psi_n)\}_n \subset \mathcal{F}\dot{H}^{\frac12} \times \mathcal{F}\dot{H}^{\frac12}$ satisfies $$\begin{aligned}
{\lVert (\Phi_n,\Psi_n)\rVert}_{\mathcal{F}\dot{H}^{\frac12} \times \mathcal{F}\dot{H}^{\frac12}} \le M\end{aligned}$$ and $$\begin{aligned}
\|(e^{it\Delta}\Phi_n,e^{\frac{1}{2}it\Delta}\Psi_n)\|_{L_t^{q,\infty}L_x^r\times L_t^{q,\infty}L_x^r}\ge \varepsilon_0\end{aligned}$$ for some $M>0$, $\varepsilon_0>0$, and $1<q,r<\infty$ with $\frac{1}{q}\in(\frac{1}{2},1)$ and $\frac{2}{q}+\frac{3}{r}=2$, then $$\begin{aligned}
\eta(\{(\Phi_n,\Psi_n)\}) \gtrsim_{M,\varepsilon_0,q,r} 1.\end{aligned}$$
To prove the proposition, we will need the following.
\[L:iStr\] It holds that $$\begin{aligned}
{\lVert e^{it\Delta}f\rVert}_{L^3_t([0,\infty);L_x^3)} \lesssim {\lVert f\rVert}_{\mathcal{F} \dot{H}^{\frac16}}^{\frac23} \sup_{N\in 2^\mathbb{Z}} \left( {\lVert e^{it \Delta} \psi_N f\rVert}_{L^3_t([0,\infty);L_x^3)} \right)^{\frac13},\end{aligned}$$ where $\psi_N$ is defined as .
By Lemma \[Square function estimate\], we have $$\begin{aligned}
{\lVert e^{it\Delta}f\rVert}_{L^{3}_x}
= {\lVert \mathcal{M}_{\frac{1}{2}}(-t)e^{it\Delta}f\rVert}_{L^{3}_x}
\sim \Bigl\|\Bigl(\sum_{N\in 2^\mathbb{Z}}\left|P_\frac{N}{2t} \mathcal{M}_\frac{1}{2}(-t)e^{it\Delta}f\right|^2\Bigr)^{1/2}\Bigr\|_{L^{3}_x}\end{aligned}$$ for $t>0$, where the implicit constant is independent of $t$ by virtue of the scaling. Denote $v_N = P_\frac{N}{2t} \mathcal{M}_\frac{1}{2}(-t)e^{it\Delta}f$ for simplicity. By a convexity argument, one has $$\begin{aligned}
\Bigl\|\Bigl(\sum_{N\in 2^\mathbb{Z}}|v_N|^2\Bigr)^\frac{1}{2}\Bigr\|_{L^3_t([0,\infty);L_x^3)}^3
&{}= \int_{[0,\infty) \times \mathbb{R}^3} \Bigl(\sum_{N\in 2^\mathbb{Z}}|v_N|^2\Bigr)^{\frac{3}{4}}
\Bigl(\sum_{M\in 2^\mathbb{Z}}|v_M|^2\Bigr)^{\frac{3}{4}}dx\, dt\\
&{}\lesssim {\sum_{M, N\in 2^\mathbb{Z}, N \le M}} \int_{[0,\infty)\times \mathbb{R}^3} |v_N|^{\frac32} |v_M|^{\frac32} dx\, dt,\end{aligned}$$ where we have used the symmetry in the last line to reduce the matter to the case $N\le M$. Take $r_1$ and $r_2$ so that $\frac{8}{3} < r_1<3 < r_2<\frac{10}{3}$ and $\frac23=\frac1{r_1} + \frac1{r_2}$. By the Hölder inequality, $$\begin{aligned}
\int_{[0,\infty) \times \mathbb{R}^3} |v_N|^{\frac32} |v_M|^{\frac32} dx\, dt
\le {\lVert v_N\rVert}_{L_t^{r_1}([0,\infty);L_x^{r_1})} {\lVert v_N\rVert}_{L_t^3([0,\infty);L_x^3)}^\frac12
{\lVert v_M\rVert}_{L_t^3([0,\infty);L_x^3)}^\frac12 {\lVert v_M\rVert}_{L_t^{r_2}([0,\infty);L_x^{r_2})}.\end{aligned}$$ Hence, $$\begin{aligned}
{\lVert e^{it\Delta}f\rVert}_{L_t^3([0,\infty);L_x^3)}^3 &{}\lesssim
\Bigl(\sup_{N\in 2^\mathbb{Z}} {\lVert v_N\rVert}_{L_t^3([0,\infty);L_x^3)}\Bigr) \sum_{M, N\in 2^\mathbb{Z}, N \le M} {\lVert v_N\rVert}_{L_t^{r_1}([0,\infty);L_x^{r_1})} {\lVert v_M\rVert}_{L_t^{r_2}([0,\infty);L_x^{r_2})}\end{aligned}$$ Remark that $$\begin{aligned}
v_N = \mathcal{F} \psi_\frac{N}{2t} \mathcal{F}^{-1} D(t) \mathcal{F} \mathcal{M}_\frac{1}{2}(t) f
= D(t) \mathcal{F} \mathcal{M}_\frac{1}{2}(t) \psi_N f = \mathcal{M}_\frac{1}{2}(t)^{-1} e^{it\Delta} \psi_N f.\end{aligned}$$ By the Strichartz estimate, $$\begin{aligned}
{\lVert v_N\rVert}_{L_t^r([0,\infty);L_x^r)}
&=\|\mathcal{M}_\frac{1}{2}(-t)e^{it\Delta}\psi_Nf\|_{L_t^r([0,\infty);L_x^r)}\lesssim\||\nabla|^\frac{10-3r}{2r}\mathcal{M}_\frac{1}{2}(-t)e^{it\Delta}\psi_Nf\|_{L_t^r([0,\infty);L_x^\frac{6r}{16-3r})}\\
&\lesssim \||t|^{-\frac{10-3r}{2r}}\|_{L_t^{\frac{2r}{10-3r},\infty}}\||t|^\frac{10-3r}{2r}|\nabla|^\frac{10-3r}{2r}\mathcal{M}_\frac{1}{2}(-t)e^{it\Delta}\psi_Nf\|_{L_t^{\frac{2r}{3r-8},r}([0,\infty);L_x^\frac{6r}{16-3r})}\\
&\lesssim {\lVert e^{it\Delta} \psi_N f\rVert}_{L^{\frac{2r}{3r-8},r}_t([0,\infty);\dot{X}_\frac{1}{2}^{\frac{10-3r}{2r},\frac{6r}{16-3r}})}\lesssim {\lVert e^{it\Delta} \psi_N f\rVert}_{L^{\frac{2r}{3r-8},2}_t([0,\infty);\dot{X}_\frac{1}{2}^{\frac{10-3r}{2r},\frac{6r}{16-3r}})}\\
&\lesssim \|\psi_Nf\|_{\mathcal{F}\dot{H}^\frac{10-3r}{2r}} =\||x|^{\frac{5}{r}-\frac{5}{3}}\psi_N|x|^\frac{1}{6}f\|_{L_x^2}
\lesssim N^{\frac{5}{r}-\frac53} {\lVert \psi_N |x|^{\frac16} f\rVert}_{L_x^2}.\end{aligned}$$ Thus, $$\begin{aligned}
&\sum_{N \le M, M,N\in 2^\mathbb{Z}} {\lVert v_N\rVert}_{L_t^{r_1}([0,\infty);L_x^{r_1})} {\lVert v_M\rVert}_{L_t^{r_2}([0,\infty);L_x^{r_2})}\\
&{}\hspace{2.0cm} \leq \sum_{R\geq1}\sum_{N\in 2^\mathbb{Z}} {\lVert v_N\rVert}_{L_t^{r_1}([0,\infty);L_x^{r_1})} {\lVert v_{NR}\rVert}_{L_t^{r_2}([0,\infty);L_x^{r_2})}\\
&{}\hspace{2.0cm} \lesssim \sum_{R\ge1} R^{-\frac{5}{r_1}+\frac53} \sum_{N\in 2^\mathbb{Z}}{\lVert \psi_N |x|^{\frac16} f\rVert}_{L_x^2} {\lVert \psi_{NR} |x|^{\frac16} f\rVert}_{L_x^2}\\
&\hspace{2.0cm} \leq \sum_{R\ge1} R^{-\frac{5}{r_1}+\frac53} \Bigl(\sum_{N\in 2^\mathbb{Z}}{\lVert \psi_N |x|^{\frac16} f\rVert}_{L_x^2}^2\Bigr)^\frac{1}{2}\Bigl( \sum_{N\in2^\mathbb{Z}}{\lVert \psi_{NR} |x|^{\frac16} f\rVert}_{L_x^2}^2\Bigr)^\frac{1}{2}\\
&{}\hspace{2.0cm} =\sum_{R\ge1} R^{-\frac{5}{r_1}+\frac53}\left(\int_{\mathbb{R}^3}\sum_{N\in2^\mathbb{Z}}\left|\psi_N(x)|x|^\frac{1}{6}f(x)\right|^2dx\right)^\frac{1}{2}\left(\int_{\mathbb{R}^3}\sum_{N\in2^\mathbb{Z}}\left|\psi_{NR}(x)|x|^\frac{1}{6}f(x)\right|^2dx\right)^\frac{1}{2}\\
&{}\hspace{2.0cm} \lesssim {\lVert f\rVert}_{\mathcal{F} \dot{H}^\frac16}^2 \sum_{R\ge1} R^{-\frac{5}{r_1}+\frac53}
\lesssim {\lVert f\rVert}_{\mathcal{F} \dot{H}^\frac16}^2.\end{aligned}$$ This completes the proof.
In what follows we denote various subsequences of $n$ again by $n$. By the pigeon hole principle, $$\begin{aligned}
{\lVert e^{i t \Delta} \Phi_n\rVert}_{L_t^{q,\infty}L_x^r} \ge \frac{\varepsilon_0}2 \quad
\text{ or }\quad
{\lVert e^{\frac{1}{2}it \Delta} \Psi_n\rVert}_{L_t^{q,\infty}L_x^r} \ge \frac{\varepsilon_0}2\end{aligned}$$ holds for infinitely many $n$. We only consider the case where the former holds for infinitely many $n$. The proof for the other case is similar.
By interpolation and boundedness lemma, there exists $\theta=\theta(q,r)>0$ such that $$\begin{aligned}
{\lVert e^{i t \Delta} \Phi_n\rVert}_{L_t^{q,\infty}L_x^r} \lesssim {\lVert e^{i t \Delta} \Phi_n\rVert}_{L_t^3([0,\infty);X^{\frac13,3}_{\frac12})}^\theta
{\lVert \Phi_n\rVert}_{\mathcal{F} \dot{H}^\frac12}^{1-\theta}.\end{aligned}$$ By means of Lemma \[L:iStr\] and the assumption, we have $$\begin{aligned}
\sup_{N\in 2^\mathbb{Z}} {\lVert e^{i t \Delta} |x|^{\frac13} \psi_N \Phi_n\rVert}_{L^3_t([0,\infty);L_x^3)}
\gtrsim_{M,\varepsilon_0,q,r} 1.\end{aligned}$$ One can choose a sequence $N_n$ so that $$\label{E:cvpf0}
{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{3}([0,\infty);L_x^3)}
\gtrsim 1.$$ Since the scaling property and Strichartz’ estimate give us $$\begin{aligned}
{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{3}([0,\tau N_n^2];L_x^3)}
&{}= N_n^{\frac53}{\lVert e^{i t \Delta} (N_n|x|)^{\frac13}\psi \Phi_n(N_n \cdot)\rVert}_{L_t^{3}([0,\tau];L_x^3)}\\
&{}\leq N_n^\frac{5}{3} \|1\|_{L_t^{12}([0,\tau])}{\lVert e^{i t \Delta} (N_n|x|)^{\frac13}\psi \Phi_n(N_n \cdot)\rVert}_{L^4_t([0,\tau];L_x^3)}\\
&{}\lesssim N_n^{\frac53} \tau^{\frac1{12}} {\lVert (N_n|x|)^{\frac13}\psi \Phi_n(N_n \cdot)\rVert}_{L_x^2}\\
&\lesssim \tau^{\frac1{12}} {\lVert \Phi_n\rVert}_{\mathcal{F} \dot{H}^{\frac12}}
\lesssim \tau^{\frac1{12}},\end{aligned}$$ one can choose $\tau_0=\tau_0(M,\varepsilon_0,q,r)>0$ small so that is improved as $$\begin{aligned}
{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{3}([\tau_0N_n^2,\infty);L_x^3)}
\gtrsim 1\end{aligned}$$ for all $n\ge1$. Hölder’s inequality gives us $$\begin{aligned}
&{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^3([\tau_0N_n^2,\infty);L_x^3)}\\
&\hspace{2.0cm} \le {\lVert |t|^{\frac32} e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{\infty}([\tau_0N_n^2,\infty);L_x^\infty)}^{\frac1{18}}
{\lVert |t|^{-\frac{3}{34}}e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{\frac{17}6}([\tau_0N_n^2,\infty);L_x^\frac{17}{6})}^{\frac{17}{18}}.\end{aligned}$$ Using the estimate $$\begin{aligned}
{\lVert |t|^{-\frac{3}{34}}e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{\frac{17}6}L_x^\frac{17}{6}}
&{}\lesssim \||t|^{-\frac{3}{34}}\|_{L^{\frac{34}{3},\infty}}{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L^{\frac{34}{9},\frac{17}6}_t L^{\frac{17}6}_{x}} \\
&\lesssim {\lVert |\nabla|^\frac{3}{34}\mathcal{M}_\frac{1}{2}(-t)e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L^{\frac{34}{9},\frac{17}6}_t L^{\frac{34}{13}}_{x}} \\
&{}\lesssim \||t|^{-\frac{3}{34}}\|_{L^{\frac{34}{3},\infty}}{\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L^{\frac{17}{3},\frac{17}{6}}_t \dot{X}^{\frac3{34},\frac{34}{13}}_{\frac12}} \\
&{}\lesssim {\lVert e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L^{\frac{17}{3},2}_t \dot{X}^{\frac3{34},\frac{34}{13}}_{\frac12}}\\
&{}\lesssim {\lVert |x|^{\frac13+\frac3{34}}\psi_{N_n} \Phi_n\rVert}_{L_x^2}\\
&{}\lesssim N_n^{-\frac4{51}},\end{aligned}$$ we reach to the estimate $$\begin{aligned}
N_n^{-\frac43} {\lVert |t|^{\frac32} e^{i t \Delta}|x|^{\frac13}\psi_{N_n} \Phi_n\rVert}_{L_t^{\infty}([\tau_0N_n^2,\infty);L_x^\infty)}
\gtrsim 1\end{aligned}$$ for all $n\ge1$. There exist $t_n \ge \tau_0 N_n^2$ and $y_n \in \mathbb{R}^3$ such that $$\begin{aligned}
N_n^{-\frac43} | t_n^{\frac32} e^{i t_n \Delta}(|x|^{\frac13}\psi_{N_n} \Phi_n)(y_n)|
\gtrsim 1. \label{E:cvpf1}\end{aligned}$$ By the integral representation of the Schrödinger group, we obtain $$\begin{aligned}
&N_n^{-\frac43}|t_n^{\frac32} e^{i t_n \Delta}(|x|^{\frac13}\psi_{N_n} \Phi_n)(y_n)| \notag \\
&{}\hspace{3.0cm} =N_n^{-\frac{4}{3}}\left|t_n^\frac{3}{2}(4\pi it_n)^{-\frac{3}{2}}\int_{\mathbb{R}^3}e^{\frac{i|x-y_n|^2}{4t_n}}|x|^\frac{1}{3}\psi_{N_n}(x)\Phi_n(x)dx\right| \notag \\
&{}\hspace{3.0cm} \lesssim N_n^{-\frac32}\left| \int_{\mathbb{R}^3} e^{-i\frac{y_n}{2t_n}\cdot x} e^{i\frac{|x|^2}{4t_n}} (N_n^{\frac16}|x|^{-\frac16} \psi_{N_n}(x)) |x|^{\frac12}\Phi_n(x) dx
\right| \notag \\
&{}\hspace{3.0cm} = \left| \int_{\mathbb{R}^3} e^{i\frac{N_n^2|x|^2}{4t_n}} (|x|^{-\frac16} \psi(x)) |x|^{\frac12} (e^{-i\frac{N_ny_n}{2t_n}\cdot x}N_n^{2}\Phi_n(N_n x))dx \label{E:cvpf2}
\right| .\end{aligned}$$ Let $$\begin{aligned}
\xi_n := - \frac{N_n y_n}{2t_n} \in \mathbb{R}^3, \quad h_n := N_n \in 2^{\mathbb{Z}}.\end{aligned}$$ Define a deformation $\mathcal{G}_n \in G$ so that $\mathcal{G}_n^{-1} = T(\xi_n)D(h_n)$.
Since $\{\mathcal{G}_n (\Phi_n,\Psi_n)\}_n$ is a bounded sequence in $\mathcal{F}\dot{H}^{\frac12} \times \mathcal{F}\dot{H}^{\frac12}$, it weakly converges to a pair $(\widetilde{\Phi}, \widetilde{\Psi})\in \mathcal{F}\dot{H}^{\frac12} \times \mathcal{F}\dot{H}^{\frac12}$ along a subsequence. It is obvious that $(\widetilde{\Phi}, \widetilde{\Psi}) \in \nu (\{(\Phi_n,\Psi_n)\})$. Notice that $0<\frac{N_n^2}{4t_n} \le \frac{1}{4\tau_0}$. Hence, by extracting a subsequence if necessary, one has $$\begin{aligned}
\int_{\mathbb{R}^3} e^{i\frac{N_n^2|x|^2}{4t_n}} (|x|^{-\frac16} \psi(x)) |x|^{\frac12} (e^{-i\frac{N_ny_n}{2t_n}\cdot x}N_n^{2}\Phi_n(N_n x)) dx
\longrightarrow
\int_{\mathbb{R}^3} e^{ia |x|^2 } (|x|^{-\frac16} \psi(x)) |x|^{\frac12} \widetilde{\Phi}(x) dx\end{aligned}$$ as $n\to\infty$, where $a\in \mathbb{R}$ is the limit of $\frac{N_n^2}{4t_n}$ along the (sub)sequence. Plugging this with and , we conclude that $$\begin{aligned}
1 \lesssim \left| \int_{\mathbb{R}^3} e^{ia |x|^2 } (|x|^{-\frac16} \psi(x)) |x|^{\frac12} \widetilde{\Phi}(x) dx\right|
\lesssim_\psi {\lVert \widetilde{\Phi}\rVert}_{\mathcal{F} \dot{H}^{\frac12}}
\le {\lVert (\widetilde{\Phi},\widetilde{\Psi})\rVert}_{\mathcal{F} \dot{H}^{\frac12}\times \mathcal{F} \dot{H}^{\frac12}}
\le \eta (\{(\Phi_n,\Psi_n)\}).\end{aligned}$$ This is the desired estimate.
Proof of Theorem \[T:l0\], Theorem \[T:case1\], and Theorem \[T:case2\] {#Proof of main theorem}
=======================================================================
In this section, we prove Theorem \[T:l0\], Theorem \[T:case1\], and Theorem \[T:case2\]. The following proof shows all these theorems.
Fix $v_0\in\mathcal{F}\dot{H}^\frac{1}{2}$. First, we consider the case $\ell_{v_0}^\dagger=\infty$. In this case, we can obtain $\ell_{v_0}^\dagger=\ell_{v_0}=\ell_0=\infty$. Indeed, we have $\infty=\ell_{v_0}^\dagger\leq\ell_{v_0}$ by Lemma \[Comparison of ell and ell\^dagger\]. On the other hand, we have $\infty=\ell_{v_0}^\dagger\leq \ell_0^\dagger\leq \ell_0$ by Lemma \[P:l0lv0\] and Lemma \[Comparison of ell and ell\^dagger\].
From now on, we assume $\ell_{v_0}^\dagger<\infty$. By definition of $\ell_{v_0}^\dagger$, we have $L_{v_0}(\ell_{v_0}^\dagger-\frac{1}{n})<\infty$ for each $n\in \mathbb{N}$, that is, $$\sup\left\{\|(u,v)\|_{W_1([0,\infty))\times W_2([0,\infty))}\middle|
\begin{array}{l}
(u,v)\text{ is the solution to \eqref{NLS} on }[0,\infty),\\[0.1cm]
v(0)=v_0,\,\,
\|u(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell_{v_0}^\dagger-\frac{1}{n}
\end{array}
\right\}<\infty.$$ We note that $T_{\max}=\infty$ because of Proposition \[Scattering criterion\]. Since $L_{v_0}(\ell)<\infty$ for any $0\leq \ell<\ell_{v_0}^\dagger$, $L_{v_0}(\ell_{v_0}^\dagger)=\infty$, and $L_{v_0}$ is non-decreasing, we can take a sequence $\{m_n\}$ of $\mathbb{N}$ such that $$\begin{aligned}
L_{v_0}\Bigl(\ell_{v_0}^\dagger-\frac{1}{m_n}\Bigr)
<L_{v_0}\Bigl(\ell_{v_0}^\dagger-\frac{1}{m_{n+1}}\Bigr)\end{aligned}$$ for each $n\in \mathbb{N}$. We take a sequence $\{u_{0,n}\}\in\mathcal{F}\dot{H}^\frac{1}{2}$ satisfying $$\begin{aligned}
\ell^\dagger_{v_0}-\frac{1}{m_n}
< \|u_{0,n}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}
\leq \ell^\dagger_{v_0}-\frac{1}{m_{n+1}} \label{023}\end{aligned}$$ and $$\begin{aligned}
L_{v_0}\Bigl(\ell_{v_0}^\dagger-\frac{1}{m_n}\Bigr)
<\|(u_n,v_n)\|_{W_1([0,\infty))\times W_2([0,\infty))}
\leq L_{v_0}\Bigl(\ell_{v_0}^\dagger-\frac{1}{m_{n+1}}\Bigr),\end{aligned}$$ where $(u_n,v_n)$ is the solution to with the initial data $(u_{0,n},v_0)$. Since $\{(u_{0,n},v_0)\}\subset \mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$ is a bounded sequence, we apply Theorem \[Linear profile decomposition\] to this sequence. Then, there exists profile $\{(\phi^j,\psi^j)\}\subset\mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$, remainder $\{(R_n^J,L_n^J)\}\subset \mathcal{F}\dot{H}^\frac{1}{2}\!\times\!\mathcal{F}\dot{H}^\frac{1}{2}$, and pairwise orthogonal families of deformations $\{\mathcal{G}_n^j=T(\xi_n^j)D(h_n^j)\}_n\subset G$ $(j=1,2,\ldots)$ such that $$\begin{aligned}
(u_{0,n},v_0)=\sum_{j=1}^J\mathcal{G}_n^j(\phi^j,\psi^j)+(R_n^J,L_n^J) \label{026}\end{aligned}$$ for any $J\geq 1$. Since $v_0$ is independent of $n$, there exists unique $j_0$ such that $\psi^{j_0}=v_0$ and $\mathcal{G}_n^{j_0}=\text{Id}$. Furthermore, the remainder for $v$-component is zero: $L_n^J=0$. Rearranging the profile $(\phi^j,\psi^j)$, we may let $j_0=1$. Then, the above decomposition reads as $$\begin{aligned}
(u_{0,n},v_0)
=(\phi^1,v_0)+\sum_{j=2}^J\mathcal{G}_n^j(\phi^j,0)+(R_n^J,0).\end{aligned}$$ From Theorem \[Linear profile decomposition\], we have Pythagorean decomposition: $$\begin{aligned}
\|u_{0,n}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2=\sum_{j=1}^J\|\phi^j\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+\|R_n^J\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2+o_n(1) \label{027}\end{aligned}$$ for each $J\geq 1$. The parameters are asymptotically orthogonal: if $j\neq k$, then $$\begin{aligned}
\frac{h_n^j}{h_n^k}+\frac{h_n^k}{h_n^j}+\frac{|\xi_n^j-\xi_n^k|}{h_n^j}\longrightarrow\infty\ \ \text{ as }\ \ n\rightarrow\infty. \label{028}\end{aligned}$$ The remainders satisfy $$\begin{aligned}
(\mathcal{G}_n^j)^{-1}R_n^J\xrightharpoonup[]{\hspace{0.4cm}}0\ \ \text{ in }\ \ \mathcal{F}\dot{H}^\frac{1}{2}\ \ \text{ as }\ \ n\rightarrow\infty\end{aligned}$$ for any $1\leq j\leq J$. $$\begin{aligned}
\lim_{J\rightarrow\infty}\limsup_{n\rightarrow\infty}\|e^{it\Delta}R_n^J\|_{L_t^{q,\infty}L_x^r}=0 \label{030}\end{aligned}$$ for any $1<q,r<\infty$ with $\frac{1}{q}\in(\frac{1}{2},1)$ and $\frac{2}{q}+\frac{3}{r}=2$.
We will prove that there exists only one $j_1$ satisfying $\phi^{j_1}\neq0$, and it satisfies $\|\phi^{j_1}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger$. From , we have $$\begin{aligned}
\sum_{j=1}^J\|\phi^j\|_{\mathcal{F}\dot{H}^\frac{1}{2}}^2\leq (\ell_{v_0}^\dagger)^2, \label{031}\end{aligned}$$ and hence, $\|\phi^j\|_{\mathcal{F}\dot{H}^\frac{1}{2}}\leq\ell_{v_0}^\dagger$ holds for any $j\geq 1$. Let $(\Phi_j,\Psi_j)$ be the solution to with a initial data $(\phi^j,\psi^j)$. We assume for contradiction that all $(\Phi_j,\Psi_j)$ scatter forward in time, that is, $$\begin{aligned}
\|(\Phi_j,\Psi_j)\|_{W_1([0,\infty))\times W_2([0,\infty))}<\infty\end{aligned}$$ is true for any $ j \ge 1$. We set $$\begin{aligned}
(\widetilde{w}_n^J,\widetilde{z}_n^J)
:=\sum_{j=1}^J\left((\Phi_j)_{[h_n^j,\xi_n^j]}(t,x),(\Psi_j)_{[h_n^j,\xi_n^j]}(t,x)\right)\end{aligned}$$ and $$\begin{aligned}
(\widetilde{u}_n^J,\widetilde{v}_n^J)
:=(\widetilde{w}_n^J,\widetilde{z}_n^J)+(e^{it\Delta}R_n^J,0),\end{aligned}$$ where $$\begin{aligned}
(\Phi_j)_{[h_n^j,\xi_n^j]}(t,x)
:={h_n^j}^2e^{ix\cdot\xi_n^j}e^{-it|\xi_n^j|^2}\Phi_j({h_n^j}^2t,h_n^j(x-2t\xi_n^j)),\end{aligned}$$ $$\begin{aligned}
(\Psi_j)_{[h_n^j,\xi_n^j]}(t,x):={h_n^j}^2e^{2ix\cdot\xi_n^j}e^{-2it|\xi_n^j|^2}\Psi_j({h_n^j}^2t,h_n^j(x-2t\xi_n^j)).\end{aligned}$$ We note that $((\Phi_j)_{[h_n^j,\xi_n^j]},(\Psi_j)_{[h_n^j,\xi_n^j]})$ is a solution to with a initial data $\mathcal{G}_n^j(\phi^j,\psi^j)$. Then, $(\widetilde{u}_n^J,\widetilde{v}_n^J)$ solves $$\begin{gathered}
i\partial_t\widetilde{u}_n^J+\Delta\widetilde{u}_n^J
=\sum_{j=1}^J\left(i\partial_t(\Phi_j)_{[h_n^j,\xi_n^j]}+\Delta(\Phi_j)_{[h_n^j,\xi_n^j]}\right)
=-2\sum_{j=1}^J(\Psi_j)_{[h_n^j,\xi_n^j]}\overline{(\Phi_j)_{[h_n^j,\xi_n^j]}},\\
i\partial_t\widetilde{v}_n^J+\frac{1}{2}\Delta\widetilde{v}_n^J
=\sum_{j=1}^J\left(i\partial_t(\Psi_j)_{[h_n^j,\xi_n^j]}+\frac{1}{2}\Delta(\Psi_j)_{[h_n^j,\xi_n^j]}\right)
=-\sum_{j=1}^J(\Phi_j)_{[h_n^j,\xi_n^j]}^2.\end{gathered}$$ We also set $$\begin{aligned}
\widetilde{e}_{1,n}^J:=i\partial_t\widetilde{u}_n^J+\Delta\widetilde{u}_n^J+2\widetilde{v}_n^J\overline{\widetilde{u}_n^J},\end{aligned}$$ $$\begin{aligned}
\widetilde{e}_{2,n}^J:=i\partial_t\widetilde{v}_n^J+\frac{1}{2}\Delta\widetilde{v}_n^J+(\widetilde{u}_n^J)^2.\end{aligned}$$ Here, we introduce the following two lemmas.
\[Lemma1 for main result\] For any $\varepsilon>0$, there exists $J_0=J_0(\varepsilon)$ such that $$\begin{aligned}
\limsup_{n\rightarrow\infty}\|(\widetilde{w}_n^J,\widetilde{z}_n^J)-(\widetilde{w}_n^{J_0},\widetilde{z}_n^{J_0})\|_{W_1([0,\infty))\times W_2([0,\infty))}\leq\varepsilon\end{aligned}$$ for any $J\geq J_0$.
\[Lemma2 for main result\] It follows that $$\begin{aligned}
\lim_{J\rightarrow\infty}\limsup_{n\rightarrow\infty}\|(\widetilde{e}_{1,n}^J,\widetilde{e}_{2,n}^J)\|_{N_1([0,\infty))\times N_2([0,\infty))}=0.\end{aligned}$$
These are shown as in [@Mas16]. Using Lemma \[Lemma1 for main result\] with $\varepsilon=1$, it follows that there exists $J_0$ such that $$\begin{aligned}
&\|(\widetilde{u}_n^J,\widetilde{v}_n^J)\|_{W_1([0,\infty))\times W_2([0,\infty))} \notag \\
&\hspace{0cm}\leq\|(\widetilde{w}_n^{J_0},\widetilde{z}_n^{J_0})\|_{W_1([0,\infty))\times W_2([0,\infty))}+\|(\widetilde{w}_n^J,\widetilde{z}_n^J)-(\widetilde{w}_n^{J_0},\widetilde{z}_n^{J_0})\|_{W_1([0,\infty))\times W_2([0,\infty))}+\|e^{it\Delta}R_n^J\|_{W_1([0,\infty))} \notag \\
&\hspace{0cm}\leq\sum_{j=1}^{J_0}\left\|\left(\Phi_j,\Psi_j\right)\right\|_{W_1([0,\infty))\times W_2([0,\infty))}+c\|R_n^J\|_{\mathcal{F}\dot{H}^\frac{1}{2}}+1 \notag \\
&\hspace{0cm}\leq\sum_{j=1}^{J_0}\left\|\left(\Phi_j,\Psi_j\right)\right\|_{W_1([0,\infty))\times W_2([0,\infty))}+c\ell_{v_0}^\dagger+1=:M \label{032}\end{aligned}$$ for any $J\geq J_0$ and $n\geq1$. Let $\varepsilon_1$ be given in Proposition \[Long time perturbation\]. Then, $$\begin{aligned}
&\left\|\left(u_{0,n}-\widetilde{u}_n^J(0),v_0-\widetilde{v}_n^J(0)\right)\right\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times \mathcal{F}\dot{H}^\frac{1}{2}}
=0. \label{033}\end{aligned}$$ Lemma \[Lemma2 for main result\] implies that there exists $J_1$ such that $$\begin{aligned}
\limsup_{n\rightarrow\infty}\left\|\left(\widetilde{e}_{1,n}^J,\widetilde{e}_{2,n}^J\right)\right\|_{N_1([0,\infty))\times N_2([0,\infty))}\leq\frac{\varepsilon_1}{4}.\end{aligned}$$ for any $J\geq J_1$. Choose $J$ with $J\geq\max\{J_0,J_1\}$. There exists $n_0$ such that $$\begin{aligned}
\left\|\left(\widetilde{e}_{1,n}^J,\widetilde{e}_{2,n}^J\right)\right\|_{N_1([0,\infty))\times N_2([0,\infty))}\leq\frac{\varepsilon_1}{2} \label{034}\end{aligned}$$ for any $n\geq n_0$. By , , , and Proposition \[Long time perturbation\], we deduce that a solution $(u_n,v_n)$ to with a initial data $(u_{0,n},v_0)$ satisfies $$\begin{aligned}
\|(u_n,v_n)\|_{W_1([0,\infty))\times W_2([0,\infty))}
\leq C(M,\varepsilon_1)
<\infty\end{aligned}$$ for any $n\geq n_0$. However, this contradicts with the definition of $(u_n,v_n)$. Therefore, there exists $j_1\ge1$ such that $$\begin{aligned}
\|(\Phi_{j_1},\Psi_{j_1})\|_{W_1([0,T_{\max}))\times W_2([0,T_{\max}))}=\infty.\end{aligned}$$ By , another characterization of $\ell_{v_0}^\dagger$ (Proposition \[Another characterization\]), and Proposition \[P:l0lv0\], we have $\|\phi^{j_1}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger$ and $\phi^{j}=0$ for all $j\neq j_1$. We encounter a dichotomy, $j_1=1$ or $j_1=2$.
Now, we suppose that $j_1=1$. Since a solution $(\Phi_1,\Psi_1)$ to with a initial data $(\phi^1,v_0)$ does not scatter, we have $\ell_{v_0}\leq\|\phi^1\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger$ by the definition of $\ell_{v_0}$. Combining this inequality and Lemma \[Comparison of ell and ell\^dagger\], we obtain $\ell_{v_0}=\ell_{v_0}^\dagger=\|\phi^1\|_{\mathcal{F}\dot{H}^\frac{1}{2}}$. This shows that $\phi^1$ is a minimizer to $\ell_{v_0}$. Moreover, it follows from Lemmas \[P:l0lv0\] and Lemma \[Comparison of ell and ell\^dagger\] that $\ell_{v_0}=\ell_{v_0}^\dagger\leq\ell_0^\dagger\leq\ell_0$. Therefore, we have the identity $\ell_{v_0}^\dagger=\min\{\ell_0,\ell_{v_0}\}$.
Let us move on to the case $j_1=2$. In this case, it follows that $(\phi^1,\psi^1)=(0,v_0)$ and $(\phi^{2},\psi^{2})=(\phi^{2},0)$. Since $(\Phi_{2},\Psi_{2})$ does not scatter, we have $\ell_0\leq\|\phi^{2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger$ by the definition of $\ell_0$. Using Lemma \[P:l0lv0\] and Lemma \[Comparison of ell and ell\^dagger\], we obtain $$\begin{aligned}
\ell_0\leq\|\phi^{2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger\leq\ell_0^\dagger\leq\ell_0.\end{aligned}$$ In particular, we have $\ell_{v_0}^\dagger=\ell_0=\|\phi^{2}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}$. This shows that $\phi^2$ is a minimizer to $\ell_{0}$. In addition, we have $$\begin{gathered}
(u_{0,n},v_0)
=\sum_{j=1,2}\mathcal{G}_n^j(\phi^j,\psi^j)+(R_n^2,0)
=(0,v_0)+\mathcal{G}_n^{2}(\phi^{2},0)+(R_n^2,0),\\
\lim_{n\rightarrow\infty}\|u_{0,n}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}
=\ell_{v_0}^\dagger,\ \ \text{ and }\ \
\lim_{n\rightarrow\infty}\|R_n^2\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=0\end{gathered}$$ by , , , and $\|\phi^{j_0}\|_{\mathcal{F}\dot{H}^\frac{1}{2}}=\ell_{v_0}^\dagger$. Remark that we have the identity $\ell_{v_0}^\dagger=\min\{\ell_0,\ell_{v_0}\}$ also in this case. In both cases, we have the identity $\ell_{v_0}^\dagger=\min\{\ell_0,\ell_{v_0}\}$, hence we have Theorem \[T:l0\]. If we assume that $\ell_0>\ell_{v_0}^\dagger$ then the second case is precluded. This is nothing but Theorem \[T:case1\].
Similarly, the assumption $\ell_{v_0}>\ell_{v_0}^\dagger$ precludes the case $j_1=1$. This shows Theorem \[T:case2\]. Indeed, the above argument applies to the minimizing sequence satisfying the assumption of Theorem \[T:case2\] and leads us to the same conclusion in the case $j_1=2$. Let $T_{\max}$ denote the maximal existence time of a solution to with a initial data $(\phi^{2},0)$. Fix $0\leq \tau <T_{\max}$. Recall that $(\Phi_j,\Psi_j)$ denotes the solution to with a initial data $(\phi^j,\psi^j)$, and $\Bigl((\Phi_j)_{[h_n^j,\xi_n^j]},(\Psi_j)_{[h_n^j,\xi_n^j]}\Bigr)$ does the solution to with a initial data $\mathcal{G}_n^j(\phi^j,\psi^j)$. We set $$\begin{aligned}
(\widetilde{u}_n,\widetilde{v}_n)
:=\sum_{j=1,2}\left((\Phi_j)_{[h_n^j,\xi_n^j]},(\Psi_j)_{[h_n^j,\xi_n^j]}\right)
=(0,e^{\frac{1}{2}it\Delta}v_0)+\left((\Phi_{2})_{[h_n^{2},\xi_n^{2}]},(\Psi_{2})_{[h_n^{2},\xi_n^{2}]}\right).\end{aligned}$$ Then, $(\widetilde{u}_n,\widetilde{v}_n)$ solves $$\begin{aligned}
i\partial_t\widetilde{u}_n+\Delta\widetilde{u}_n
&=\sum_{j=1,2}\left(i\partial_t(\Phi_j)_{[h_n^j,\xi_n^j]}+\Delta(\Phi_j)_{[h_n^j,\xi_n^j]}\right)
=-2(\Psi_{2})_{[h_n^{2},\xi_n^{2}]}\overline{(\Phi_{2})_{[h_n^{2},\xi_n^{2}]}},\\
i\partial_t\widetilde{v}_n+\frac{1}{2}\Delta\widetilde{v}_n
&=\sum_{j=1,2}\left(i\partial_t(\Psi_j)_{[h_n^j,\xi_n^j]}+\frac{1}{2}\Delta(\Psi_j)_{[h_n^j,\xi_n^j]}\right)
=-(\Phi_2)_{[h_n^2,\xi_n^2]}^2.\end{aligned}$$ We also set $$\begin{gathered}
\widetilde{e}_{1,n}
:=i\partial_t\widetilde{u}_n+\Delta\widetilde{u}_n+2\widetilde{v}_n\overline{\widetilde{u}_n}
=2(\Psi_{1})_{[h_n^1,\xi_n^1]}\overline{(\Phi_{2})_{[h_n^{2},\xi_n^{2}]}},\\
\widetilde{e}_{2,n}
:=i\partial_t\widetilde{v}_n+\frac{1}{2}\Delta\widetilde{v}_n+(\widetilde{u}_n)^2
=0.\end{gathered}$$ We check the assumptions of Proposition \[Long time perturbation\]. One has $$\begin{aligned}
&\|(\widetilde{u}_n,\widetilde{v}_n)\|_{W_1([0,\tau/(h_n^{2})^2))\times W_2([0,\tau/(h_n^{2})^2))}\\
& \qquad\leq \|(0,e^{\frac{1}{2}it\Delta}v_0)\|_{W_1([0,\infty))\times W_2([0,\infty))}+\|(\Phi_{j_0},\Psi_{j_0})\|_{W_1([0,\tau))\times W_2([0,\tau))} =:M<\infty,\end{aligned}$$ $$\begin{aligned}
\|(u_{0,n},v_0)-(\widetilde{u}_n(0),\widetilde{v}_n(0))\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times \mathcal{F}\dot{H}^\frac{1}{2}}
=\|(R_n^2,0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}\times \mathcal{F}\dot{H}^\frac{1}{2}}
\longrightarrow0\ \ \text{ as }\ \ n\rightarrow\infty,\end{aligned}$$ and $$\begin{aligned}
\|(\widetilde{e}_{1,n},\widetilde{e}_{2,n})\|_{N_1([0,\tau/(h_n^{2})^2))\times N_2([0,\tau/(h_n^{2})^2))}
= \|\widetilde{e}_{1,n}\|_{N_1([0,\tau/(h_n^{2})^2))}
\longrightarrow0\ \ \text{ as }\ \ n\rightarrow\infty,\end{aligned}$$ where the last estimate is shown as in the same spirit of Lemma \[Lemma2 for main result\] with a help of the first estimate. Therefore, we obtain $$\begin{aligned}
\left\|(u_n,v_n)-(0,e^{\frac{1}{2}it\Delta}v_0)-\Bigl((\Phi_{2})_{[h_n^{2},\xi_n^{2}]},(\Psi_{2})_{[h_n^{2},\xi_n^{2}]}\Bigr)\right\|_{L_t^\infty([0,\tau/(h_n^{2})^2);\dot{X}_{1/2}^{1/2})\times L_t^\infty([0,\tau/(h_n^{2})^2);\dot{X}_1^{1/2})}
\longrightarrow0\end{aligned}$$ as $n\rightarrow\infty$.
We next consider the optimizing problem $\ell_f$ defined in .
\[T:ellfminimizer\] Let $f(x,y)$ be a function on $[0,\infty)\times [0,\infty)$ satisfying the following three conditions:
- Strictly increasing with respect to the both variables, i.e., $$0\le x_1 \le x_2 ,\, 0\le y_1 \le y_2 \Longrightarrow f(x_1,y_1) \le f(x_2,y_2)$$ and the equality holds only if $x_1=x_2$ and $y_1=y_2$.
- Continuous, i.e., for any $(x_0,y_0)\in [0,\infty)\times [0,\infty)$, $$\lim_{[0,\infty)\times [0,\infty) \ni (x,y) \to (x_0,y_0) } f(x, y) = f(x_0,y_0).$$
- $f(0,0)=0$.
Let $\ell_f$ be defined in . Define $$\widetilde{\ell}_f := \inf_{v_0 \in \mathcal{F} \dot{H}^{\frac12}} f(\ell_{v_0}, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}).$$ Then, it follows that $$\ell_f= \widetilde{\ell}_f= \inf_{v_0 \in \mathcal{F} \dot{H}^{\frac12}} f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}).$$ Furthermore, there exists a minimizer $(u^{(f)}(t),v^{(f)}(t))$ to $\ell_{f}$ such that
1. $f({\lVert u^{(f)}(0)\rVert}_{\mathcal{F} \dot{H}^{\frac12}}, {\lVert v^{(f)}(0)\rVert}_{\mathcal{F} \dot{H}^{\frac12}}) = \ell_{f}$;
2. $(u^{(f)}(t),v^{(f)}(t))$ does not scatter forward in time;
3. ${\lVert u^{(f)}(0)\rVert}_{\mathcal{F} \dot{H}^{\frac12}}=\ell_{v^{(f)}(0)}$.
The minimizer is not a ground state.
Let us first show $\ell_f \ge \widetilde{\ell}_f$. By definition of $\ell_{v_0}$ and the fact that $f$ is increasing in $x$, one sees that the inequality $$f(\ell_{v_0}, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}) \le f({\lVert u_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}})$$ is true for any $(u_0,v_0) \notin \mathcal{S}_+$. Taking the infimum over $(u_0,v_0) \notin \mathcal{S}_+$, we obtain $\widetilde{\ell}_{f} \le \ell_f$. Let us introduce $\ell^\dagger_f$ as follows: $$L_{f}(\ell):=\sup\left\{\|(u,v)\|_{W_1([0,T_{\max}))\times W_2([0,T_{\max}))} :
\begin{array}{l}
(u,v)\text{ is the solution to \eqref{NLS} on }[0,T_{\max}),\\[0.1cm]
f( \|u(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}}, \|v(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}}) \leq\ell
\end{array}
\right\},$$ and $$\label{D:ellfd}
\ell_{f}^\dagger:=\sup\{\ell:L_{f}(\ell)<\infty\} \in (0,\infty].$$ Our next goal is to show $\ell_f=\ell_f^\dagger$. Since $f(0,0)=0$ and $f$ is increasing with respect to the both variables, we see that $$\{(u(0),v(0))\in \mathcal{F}\dot{H}^\frac{1}{2} \times \mathcal{F}\dot{H}^\frac{1}{2} : f(\|u(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}},\|v(0)\|_{\mathcal{F}\dot{H}^\frac{1}{2}}) \le \ell \}$$ is a small neighborhood of $(0,0)$ for small $\ell>0$. By the small data theory, we have $L_f(\ell)\lesssim_f 1$ for small $\ell>0$, showing that $\ell_f^\dagger>0$. Mimicking the argument in Proposition \[Continuity of L\], we see that $L_f(\ell)$ is a non-decreasing continuous extended function defined on $[0,\infty)$, thanks to the strictly increasing property of $f$ in both valuables. We also have $$\ell_{f}^\dagger=\inf\{\ell:L_{f}(\ell)=\infty\}$$ and $\ell_{f}^\dagger \le \ell_f$ as in the proof of Propositions \[Another characterization\] and Lemma \[Comparison of ell and ell\^dagger\].
We now show the other direction $\ell_{f}^\dagger \ge \ell_f$. Take an optimizing sequence for $\ell_f^\dagger$. Then, by a similar argument to the proof of Theorem \[T:l0\], we obtain a minimizer $(u_c(t),v_c(t))$ to $\ell_f^\dagger$, which completes the proof of $\ell_f = \ell_f^\dagger$. We omit the details of the proof but point out several different respects compared with an optimizing sequence for $\ell_{v_0}^\dagger$. First of all, the second component $v_{0,n}$ of the optimizing sequence may vary in $n$. As a result, we do not have a priori information about the second component in the profile decompositions, hence the decomposition takes the form $$\begin{aligned}
(u_{0,n},v_{0,n})
=\sum_{j=1}^J\mathcal{G}_n^j(\phi^j,\psi^j)+(R_n^J,L_n^J).
$$ Here we remark that since $({\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12},{\lVert v_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12})$ belongs to a compact set, say $\{ f(x,y) \le \ell_f^\dagger +1 \}$, one may suppose that it converges to a point $(x_\infty,y_\infty)$ such that $f(x_\infty,y_\infty)=\ell_f^\dagger$ along a subsequence.
A contradiction argument shows there exists at most one $j$ such that $(\phi^j,\psi^j) \notin S_+$. We may let $j=1$. Then, the second difference is that we are also able to show that the number of nonzero profile is at most one. This is because if $(\phi^2,\psi^2)\neq(0,0)$ then we have $${\lVert \phi^1\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 + {\lVert \phi^2\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 \le \lim_{n\to\infty} {\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 = x_\infty^2$$ and $${\lVert \psi^1\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 + {\lVert \psi^2\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 \le \lim_{n\to\infty} {\lVert v_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12}^2 = y_\infty^2$$ by the Pythagorean decomposition. This shows ${\lVert \phi^1\rVert}_{\mathcal{F} \dot{H}^\frac12}\le x_\infty$ and ${\lVert \psi^1\rVert}_{\mathcal{F} \dot{H}^\frac12}\le y_\infty$. Since one of the above equality fails when $(\phi^2,\psi^2)\neq(0,0)$, the strictly increasing property of $f$ shows $$f({\lVert \phi^1\rVert}_{\mathcal{F} \dot{H}^\frac12},{\lVert \psi^1\rVert}_{\mathcal{F} \dot{H}^\frac12})
< f(x_\infty,y_\infty) = \ell_f^\dagger.$$ However, as $(\phi^1,\psi^1) \notin S_+$, the left hand side is not less than $\ell_f$. Thus, we obtain $\ell_f < \ell_f^\dagger$, a contradiction. We conclude that $\ell_f^\dagger = \ell_f$ and a solution $(u^{(f)}(t),v^{(f)}(t))$ which satisfies the initial condition $(u^{(f)}(0),v^{(f)}(0))=(\phi^1,\psi^1)$ is a desired minimizer to $\ell_f$.
To complete the proof, it suffices to show that $$\ell_f^\dagger \le \inf_{v_0 \in \mathcal{F} \dot{H}^{\frac12}} f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}})$$ because we have already shown $\ell_f^\dagger=\ell_f\ge \widetilde{\ell}_f$ and because $\widetilde{\ell}_f$ is obviously greater than or equal to the right hand side in view of Theorem \[T:l0\] and the increasing property of $f$. Fix $v_0 \in \mathcal{F} \dot{H}^\frac12$. By , we can pick a sequence $\{u_{0,n}\}_{n} \subset \mathcal{F} \dot{H}^\frac12$ so that ${\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12} \le \ell_{v_0}^\dagger$ and the corresponding solution $(u_n(t),v_n(t))$ with the data $(u_n(0),v_n(0))=(u_{0,n},v_0)$ satisfies $${\lVert (u_n,v_n)\rVert}_{W_1([0,T_{\max}))\times W_1([0,T_{\max}))} \ge n .$$ Further, it follows from the increasing property of $f$ that $$f({\lVert u_{0,n}\rVert}_{\mathcal{F} \dot{H}^\frac12}, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}) \le f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}})$$ for all $n\ge0$. The existence of the above $\{u_{0,n}\}_{n}$ implies that $$L_f( f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}) )=\infty.$$ The other characterization of $\ell_f^\dagger$ then gives us $$\ell_f^\dagger \le f(\ell_{v_0}^\dagger, {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}}).$$ Since $v_0$ is arbitrary, we obtain the result.
Proof of corollaries of Theorem \[Nonpositive energy\] {#Proof of corollaries}
======================================================
We have proven Theorem \[Nonpositive energy\] in Subsection \[subsec:Nonpositive energy\]. Let us show its corollaries.
For given $v_0 \in \mathcal{F} \dot{H}^{\frac12} \cap H^1$ with $v_0\neq0$, we take $$u_0 = v_0(x)^{\frac12} |v_0 (x)|^{\frac12} \in \mathcal{F} \dot{H}^{\frac12} \cap H^1.$$ Then, we have $$E[ c^{\frac12} d u_0,c v_0] \le c d^2 {\lVert \nabla v_0\rVert}_{L^2}^2 + \frac{c^2}2 {\lVert \nabla v_0\rVert}_{L^2}^2
-2 c^{2}d^2{\lVert v_0\rVert}_{L^3}^3$$ for $c>0$ and $d={\lVert \nabla v_0\rVert}_{L^2} {\lVert v_0\rVert}_{L^3}^{-3/2}$. There exists $c_0=c_0(v_0)>0$ such that the right side is negative for any $c\ge c_0$. For such $c$, the corresponding solution does not scatter by virtue of Theorem \[Nonpositive energy\]. This also shows the bound $$\ell_{c v_0} \le {\lVert c^{\frac12}d u_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}} =
c^{\frac12} {\lVert v_0\rVert}_{\mathcal{F} \dot{H}^{\frac12}} {\lVert \nabla v_0\rVert}_{L^2} {\lVert v_0\rVert}_{L^3}^{-3/2}.$$ We have the desired result.
We have $$\begin{aligned}
-\Delta\varphi-2({\operatorname{Re}}e^{i\theta} v_0)\varphi= \tilde{e}\varphi.\end{aligned}$$ Remark that $\varphi$ is real-valued. Multiplying this identity by $\varphi$, and integrating, we have $$\begin{aligned}
(-\Delta\varphi,\varphi)_{L^2}-(2({\operatorname{Re}}e^{i\theta} v_0)\varphi,\varphi)_{L^2}=\tilde{e}(\varphi,\varphi)_{L^2}.\end{aligned}$$ This can be rearranged as $$\begin{aligned}
\|\nabla\varphi\|_{L^2}^2 + 2{\operatorname{Re}}\int (-e^{-i\theta} \varphi^2) \overline{v_0} dx =\tilde{e} \|\varphi\|_{L^2}^2.\end{aligned}$$ Here, we take $u_0= e^{-i\theta/2} \varphi$. Then, $$\begin{aligned}
E[cu_0,v_0]=c^2\|\nabla\varphi\|_{L^2}^2+\frac{1}{2}\|\nabla v_0\|_{L^2}^2- 2{\operatorname{Re}}\int (c^2e^{-i\theta} \varphi^2) \overline{v_0} dx
=c^2\tilde{e} \|\varphi\|_{L^2}^2+\frac{1}{2}\|\nabla v_0\|_{L^2}^2.\end{aligned}$$ From $\tilde{e}<0$, the choice $c^2=\frac{\|\nabla v_0\|_{L^2}^2}{2|\tilde{e}|\|\varphi\|_{L^2}^2}$ gives us $E(u_0,v_0)=0$. Therefore, $(cu_0,v_0)\notin S_+$ by Theorem \[Nonpositive energy\]. This also implies the bound $$\begin{aligned}
\ell_{v_0}\leq \|cu_0\|_{\mathcal{F}\dot{H}^\frac{1}{2}}= \frac{\|\varphi\|_{\mathcal{F}\dot{H}^\frac{1}{2}}}{\sqrt{2|\tilde{e}|}\|\varphi\|_{L^2}}\|\nabla v_0\|_{L^2}.\end{aligned}$$ We complete the proof.
Acknowledgments {#acknowledgments .unnumbered}
---------------
M.H. was supported by JSPS KAKENHI Grant Number JP19J13300. S.M. was supported by JSPS KAKENHI Grant Numbers JP17K14219, JP17H02854, JP17H02851, and JP18KK0386.
[99]{} J. Bergh and J. Löfström, Interpolation spaces. An introduction. Grundlehren der Mathematischen Wissenschaften, No. 223. Springer-Verlag, Berlin-New York, 1976. x+207 pp. MR0482275 M. Christ and M. Weinstein, Dispersion of small amplitude solutions of the generalized Korteweg-de Vries equation. J. Funct. Anal. **100** (1991), no. 1, 87–109. MR1124294 V. D. Dinh, Existence, stability of standing waves and the characterization of finite time blow-up solutions for a system NLS with quadratic interaction. Nonlinear Anal. **190** (2020), 111589, 39 pp. MR3991960 B. Dodson, Global well-posedness and scattering for the mass critical nonlinear Schrödinger equation with mass below the mass of the ground state. Adv. Math. **285** (2015), 15891618. MR3406535 D. Fang, J. Xie, and T. Cazenave, Scattering for the focusing energy-subcritical nonlinear Schrödinger equation. Sci. China Math. **54** (2011), no. 10, 20372062. MR2838120 J. Ginibre, T. Ozawa, and G. Velo, On the existence of the wave operators for a class of nonlinear Schrödinger equations. Ann. Inst. H. Poincar Phys. Théor. **60** (1994), no. 2, 211239. MR1270296 J. Ginibre and G. Velo, Smoothing properties and retarded estimates for some dispersive evolution equations. Comm. Math. Phys. **144** (1992), no. 1, 163–188. MR1151250 M. Hamano, Global dynamics below the ground state for the quadratic Schrödinger system in 5d, preprint, arXiv:1805.12245 R. A. Hunt, On $L(p,q)$ spaces. Enseignement Math. (2) **12** (1966) 249–276. MR0223874 T. Inui, N. Kishimoto, and K. Nishimura, Scattering for a mass critical NLS system below the ground state with and without mass-resonance condition, preprint, arXiv:1810.07904 T. Kato, An $L^{q,r}$-theory for nonlinear Schrödinger equations, Spectral and scattering theory and applications, Adv. Stud. Pure Math. **23**, Math. Soc. Japan, Tokyo, (1994), 223–238. MR1275405 M. Keel and T. Tao, Endpoint Strichartz estimates. Amer. J. Math. **120** (1998), no. 5, 955–980. MR1646048 C. E. Kenig and F. Merle, Global well-posedness, scattering and blow-up for the energy-critical, focusing, non-linear Schrödinger equation in the radial case. Invent. Math. **166** (2006), no. 3, 645675. MR2257393 C. E. Kenig, G. Ponce, and L. Vega, Well-posedness and scattering results for the generalized Korteweg-de Vries equation via the contraction principle, Comm. Pure Appl. Math. **46** (1993), no. 4, 527–620. MR1211741 R. Killip, S. Masaki, J. Murphy, and M. Visan, Large data mass-subcritical NLS: critical weighted bounds imply scattering. NoDEA Nonlinear Differential Equations Appl. **24** (2017), no. 4, Art. 38, 33 pp. MR3663612 R. Killip, T. Tao, and M. Visan, The cubic nonlinear Schrödinger equation in two dimensions with radial data. J. Eur. Math. Soc. (JEMS) **11** (2009), no. 6, 12031258. MR2557134 R. Killip, M. Visan, and X. Zhang, The mass-critical nonlinear Schrödinger equation with radial data in dimensions three and higher. Anal. PDE **1** (2008), no. 2, 229266. MR2472890 S. Masaki, A sharp scattering condition for focusing mass-subcritical nonlinear Schrödinger equation. Commun. Pure Appl. Anal. **14** (2015), no. 4, 1481–1531. MR3359531 S. Masaki, On minimal nonscattering solution for focusing mass-subcritical nonlinear Schrödinger equation. Comm. Partial Differential Equations **42** (2017), no. 4, 626–653. MR3642096 S. Masaki, Two minimization problems on non-scattering solutions to mass-subcritical nonlinear Schrödinger equation, preprint, arXiv:1605.09234. S. Masaki and J. Segata, Existence of a minimal non-scattering solution to the mass-subcritical generalized Korteweg-de Vries equation. Ann. Inst. H. Poincaré Anal. Non Linéaire **35** (2018), no. 2, 283–326. MR3765544 K. Nakanishi and T. Ozawa, Remarks on scattering for nonlinear Schrodinger equations. NoDEA Nonlinear Differential Equations Appl. **9** (2002), no. 1, 45–68. MR1891695 R. O’Neil, Convolution operators and L(p,q) spaces. Duke Math. J. **30** (1963), 129–142. MR0146673 R. Strichartz, Restrictions of Fourier transforms to quadratic surfaces and decay of solutions of wave equations. Duke Math. J. **44** (1977), no. 3, 705–714. MR0512086 M. Visan, The defocusing energy-critical nonlinear Schrödinger equation in higher dimensions. Duke Math. J. **138** (2007), no. 2, 281–374. MR2318286 M. Visan, Global well-posedness and scattering for the defocusing cubic nonlinear Schrödinger equation in four dimensions. Int. Math. Res. Not. IMRN 2012, no. 5, 1037–1067. MR2899959
[^1]: Suppose that we have a criteria “if $(u,v)$ satisfies a condition $P$ then $(u,v)$ scatters." Then, it actually reads as “if there exists $\lambda>0$ such that $(u_{[\lambda]},v_{[\lambda]})$ satisfies a condition $P$ then $(u,v)$ scatters." The latter criteria is scaling invariant in such a sense that the validity of its assumption is left invariant under the scaling.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The Progressive-X algorithm, Prog-X in short, is proposed for geometric multi-model fitting. The method interleaves sampling and consolidation of the current data interpretation via repetitive hypothesis proposal, fast rejection, and integration of the new hypothesis into the kept instance set by labeling energy minimization. Due to exploring the data progressively, the method has several beneficial properties compared with the state-of-the-art. First, a clear criterion, adopted from RANSAC, controls the termination and stops the algorithm when the probability of finding a new model with a reasonable number of inliers falls below a threshold. Second, Prog-X is an any-time algorithm. Thus, whenever is interrupted, e.g. due to a time limit, the returned instances cover real and, likely, the most dominant ones. The method is superior to the state-of-the-art in terms of accuracy in both synthetic experiments and on publicly available real-world datasets for homography, two-view motion, and motion segmentation.'
author:
- |
Daniel Barath$^{12}$, and Jiri Matas$^{1}$\
$^1$ Centre for Machine Perception, Department of Cybernetics\
Czech Technical University, Prague, Czech Republic\
$^2$ Machine Perception Research Laboratory, MTA SZTAKI, Budapest, Hungary\
[barath.daniel@sztaki.mta.hu]{}
bibliography:
- 'egbib.bib'
title: 'Progressive-X: Efficient, Anytime, Multi-Model Fitting Algorithm'
---
Introduction
============
The multi-class multi-model fitting problem is to interpret a set of input points as the mixture of noisy observations originating from multiple model instances which are not necessarily of the same class. Examples of this problem are the estimation of $k$ planes and $l$ spheres in a 3D point cloud; multiple line segments and circles in 2D edge maps; a number of homographies or fundamental matrices from point correspondences; or multiple motions in videos. Robustness is achieved by considering the assignment to one or multiple outlier classes.
![The Prog-X multi-model fitting pipeline gets a point set as input (a). It then repeats three steps. *First*, proposal of putative hypotheses (b) considering the active instance set (c), i.e. the compound model. In (b), the blue points are inliers of the putative instance, and the green ones are that of the compound one. *Second*, fast rejection of redundant hypotheses. *Third*, optimization by integrating new hypotheses, re-assigning points, maintaining model parameters, and removing unnecessary instances. In (c) and (d), the color codes the assignment to instances. []{data-label="fig:approach_new"}](assets/Progressive-X_Diagram.pdf){width="0.99\columnwidth"}
Multi-model fitting has been studied since the early sixties. The Hough-transform [@vc1962method; @illingworth1988survey] perhaps is the first popular method for finding multiple instances of a single class [@guil1997lower; @matas2000robust; @rosin1993ellipse; @xu1990new]. The RANSAC [@fischler1981random] algorithm was extended as well to deal with multiple instances. Sequential RANSAC [@vincent2001detecting; @kanazawa2004detection] detects instances one after another by searching in the point set from which the inliers of the detected instances have been removed. The greedy approach makes RANSAC a powerful tool for finding a single instance, is a drawback in multi-instance setting. Points are assigned not to the best but to the first instance, typically the one with the largest support, for which they cannot be deemed outliers. MultiRANSAC [@zuliani2005multiransac] forms compound hypotheses about $n$ instances. Besides requiring the number $n$ of the instances to be known a priori, the processing time is high, since samples of size $n$ times $m$ in are drawn in each iteration, where $m$ is the number of points required for estimating a model instance and finding an all-inlier simple is rare. Nevertheless, RANSAC-based approaches have a desirable property of interleaving the hypothesis generation and verification steps. Moreover, they have a justifiable termination criterion based on the inlier-outlier ratio in the data which provides a probabilistic guarantee of finding the best instance.
[0.99]{} {width="0.49\columnwidth"} {width="0.49\columnwidth"}
[0.99]{} {width="0.49\columnwidth"} {width="0.49\columnwidth"}
\
[0.99]{} {width="0.49\columnwidth"} {width="0.49\columnwidth"}
[0.99]{} {width="0.49\columnwidth"} {width="0.49\columnwidth"}
Most recent approaches for multi-model fitting [@wongiccv2011; @isack2012energy; @pham2014interacting; @magri2014t; @magri2015robust; @wang2015mode; @magri2016multiple; @barath2018multix; @amayo2018geometric] follow a two-step procedure, first, generating a number of instances using RANSAC-like hypothesis generation. Second, a subset of the generated hypotheses is selected interpreting the input data points the most. This selection is done in various ways. For instance, a popular group of methods [@isack2012energy; @pham2014interacting; @barath2018multix; @amayo2018geometric] optimizes point-to-instance assignments by energy minimization using graph labeling techniques [@boykov2004experimental]. The energy originates from point-to-instance residuals, label costs [@delong2012minimizing], geometric priors interacting among the models [@pham2014interacting], and from the spatial coherence of the points. Another group of methods uses preference analysis based on the distribution of the residuals of data points. [@zhang2007nonparametric; @magri2014t; @magri2015robust; @magri2016multiple]. Also, there are techniques [@wang2015mode; @wang2018searching] approaching the problem as a hyper-graph partitioning where the instances are vertices, and the points are hyper-edges.
*The common part of these algorithms* is the initialization step when model *instances are generated blindly*, having no information about the data. As a consequence, it must be decided by the user whether to consider the worst case scenario and, thus, generate an unnecessarily high number of instances; or to use some rule of thumb, e.g. to generate twice the point number hypotheses. In practice, this is what is usually done. It, however, offers no guarantees for covering all the desired instances. The first method recognizing this problem was Multi-X [@barath2018multix]. It added a new step to the optimization procedure to reduce the number of instances by replacing sets of labels by the corresponding density modes in the model parameter domain. Even though this step allows the generation of more initial instances than before without high computational demand, it still does not provide guarantees, especially, in case of low inlier ratio.
The main contribution of the proposed Prog-X is its any-time nature which originates from interleaving sampling and consolidation of the current data interpretation. This is done via repeated instance proposal minding the already proposed ones; fast rejection of redundant hypotheses; and integration of the new hypothesis into the kept instance set by energy minimization (see Fig. \[fig:approach\_new\] for the basic concept). Moreover, Prog-X adopts the probabilistic guaranties of RANSAC by progressively exploring the data. We use a general instance-to-instance metric based on set overlaps which can be efficiently estimated by the min-hash method [@broder1997resemblance] and modify the model quality function of RANSAC considering the existence of multiple model instances. The method is tested both in synthetic experiments and on publicly available real-world datasets. It is superior to the state-of-the-art in terms of accuracy for homography, two-view motion, motion clustering, and simultaneous plane and cylinder fitting.
Terminology and notation
========================
In this section, we discuss the most important concepts used in this paper. For the sake of generality, we consider multi-class multi-model fitting, thus, aiming to find multiple model instances not necessarily of the same model class. We adopt the notation from [@barath2018multix].
**Multi-class multi-instance model fitting.** First, we show the problem through examples. A simple problem is finding a pair of *line instances* $h_1, h_2 \in \mathcal{H}_l$ interpreting a set of 2D points ${\mathcal{P}}\subseteq \mathbb{R}^2$. Line class $\mathcal{H}_l$ is the space of lines $\mathcal{H}_l = \{(\theta_l, \phi_l, \tau_l), \theta_l = [\alpha \quad c]^{\text{T}}\}$ equipped with a distance function $\phi_l(\theta_l, p) = |\cos(\alpha) x + \sin(\alpha) y + c|\;$ ($p = [x \quad y]^{\text{T}}\in {\mathcal{P}}$) and a function $\tau_l(p_1, ..., p_{m_l}) = \theta_l$ for estimating $\theta_l$ from $m_l \in \mathbb{N}$ data points. *Multi-line fitting* is the problem of finding multiple line instances $\{h_1, h_2, ...\} \subseteq \mathcal{H}_l$, while the *multi-class* case is extracting a subset $\mathcal{H} \subseteq \mathcal{H}_\forall$, where $\mathcal{H}_\forall = \mathcal{H}_l \cup \mathcal{H}_c \cup \mathcal{H}_. \cup \cdots$. The set $\mathcal{H}_\forall$ is the space of all classes including that of lines and circles. Also, the formulation includes the outlier class $\mathcal{H}_o = \{(\theta_o, \phi_o, \tau_o), \theta_o = \varnothing \}$ where each instance has a constant distance to all points $\phi_o(\theta_o, p) = k$, $k~\in~\mathbb{R}^+$ and $\tau_o() = \varnothing$. The *objective of multi-instance multi-class model fitting* is to determine a set of instances $\{ h_i \}_{i = 1}^{n_i} \subseteq \mathcal{H}_\forall$ and labeling $L \in {\mathcal{P}}\rightarrow \mathcal{H}$ assigning each point $p \in {\mathcal{P}}$ to an instance from $\{ h_i \}_{i = 1}^{n_i}$ minimizing energy $E$.
Progressive multi-model fitting
===============================
In this section, a new pipeline is proposed for multi-model fitting. Before describing each step in depth, a few definitions are discussed.
![Examples of the compound model, represented by the union of the distance fields, generated by (left) two lines and (right) a line and circle. Best viewed in color. []{data-label="fig:compound_distance_field"}](assets/two_lines_example.pdf){width="0.89\columnwidth"}
A [*putative model instance*]{} ${h}_p$ is temporary, generated by the proposal engine, not activated to take part in the optimization procedure.
An [*active model instance*]{} ${h}_a$ is an instance whose parameters and support are updated in the optimization procedure.
[*Compound model instance*]{}. Given a set of activate instances $\{ {h}_{a,i} \}_{i = 1}^{n_{a}}$, where $n_{a}$ is the cardinality of the set, the compound model instance ${h_{\cup}}\in \mathcal{H}_\forall$ is defined as the union of the distance fields each generated by an individual active instance ${h}_{a,i}$ ($i \in [1, n_{a}]$). The distance of a point ${\mathbf{p}}$ from ${h_{\cup}}$ is ${\phi}({h_{\cup}}, {\mathbf{p}}) = \min_{i = 1}^{n_{a}} {\phi}({h}_{a,i}, {\mathbf{p}})$.
Examples of compound instances are shown in Fig. \[fig:compound\_distance\_field\]. The color codes the distance (blue – close, red – far). The left plot shows the union of the distance fields generated by two lines. The right one shows that of a circle and a line.
*Proposal engine* $\Sigma : {h_{\cup}}\times \mathcal{P}^* \to \mathcal{H}_\forall^*$ is a function generating a putative instance from the data, using the compound model. Operator $^*$ denotes the power set.
Therefore, function $\Sigma$ gets the compound model and the set of points and outputs one or multiple proposals.
*Multi-instance optimization* procedure $\Theta : \{ {h}_{a,i} \}_{i = 1}^{n_{a}} \times \{ {h}_p \}_{i = 1}^{n_p} \times \mathcal{P}^* \to \{ \widehat{{h}}_{a,i} \}_{i = 1}^{\widehat{n}_a} \times L$ is a function getting the active, the putative instances and the data as input. It returns a set of active instances and a labeling $L : \mathcal{P}^* \to {h}_a^*$ which assigns each point to a single active instance.
Function $\Theta$ gets the active instances $\{ {h}_{a,i} \}_{i = 1}^{n_{a}}$, set of proposals $\{ {h}_p \}_{i = 1}^{n_p}$ and input points. It returns the optimized active instances and the labeling. Putative instances can be activated or rejected. Also, activate instances may be deactivated and removed here. The set of active instances always contain an instance of the outlier class.
Proposal engine
---------------
The *proposal engine* proposes a yet unseen instance in every iteration. For the engine, we choose a recent variant of RANSAC [@fischler1981random]: Graph-Cut RANSAC [@barath2018graph] since it is state-of-the-art and its implementation is publicly available[^1]. Due to assuming local structures, similarly as in [@barath2018multix], we choose NAPSAC [@nasuto2002napsac] to be the sampler inside GC-RANSAC. The main objective of the proposal engine is to propose unseen instances, i.e. the ones which are possibly not among to active ones in $\{ {h}_{a,i} \}_{i = 1}^{n_{a}}$. A straightforward choice for achieving this goal is to prefer instances having a reasonable number of *points not shared with the compound instance* ${h_{\cup}}$. Therefore, we propose a new quality function $\widehat{Q}: \mathcal{H}_\forall \times \mathcal{H}_\forall^* \times \mathcal{P}^* \times \mathbb{R} \to \mathbb{R}$, for RANSAC, measuring the score of a model instance originating from $\mathcal{H}_\forall$ w.r.t. the compound instance (from $\mathcal{H}_\forall^*$), data and a manually set threshold. For the sake of easier understanding, we start from one of the simplest quality functions, i.e. the inlier counting of RANSAC. It is as follows: $
Q_\text{R}({h}, {\mathcal{P}}, \epsilon) = \sum_{{\mathbf{p}}\in {\mathcal{P}}} [{\phi}({h}, {\mathbf{p}}) < \epsilon],
$ where $[.]$ is the Iverson bracket which is equal to one if the condition inside holds and zero otherwise. Based on $Q_\text{R}$, the modified quality function which takes ${h_{\cup}}$ into account and, thus, does not count the shared points is the following: $$\widehat{Q}_\text{R}({h}, {\mathcal{P}}, \epsilon) = \sum_{{\mathbf{p}}\in {\mathcal{P}}} [{\phi}({h}, {\mathbf{p}}) < \epsilon \wedge {\phi}({h_{\cup}}, {\mathbf{p}}) \geq \epsilon].$$ The condition inside holds if and only if the distance of point ${\mathbf{p}}$ from instance ${h}$ is smaller than $\epsilon$ and, at the same time, its distance from the compound model ${h_{\cup}}$ is greater than $\epsilon$. Therefore, $\widehat{Q}_\text{R}$ counts the points which are not inliers of the compound instance but inliers of the new one.
It nevertheless turned out that, in practice, the truncated quality function of MSAC [@torr2002bayesian] is superior to the inlier counting of RANSAC in terms of accuracy and sensitivity of the user-defined threshold. It is as follows: $$Q_\text{M}({h}, {\mathcal{P}}, \epsilon) = |{\mathcal{P}}| - \sum_{{\mathbf{p}}\in {\mathcal{P}}} \min\left(1, \frac{{\phi}({h}, {\mathbf{p}})^2}{\gamma(\epsilon)^2}\right), \; \gamma(\epsilon) = \frac{3}{2} \epsilon.$$ Considering the previously described objective, $Q_\text{M}$ is modified as follows to reduce the score of the points which are shared with the compound model: $$\begin{aligned}
\label{eq:new_quality_function}
\begin{split}
\widehat{Q}_\text{M}({h}, {\mathcal{P}}, \epsilon) = |{\mathcal{P}}| - \\
\sum_{{\mathbf{p}}\in {\mathcal{P}}} \min\left(1, \max\left( \frac{{\phi}({h}, {\mathbf{p}})^2}{\gamma(\epsilon)^2}, 1 - \frac{{\phi}({h_{\cup}}, {\mathbf{p}})^2}{\gamma(\epsilon)^2}\right)\right).
\end{split}\end{aligned}$$ Consequently, for point ${\mathbf{p}}$, the implied score is zero (i) if ${\mathbf{p}}$ is close to both the proposal and the compound instances, and (ii) if it is far from the proposal.
Summarizing this section, we apply GC-RANSAC as a proposal engine with NAPSAC sampling and quality function $\widehat{Q}_\text{M}$ to propose instances one-by-one.
Proposal validation
-------------------
The *validation* is an intermediate step between the proposal and the optimization to decide if an instance should be involved in the optimization. To do so, an instance-to-instance distance has to be defined measuring the similarity of the proposed and compound instances. If the distance is low, the proposal is likely to be an already seen one and, thus, is unnecessary to optimize. In [@barath2018multix], the instances are represented by problem-specific sequences of points and This approach, leads to the question of how to represent models by points. In general, the answer is not trivial and the representation affects the outcome significantly. There, is a straightforward solution for representing instances by point sets. In [@toldo2008robust], the models are described by their preference sets, and the similarity of two instances is defined via their Jaccard score. The preference set of instance ${h}$ is $\mathbf{P}_{h}\in \{0, 1\}^{|{\mathcal{P}}|}$, where $\mathbf{P}_{{h},j}$ is one if the $j$th point is an inlier of ${h}$, otherwise zero ($j \in [1, |{\mathcal{P}}|]$). The proposed criterion of accepting a putative instance is $$\text{J}({h}, {h_{\cup}}) = \frac{|\mathbf{P}_{{h}} \cap \mathbf{P}_{{h_{\cup}}}|}{|\mathbf{P}_{{h}} \cup \mathbf{P}_{{h_{\cup}}}|} > \epsilon_{\text{S}},$$ where $\text{J}$ holds if the Jaccard similarity is higher than a manually set threshold $\epsilon_{\text{S}} \in [0, 1]$ and $\text{J}$ is false otherwise. We choose Jaccard similarity instead of Tanimoto distance [@tanimoto1958elementary; @magri2014t] since representing the instances by sets offers a straightforward way of speeding up the procedure.
Computing the Jaccard similarity in case of having thousands of points is a computationally demanding operation. Luckily, we are mostly interested in recognizing if the overlap of the instances is significant. The min-hash algorithm [@broder1997resemblance] is a straightforward choice for making the processing time of the similarity calculation fast and independent on the number of points. *Therefore, the validation step runs in constant time.*
Multi-instance optimization
---------------------------
The objective of this step is to optimize the set of active model instances whenever a new putative one comes and to decide if this new instance shall be activated or should be rejected. Due to aiming at the most general case, i.e. having multiple classes, there are just a few algorithms [@isack2012energy; @delong2012minimizing; @pham2014interacting; @barath2018multix] which can approach the problem without requiring non-trivial modifications. These algorithms are based on labeling energy-minimization. In general, the major issue of labeling algorithms is their computational complexity, especially, in the case of large label space. In our case, due to proposing instances one-by-one, the label space is always kept small and, therefore, the time spent on the labeling is not significant.
Multi-X [@barath2018multix] could be a justifiable choice. However, in our case, it is simplified to the PEARL algorithm [@isack2012energy; @delong2012fast] since its major contribution is a move in the label space replacing a label set with the corresponding density mode. When having just a few labels, this move is not needed. Thus, we choose PEARL as the optimization procedure.[^2]
Termination criterion
---------------------
In this section, we propose a criterion to determine when to stop with proposing new instances. The adaptive termination criterion of RANSAC is based on $$1 - \mu \leq \left(1 - \left(|\mathcal{I}| / |{\mathcal{P}}| \right)^m \right)^k,
\label{eq:orig_termination}$$ where $\mu$ is the required confidence in the results typically set to $0.95$ or $0.99$; $k$ is the number of iterations; $m$ is the size of the minimal sample; $|\mathcal{I}|$ and $|{\mathcal{P}}|$ are the number of inliers and points, respectively. In RANSAC, Eq. \[eq:orig\_termination\] is formulated to determine $k$, i.e. the number of iterations required, using the current inlier ratio. We, instead, formulate Eq. \[eq:orig\_termination\] to have an estimate of the maximum number of inliers independent on the compound model as follows: $
(|{\mathcal{P}}| - |{h_{\cup}}|) \sqrt[m]{1 - \sqrt[k]{1 - \mu}} \geq |\mathcal{I}|
$ where $|{h_{\cup}}|$ is the inlier number of the compound model. Therefore, $\overline{I}({\mathcal{P}}, {h_{\cup}}, m, k, \mu) = (|{\mathcal{P}}| - |{h_{\cup}}|) \sqrt[m]{1 - \sqrt[k]{1 - \mu}}$ is an upper limit for the inlier number of a not yet found instance with confidence $\mu$ in the $k$th iteration.
It can be easily seen that to distinguish an instance, at least $m + 1$ points have to support it, where $m$ is the size of a minimal sample. Therefore, the algorithm updates $\overline{I}$ in every iteration and terminates if $\overline{I} < m + 1$. *This constraint guaranties* that when the algorithm terminates, the probability of having an unseen model with at least $m + 1$ inliers is smaller than $1 - \mu$.
Note that in practice, it is more convenient to set this limit on the basis of the optimization. For instance, if the optimization does not accept instances having fewer than $20$ inliers, it does not make sense to propose ones with fewer.
Experimental results
====================
In this section, we evaluate the proposed Prog-X method on various computer vision problems that are: 2D line, homography, two-view motion, motion estimation, and simultaneous plane and cylinder fitting. The reported results are the averages over five runs and obtained by using fixed parameters. See Table \[tab:parameter\_table\] for the parameters of Prog-X.
Synthesized tests
-----------------
In this section Prog-X is tested in a synthetic environment. We chose line fitting to 2D points and downloaded the dataset used in [@toldo2008robust]. It consists of three scenes: , and (see Fig. \[fig:gt\_star\_dataset\]). The outlier ratio in all test cases is $0.5$ thus having equal number of inliers and outliers.
[0.9]{} {width="0.14\columnwidth"} {width="0.14\columnwidth"} {width="0.14\columnwidth"}
\
[0.02]{}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
[0.15]{} {width="1.0\columnwidth"} {width="1.0\columnwidth"} {width="1.0\columnwidth"}
-- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- -------
\# fn \# fp \# fn \# fp \# fn \# fp \# fn \# fp \# fn \# fp \# fn \# fp
**1** **0** **1** 4 2 5 2 2 4 4 4 2
**0** **0** **0** **0** **0** **0** **0** **0** **0** **0** 2 **0**
**0** **0** **0** 2 **0** 3 5 5 1 1 9 **0**
**1** **0** **1** 6 2 8 7 7 5 5 15 2
-- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- ------- -------
**Comparison of multi-model fitting algorithms.** We compare Prog-X, Multi-X[^3], [@barath2018multix] RansaCov [@magri2016multiple], RPA [@magri2015robust], T-Linkage[^4] [@magri2014t], and PEARL [@boykov2001fast] on line fitting in this sections. All methods were applied five times to all scenes with fixed parameters. For T-Linkage, RPA and RansaCov, we used the parameters which the authors proposed. We tuned Prog-X, Multi-X, and PEARL to achieve the most accurate average results.
The *worst results* in five runs are visualized in Fig. \[fig:linefitting\_visual\]. Plot (a) shows the ground truth lines and point-to-line assignments. The points of each cluster are drawn by color. The number of false negative, i.e. a line which is not found, and false positive, i.e. a found line which is not in the ground truth set, instances are reported in Table \[tab:linefitting\_table\]. It can be seen, that the proposed method leads to the most accurate results. It finds all but one lines and does not return false positives.
**Any-time property.** To demonstrate the any-time property of Prog-X, we applied the methods which minimize an energy function iteration-by-iteration, i.e. Prog-X, PEARL [@boykov2001fast] and Multi-X [@barath2018multix], to the scene and reported their states in each iteration. All methods were applied five times and, for each, the run with the worst outcome was selected.
[0.45]{} ![*The results iteration-by-iteration.* The compared methods minimizing an energy function iteratively: the proposed Prog-X (red), Multi-X [@barath2018multix] (blue) and PEARL [@boykov2001fast] (orange). **(a)** The number of stored instances are shown (vertical axis) as a function of the iteration number (horizontal). The values in the $1$st iteration are $2\;200$ for PEARL and $550$ for Multi-X. **(b)** The ratio (vertical) of the number of ground truth models covered by an instance and the number of all instances stored are reported. **(c)** The processing times (vertical; in ms) are shown for each iteration (horizontal). The times in the $1$st iteration are $6\;904$ ms for PEARL and $1\;949$ ms for Multi-X. In total, PEARL required $9\;880$ ms, Multi-X $2\;065$ ms and Prog-X $1\;534$ ms. **(d)** The energy divided by the energy in the $1$st iteration (horizontal) is is shown. []{data-label="fig:anytime_plots"}](assets/it_by_it_inst_number.pdf "fig:"){width="0.99\columnwidth"}
[0.45]{} ![*The results iteration-by-iteration.* The compared methods minimizing an energy function iteratively: the proposed Prog-X (red), Multi-X [@barath2018multix] (blue) and PEARL [@boykov2001fast] (orange). **(a)** The number of stored instances are shown (vertical axis) as a function of the iteration number (horizontal). The values in the $1$st iteration are $2\;200$ for PEARL and $550$ for Multi-X. **(b)** The ratio (vertical) of the number of ground truth models covered by an instance and the number of all instances stored are reported. **(c)** The processing times (vertical; in ms) are shown for each iteration (horizontal). The times in the $1$st iteration are $6\;904$ ms for PEARL and $1\;949$ ms for Multi-X. In total, PEARL required $9\;880$ ms, Multi-X $2\;065$ ms and Prog-X $1\;534$ ms. **(d)** The energy divided by the energy in the $1$st iteration (horizontal) is is shown. []{data-label="fig:anytime_plots"}](assets/it_by_it_good_per_all.pdf "fig:"){width="0.99\columnwidth"}
[0.45]{} ![*The results iteration-by-iteration.* The compared methods minimizing an energy function iteratively: the proposed Prog-X (red), Multi-X [@barath2018multix] (blue) and PEARL [@boykov2001fast] (orange). **(a)** The number of stored instances are shown (vertical axis) as a function of the iteration number (horizontal). The values in the $1$st iteration are $2\;200$ for PEARL and $550$ for Multi-X. **(b)** The ratio (vertical) of the number of ground truth models covered by an instance and the number of all instances stored are reported. **(c)** The processing times (vertical; in ms) are shown for each iteration (horizontal). The times in the $1$st iteration are $6\;904$ ms for PEARL and $1\;949$ ms for Multi-X. In total, PEARL required $9\;880$ ms, Multi-X $2\;065$ ms and Prog-X $1\;534$ ms. **(d)** The energy divided by the energy in the $1$st iteration (horizontal) is is shown. []{data-label="fig:anytime_plots"}](assets/it_by_it_time.pdf "fig:"){width="0.99\columnwidth"}
[0.45]{} ![*The results iteration-by-iteration.* The compared methods minimizing an energy function iteratively: the proposed Prog-X (red), Multi-X [@barath2018multix] (blue) and PEARL [@boykov2001fast] (orange). **(a)** The number of stored instances are shown (vertical axis) as a function of the iteration number (horizontal). The values in the $1$st iteration are $2\;200$ for PEARL and $550$ for Multi-X. **(b)** The ratio (vertical) of the number of ground truth models covered by an instance and the number of all instances stored are reported. **(c)** The processing times (vertical; in ms) are shown for each iteration (horizontal). The times in the $1$st iteration are $6\;904$ ms for PEARL and $1\;949$ ms for Multi-X. In total, PEARL required $9\;880$ ms, Multi-X $2\;065$ ms and Prog-X $1\;534$ ms. **(d)** The energy divided by the energy in the $1$st iteration (horizontal) is is shown. []{data-label="fig:anytime_plots"}](assets/it_by_it_relative_energy.pdf "fig:"){width="0.99\columnwidth"}
Fig. \[fig:anytime\_plots\] plots the evolution of the properties iteration-by-iteration. In (**a**), the numbers of stored instances (vertical axis) are plotted as the function of the iteration number (horizontal). It can be seen that PEARL and Multi-X have an unnecessarily high number of instances stored in the first iteration. Then, from the second one, it drops significantly and remains almost constant. For Prog-X, the number of instances increases by one in each iteration. The ground truth number was $11$ in this scene. In (**b**), it is simulated what would happen if a method is stopped before it terminates. In this case, all methods have a set of instances stored. In the plot, the ratio (vertical axis) of the number of desired instances kept and the number of all stored instances is shown as the function of the iteration number (horizontal). A desired instance is one which overlaps with a ground truth one. A ground truth instance can be covered only by one instance. The number of desired instances is, thus, at most the number of ground truth models in the data. It can be seen that PEARL and Multi-X are significantly worse than Prog-X in the first iteration since the number of stored instances is far bigger than the number of desired ones. Even if all the ground truth ones are covered, there are many false positives and, thus, the instances are not usable without further optimization. This ratio for Prog-X is one in all iterations and, as a consequence, *it can be stopped any time and still returns solely desired instances*. In (**c**), the processing times in milliseconds (vertical axis) are reported for each iteration (horizontal). PEARL and Multi-X spend seconds in the first iteration and, then, their processing times drop significantly. This is the expected behavior knowing that their most time-consuming operation is the $\alpha$-expansion algorithm which has to manage a fairly large label space in the first iterations. For Prog-X, the processing time is almost constant since it alternates between the proposal and optimization step with slightly increasing label space. **Summarizing** (a–c), PEARL and Multi-X are not any-time algorithms. To (a) and (b), they can be stopped after the first iteration without significant deterioration in the accuracy. However, to (c), the first iteration requires more processing time than the rest in total. Prog-X can be interrupted at any time; all of the stored instances cover ground truth ones.
Plot (d) shows the change of the energy (vertical axis) iteration-by-iteration (horizontal). We divided the energy in each iteration by the energy in the first one. It can be seen that Prog-X leads to a significantly higher reduction in the energy than the competitor algorithms.
**Comparison with increasing outlier ratio.** To evaluate, how the outlier ratio affects the outcome of multi-model fitting algorithms, we first kept solely the noisy inliers from the scene. Then outliers, i.e. points uniformly distributed in the scene, were added to achieve a given outlier ratio $\nu$. Fig. \[fig:increasing\_outlier\_ratio\] reports (a) the processing time (in milliseconds; vertical axis) and (b) the difference between the ground truth and the obtained instance number (vertical) as the function of the outlier ratio (horizontal). It can be seen that the proposed Prog-X leads to similar processing time as Multi-X but the returned number of instances is significantly closer to the desired one. Consequently, it is the least sensitive to the outlier ratio.
[0.45]{} ![(**a**) The average (of $1\;000$ runs) processing time (in msecs; vertical axis) and (**b**) the difference of ground truth and returned instance numbers (vertical) plotted as the function of the outlier ratio (horizontal) on . In each run, random, uniformly distributed, outliers are generated and the coordinates of the original noisy inliers are perturbed by zero-mean Gaussian-noise with $\sigma = 1.0$ pixels. []{data-label="fig:increasing_outlier_ratio"}](assets/star11_time.pdf "fig:"){width="0.99\columnwidth"}
[0.45]{} ![(**a**) The average (of $1\;000$ runs) processing time (in msecs; vertical axis) and (**b**) the difference of ground truth and returned instance numbers (vertical) plotted as the function of the outlier ratio (horizontal) on . In each run, random, uniformly distributed, outliers are generated and the coordinates of the original noisy inliers are perturbed by zero-mean Gaussian-noise with $\sigma = 1.0$ pixels. []{data-label="fig:increasing_outlier_ratio"}](assets/star11_instance_ratio.pdf "fig:"){width="0.99\columnwidth"}
{width="1.00\columnwidth"}\
{width="1.90\columnwidth"} {width="1.51\columnwidth"}
Real world problems
-------------------
To evaluate the proposed method on real-world problems, we downloaded a number of publicly available datasets. The error is the misclassification error (ME), i.e. the ratio of points assigned to the wrong cluster. In Fig. \[fig:real\_results\], the first images are shown of image pairs which are some of the worst (left plots) and best (right) results of Prog-X running five times on each scene with a fixed parameters (reported in Table \[tab:parameter\_table\]). The shown percentages are the misclassification errors. The white points are labeled as outliers, while the ones with color are assigned to an instance. It can be seen the error originates mostly from missing instances and, thus, the returned ones are usable.
**Two-view motion** fitting is evaluated on the AdelaideRMF motion dataset consisting of $21$ image pairs of different sizes and correspondences manually assigned to motion clusters. In this case, multiple fundamental matrices were fit. In the proposal step, the 7-point algorithm [@hartley2003multiple] was applied for fitting to a minimal sample and the normalized 8-point method [@hartley1997defense] for the polishing steps on non-minimal samples. The average (of 5 runs with fixed parameters) errors and their standard deviations are shown in the first block of Table \[tab:summary\_table\] for each method (from the $2$nd to the $4$th columns). Prog-X is superior to the competitors in all investigated properties. Detailed results for each scene are shown in the top-left plot of Fig. \[fig:detailed\_results\]. It can be seen that Prog-X is always the most or the second most accurate method.
**Homography** fitting is evaluated on the AdelaideRMF homography dataset [@wongiccv2011] used in most recent publications. AdelaideRMF consists of $19$ image pairs of different resolutions with ground truth point correspondences assigned manually to homographies. In the proposal step, the normalized 4-point algorithm [@hartley2003multiple] was used for fitting to a minimal sample and also for the polishing steps. The results are shown in the second block of Table \[tab:summary\_table\] (from the $5$th to the $7$th columns). Prog-X leads to the lowest average error and standard deviation. Similar results can be seen in the top-right plot of Fig. \[fig:detailed\_results\] as before: Prog-X always leads to the lowest or second lowest misclassification errors.
**Motion** segmentation is tested on the $155$ videos of the Hopkins dataset [@tron2007benchmark]. The dataset consists of $155$ sequences divided into three categories: , , and sequences. The trajectories are inherently corrupted by noise, but no outliers are present. Motion segmentation in videos is the retrieval of sets of points undergoing rigid motions contained in a dynamic scene captured by a moving camera. It can be considered as a subspace segmentation under the assumption of affine cameras. For affine cameras, all feature trajectories associated with a single moving object lie in a 4D linear subspace in $\mathbb{R}^{2F}$, where $F$ is the number of frames [@tron2007benchmark]. The results are shown in the third block of Table \[tab:summary\_table\] (from the $8$th to the $10$th columns). Prog-X leads to the lowest average errors. Its standard deviation is the second lowest. Detailed results on the sequences of are put in the bottom of Fig. \[fig:detailed\_results\]. **Plane and cylinder** fitting is evaluated on the dataset from [@barath2018multix]. It consists of LiDAR point clouds of traffic signs, their columns and the neighboring points. Points were manually assigned to signs (planes) and columns (cylinders). The proposal step of Prog-X alternately proposes cylinders and planes. The results are in the last three columns of Table \[tab:summary\_table\]. Prog-X obtains the most accurate results, but, most importantly, it is *three orders of magnitude faster* than the second fastest method. The reason is the large number of points in the scenes (from $1\;260$ up to $52\;445$). The processing time of Prog-X, due to being dominated by a number of RANSAC runs, depends linearly on the point number. All of the other methods depend approximately quadratically on $n$. In RPA, T-Linkage and RansaCov, the preference calculation has $\mathcal{O}(n^2)$ complexity. In Multi-X and PEARL, if at least $n$ initial instances are generated, the label space is already too large for the inner $\alpha$-expansion to finish early.
Conclusion
==========
The Prog-X algorithm is proposed for geometric multi-class multi-model fitting. The method is fast and superior to the state-of-the-art in terms of accuracy in a synthetic environment and on publicly available real-world datasets for homography, two-motion, and motion segmentation. Additionally, it is an any-time algorithm. Therefore, whenever is interrupted, e.g. due to a strict time limit, the returned instances cover real and, likely, the most dominant ones. The termination criterion, adopted from RANSAC, makes Prog-X robust to the inlier-outlier ratio.
[^1]: <https://github.com/danini/graph-cut-ransac>
[^2]: <https://github.com/nsubtil/gco-v3.0>
[^3]: <https://github.com/danini/multi-x>
[^4]: <http://www.diegm.uniud.it/fusiello/demo/>
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'All real and virtual infrared singularities in the standard analysis of the perturbative Quantum Electrodynamics (like that of Yennie–Frautschi–Suura) are associated with photon emissions from the external legs in the scattering process. External particles are stable, with the zero decay width. Such singularities are well understood at any perturbative order and are resummed. The case of production and decay of the semi-stable [*neutral*]{} particles, like the $Z$-boson or the $\tau$-lepton, with the narrow decay width, $\Gamma/M \ll 1$, is also well understood at any perturbative order and soft-photon resummation can be done. For an absent or loose upper cut-off on the total photon energy $\omega$, production and decay processes of the semi-stable (neutral) particles decouple approximately and can be considered quasi-independently. In particular, the soft-photon resummation can be done separately for the production and the decay, treating a semi-stable (neutral) particle as stable. QED interference contributions between the production and decay stages are suppressed by the $\Gamma/M$ factor. If experimental precision $\omega$ is comparable with or better than $\Gamma/M$, these interferences have to be included. In the case of $\omega \ll \Gamma$ decoupling of production and decay does not work any more and the role of semi-stable particles is reduced to the same role as that of other internal off-shell particles. So far, consistent treatment of the soft photon resummation for semi-stable [*charged*]{} particles like the $W^\pm$ bosons is not available in the literature, and the aim of this work is to present a solution to this problem. Generally, this should be feasible because the underlying physics is the same as in the case of neutral semi-stable resonances – in the limit of $\Gamma/M \ll 1$ the production and decay processes for charged particles also necessarily decouple due to long lifetime of intermediate particles. Technical problems to be solved in this work are related to the fact that semi-stable charged particle are able to emit photons. Practical importance of the presented technique to the $e^+e^-\to W^+W^-$ process at the Future electron–positron Circular Collider (FCC-ee) is underlined.'
---
=1
[**IFJPAN-IV-2019-5**]{}
[**QED Exponentiation for quasi-stable charged particles: the $e^-e^+\to W^-W^+$ process${^\dag}$**]{}
[**S. Jadach$^{a}$,**]{} [**W. Płaczek$^{b}$**]{} [*and*]{} [**M. Skrzypek$^{a}$**]{}\
[*$^a$ Institute of Nuclear Physics, Polish Academy of Sciences,\
ul. Radzikowskiego 152, 31-342 Kraków, Poland*]{}\
[*$^b$ Institute of Applied Computer Science, Jagiellonian University,\
ul. Łojasiewicza 11, 30-348 Kraków, Poland*]{}
- This work is partly supported by the Polish National Science Center grant 2016/23/B/ST2/03927 and the CERN FCC Design Study Programme.
[ **IFJPAN-IV-2019-5** ]{}
Introduction
============
The Standard Model Electroweak (EW) Field Theory was confirmed as the correct physics theory of electromagnetic and weak interactions between elementary particles by precision measurements of the LEP experiments [@Alcaraz:2007ri; @Schael:2013ita]. The LEP data were precise enough to test all important dynamical properties of the EW theory, such as quantum loop effects, consequences of the renormalisation, multiple photon emission, etc. In particular, EW gauge cancellations and quantum loop effects were verified experimentally at LEP in the $e^+e^- \to W^+W^-$ process at the precision tag for the total cross section at the level of $0.3$–$05\%$. The mass of the $W$-boson was also measured directly with the precision of $33\,$MeV.
The electron-positron Future Circular Collider (FCC-ee) [@Abada:2019lih; @Abada:2019zxq], considered as the future project at CERN, will be able to produce the number of $W$-boson pairs by a factor of $10^3$ higher than at LEP. This will serve to determine total $WW$ cross section, the mass and width of $W$ with the unprecedented precision and search for any anomalous phenomena beyond the Standard Model (SM) of the EW and strong interactions. Obviously, analysing FCC-ee data will also require new SM perturbative calculations for the $e^+e^- \to W^+W^-$ process, much more precise than these available at the LEP era [@Blondel:2018mad; @Blondel:2019qlh] The precision tag expected in FCC-ee experiments is at the level of about $0.01\%$, a factor of $10$ better than at LEP. This will require to go beyond the state of the art of the LEP era in the calculations of the SM predictions for the $e^+e^- \to W^+W^-$ or $e^+e^- \to 4f$ processes.
For general discussion of the theoretical issues in the $W$-pair production process the reader should consult the reviews of refs. [@Grunewald:2000ju; @Placzek:2002ft]. In particular, the delicate question of the EW gauge invariance for the Dyson summation leading to imaginary part of the $W$ and $Z$ propagators is covered there.
Here we shall focus on the important QED part of the EW/SM perturbative corrections to the $W$-pair production process. More precisely, on this part of the QED corrections which is related to soft and collinear (SC) singularities for real and virtual photon emissions on the external legs[^1]. According to the accumulated knowledge on the SC photonic contribution, it is quite clear that they factorise either at the amplitude level, or for the differential distributions and can be calculated separately to a much higher order than the remaining genuine EW corrections[^2]. This is very convenient, because SC contributions are much bigger numerically than genuine EW corrections, simpler to calculate, and can be resummed to the infinite order. Once separation of the QED and EW parts is established, resummation of some higher-order contributions in each of these two classes can be done independently. The important nontrivial final step is then merging/matching them in the final results.
There is little doubt that the factorisation and resummation of the QED soft/collinear corrections is the key to the success in the high-precision calculations of the SM predictions for the $W$-pair production process at FCC-ee.
There are four classes of QED corrections to the $W$-pair production and decay process: initial-state corrections (ISR), final-state corrections (FSR) in the decays of two $W^\pm$, final-state Coulomb corrections (FSC) and the so-called non-factorisable interferences (NFI) between the production and the decays (IFI) and between two $W^\pm$ decays (FFI). The IFI corrections are suppressed due to relatively long lifetime of $W$’s and FFI due to large space separation. The effects due to ISR are numerically the biggest but also easier to control, while the FSR effects can be also quite sizeable for typical experimental cut-offs.
The IFI and FFI interferences are small, suppressed by the factor $\Gamma_W/M_W$ away from the $WW$-production threshold, strongly cut-off dependent and algebraically most complicated. At LEP they could be neglected but for the FCC-ee precision they have to be handled with great care! The relative narrowness of the $W$ boson resonance not only causes suppression of the QED interferences, but also provides for the expansion in terms of $\Gamma_W/M_W\sim$ [${\cal O}(\alpha)$]{} of the matrix element of the $e^+e^- \to 4f$ process into the numerically biggest and physically most interesting double-resonant $e^+e^- \to W^+W^-$ part, and less important single-resonant and non-resonant background parts. In the following we shall refer to them as the double-pole (DP), single-pole (SP) and non-pole (NP) contributions, as it was common in the LEP-era literature.
The above pole expansion (POE) in the powers of $\Gamma_W/M_W$, disentangling the DP, SP and NP components at the scattering amplitude-level is very useful because it allows for each of these three components to calculate the genuine EW corrections at a different perturbative order and to perform resummation of the QED soft/collinear contributions at a different sophistication level. In the final stage of the calculation, the best way is to sum POE contributions coherently at the amplitudes level, before summing over spin and taking modulus squared, rather than for differential cross sections, thus avoiding proliferation of many interference terms.
At the time of LEP experiments, two solutions based on the pole expansion were worked out, in which the [${\cal O}(\alpha^1)$]{} EW corrections were complete only for the DP component $e^+e^- \to W^+W^-$ of the $e^+e^- \to 4f$ process. One of them, nicknamed [[KandY]{}]{} [@Jadach:2001mp; @Jadach:2000kw], was based on the combination of [[YFSWW]{}]{}3[@Jadach:1996hi; @Jadach:2001uu] Monte Carlo (MC)[^3] for the $e^+e^- \to W^+W^-$ and $W^\pm$-decay processes with another MC program [[KORALW]{}]{}[@Jadach:1998gi] for the remaining background. The multiphoton emission for ISR, including higher orders, was implemented using the soft-photon resummation inspired by the Yennie–Frautschi–Suura (YFS) work[@Yennie:1961ad]. The QED FSR was added in $W$ decays using the [[PHOTOS]{}]{} program [@Jadach:1993hs; @Barberio:1990ms]. Another POE-based solution was that of [[RACOONWW]{}]{} [@Denner:2000bj; @Denner:2002cg], also with the complete EW [${\cal O}(\alpha^1)$]{} corrections implemented only for the signal $e^+e^- \to W^+W^-$ process and not for the background part.
Implementation of QED corrections in [[RACOONWW]{}]{} was very different from that in [[KandY]{}]{}. On the one hand, [[RACOONWW]{}]{} was using exact matrix element for the entire $e^+e^- \to 4f\gamma$ process but it was lacking sophisticated soft photon resummation of the [[KandY]{}]{}. For more detailed comparison of the two approaches see the review of ref. [@Placzek:2002ft] or more recent of ref.[@Jadach:2019bye]. Both approaches were instrumental in the analysis of the LEP data for the $e^+e^- \to W^+W^-$ process [@Schael:2013ita], where the gauge cancellations and the quantum effects of the EW theory were tested experimentally for the first time.
Both approaches, [[KandY]{}]{} and [[RACOONWW]{}]{}, neglect terms of [${\cal O}(\alpha\Gamma_W/M_W)$]{}. The QED NFI interferences between $W$ production and decays were either neglected completely ([[KandY]{}]{}) or included in the soft-photon approximation ([[RACOONWW]{}]{}) without resummation. The overall precision of these calculations was about $0.3$–$0.5\%$. The FCC-ee experiments will require new calculations with the precision tag below $0.1\%$, thus adding missing $\alpha\Gamma_W/M_W$ corrections, electroweak corrections to the DP component, a more advanced QED factorisation/resummation scheme, subleading QED corrections and more will be needed [@Blondel:2018mad; @Blondel:2019qlh]. In particular, inclusion the QED NFI corrections in the fully exclusive way[^4], taking into account the $\Gamma_W/M_W$ suppression, will be necessary.
The aim of the present work is to work out a new methodology of the soft photon resummation including NFI corrections for charged unstable particles, similarly as it was done for the production and decay of the narrow neutral $Z$-boson in the process $e^+e^-\to f\bar{f}+n\gamma$ with a built-in $\Gamma_Z/M_Z$ suppression for the QED initial-final interferences (IFI) at any perturbative order [@Jadach:1999vf; @Jadach:2000ir]. This method was already tested for the $Z$ resonance in the Monte Carlo event generator [[KKMC]{}]{}[@Jadach:1999vf]. Its matrix element is built according to the so-called coherent exclusive exponentiation (CEEX) scheme, in which factorisation of the infrared (IR) divergences is done entirely at the amplitude level (before squaring and spin-summing). The older version of the exclusive exponentiation (EEX) of refs.[@Jadach:1988gb; @Jadach:1991dm] was done at the level of differential distributions for the same $e^+e^-\to f\bar{f}+n\gamma$ process and features multiphoton resummation of ISR and FSR. Both approaches, CEEX and EEX, are inspired by the pioneering work of Yennie–Frautschi–Suura [@Yennie:1961ad].
In the present work we shall generalise the CEEX scheme to the case of any number of narrow [*charged*]{} intermediate resonances, like the $W$-boson – the scheme is however quite general and applies to any charged resonance of any spin. The new CEEX scheme provides exclusive (unintegrated) description for multiple real photons of any energy, for $E_\gamma\sim \Gamma_W$, $E_\gamma \ll \Gamma_W$ and $E_\gamma\sim \sqrt{s}$, with all QED interferences between production and decays properly accounted for. Multiple real and virtual photon emission from all external stable particles and the intermediate [*semi-stable*]{} charged resonance will be described correctly in the soft photon limit and summed up to the infinite order. As in the case of CEEX of refs. [@Jadach:1999vf; @Jadach:2000ir], its present extension will provide for a well-defined methodology of incorporating [*non-soft*]{} contributions[^5] (including the genuine EW corrections) calculated up to a finite perturbative order into multiphoton amplitudes of the soft-photon resummation scheme. In particular, sizeable but easier to calculate QED non-soft collinear contributions can also be included easily up to an arbitrarily high order.
The consistent resummation of the apparently IR-divergent contribution due to photon emissions from the semi-stable intermediate charged particle (narrow resonances) in the perturbative expansion is a non-trivial issue. Let us first consider $\Gamma\to 0$ limit. The best illustrative example is that of the $\tau^\pm$-pair production and decay in the $e^+e^-$ annihilation where a time scale of the $\tau$-pair formation (production process) is shorter than the $\tau$ lifetime by at least a factor of $\Gamma_\tau/m_\tau \simeq 3 \cdot 10^{-12}$, hence photons emitted in these two stages get completely decoupled and the QED effects in the production and the decay can be implemented separately [@Jadach:1984iy; @Jadach:1991ws; @Jadach:1999vf].
The situation in the $W$-pair production is similar but the suppression factor $\Gamma_W/M_W\simeq 0.026$ is not that small. The QED interferences are therefore expected to be of the order of $\alpha\Gamma_W/M_W\simeq 2\cdot 10^{-4}$. In LEP experiment this size could be neglected, but for the FCC-ee precision, effects of this size have to be calculated and taken into account. Moreover, such interferences depend on kinematical cut-offs – from the experience with the $Z$-boson case we know that they may grow by a factor of $2$–$5$ even for relatively mild cut-offs on photon energies. Also, in the case when photon energy resolution $\omega$ of the detector approaches $2\,$GeV, which is the case for FCC-ee detectors, photon emission from FSR in the production process and from $W$ decays cannot be separated and treated in the soft photon approximation, consequently the off-shell $W$’s have to be treated the same way as other internal exchanges in the $e^+e^- \to 4f$ process.
Our aim is to construct a variant of CEEX spin amplitudes in which we profit as much as possible from the smallness of $\Gamma_W/M_W$ and the classic YFS soft-photon limit for the entire $e^+e^- \to 4f$ process is correctly reproduced for $\omega \ll \Gamma_W$. The basic technical problem will be that if we want to treat $W$’s as stable particles in the $W$-pair production process with the zero width, then amplitudes of photon emission from $W$ must be IR-singular, while for the semi-stable $W$’s they are not (the $W$ width acts as a IR regulator). Our aim is to reconcile these two contradictory situations in a single algebraic framework.
In the [[YFSWW]{}]{}3 program, photon emission in the $e^+e^- \to W^+W^- \to 4f$ process was treated in a similar way as in the above $\tau$-pair production and decay, except that $W$ invariant masses were not fixed but modelled according to the Breit–Wigner shape. The QED matrix element in [[YFSWW]{}]{}3 for $e^+e^-\to W^+W^-$ with the soft photon resummation is of the EEX type, including ISR, FSR and IFI. Decays of $W^\pm$s are supplemented with additional photons using [[PHOTOS]{}]{}. However, it could be easily replaced with the multiphoton MC implementation of the EEX of the [[WINHAC]{}]{} program [@Placzek:2003zg]. Once EEX implementation is available in the $W$-pair process for the production and decays, the new CEEX matrix element developed in the present work can be introduced using an additional multiplicative MC weight[^6], without any change in the underlying MC program. The above would be the solution for the resummed QED corrections of the DP part of the $e^+e^- \to 4f$ process. The [${\cal O}(\alpha^1)$]{} and [${\cal O}(\alpha^2)$]{} genuine EW corrections can be added in the on-shell approximation within the CEEX matrix element in the similar way as it was done for the $e^+e^-\to 2f$ process in refs. [@Jadach:1999vf; @Jadach:2000ir]. So far only [${\cal O}(\alpha^1)$]{} EW corrections are available. In order to exploit fully FCC-ee data, the [${\cal O}(\alpha^2)$]{} EW corrections will be needed. As pointed out in ref. [@Blondel:2018mad], the clear and clean separation of the QED and the genuine EW correction at any perturbative order is a useful built-in feature of the CEEX factorisation/resummation scheme.
The single-pole group of diagrams of the $e^+e^- \to 4f$ process process is separated at the amplitude level in the CEEX scheme. It would be enough to include the genuine EW corrections to the SP part at [${\cal O}(\alpha^1)$]{}. They are in principle known, because they are part of the [${\cal O}(\alpha^1)$]{} corrections to $e^+e^- \to 4f$ process in refs. [@Denner:2005es; @Denner:2005fg], although it may be not simple to disentangle them from the rest of the existing calculations. For the non-pole part of the $e^+e^- \to 4f$ process it would probably be enough to take it at the tree-level as far as the genuine EW corrections are concerned and take care of the QED corrections only, either in the CEEX or EEX scheme.
In this work, the CEEX scheme will be defined only for the DP part leaving the easier SP and NP variants for the future development. On the other hand, we shall also discuss in a more detail the explicit algebraic relation between the CEEX scheme and the EEX scheme of the [[YFSWW]{}]{}3. This will provide better understanding of the theoretical foundation of the existing EEX scheme of the [[YFSWW]{}]{}3. The main result of this work is, however, that it provides an important building block for the future high-precision calculations for the $W$ pair production process, and also for any other process with narrow charged resonances.
Close to the $WW$ threshold, where the $W$ mass is planned in the FCC-ee experiment to be measured with the $\leq 0.5\,$MeV precision (using the total cross section [@Abada:2019lih; @Abada:2019zxq]), the problem is that the pole expansion for the non-QED part of the scattering matrix element is not efficient any more. The partial suppression of the QED IFI and FFI corrections will still work close to threshold as long as resonant curves of $W$’s are not fully “distroyed” by the threshold cut-off. However, as shown in works based on the effective field theory (EFT) [@Beneke:2007zg; @Actis:2008rb], near the threshold one may exploit expansion in the Lorentz velocity $\beta=\sqrt{s-4M_W^2}/2M_W \ll 1$ of the $W$’s in order to reduce substantially a number of diagrams, such that higher-order EW and QED corrections are again within the reach of practical evaluations. This kind of expansion should be exploited in the standard diagrammatic approach as well.
Summarising, a combination of the pole expansion and of the QED exclusive exponentiation has already proven to be an economical solution for precision calculations of the SM prediction for the $W$-pair production process at LEP and is the best candidate for the further development in future electron-positron collider projects, especially for FCC-ee. The inclusion of the QED interferences between the $W$ production and decays, and of other missing corrections of the order of $\alpha\Gamma_W/M_W$ will require applying a more sophisticated soft/collinear photon factorisation and resummation scheme, combined with POE. We propose here a new solution based on the coherent exclusive exponentiation, CEEX, in which resummation of the infrared (IR) divergences is done entirely at the amplitude level. The interesting feature of this new scheme is that the $\Gamma_W/M_W$ suppression of the QED interferences between production and decay is a built in feature valid in any order and at any photon energy scale/resolution, all over the entire multiphoton phase space. The new scheme is similar to the CEEX scheme previously formulated and successfully applied to the case of the neutral intermediate resonances (the $Z$-boson).
One should not give up on the more traditional EEX schemes, however. We shall discuss briefly alternative solutions within the traditional EEX schemes (extensions of EEX of [[YFSWW]{}]{}3). We shall also examine approximations or simplification done in the transition from the CEEX to EEX schemes, and between various variants of them.
Concluding, this work provides an important building block for the future high-precision Standard Model calculations for the $W$-pair production process at the future $e^+e^-$ colliders.
The paper is organised as follows. In Section 2 we describe the pole expansion for the $W$-pair production process. Section 3 is devoted to a general discussion of various kinds of the exclusive QED exponentiation and a problem of photon emission from an intermediate semi-stable charged particle. In Section 4 we present details on the CEEX scheme for the process $e^+e^- \to 4f$ involving intermediate resonant $W$-bosons. Relations between the CEEX and EEX schemes are discussed in Section 5. Section 6 contains summary and outlook of our work. Finally, detailed derivations of factoring multiphoton radiation from an intermediate semi-stable charged particle, resummation of real-photon emissions and the virtual YFS form-factor for the pertinent process are given in Appendices A, B and C, respectively.
Shorter version of this work was reported in the conference materials of Ref. [@Jadach:2019yhw].
Pole expansion for $W$-pair production
======================================
As pointed out by R. Stuart [@Stuart:1995zr], it is always possible to decompose the matrix element into a combination of Lorentz covariant tensors and Lorentz invariant functions. If unstable particles are involved in a process, one can then perform a Laurent expansion about complex poles corresponding to those unstable particles. However, only the Lorentz invariant functions (mathematically, analytic functions of complex variables) are subject to this expansion, while the Lorentz covariant and spinor structure of the matrix element should remain untouched. In the so-called leading-pole approximation (LPA) one retains only the leading terms in the above expansion, neglecting the rest of the Laurent series. As discussed in Ref. [@Stuart:1995zr], the whole procedure does not violate gauge invariance of the matrix element. This is guaranteed by the fact that all terms in the pole expansion are independent of each other, e.g. in the case of two unstable particles, the doubly-resonant terms are independent of the singly-resonant and non-resonant ones, therefore there cannot be gauge cancellations between those terms. In Ref. [@Stuart:1995zr], the process of $Z$-pair production and decay was presented as an example.
Here, we discuss the process of $W$-pair production and decay: $$e^-(p_1) + e^+(p_2) \longrightarrow W^-(Q_1) + W^+(Q_2)
\longrightarrow f_1(q_1) + \bar{f}_2(q_2) + f_3(q_3) + \bar{f}_4(q_4),
\label{process}$$ where $W^-$ decays into $f_1,\bar{f}_2$ and $W^+$ into $f_3,\bar{f}_4$. At the lowest order, the minimum gauge invariant subset of Feynman diagrams needed for this process is the so-called CC11-class of graphs. It includes apart from doubly-resonant $WW$ graphs (the so-called CC03) also singly-resonant $W$ graphs. Below we discuss how to apply the pole expansion this process.
Since we are interested only in LPA (a double-pole approximation in this case) we start from extracting a part of the full matrix element that can give rise to doubly-resonant contributions (the rest will drop in LPA anyway). It can be written as follows: $$\begin{aligned}
{\cal M} &=& \sum_i \left[\bar{v}_e(p_2) T^i_{\mu\nu} u_e(p_1) \right]
M_i(s,t,s_1,s_2) \nonumber \\
& & \times D_W^{-1}(s_2)
\left[\bar{u}_{f_3}(q_3) \gamma^{\mu} V_{Wf}(s_2)
\omega_- v_{f_4}(q_4) \right] \nonumber \\
& & \times D_W^{-1}(s_1)
\left[\bar{u}_{f_1}(q_1) \gamma^{\nu} V_{Wf}(s_1)
\omega_- v_{f_1}(q_1) \right] \;,
\label{2res-me}\end{aligned}$$ where $$D_W(s) = s - M_W^2 + \Pi_W(s)
\label{Wpropag}$$ is a Dyson-resumed $W$ propagator with $\Pi_W(s)$ being the $W$ self-energy correction. In the above we have used the following notation: $$\begin{aligned}
& & \omega_- = \frac{1}{2}(1 - \gamma_5), \nonumber \\
& & s_1 = Q_1^2,\; s_2 = Q_2^2,
\label{notat1} \\
& & Q_1 = q_1 + q_2, \; Q_2 = q_3 + q_4. \nonumber\end{aligned}$$ $T^i_{\mu\nu}$ are the Lorentz covariant tensors spanning the tensor structure of the matrix element, while $M_i,\,\Pi_W,\,V_{Wf}$ are Lorentz scalars that are analytic functions of independent Lorentz invariants of the process. These functions then undergo the Laurent expansion about the complex poles corresponding to a finite-range propagation of two $W$’s. Keeping only the leading terms in the above expansion, we end up with the LPA matrix element [@Jadach:2000kw; @Jadach:2001cz] $$\begin{aligned}
{\cal M}_{\rm LPA} &=& \sum_i \left[\bar{v}_e(p_2) T^i_{\mu\nu} u_e(p_1) \right]
M_i(s,t,s_p,s_p) \nonumber \\
& & \times \frac{F_W(s_p)}{s_2-s_p} \left[\bar{u}_{f_3}(q_3)
\gamma^{\mu} V_{Wf}(s_p) \omega_- v_{f_4}(q_4)\right]\nonumber \\
& & \times \frac{F_W(s_p)}{s_1-s_p} \left[\bar{u}_{f_1}(q_1)
\gamma^{\nu} V_{Wf}(s_p) \omega_- v_{f_1}(q_1)\right]\;,
\label{LPA-me}\end{aligned}$$ where the pole position $s_p$ is a solution to the equation $$s - M_W^2 + \Pi_W(s) = 0,\quad F_W(s_p) = [1 + \Pi^{'}_W(s_p)]^{-1}.
\label{spequ}$$ At the lowest order the Lorentz tensors read $$\begin{aligned}
T_{\mu\nu}^{1,2} &=& \gamma^{\lambda}\Gamma_{\lambda\mu\nu}(Q,Q_1,Q_2),
\label{T12} \\
T_{\mu\nu}^3 &=& \gamma_{\mu}(\not\!p_2 - \not\!Q_2) \gamma_{\nu},
\label{T3} \end{aligned}$$ and the Lorentz scalars are $$\begin{aligned}
M_1 &=& e^2 \frac{1}{s},
\label{M1} \\
M_2 &=&-e^2 \frac{s_W}{c_W}[v_e - a_e\gamma_5]
\frac{1}{s-M_Z^2 + iM_Z\Gamma_Z},
\label{M2} \\
M_3 &=& \frac{e^2}{s_W^2}\frac{1}{t} ,
\label{M3} \\
V_{Wf} &=& \frac{eU_{ij}\sqrt{N_c}}{2s_W^2},
\label{Vwf} \end{aligned}$$ where $Q = p_1+p_2,\; s=Q^2,\;t=(p_2-Q_2)^2$, $U_{ij}$ is the CKM matrix element, $N_c$ is the QCD colour factor, $s_W = \sin\theta_W$, $c_W = \cos\theta_W$, $v_e$ and $a_e$ are the vector and axial couplings of a $Z$ boson to electrons, $\Gamma_{\lambda\mu\nu}$ is the $VWW$ coupling ($V=\gamma,Z$): $$\Gamma_{\lambda\mu\nu}(Q,Q_1,Q_2) = (Q+Q_1)_{\nu}g_{\lambda\mu}
+ (Q_2-Q_1)_{\lambda}g_{\mu\nu}
- (Q+Q_2)_{\mu}g_{\nu\lambda}.
\label{TGC}$$ In the scalar function $M_2$ we have also applied LPA to the intermediate $Z$-boson. It is done in a similar way as for $W$’s. The $W$-pole position, up to ${\cal O}(\alpha^2)$, is given by $$s_p = M_W^2 - iM_W\Gamma_W + {\cal O}(\alpha^2),
\label{spappr}$$ where $M_W,\,\Gamma_W$ are the usual on-shell scheme $W$ mass and width, and $F_W=1$. One can easily check that at the lowest order this LPA matrix element has the same form as the CC03 matrix element calculated in ’t Hooft–Feynman gauge and in the constant $W$-width scheme. It was noticed in Ref. [@Beenakker:1994vn] that when the CC03 matrix element is calculated in the axial gauge also singly-resonant terms appear. This indicates that the singly-resonant graphs are needed to guarantee gauge invariance of the matrix element, i.e. that CC03 itself is not gauge-invariant, but one has to take at least CC11 for hadronic, CC10 for semi-leptonic and CC09 for leptonic final states. In the LPA approach described above it does not matter what gauge is used in the calculations. We start from the gauge-invariant matrix element and then apply the pole expansion. In the resulting LPA matrix element all non-double-pole terms drop out.
One of the complications that arises when going to higher orders is the fact that $W$’s are electrically charged and therefore radiate photons. When a real or virtual photon is emitted from the $W$ one has more than just two $W$ propagators in the matrix element and the question is how to apply the pole expansion in such a case. Here, however, one can exploit a partial-fraction decomposition of a product of two propagators, namely: $$\frac{1}{Q^2 - M^2}\frac{1}{{Q'}^2 - M^2}
\equiv \frac{1}{2kQ' + k^2}\frac{1}{{Q'}^2 - M^2}
- \frac{1}{Q^2 - M^2}\frac{1}{2kQ - k^2},
\label{pfd-prop}$$ where $M^2=M^2_W+i\Gamma_W M_W$, $Q,\,Q'=Q-k$ are the $W$ four-momenta before and after radiation of a photon of the four-momentum $k$, respectively[^7]. So, a product of two propagators can be replaced by a sum of single propagators multiplied by eikonal factors. This corresponds to splitting the photon radiation into the radiation in the $W$-production stage and the radiation in the $W$-decay stage. These two stages are separated by the finite-range $W$ propagation. The above decomposition can be applied both to the real and virtual photon emissions. In the case of the real photons the radiation amplitude splits into the sum of the amplitudes corresponding to photon emission in the $WW$-production and two $W$-decays. At the level of the cross section this results in the sum of contributions corresponding to the photon radiation at each stage of the process – the factorisable corrections, and the contributions corresponding to interferences between various stages – the non-factorisable corrections. Similarly, for the virtual corrections, the contributions with photons attached to the same stage give rise to the factorisable corrections, while the ones where photons interconnect different stages of the process contribute to the non-factorisable corrections. In this way all radiative corrections can be split in a gauge-invariant way into the factorisable and non-factorisable ones.
Since the non-factorisable corrections were negligible for the main LEP2 observables one could drop them[^8] and concentrate only on the factorisable ones. For factorisable corrections one can employ the existing calculation for the on-shell $WW$-production and the on-shell $W$-decay. Our aim is to treat the QED corrections according to the YFS exclusive exponentiation procedure and also apply the LPA, described above, in order to obtain the gauge-invariant formulation. How to do this? Extraction of infra-red (IR) contributions for both real and virtual photons can be done in a gauge-invariant way according to the YFS theory for each of the stages separately. These contributions are then sum up to infinite order and result in the so-called YFS form-factor. This means that the YFS form-factors and the IR real-photon $\tilde{S}$-factors involving $W$’s do not have to be taken on-pole but can be calculated like for stable particles. After having done this we can apply the pole expansion to the IR-residuals – the YFS $\bar{\beta}$-functions. We proceed in the way described at the beginning of this section and retain only the leading-pole (double-pole) terms. The ${\cal O}(\alpha)$ LPA matrix element for the real photon contribution reduces, similarly to the lowest order, to the form that can be obtained from the doubly-resonant Feynman graphs with single-photon emission in the ’t Hooft–Feynman gauge. The ${\cal O}(\alpha)$ virtual correction form-factors should, in principle, be evaluated on the complex pole. This would require an analytic continuation of the usual one-loop results to the second Riemann sheet (this may be a technical problem). However, for the aimed LPA accuracy, it is sufficient to use the approximation $s_p \simeq M_W^2$. This would correspond to neglecting terms of ${\cal O}(\frac{\alpha}{\pi}\frac{\Gamma_W}{M_W})$. More details about implementation of the ${\cal O}(\alpha)$ corrections in the $WW$-production process in the MC event generator [[YFSWW]{}]{}3 can be found in Ref. [@Jadach:1996hi].
General discussion {#sec:GenDis}
==================
In this section we collect discussion on various aspects of the photon radiation in the $W$ pair production process, in particular we discuss various exponentiation schemes preparing grounds for defining them explicitly in the following sections. We define more precisely our aims, discuss various constraints, introduce notation and terminology.
The fact that $W$’s are narrow resonances and behave like almost stable particles is of great practical importance for the evaluation of the radiative corrections, because it provides an additional small parameter $\Gamma_W/M_W$ which can be used as an expansion parameter, leading to reduction of the complexity of calculations of radiative corrections. As a result, the dominant double-resonant part of the process (\[eeWW4f\]) can be well approximated as three independent processes: one production process and two decay processes. For the double resonant part it is possible to use simpler on-shell radiative corrections, while for the single-resonant part we may stay at the Born-level or use some crude leading-order (LO) approximations for the radiative corrections. Of course, we have to have at our disposal a method of splitting the Born amplitude and the amplitude with the radiative corrections into the double- and single-resonant parts, without breaking gauge invariance and other elementary principles. The pole expansion (POE) seems to be the best method available. Once POE is used for $W$-pair production process to isolate the double-pole (DP), single-pole (SP) and non-pole (NP) parts, photon emission from the intermediate unstable $W$’s has to be reorganised in a consistent way. In addition, it would be desirable to sum up photon emission from $W$’s to infinite order (exponentiate), for instance using one of EEX or CEEX schemes.
In the following we shall characterise various methods of the known soft photon resummation and then characterise problems related to soft-photon emission from charged semi-stable intermediate particles (resonances), like the $W$-bosons.
Various kinds of exclusive exponentiation {#subsec:VarEex}
-----------------------------------------
[|c|c|c|c|c|]{} Resummation & Formalism & NFI interf. & Implementations & Order\
\
EEX$_{B}$ &[@Jadach:1988rr; @Yennie:1961ad] & – & [YFS1]{} & [${\cal O}(\alpha^1)$]{}\
CEEX$_{B}$ & None & – & None & –\
\
EEX$_{R}$ &[@Jadach:1991dm] & No & [YFS3, KORALZ]{} & [${\cal O}(\alpha^3)$]{}\
CEEX$_{R}$ &[@Jadach:1999vf; @Jadach:2000ir] & Yes & [[KKMC]{}]{} & [${\cal O}(\alpha^2)$]{}\
\
EEX$_{R}$ &[@Jadach:1996hi; @Jadach:2001uu] & No & [YFSWW3]{} & [${\cal O}(\alpha^3)$]{}\
CEEX$_{R}$ & This work & Yes & None & –\
Generally, there are two kinds of exclusive exponentiation schemes: (1) the older one, which we call EEX, in which isolation of IR singularities due to [*real*]{} photons is done for differential distributions (probabilities), as in the classic work of Yennie–Frautschi–Suura (YFS) [@Yennie:1961ad], and (2) the newer one of refs. [@Jadach:1998jb; @Jadach:2000ir; @Jadach:1999vf], referred to as CEEX, in which the same isolation of the real photon IR singularities is done for the amplitudes themselves, that is before squaring and spin-summing them. CEEX has a number of advantages over EEX. The price to pay is that it can be more complicated in the implementation and slower in the numerical evaluation.
Since we are interested mainly in the exclusive exponentiation for the processes with the narrow resonances, it is worth to note that, within EEX and CEEX families, there are two distinct subgroups of implementations which differ rather strongly in the treatment of the narrow resonances (or of sharp $t$-channel peaks). The key difference is in the treatment of the shift of the energy-momentum in the propagator of the resonance due to emission of the real or virtual photons. Let us, for the purpose of this work, call this effect a “recoil effect” or shortly a “recoil”. Within the EEX family there is a baseline variant based on the original YFS work [@Yennie:1961ad], in which the recoil is realised in an order-by-order way. Let us denote them with EEX$_{B}$. Examples of the EEX$_{B}$ variants are: the unpublished MC code [YFS1]{} described in ref. [@Jadach:1988rr] and [[BHLUMI]{}]{} 1.x of ref. [@Jadach:1988ec][^9]. In EEX$_{B}$ the recoil is absent completely at the level of . Then, it is gradually introduced in an order-by-order manner, through the so-called IR-finite $\bbeta$-functions. For instance, in the exact recoil in the differential distribution is realised due to two hard real photons – if there is a third “spectator” hard photon, then its contribution to resonance propagator is simply ignored. The problem is that, from the point of view of the strong variation of the resonance propagator, a photon with the energy of the order of the resonance width $\Gamma$ is already hard! This is why EEX$_{B}$ can be disastrous for narrow resonances, where in order to realise the recoil, it would be mandatory to jump immediately to very high perturbative orders, otherwise the perturbative convergence for the QED corrections would be miserable. EEX$_{B}$ can be a convenient and natural choice if there are no resonances at all.
In the second class of the EEX scenarios, the recoil in the resonance propagator (or sharp $t$-channel exchange) is a built-in feature of the scheme, which is present already in . Let us call such a scheme EEX$_{R}$. It is realised for the first time in the [YFS3]{} event generator [@Jadach:1991dm] and later on included in the [[KORALZ]{}]{} [@Jadach:1993yv], [[KKMC]{}]{} [@Jadach:1999vf] programs and finally in the [[YFSWW]{}]{}3 program [@Jadach:1996hi; @Jadach:2001uu]. The analogous scheme for a process dominated by the $t$-channel was implemented in the [[BHLUMI]{}]{} MC program [@Jadach:1991by; @Jadach:1996is]. In EEX$_{R}$, the total energy-momentum in the resonance propagator (or $t$-channel exchange) includes the contribution from all real photons emitted prior to resonance formation ($t$-channel exchange). This means that for each photon we have to know whether it belongs to resonance production or decay process (ISR or FSR). This is possible because in this scenario one always neglects completely and irreversibly the QED interferences between the ISR and FSR[^10]. Neglecting these interferences may be not so harmful as compared to experimental precision, because they are suppressed by the $\Gamma/M$ factor. The EEX$_{R}$ is obviously very well suited for narrow resonances, as long as we can afford neglecting interference corrections, and we do not attempt to examine experimentally spectra of photons with energies $E_\gamma\simeq \Gamma$.
In the CEEX family of exponentiations there are analogous two sub-classes: either the recoil is implemented in the infinite order (CEEX$_{R}$) or in the order-by-order manner (CEEX$_{B}$). One great advantage of CEEX is that, in the process with the resonant component and the non-resonant background, one may apply CEEX$_{R}$ to the resonant part of the amplitude and CEEX$_{B}$ to the background and add the two coherently afterwards.
Let us comment on the relation of the above schemes to the classic YFS work and the relation of EEX$_{R}$ to other ones. All the above exponentiation schemes are inspired by the classic YFS work [@Yennie:1961ad] in one way or another. However, it is in fact only the EEX$_{B}$ scheme which was formulated explicitly in the original YFS work. CEEX is a non-trivial extension of the YFS exponentiation scheme, see ref. [@Jadach:2000ir] for more discussion. So far, there is no implementation of the CEEX$_{B}$ scheme, while more sophisticated CEEX$_{R}$ is successfully implemented in [[KKMC]{}]{} [@Jadach:1999vf] program for the neutral semi-stable $Z$ boson production and decay in the electron–positron annihilation and recently in the proton–proton collision [@Jadach:2017sqv].
The above inventory of all schemes of the exclusive QED exponentiations and their implementations are summarised in table \[tab:table1\].
Finally, let us note that there is another variant of the EEX$_{R}$ scheme implemented in the [[BHWIDE]{}]{} program of ref. [@Jadach:1995nk], featuring partial implementation of the QED NFI interferences for semi-stable neutral boson exchanges. It will be discussed in the following whether this kind of scheme could be extended to include the QED NFI interferences for the charged semi-stable $W$-boson.
Photons from intermediate semi-stable charged particle {#subsec:PhoInChPar}
------------------------------------------------------
Let us present an introductory discussion on the photon emission from the intermediate charged unstable $W$’s.
In order to better grasp physics of the photon emission from unstable charged particles, let us consider one more time the case of $e^+e^-\to \tau^+\tau^- +n\gamma,\;\; \tau^\pm \to X^\pm$ process. In this case, with $\Gamma_\tau/m_\tau= 2.27\cdot 10^{-12}$, the production and decay processes are well separated in time due to this factor. For instance, the formation time of the $\tau$-pair at $\sqrt{s}=100\,$GeV is $\sim 10^{-24}\,$sec while the $\tau$ lifetime is much longer, $2.9\cdot 10^{-13}\,$sec. This is why the ISR photons emitted from initial beams have no chance to interfere with these of the $\tau$ decays. The FSR photons emitted from the outgoing ultrarelativistic $\tau$’s are quite copiously, because $\ln(s/m_\tau^2)=8.06$, but still, the emissions of the FSR photons and photons in the decays are time-separated by the factor of $\Gamma_\tau/m_\tau= 2.27\cdot 10^{-12}$[^11]. The suppression of the interferences between photon emission from two decays is even stronger, by the factor $\Gamma_\tau/\sqrt{s} \sim 10^{-14}$. Consequently, all practical calculation for QED effects in the $\tau$-pair production and decay process from the production threshold onwards were implemented in the Monte Carlo programs independently for the production and decay parts [@Jadach:1984iy; @Jadach:1991ws; @Jadach:1999vf]. The $\tau$-leptons in the production process are treated in the perturbative/diagrammatic QED calculations and in the phase-space integration as stable particles [*with the fixed mass and the zero decay width*]{}. Photon emission from the unstable intermediate $\tau$’s is of course exponentiated – the same way in the decay parts. Can the above production-decay separation break down? Yes, if the energy resolution in the photon energy (a cut on photon energies) is smaller than the $\tau$ width, that is below $0.003\,$eV, which is experimentally unfeasible.
![ Kinematics of the four-fermion production process with multiple photons. []{data-label="fig:ee4f"}](ee4f.pdf){height="70mm" width="110mm"}
In order to see that the problem of the photon emission from the unstable intermediate $W$’s is not a completely trivial, let us recall a well-known elementary fact [@Yennie:1961ad]: the emission of photons from the stable initial beams and four final fermions can be factorised into a product of the soft factors $\prod_i J_{6f}^{\mu_i}(k_i)$ with the total electric current for all six [*external*]{} particles: $$\label{4fcurrent}
J_{6f}^\mu(k) = \hat{J}_{a}^\mu(k) +\hat{J}_{b}^\mu(k) +\hat{J}_{c}^\mu(k)
+\hat{J}_{d}^\mu(k) +\hat{J}_{e}^\mu(k) +\hat{J}_{f}^\mu(k),$$ where $$\hat{J}_{x}^\mu(k) = \theta_x Q_x
\frac{2p_x^\mu\theta_x +k}{k^2+2k\cdot p_x\theta_x +i\varepsilon},$$ $p_x$ and $Q_x$ are the momentum and charge (in the units of positron charge) of the emitter particle $x$, and $\theta_x=+1,-1$ for the initial- and final-state particle, respectively. For the virtual photons there might be contractions among the pairs of the currents $J_{6f}^{\mu_i}(k_i)$ and $J_{6f}^{\mu_j}(k_j)$, see next sections for the explicit formulation. Fig. \[fig:ee4f\] provides a visual representation of the process of four-fermion production in electron–positron collisions. All possible contractions (loops) for the virtual photons are not explicitly marked there.
Strictly speaking, in the orthodox YFS scheme [@Yennie:1961ad], the emissions from the intermediate $W$’s [*should not*]{} be included in the IR soft factors, because $W$’s are internal exchanges and the corresponding emission does not contribute any IR singularity. This is true, not only because each $W$ resonance is off-shell ($p_W^2\neq M_W^2$), but also because photons with energy below $W$ width, $E_\gamma \ll \Gamma_W$, emitted according to the above $J_{6f}^\mu$, “know nothing” about $W$’s[^12]. The reason is that, $W$’s live too shortly to affect the distributions of such very soft (long-wavelength) photons.
![ Kinematics of the double-resonant process. []{data-label="fig:eeWW4f"}](eeWW4f.pdf){height="80mm" width="130mm"}
On the other hand, looking into the example of the $\tau$-pair production and decay, the emission of soft and hard photons out of $W$’s definitely makes a lot of sense. However, in the case of the $W$-pair, the time separation of the production and decay stages is not that extremely long – this is why it is desirable to implement [*smooth*]{} analytical transition from the situation in which emission of photons with $E_\gamma < \Gamma_W$ is governed solely by the $J_{6f}^\mu$ currents to a situation in which the emission of photons with $E_\gamma > \Gamma_W$ gets a well-defined contribution from the intermediate $W$’s. The above situation is visualised in fig. \[fig:eeWW4f\] which describes a double-resonant process $$\label{eeWW4f}
\begin{split}
&e^-(p_{a}) +e^+(p_{b}) \to W^-(p_{g}) +W^+(p_{h})+n\gamma,\\
&W^-(p_{g}) \to f_{c}(p_c) +\bar{f}_{d}(p_d)+n\gamma,\quad
W^+(p_{h}) \to f_{e}(p_e) +\bar{f}_{f}(p_f)+n\gamma,
\end{split}$$ where we understand again that we may also contract any pair of the photon lines into a virtual photon exchange (loop). Here and in the following we use the following short-hand notation: $$p_{ab}=p_a+p_b,\quad p_{cd}=p_c+p_d,\quad p_{ef}=p_e+p_f.$$
The key point is a very special way in which the recoil is implemented in the resonance propagators. To understand this problem better, let us consider first the case with one real photon $n=1$ in the two soft limit regimes: (i) semi-soft, $k^0 \sim \Gamma \ll \sqrt{s}$ and (ii) true-soft, $k^0\ll \Gamma_W \ll\sqrt{s}$. The true-soft case (ii) is the case of the standard YFS, in which we have $$\label{eq:current-yfs61}
\begin{aligned}
&{{\EuScript{M}}^{(0)}}^{\mu_1}(k_1) \simeq
Const\; \frac{1}{p_{cd}^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\\
& \times \Bigg\{
Q_a \frac{2p_a^\mu}{2p_a k_1}
+Q_b \frac{2p_b^\mu}{2p_b k_1}
-Q_c \frac{2p_c^\mu}{2p_c k_1}
-Q_d \frac{2p_d^\mu}{2p_d k_1}
-Q_e \frac{2p_e^\mu}{2p_e k_1}
-Q_f \frac{2p_d^\mu}{2p_f k_1}
\Bigg\}.
\end{aligned}$$
In eq. (\[eq:current-yfs61\]) there is no emission from any internal $W$ line and no dependence in the resonance propagators due to photon emission. In the semi-soft regime (i) we have to restore such a dependence in the resonance propagators, that is take into account the recoil. This [*cannot*]{} be done without introducing photon emission from the intermediate charged resonance into the total electromagnetic current (unless we drop the NFI corrections altogether, as we already discussed). In order to see this point more clearly, let us write down a naive extension of the formula of eq. (\[eq:current-yfs61\]) in the [*complete analogy*]{} with the CEEX for the neutral resonances, like the $Z$-boson: $$\begin{aligned}
{{\EuScript{M}}^{(0)}_1}^{\mu}(k_1) &\simeq
\frac{1}{p_{cd}^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\Bigg\{ Q_a \frac{2p_a^\mu}{2p_a k_1} +Q_b \frac{2p_b^\mu}{2p_b k_1} \Bigg\}
\\&
+\frac{1}{(p_{cd}+k_1)^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\Bigg\{ -Q_c \frac{2p_c^\mu}{2p_c k_1} -Q_d \frac{2p_d^\mu}{2p_d k_1} \Bigg\}
\\&
+\frac{1}{p_{cd}^2-M^2}\; \frac{1}{(p_{ef}+k_1)^2-M^2}
\Bigg\{ -Q_e \frac{2p_e^\mu}{2p_e k_1} -Q_f \frac{2p_d^\mu}{2p_f k_1} \Bigg\}.
\end{aligned}$$ The above extension is, however, [*useless*]{}, because it is not gauge invariant. We have to restore emission from the internal $W$ in order to cure the gauge invariance, while maintaining recoil in the resonance propagator!
We therefore [*restore*]{} photon emission from the internal $W$ in the soft photon approximation (starting from Feynman diagrams) and next, factorise it into the product of the emission factors using the identity (\[eq:ident1phot\]) given in Appendix A. This identity also shows why it is necessary to sum up [*coherently*]{} over two photon assignments, either to $W$ in the production or to $W$ in the decay.
For the single real semi-soft photon under consideration, we obtain immediately the following gauge-invariant amplitude being the sum of three parts, each of them gauge invariant by itself[^13]: $$\label{eq:real-single}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_1}^{\mu}(k_1) \simeq
\\&
\frac{1}{p_{cd}^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\Bigg\{ Q_a \frac{2p_a^\mu}{2p_a k_1} +Q_b \frac{2p_b^\mu}{2p_b k_1}
-Q_g \frac{2p_g^\mu}{2p_g k_1} -Q_h \frac{2p_h^\mu}{2p_h k_1}
\Bigg\}
\\&
+\frac{1}{(p_{cd}+k_1)^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\Bigg\{ Q_g \frac{2p_g^\mu}{2p_g k_1}
-Q_c \frac{2p_c^\mu}{2p_c k_1} -Q_d \frac{2p_d^\mu}{2p_d k_1} \Bigg\}
\\&
+\frac{1}{p_{cd}^2-M^2}\; \frac{1}{(p_{ef}+k_1)^2-M^2}
\Bigg\{ Q_h \frac{2p_h^\mu}{2p_h k_1}
-Q_e \frac{2p_e^\mu}{2p_e k_1} -Q_f \frac{2p_f^\mu}{2p_f k_1} \Bigg\}
\\&
= \frac{1}{p_{cd}^2-M^2}\; \frac{1}{p_{ef}^2-M^2}
\Bigg\{ j^\mu_P
+\frac{p_{cd}^2-M^2 }{(p_{cd}+k_1)^2-M^2}\; j^\mu_{D_1}
+\frac{p_{ef}^2-M^2 }{(p_{ef}+k_1)^2-M^2}\; j^\mu_{D_2} \Bigg\}
\\&
= \sum_{\wp=(P,D_1,D_2)}^{3}
\frac{1}{p_{G}^2-M_W^2}
\frac{1}{p_{H}^2-M_W^2}
j_{\wp}^{\mu},
\end{aligned}$$ where $p_g=p_c+p_d+k_1$ and $p_h=p_e+p_f+k_1$. In the last line we have used $$\label{eq:current}
\begin{aligned}
& p_G=p_c+p_d+K_{D_1},~~p_H=p_e+p_f+K_{D_2},~~K_X =\sum_{i\in X}k_i
\\ & j_{P}^{\mu_i}=
\frac{2 p_{a}^{\mu_i}}{2p_{a} k_i}
+\frac{2 p_{b}^{\mu_i}}{2p_{b} k_i}
-\frac{2 p_{G}^{\mu_{i}}}{2p_{G} k_i}
-\frac{2 p_{H}^{\mu_{i}}}{2p_{H} k_i},
\\ &
j_{D_1}^{\mu_i}=
\frac{2 p_{G}^{\mu_{i}}}
{2p_{G} k_i}
-\frac{2 p_{c}^{\mu_i}}{2p_{c} k_i}
-\frac{2 p_{d}^{\mu_i}}{2p_{d} k_i},
~~j_{D_2}^{\mu_i}=
\frac{2 p_{H}^{\mu_{i}}}
{2p_{H} k_i}
-\frac{2 p_{e}^{\mu_i}}{2p_{e} k_i}
-\frac{2 p_{f}^{\mu_i}}{2p_{f} k_i}.
\end{aligned}$$ We keep in mind that in general $p_{g,h}^2\neq M^2$. The strange looking notation in the last line with the sum over partitions assigning photon to production or decays is done for the purpose of easy generalisation to the $n$-emissions case.
The single-photon amplitude of eq. (\[eq:real-single\]) coincides precisely (up to fermion spinors) with the $n=1$ case of the multiphoton amplitude of eq. (\[eq:ccex-alf0\]) in the next section. It features a proper dependence of the resonance propagators on the photon momentum in the entire photon energy region $k^0\ll \sqrt{s}$, including $k^0 \sim \Gamma_W$, and interpolates smoothly with the classic YFS formula of eq. (\[eq:current-yfs61\]), in the limit $k^0 \ll \Gamma_W$. The same will be true for the amplitude of eq. (\[eq:ccex-alf0\]) in a more general case of $n>1$.
Let us close this section with the multiple-photon extension of the formula (\[eq:real-single\]) with the notation of (\[eq:current\]) (details of its derivation can be found in Appendix \[app:b\]): $$\label{eq:multi-WW-real4}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N)
\simeq
\sum_{\wp=(P,D_1,D_2)^N}^{3^N}
\frac{1}{p_{G}^2-M_W^2}
\frac{1}{p_{H}^2-M_W^2}
\prod_{i=1}^{N}
j_{\wp_i}^{\mu_i}.
\end{aligned}$$
CEEX scheme for charged unstable emitters
=========================================
In the following we shall implicitly assume that IR-singularities are regularised with the photon mass $m_\gamma$. The exact IR cancellations between the real photons phase-space integrals $\int_{m_\gamma} d\Phi$ and the virtual form-factor $\alpha B(m_\gamma)$ work very schematically as follows: $$\sigma = \sum_{n=0}^\infty \frac{1}{n!}
\int_{m_\gamma} d\Phi_{4+n}(k_1\dots k_n)
\sum_{spin} |e^{\alpha B(m_\gamma)} {\mathfrak{M}}(k_1\dots k_n)|^2.$$ One may, of course, introduce the traditional IR-cut $E_\gamma > E_{\min}$ for all real photons, see refs. [@Yennie:1961ad; @Jadach:2000ir] for details. This we shall not do in the following, because it would obscure notation and is in fact unnecessary (even in the MC realisation we could stick to the $m_\gamma$ regulator).
In the following we shall present the formalism of CEEX for $e^-e^+\to W^+W^-,\;\; W^\pm \to X^\pm$. However, this formalism is quite general and applies also to the single-$W^\pm$ production and decay (also in hadron–hadron collisions) and also to any other process with any unstable intermediate charged particles of arbitrary spin.
Non-resonant variant of CEEX for $e^-e^+\to 4f$
------------------------------------------------
Let us start from defining CEEX for the $e^-e^+\to 4f$ process with the simplest possible variant of CEEX, in which the exponentiation procedure is [*not*]{} influenced by the presence of any narrow charged resonances in the Born matrix element $M^{(0)}$. This CEEX$_{B}$ scheme (according to the notation introduced in Subsection \[subsec:VarEex\]) can be used for the non-resonant background in the $e^-e^+\to 4f$ process. It is a kind of warm-up example in which we introduce some notation and terminology employed in the following.
Suppressing momenta and spin indices of the fermions, the and $n$-photon spin amplitudes can be written in a straightforward way $$\label{eq:ceexA-master}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_n}^{\mu_1,\mu_2,...\mu_n}(k_1,k_2,...k_n)=
\frac{1}{n!} e^{\alpha B_6^{\rm YFS}}\;
\hbeta_0^{(0)} \prod_{i=1}^n j^{\mu_i}(k_i),\qquad \hbeta_0^{(0)}=M^{(0)},
\\
&{{\EuScript{M}}^{(1)}_n}^{\mu_1,\mu_2,...\mu_n}(k_1,k_2,...k_n)=
\frac{1}{n!} e^{\alpha B_6^{\rm YFS}} \Bigg\{
\hbeta_0^{(1)} \prod_{i=1}^n j^{\mu_i}(k_i)
+\sum_{j=1}^n \hbeta_1^{(1)}{}^{\mu_j}(k_j) \prod_{i\neq j} j^{\mu_i}(k_i)
\Bigg\},
\end{aligned}$$ where the total electric current $$\label{eq:real-curr}
j^{\mu}(k_i) = ie \sum_{X=a,b,c,d,e,f} \hat{j}_X^\mu(k_i),\qquad
\hat{j}_X^\mu(k_i) \equiv Q_X \theta_X \frac{2p_X^\mu}{2p_X k_i},$$ sums contributions from all six external fermions $X=a,b,...f$, see fig. \[fig:ee4f\], and $\theta_X=+1$ for the incoming particle $X$ (in the initial state), $\theta_X=-1$ for the outgoing particle $X$ (in the final state). No emission from $W$’s is seen in $j^{\mu}$. The IF-finite $\beta$-functions are defined in the usual way $$\begin{aligned}
&\hbeta^{(1)}_0
= \left[ e^{-\alpha B_6^{\rm YFS}} M^{(0)}_0 \right]_{ {\cal O}(\alpha^1)},
\\
& {\hbeta^{(1)}_1}{}^\mu(k)
= {M^{(1)}_1}^\mu(k) -j^\mu(k) M^{(0)}_0.
\end{aligned}$$
The UV-finite, IR-divergent, gauge-invariant YFS form-factor is defined in the standard way, see also Appendix C: $$\label{eq:B6YFS}
\begin{aligned}
&B_6^{\rm YFS} = \int \frac{i}{(2\pi)^3}
\frac{d^4 k}{ k^2-m_\gamma^2 +i\varepsilon}\;
J^\mu(k) \circ J_\mu(k),
\\&
J^\mu(k)= \sum_{X=a,b,c,d,e,f} \hat{J}_{X}^\mu(k),\qquad
\hat{J}_{X}^\mu(k) \equiv
Q_X \theta_X\;
\frac{ 2p_X^\mu\theta_X +k^\mu}{k^2+ 2p_X k\theta_X +i\varepsilon},
\end{aligned}$$ where $\theta_X$ is defined as above, and we use the following short-hand notation: $$\label{eq:virt-curr}
\begin{aligned}
&S(k)= J(k) \circ J(k) =\sum_{ {X=a,b,c,d,e,f}\atop{Y=a,b,c,d,e,f} } J_X(k) \circ J_Y(k),
\\&
J_X(k) \circ J_Y(k) \equiv J_X(k) \cdot J_Y(-k),\; {\rm for}\; X\neq Y,\quad
\\&
J_X(k) \circ J_X(k) \equiv J_X(k) \cdot J_X(k).
\end{aligned}$$ As we see, $B_6^{\rm YFS}$ sums up the contributions from all six external fermions. IR-cancellations occur after squaring, spin-summing and integrating over the phase space, in a way which was shown using several methods in refs. [@Yennie:1961ad; @Jadach:2000ir]
Resonant variant of CEEX for $e^-e^+\to 4f$
--------------------------------------------
In the following we shall discuss the variant of CEEX for $e^-e^+\to 4f$ in which the recoil in resonance propagators is realised at any perturbative order and the $\Gamma_W/M_W$ suppression of the NFI contributions is a natural, built-in feature, valid in every perturbative order , $r=0,1,2,...$. In order to formulate such a scheme completely, one has to re-consider the isolation of IR-singular photon-emission factor to infinite order from the internal $W$ lines, [*going beyond the scope of the classic scheme of YFS [@Yennie:1961ad]*]{}. The important element of the isolation of the apparent IR-singularities due to emission of photons from the resonant charged particles is the reorganisation of the product of the internal propagators, derived in Appendix A. The virtual exponential form-factor has also a more complicated structure and is re-derived in Appendix C. Our derivation of the CEEX amplitudes is based on rearrangement of the infinite perturbative expansion in terms of Feynman diagrams, as in refs [@Yennie:1961ad; @Jadach:2000ir] and the use of the pole-expansion[^14]. Although our aim are the amplitudes, the main features of the scheme can already be defined and discussed for the simpler case, which will be discussed first. The extension of the presented technique to with the complete non-soft second-order photonic corrections and the genuine EW corrections is straightforward.
### Introductory double-pole CEEX {#sssect:ceex0}
Let us assume that for the $e^-e^+\to 4f$ process depicted in fig. \[fig:ee4f\] we have at our disposal the Born matrix element ${\cal M}^{(0)}_0$ which we expand into the non-pole part ${M}^{(0)}_0()$, the single-pole part $M^{(0)}_0(Q)$ and the double-pole part $M^{(0)}_0(Q,R)$, where $Q$ and $R$ are four-momenta in the $W$ propagators $${{\cal M}^{(0)}_0}^\mu()={M^{(0)}_0}^\mu() +{M^{(0)}_0}^\mu(Q)+{M^{(0)}_0}^\mu(Q,R),$$ The same pole-expansion is done for the exact single-photon spin amplitudes $${{\cal M}^{(1)}_1}^\mu(k)={M^{(1)}_1}^\mu(k)+{M^{(1)}_1}^\mu(Q,k)+{M^{(1)}_1}^\mu(Q,R,k),$$ where $k$ is the photon four-momentum and the index $\mu$ is understood to be contracted with the photon polarisation vector. The one-loop corrected complete spin amplitudes in the POE we denote as $M^{(1)}_0()$, $M^{(1)}_0(Q^2)$ and $M^{(1)}_0(Q,R)$: $${{\cal M}^{(1)}_0}^\mu={M^{(1)}_0}^\mu() +{M^{(1)}_0}^\mu(Q)+{M^{(1)}_0}^\mu(Q,R).$$
Let us focus now on the double-resonant part of the amplitudes ${M^{(0)}_0}^\mu(Q,R)$ and ${M^{(1)}_0}^\mu(Q,R)$. The single-resonant part is completely analogous (we shall list the differences) and the non-resonant case has already been discussed in the previous subsection.
The CEEX spin amplitudes for $n$ photons can be derived as the following gauge-invariant subset of the complete perturbative series $$\label{eq:ccex-alf0}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_n}^{\mu_1,\mu_2,...,\mu_n}(k_1,k_2,...,k_n)
=\sum_{\wp\in\{P,D_1,D_2\}^n}\!\!\!\!\!\!
e^{\alpha B_{10}(U_\wp,V_\wp)}\;
\hbeta^{(0)}_0\left(U_\wp,V_\wp\right)
\prod_{i=1}^n \; j_{\{\wp_i\}}^{\mu_i}(k_i),\\
& U_\wp= p_c+p_d+\sum_{\wp_i=D_1}k_i,\quad
V_\wp= p_e+p_f+\sum_{\wp_i=D_2}k_i.
\end{aligned}$$ Here, the fermion four-momenta $p_A$ and helicities $\lambda_A,A=a,b,c,d,e,f$ are suppressed. Photons are grouped into three sets: production, first decay and second decay, denoted as $P,D_1,D_2$. The [*coherent*]{} sum is taken over all $3^n$ assignments of a photon to 3 stages of the process. Each assignment is represented by the vector $(\wp_1,...,\wp_n)$ whose components are taking three possible values $\wp_j= P,D_1,D_2$. The cornerstone of this construction are three gauge invariant electric currents $$\begin{aligned}
&j^{\mu}_P(k_i) = ie \sum_{X=a,b,g,h} \hat{j}^{\mu}_X(k),\quad
j^{\mu}_{D_1}(k_i) = ie \sum_{X=g,c,d} \hat{j}^{\mu}_X(k),\quad
j^{\mu}_{D_2}(k_i) = ie \sum_{X=h,e,f} \hat{j}^{\mu}_X(k),\quad\\
&p_g=U_\wp,\quad p_h=V_\wp,
\end{aligned}$$ defined in terms of elementary currents $\hat{j}_X(k)$ of eq. (\[eq:real-curr\]). They include also $\hat{j}$’s for two $W$’s, see eq. (\[eq:real-curr\]). The essential steps in derivation of the CEEX formula of eq. (\[eq:ccex-alf0\]) are given in Appendices A, B and C.
The dependence of the amplitude in eq. (\[eq:ccex-alf0\]) on the four-momenta was already analysed in the case of the single real photon in the previous section. The case of many real photons is completely analogous. Let us turn now our attention to a more interesting case of multiple virtual photos which contribute to the virtual form-factor $\exp(B_{10})$.
The virtual IR-singularities factorise off in eq. (\[eq:ccex-alf0\]) into the factor $\exp(B_{10})$. Let us recall that our aim is to reproduce the $\Gamma/M$ suppression of the NFI corrections already at the level. It would be incorrect to employ here the classic YFS form-factor $B_6^{\rm YFS}$ of eq. (\[eq:B6YFS\]). This choice would render eq. (\[eq:ccex-alf0\]) IR-finite, however, it would fail to resum the $\alpha\ln(\Gamma/M)$ contributions and miss the $\Gamma/M$ suppression of NFI corrections, at the level. How to see it? One may check it by explicit analytical calculation, similar to the one performed in ref. [@Jadach:2000ir], or numerically. Quite generally, the reason for the above failure is that the effective energy scale for NFI is not $\sqrt{s}$ but $\Gamma_W$. The NFI contributions for the real photon energies above $\Gamma_W$ are suppressed strongly by the resonance propagator. However, this works for the real but not for virtual photons in $B_6^{YFS}$, hence the energy scale for virtual photons is necessarily $\sqrt{s}$. The mismatch between the scale for real and virtual photon will cause the NFI contribution to blow up at the by orders of magnitude, and even for they may be far from the reality.
The remedy for the above problem is well known for the neutral resonances [@Greco:1975rm; @Greco:1980mh; @Jadach:2000ir] and also can be deduced from the calculation (without exponentiation) of the NFI term for the charged resonance of $W$, see refs. [@Fadin:1993dz; @Chapovsky:1999kv; @Dittmaier:1999mb]. The modified CEEX form-factor which should be used in eq. (\[eq:ccex-alf0\]) is the following: $$\label{eq:B10}
\begin{aligned}
&B_{10}(p_{cd},p_{ef})=
\int \frac{i}{(2\pi)^3}\; \frac{d^4k}{k^2-\lambda^2+i\varepsilon}\;
\\&\qquad
\bigg\{ J_P(k) \circ J_P(k) +J_{D_1}(k)\circ J_{D_1}(k) +J_{D_2}(k) \circ J_{D_2}(k)
\\&\qquad
+\frac{p_{cd}^2-M^2 }{(p_{cd}-k)^2-M^2}\; 2 J_P(k)\circ J_{D_1}(k)
+\frac{p_{ef}^2-M^2 }{(p_{ef}-k)^2-M^2}\; 2 J_P(k)\circ J_{D_2}(k)
\\&\qquad
+\frac{p_{cd}^2-M^2 }{(p_{cd}+k)^2-M^2}
\;\frac{p_{ef}^2-M^2 }{(p_{ef}-k)^2-M^2}\;
2 J_{D_1}(k)\circ J_{D_2}(k)
\bigg\},
\end{aligned}$$ where $$\begin{aligned}
&J^{\mu}_P(k) = \sum_{X=a,b,g,h} \hat{J}_X^\mu(k),\qquad
J^{\mu}_{D_1}(k) = \sum_{X=g,c,d} \hat{J}_X^\mu(k),\qquad
J^{\mu}_{D_2}(k) = \sum_{X=h,e,f} \hat{J}_X^\mu(k),\\
&p_g = p_{cd}+K_1,\quad p_h = p_{ef}+K_2,
\end{aligned}$$ see eq. (\[eq:virt-curr\]) for definition of elementary virtual current $\hat{J}_X$ and of its circle-products. In eq. (\[eq:ccex-alf0\]) the four-momenta $U_\wp,V_\wp$ in $B_{10}(U_\wp,V_\wp)$ should be identified with $p_{cd}+K_1$ and $p_{ef}+K_2$ in eq. (\[eq:B10\]), where $K_1$ and $K_2$ are total four momenta of all [*real*]{} photons in the two decay processes. Note that the above form-factor is gauge invariant and UV-finite. Moreover, each of its six components is also separately gauge invariant and UV-finite. Almost all its components are already available in the literature. We have omitted from discussion the important Coulomb effect, see ref. [@Chapovsky:1999kv; @Actis:2008rb] for more details.
The index $10$ in $B_{10}$ reflects the fact that we have $10$ emission currents in $B_{10}$: $6$ for fermions and $4$ for $W$’s – that is $2$ for $W$’s in the production and $2$ for $W$’s in the decays.
Heuristic derivation of the above CEEX form-factor, directly from the Feynman diagram, is done in Appendix C using similar techniques as in Subsection 3.2.2 of ref. [@Jadach:2000ir]. In this derivation one may see explicitly why the first three components for the production and decays are exactly like in the standard YFS scheme, while three interferences are modified.
### The CEEX for double-pole component {#sssect:dpceex1}
The construction of for the $e^+e^-\to 4f$ process of the previous subsection was based, on one hand, on the gauge invariant POE of the Born spin amplitudes into the double-, single- and non-pole parts and, on the other hand, on the soft photon approximation in which real and virtual photon emission/absorption is represented as a product of the universal (spin-independent) factors, taking care of the recoil in all resonance propagators.
We intend now to extend the above scheme in such a way that the complete to the $e^+e^-\to 4f$ process are or can be included. The immediate question is to what extent POE into the double-, single- and non-pole parts can be kept at all at ?
Concerning POE at , we assume that both the amplitudes: $M_1^{(1)\mu}(k)$ with the emission of an additional single photon and $M_0^{(1)}$ with the complete one-loop corrections can be pole-expanded into the double-, single- and non-pole parts[^15]. Obviously, this can be done in many ways. Essentially it can be done (in principle) because the two propagators for the [*internal*]{} $W$ line due to photon emission can always be replaced by a sum of “two poles” using the identity of eq. (\[eq:ident1phot\]). Each of these terms can be made gauge invariant by taking a residue value for the entire expression multiplying the pole term, or more selectively, in its scalar part. This can be done (in principle) for both the amplitudes $M_1^{(1)\mu}(k)$ and $M_0^{(1)}$ representing the exact results of the Feynman diagrams at . The soft-photon-approximated universal part is already included in the calculation due to the exponentiation, in the same way as at .
The double-pole CEEX amplitude, including terms of due to the NFI interferences, reads as follows: $$\label{eq:ceex1b-resonanat}
\begin{aligned}
&{{\EuScript{M}}^{(1)}_n}^{\mu_1,\mu_2,...,\mu_n}(k_1,k_2,...,k_n)_{\rm DP}
=\sum_{\wp\in\{P,D_1,D_2\}^n} \!\!\!\!
e^{\alpha B_{10}(U_\wp,V_\wp)}
\hbeta^{(1)}_0\left(U_\wp,V_\wp\right)
\prod_{i=1}^n \; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
\\&\qquad
+ \sum_{j=1}^n\;\;
\sum_{\wp\in\{P,D_1,D_2\}^{n-1}} \!\!\!\!
e^{\alpha B_{10}(U_\wp,V_\wp)}
\hbeta^{(1)\mu_j}_{1\{\wp_j\}}\left(U_\wp,V_\wp,k_j\right)
\prod_{i\neq j}\; j_{\{\wp_i\}}^{\mu_i}(k_i),
\end{aligned}$$ where $$\hbeta^{(1)\mu}_{1}(U,V,k)
=M^{(1)\mu}_{1}(U,V,k)
-\sum_{\wp=P,D1,D2} j_{\wp}^\mu(k) M^{(0)}_0(U_{\wp},V_{\wp}).$$ The IR-finite $\hbeta_0$-functions is here defined as follows: $$\hbeta^{(1)}_0(U,V)
= \left[ e^{-\alpha B_{10}(U,V)} M^{(1)}_0(U,V)
\right]_{ {\cal O}(\alpha^1)} = M^{(1)}_0(U,V) - B_{10}(U,V) M^{(0)}_0(U,V),$$ where $B_{10}(U,V)$ is the complete variant of eq. (\[eq:B10\]) and the one-loop corrections in the double-pole $M^{(1)}_0(U,V)$ have to be complete at the , including terms of . Special care should be taken in order to preserve gauge invariance. Infrared regulation using $m_\gamma$ or any other method may be employed in the intermediate steps, but the final $B_{10}(U,V)$ will be IR-finite.
Needless to say that in the above expressions, as usual in all resummation schemes, one has to provide a recipe for extrapolating the results, originally defined in the phase space with zero or one real photon, to the phase space enriched with many additional “spectator” photons[^16]. The uncertainty due to freedom in this extrapolation is of the class.
### CEEX for single-pole component
The above implementation of CEEX for the DP component of the QED corrections are complete including corrections due the NFI interferences. However, the corrections arise also from the entire QED correction to a single-pole component (which by itself is of ). It is therefore necessary to define CEEX for the SP part. In addition, CEEX for the SP process is also of the vital importance for the $q\bar{q}\to W \to f\bar{f}$ process at hadron colliders, such as the LHC.
On the other hand, the non-pole (background) part, which is of ), may included without QED corrections or any kind of implementation of QED corrections, for instance using the simple baseline CEEX version of Subsection \[sssect:ceex0\].
The CEEX single-pole and double-pole spin amplitudes will be combined [*additively*]{} as follows[^17]: $${{\EuScript{M}}^{(1),\mu_1,\dots,\mu_n}_{n}}(k_1,k_2,...,k_n)_{\rm DSP}
={{\EuScript{M}}^{(1),\mu_1,\dots,\mu_n}_{n}}(k_1,k_2,...,k_n)_{\rm SP}
+{{\EuScript{M}}^{(1),\mu_1,\dots,\mu_n}_{n}}(k_1,k_2,...,k_n)_{\rm DP}.$$
The single-pole ${{\EuScript{M}}^{(1)}_{n}(\ldots)_{\rm SP}}$ amplitude is constructed analogously as in eq. (\[eq:ccex-alf0\]). The differences are that: (i) the current $j^\mu_P$ in the production process $e^+e^-\to f_c+\bar{f}_d +W^+$ has five components instead of four, (ii) the function $B_8$ replaces $B_{10}$, the $B_8$ has less components, in particular one interference term instead of three, (ii) the sum over photon assignment is reduced to the sum over the set $\{\wp\}=(P,D_1)^n$ corresponding to $2^n$ assignments: $$\label{eq:ceex1c-resonanat}
\begin{aligned}
&{{\EuScript{M}}^{(1)}_n}^{\mu_1,\mu_2,...,\mu_n}(k_1,k_2,...,k_n)_{SP}
=\sum_{\wp\in\{P,D_1\}^n} \!\!\!\!
e^{\alpha B_{8}(U_\wp)}
\hbeta^{(1)}_0\left(U_\wp\right)
\prod_{i=1}^n \; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
\\&\qquad\qquad\qquad
+ \sum_{j=1}^n\;
\sum_{\wp\in\{P,D_1\}^{n-1}} \!\!\!\!
e^{\alpha B_{8}(U_\wp)}
\hbeta^{(1)\mu_j}_{1\{\wp_j\}}\left(U_\wp,k_j\right)
\prod_{i\neq j}\; j_{\{\wp_i\}}^{\mu_i}(k_i),
\end{aligned}$$ where $$\hbeta^{(1)\mu}_{1}(U,k)
=M^{(1)\mu}_{1}(U,k) -\sum_{\wp=P,D1} j_{\wp}^\mu(k) M^{(0)}_0(U_{\wp}).$$ The IR-finite $\hbeta_0$-functions is defined here as follows: $$\hbeta^{(1)}_0(U)
= \left[ e^{-\alpha B_{8}(U)} M^{(1)}_0(U)
\right]_{ {\cal O}(\alpha^1)} = M^{(1)}_0(U) - B_{8}(U) M^{(0)}_0(U),$$ where $M^{(1)}_0(U)$ is the single-pole part in the Born amplitude of the $e^+e^-\to 4f$ process and the one-loop corrected single-pole $M^{(1)}_0(U)$ amplitude is complete at . The $B_{8}(U)$ function is the following variant of that in eq. (\[eq:B10\]): $$\label{eq:B8}
\begin{aligned}
&B_{8}(p_{cd})=
ie \int \frac{i}{(2\pi)^3}\; \frac{d^4k}{k^2-\lambda^2+i\varepsilon}\;
\\&\qquad
\bigg\{ J_P(k) \circ J_P(k) +J_{D_1}(k)\circ J_{D_1}(k)
+\frac{p_{cd}^2-M^2 }{(p_{cd}+k)^2-M^2}\; 2 J_P(k)\circ J_{D_1}(k)
\bigg\}.
\end{aligned}$$
The above CEEX for the single-pole part of the $e^+e^-\to 4f$ process implemented in ${\EuScript{M}}^{(1)}_{n}(\ldots)_{\rm SP}$ provides, together with the double-pole CEEX amplitude ${\EuScript{M}}^{(1)}_{n}(\ldots)_{\rm DP}$ of the previous section, the complete QED corrections at the order of , and for the $e^+e^-\to 4f$ process. Let us keep in mind that the definition of the terms in ${\EuScript{M}}^{(1)}_{n}(\ldots)_{\rm SP}$ and ${\EuScript{M}}^{(1)}_{n}(\ldots)_{\rm DP}$ depends on the exact definition of the SP and DP components in POE. Only the sum of them is uniquely defined – more precisely up to the terms of .
In the above formalism, the fermions labeled $e$ and $f$ do not form the resonance. In the case of the single-$W$ production in the quark–antiquark annihilation in hadron–hadron collision, the same formalism applies but the particles $e$ and $f$ are just absent.
### Approximate version of CEEX
Let us also consider one simpler case of the CEEX matrix element, with the incomplete corrections. It may be of some practical significance for applications with limited precision and will be described for the DP part only.
In this alternative scheme, the part is kept the same as in the full version of the CEEX scheme for the DP part of Subsection \[sssect:dpceex1\]. The main difference is in the simplification of the non-soft remnants, in which the non-factorisable QED interferences between the production and the decays are downgraded to the soft-photon approximation.
In such an approximation, the non-soft corrections are calculated separately for the production and two decay processes, and they contribute separately and additively to both real $\hbeta^{(1)\mu_1}$ and virtual $\hbeta^{(1)\mu}_0$: $$\hbeta^{(1)\mu}_1(U,V,k) = \sum_{X=P,D_1,D_2} \hbeta^{(1)\mu}_{1,X}(U,V,k),\quad
\hbeta^{(1)}_0(U,V) = \sum_{X=P,D_1,D_2} \hbeta^{(1)}_{0,X}(U,V),$$ where $U=p_{cd}$, $V=p_{ef}$. For instance, the non-soft contributions from penta-box diagrams in the NFI class are neglected completely in the $\hbeta^{(1)}_0(U,V)$, because their soft part (including resonance effects) is already included in the $B_{10}(U,V)$ function. The single real photon emission spin amplitudes factorise into the production and decay parts $$\begin{aligned}
&{\EuScript{M}}_1^{(1)\mu} (k)
={\EuScript{M}}_{1,P}^{(1)\mu}(k)\; {\EuScript{M}}_{0,D_1}^{(0)}\; {\EuScript{M}}_{0,D_2}^{(0)}
+{\EuScript{M}}_{1,P}^{(0)}\; {\EuScript{M}}_{1,D_1}^{(1)\mu}(k)\; {\EuScript{M}}_{0,D_2}^{(0)}
+{\EuScript{M}}_{1,P}^{(0)}\; {\EuScript{M}}_{0,D_1}^{(0)}\; {\EuScript{M}}_{1,D_2}^{(1)\mu}(k)\;
\\&
= {\EuScript{M}}_{0,P}^{(0)} \big[ j_{P}^\mu(k) {\EuScript{M}}_{0,D_1}^{(0)} {\EuScript{M}}_{0,D_2}^{(0)}
+ {\EuScript{M}}_{0,D_1}^{(0)}(k) j_{D_1}^\mu(k) {\EuScript{M}}_{0,D_2}^{(0)}
+ {\EuScript{M}}_{0,D_1}^{(0)} {\EuScript{M}}_{0,D_2}^{(0)}(k) j_{D_2}^\mu(k))\big]
\\&~~~
+ \tbeta^{(1)\mu}_P(k) {\EuScript{M}}_{0,D_1}^{(0)} {\EuScript{M}}_{0,D_2}^{(0)}
+ {\EuScript{M}}_{1,P}^{(0)}\; \tbeta^{(1)\mu}_{D_1}(k) {\EuScript{M}}_{0,D_2}^{(0)}
+ {\EuScript{M}}_{1,P}^{(0)}\; {\EuScript{M}}_{0,D_1}^{(0)} \tbeta^{(1)\mu}_{D_2}(k)
\\&
= \hbeta^{(0)}_0\left(U, V \right) j_{P}^\mu(k)
+\hbeta^{(0)}_0\left(U+k,V \right) j_{D_1}^\mu(k)
+\hbeta^{(0)}_0\left(U,V+k \right) j_{D_2}^\mu(k)
\\&~~~
+\hbeta^{(1)\mu}_{1P}(k) +\hbeta^{(1)\mu}_{1D_1}(k)+\hbeta^{(1)\mu}_{1D_2}(k),
\end{aligned}$$ where $\tbeta^{(1)\mu}_X(k),\;\; X=P,D_1,D_2$ are the CEEX elements for the production and the decays separately, and we have adopted a convention that the $W$ propagator is included in the lowest order decay amplitude ${\EuScript{M}}_{0,D_i}^{(0)}$. An additional argument $(k)$ in ${\EuScript{M}}_{0,D_i}^{(0)}(k)$ marks that this $W$ propagator includes the momentum $k$ of the photon emitted in the decay.
The resulting variant of the CEEX amplitude reads as follows: $$\label{eq:ceex1-approx}
\begin{aligned}
&{{\EuScript{M}}^{(1)}_n}^{\mu_1,\mu_2,...,\mu_n}(k_1,k_2,...,k_n)
\\&
=\!\!\! \sum_{\wp\in\{P,D_1,D_2\}^n} \!\!\!\!\!
e^{\alpha B_{10}(U_\wp,V_\wp)}
\Bigg\{
\hbeta^{(1)}_0\left(U_\wp,V_\wp\right)
\prod_{i=1}^n \; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
+\sum_{j=1}^n
\hbeta^{(1)\mu_j}_{1\{\wp_j\}}\left(U_\wp,V_\wp,k_j\right)
\prod_{i\neq j}\; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
\Bigg\}.
\end{aligned}$$ The important difference with respect to the previous case is that due to the splitting of $\hbeta^{(1)}$ into the production and decay parts, the photon $k_j$ entering $\hbeta^{(1)}$ is included into the sum over the photon assignments.
### Higher order upgrades and inclusion of genuine electroweak corrections
The upgrade of the CEEX amplitudes from to is straightforward, following the same path as in the analogous case of the QED CEEX scheme implemented in the [[KKMC]{}]{} project[@Jadach:1999vf; @Jadach:2000ir]. The CEEX scheme offers great flexibility, allowing to truncate a perturbative series at a different order for ISR, FSR, IFI and IFF. This may be exploited in a convenient staging of construction of the respective numerical Monte Carlo program. In particular, for the ISR corrections it would be good to include the LO corrections. From the experience of the [[KKMC]{}]{} project we know that calculations of the CEEX matrix element may be slow, due to the need of summations over the assignments of photons among production and decays. However, most of numerical contributions from these photon assignments are numerically negligible and one may invent methods of the effective forecasting which assignments can be omitted from the evaluation. This would speed up significantly numerical MC calculations[^18].
In the present work we concentrate on the QED part of the SM calculations for the $e^+e^-\to W^+W^-$ process. Is it possible to factorise and treat separately the QED part from the rest of the SM corrections, the genuine EW corrections? The answer is positive because the soft-photon factorisation for both the real and virtual photons is well established in the framework of perturbative calculations [@Yennie:1961ad]. The remaining genuine EW $r=1,2$ corrections are located in the IR-finite remnants $\hbeta^{(r)\mu}_0$, $\hbeta^{(r)\mu}_1(k), \hbeta^{(r)\mu_1\mu_2}_2(k_1,k_2)$. It is only important to remember that the CEEX scheme works at the amplitude level and in the calculation of the loop corrections leading to $\hbeta^{(r)\mu}_0$ or $\hbeta^{(r)\mu}_1(k)$, all the IR divergences are removed by means of subtracting the $B_{10}$ function – adding the real emissions à la Bloch–Norsieck in order to obtain finite results is a methodological mistake! Because of that it is much easier to manage the genuine EW corrections in the CEEX scheme of any perturbative order than in any other scheme, especially beyond .
In the [[KandY]{}]{} ([[YFSWW]{}]{}3) calculations of the LEP era, the genuine EW corrections were included in $\hbeta^{(1)\mu}_0$ for the DP production part of the process (similarly as in [[RACOONWW]{}]{}). In order to match a very high precision of the FCC-ee experiments, it will be necessary to introduce the corrections in $\hbeta^{(2)\mu}_0$ and $\hbeta^{(1)\mu}_1(k)$ of the DP component. They are not available yet. In addition, it will be needed to introduce the EW corrections in $\hbeta^{(1)\mu}_0$ of the SP component. This subgroup of corrections can, in principle, be extracted from the existing EW calculations for the entire $e^+e^-\to 4f$ process of ref.[@Denner:2005es; @Denner:2005fg].
Relations between CEEX and EEX schemes
======================================
Tracing exact relations between various CEEX and EEX schemes is quite important for at least two reasons. The EEX implementation of the exclusive exponentiation in [[YFSWW]{}]{}3 is the only existing one for the $e^+e^-\to 4f$ process, so it is desirable to show that it can be embedded in the CEEX scheme as a kind of a well-defined approximation. It will also help to better understand the physics of photon emission from unstable charged intermediate particles and the inherent limitations of the EEX exponentiation scheme in [[YFSWW]{}]{}3, in particular clarifying the question: what is exactly the mechanism of neglecting the NFI interferences in EEX of [[YFSWW]{}]{}3?
Another important reason is that it would be desirable to implement the CEEX matrix element using a MC correction weight on top of the same baseline MC distributions, which is implemented in the MC event generator for the EEX matrix element. This strategy was successfully exploited in the [[KKMC]{}]{} program and also in the [[KandY]{}]{} hybrid Monte Carlo. For these reasons it is interesting to establish the relation between the CEEX and EEX distributions all over the entire multiphoton phase space.
From CEEX$_R$ to EEX$_R$ algebraically
--------------------------------------
As we have already indicated in the introduction, the EEX differential distributions for the process $e^-e^+\to W^-W^+,\; W^\pm\to f\bar{f}$, can be obtained as a limiting case of the CEEX scheme for the process $e^-e^+\to 4f$, defined in this paper. Let us do it in the following. This is analogous to the derivation of EEX of [[KORALZ]{}]{}out of the CEEX amplitudes given in Section 4 of ref. [@Jadach:2000ir][^19]. The transition to EEX of [[YFSWW]{}]{}3 requires a few additional steps described in the next subsection.
As a starting point we take an approximate variant of CEEX of eq. (\[eq:ceex1-approx\]), which is obtained from the exact one of eq. (\[eq:ceex1b-resonanat\]) by means of neglecting some non-IR interference NFI terms: $$\begin{aligned}
&\sigma = \frac{1}{flux}\; \sum_{n=0}^\infty \frac{1}{n!}
\int dLips_{4+n}(p_a+p_b;p_c,p_d,p_e,p_f,k_1...k_n)
\\&\times
\!\!\! \sum_{\wp\in\{P,D_1,D_2\}^n} \!\!\!\!\!
e^{\alpha B_{10}(U_\wp,V_\wp)}
\Bigg\{
\hbeta^{(1)}_0\left(U_\wp,V_\wp\right)
\prod_{i=1}^n \; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
+\sum_{j=1}^n
\hbeta^{(1)\mu_j}_{1\{\wp_j\}}\left(U_\wp,V_\wp,k_j\right)
\prod_{i\neq j}\; j_{\{\wp_i\}}^{\mu_i}(k_i)\;
\Bigg\}
\\&\times
\!\!\!\sum_{\wp'\in\{P,D_1,D_2\}^n} \!\!\!\!\!
e^{\alpha B_{10}^*(U_{\wp'},V_{\wp'})}
\Bigg\{
\hbeta^{(1)}_0\left(U_{\wp'},V_{\wp'}\right)
\prod_{i=1}^n \; j_{\{\wp'_i\}}^{\mu_i}(k_i)\;
+\sum_{j=1}^n
\hbeta^{(1)\mu_j}_{1\{\wp'_j\}}\left(U_{\wp'},V_{\wp'},k_j\right)
\prod_{i\neq j}\; j_{\{\wp'_i\}}^{\mu_i}(k_i)\;
\Bigg\}^*,
\end{aligned}$$ where $U_\wp=p_{cd}+\sum_{\wp_i=D_1} k_{i}$ and $V_\wp=p_{ef}+\sum_{\wp_i=D_2} k_{i}$.
The consistent method of omitting all of the remaining QED NFI interferences between the production and two decays requires omitting from the double sum over photon assignments all non-diagonal terms, $\wp\neq\wp'$, and the interference terms in $B_{10}$. After doing that the above omission the sum over photons can be reorganised into a product of three separate sums, one for the production and two for the decays. In this way we get the following EEX expression: $$\label{eq:YFS3raw}
\begin{aligned}
&\sigma
= \frac{1}{flux}\; \sum_{n=0}^\infty\; \frac{1}{n!}\;
\int dLips_{4+n}(p_a+p_b;p_c,p_d,p_e,p_f,k_{1}...k_{n})
\\&\times
\!\!\sum_{\wp\in\{P,D_1,D_2\}^n}
e^{2\alpha \Re B_{PDD}(U_\wp,V_\wp)}
\prod_{i=1}^n \; |j_{\{\wp_i\}}^{\mu_i}(k_i)|^2\;
\Bigg\{
|\hbeta^{(1)}_0\left(U_\wp,V_\wp \right)|^2
\\&+
\sum_{j=1}^{n} \left(
2\Re \big( \hbeta^{(1)}_{1\{\wp_j\}}(U_\wp,V_\wp,k_j)\cdot j_{\{\wp_j\}}(k_{j})^* \big)
+ |\hbeta^{(1)}_{1\{\wp_j\}}(U_\wp,V_\wp,k_j)|^2
\right)
|j_{\{\wp_j\}}(k_{j})|^{-2}\;
\Bigg\}.
\end{aligned}$$ In the above expression the YFS form-factor $e^{2\alpha \Re B_{10}}$ factorises into the product of independent form-factors for the production and two decay processes: $$\label{eq:ReBPDD}
e^{2\alpha \Re B_{PDD}}=e^{2\alpha \Re B_P} e^{2\alpha \Re B_{D_1}} e^{2\alpha \Re B_{D_2}}.$$ Eq. (\[eq:YFS3raw\]) can be rewritten in a more traditional EEX notation as follows: $$\label{eq:YFS3trad}
\begin{aligned}
&\sigma
= \frac{1}{flux}\; \sum_{n=0}^\infty\; \frac{1}{n!}\;
\int dLips_{4+n}(p_a+p_b;p_c,p_d,p_e,p_f,k_{1}...k_{n})
\!\!\sum_{\wp\in\{P,D_1,D_2\}^n}
\\&
e^{2\alpha \Re B_{PDD}(U_\wp,V_\wp)}
\prod_{i=1}^n \; \tilde{S}_{\{\wp_i\}}(k_i)\;
\Bigg\{
\bbeta^{(1)}_0\left(U_\wp,V_\wp \right)
+\sum_{j=1}^{n} \bbeta^{(1)}_{1\{\wp_j\}}(U_\wp,V_\wp,k_j)
\big[\tilde{S}_{\{\wp_j\}}(k_{j})\big]^{-1}\;
\Bigg\},
\end{aligned}$$ where $$\tilde{S}_X(k)= |j_X^\mu(k)|^2,\; X=P,D_1,D_2.$$ Note that in the above expression for each photon assignment we perfectly know the four momentum in each $W$ propagator – simply because each photon is associated with the production or one of the two decays.
In fact eq. (\[eq:YFS3raw\]) looks like three separate EEX exponentiation schemes for the three subprocesses. They talk to each other only through total energy conservation and spin correlations[^20]. This can be seen manifestly even more clearly when, for the purpose of the MC implementation, eq. (\[eq:YFS3trad\]) is transformed into the following form in which the $\tilde{S}$-factors for the production and the decays are factorised. For $n$ photons in the overall sum over $3^n$ assignments of the photons $\{P,D_1,D_2\}^n$ there are groups (partitions) of $\frac{n!}{n_0!n_1!n_2!}$ choices, with $n_0$ photons in the production, $n_1$ photons in the first decay and $n_2$ photons in the second decay, $n_0+n_1+n_2=n$. The assignments in each partition are related by the permutation of the photons within the partition. We may replace in eq. (\[eq:YFS3trad\]) the whole such a partition just by one permutation member, getting the following expression: $$\label{eq:YFS3mc}
\begin{aligned}
&\sigma
= \sum_{n_0=0}^\infty \sum_{n_1=0}^\infty \sum_{n_2=0}^\infty
\int dLips_{4+n_0+n_1+n_2}(p_a+p_b;p_c,p_d,p_e,p_f,k_{1}...k_{n_2})
\\&\times
\frac{1}{n_0!} \prod_{i_1=0}^{n_0} \; \tilde{S}_{P}(k_{i_0})\;
\frac{1}{n_1!} \prod_{i_1=1}^{n_1} \; \tilde{S}_{D_1}(k_{i_1})\;
\frac{1}{n_2!} \prod_{i_2=1}^{n_2} \; \tilde{S}_{D_2}(k_{i_2})\;
\\&\times
e^{2\alpha \Re B_{PDD}(U_1, V_2)}
\Bigg\{
\bbeta^{(1)}_0\left(U_1, V_2 \right)
+\sum_{j=1}^{n_0} \bbeta^{(1)}_{1\{P\}}(U_1,V_2,k_j)\; \tilde{S}_{P}(k_{j})^{-1}\;
\\&
+\sum_{j=1}^{n_1} \bbeta^{(1)}_{1\{D_1\}}(U_1,V_2,k_j)\; \tilde{S}_{D_1}(k_{j})^{-1}
+\sum_{j=1}^{n_2} \bbeta^{(1)}_{1\{D_2\}}(U_1,V_2,k_j)\; \tilde{S}_{D_2}(k_{j})^{-1}
\Bigg\},
\end{aligned}$$ where $U_1=p_{cd}+\sum_{i_1=0}^{n_1} k_{i_1}$ and $V_2=p_{ef}+\sum_{i_2=0}^{n_2} k_{i_2}$. One can always come back to the configuration of eq. (\[eq:YFS3trad\]) by means of symmetrisation over photons. In MC computations, the sum over photons is “randomised” in a natural way and only one partition member is generated at a time, (using effectively eq. (\[eq:YFS3mc\])) so the fact that the basic distribution for EEX$_{R}$ is that of eq. (\[eq:YFS3trad\]) can be easily overlooked, see also discussion in [@Jadach:1999vf].
From the above algebra we see in a detail how EEX$_R$ can be embedded in a natural way in the full CEEX$_R$, defined in the previous section.
Last step towards EEX$_R$ of [[YFSWW]{}]{}3
-------------------------------------------
The EEX of eq. (\[eq:YFS3mc\]) is not exactly that of EEX of [[YFSWW]{}]{}3and [[KandY]{}]{}, as described in refs. [@Jadach:2001uu; @Jadach:2001mp]. Let us discuss the remaining differences. The most important difference is that QED matrix element for the $W$-boson decay in [[YFSWW]{}]{}3is implemented using the [[PHOTOS]{}]{} program whose has matrix element is not in the EEX scheme, although very close to it. At the precision of the LEP experiments this was the acceptable and economic solution. There would be no problems with replacing [[PHOTOS]{}]{} with the true EEX implementation for the $W$ decays because such an implementation is already available in the [[WINHAC]{}]{} program developed for the single-$W$ production at hadron colliders [@Placzek:2003zg].
The implementation of the EEX matrix elements for the production process in [[YFSWW]{}]{}3is described in fine detail in ref. [@Jadach:2001uu]. It is based on the [YFS3]{} event generator [@Jadach:1991dm] for the $e^+e^-\to 2f$ process in which the final-state massive fermions are replaced with $W$’s. The [YFS3]{} program does not include the QED initial-final state interferences (IFI) between initial $e^\pm$ and final particles. Such interferences (present in EEX of eq. (\[eq:YFS3mc\]) were also added in [[YFSWW]{}]{}3 using the reweighting technique of the [[BHWIDE]{}]{} program [@Jadach:1995nk].
From EEX$_R$ to CEEX$_R$ in MC implementation
---------------------------------------------
The upgrade from EEX of eq. (\[eq:YFS3mc\]) to CEEX in the MC implementation is feasible and well defined. In the Monte Carlo program implementing EEX, one usually generates MC events according to some baseline distribution[^21] and the final correcting weight introduces fine details of the EEX matrix element. The CEEX matrix elements can be implemented by reweighting events generated according to the same baseline distributions as in the EEX case, just by replacing the EEX final MC correcting weight with that of CEEX, without any changes in the baseline MC. This kind of flexible and economic solution was already applied in the [[KKMC]{}]{} program [@Jadach:1999vf]. Similarly as in [[KKMC]{}]{}, the MC weight correcting from EEX to CEEX will be not bound from the above. There are several solutions for this purely technical problem.
Photon distributions around $E_\gamma\sim\Gamma$
------------------------------------------------
Let us finally comments on two apparent deficiencies of the EEX$_{R}$ scheme:
- lack of transmutation of photon distributions around $E_\gamma\sim\Gamma$,
- excess of photon-multiplicity for very soft photons, $E_\gamma\le\Gamma$.
The phenomenon of “transmutation of photon distributions” occurs when photon energy changes from the “semisoft region” $\Gamma < E_\gamma\ll E_{beam}$ down to “true soft region” $E_\gamma < \Gamma$. In the true-soft region photon distributions do not reflect the existence of the the single charged object, the resonance – they reflect, instead, momenta and charges of all its decay products. For these long range photons, the resonance itself just lives too shortly to be “felt”. On the other hand, the semi-soft photons with shorter wavelength can see the resonance as a distinct object – its presence is imprinted in the distributions of photon energy and angles. In fact, it is the interference between the production-current $j^\mu_{P}$ and the decay-current $j^\mu_{D1,D2}$ which enforces the transition in the photon distributions. This effect can be also seen explicitly in the instrumental identity of eq. (\[eq:ident1phot\]), or in the explicit one-photon emission amplitude of eq. (\[eq:real-single\]). The absence of this interference in EEX, where all the NFI interferences are neglected, causes that in EEX (of [[YFSWW]{}]{}3) the above beautiful transmutation phenomenon cannot be present[^22].
The lack of the above interferences causes also certain unphysical effect for very soft photons. As we know, in the real world (and in CEEX) there is no IR singularity (neither real nor virtual) for the photon emission from the internal $W$ line, see eq. (\[eq:real-single\]), while in EEX there is such (real and virtual), as seen explicitly in eq. (\[eq:YFS3mc\]). How to explain this paradox? Is this something dangerous? The artificial IR divergence in EEX is not dangerous as long as we are at the precision level for the distributions which are inclusive enough, such that we do not examine multiplicities and angular spectra of the photons with $E_\gamma < \Gamma$. Extra unphysical photons in this energy range do not contribute to integrated cross section, because their contribution is countered immediately by the virtual form-factor. They will however affect multiplicity of such very soft photons.
The good agreement of the soft photon spectra between [[YFSWW]{}]{}3 and [[RACOONWW]{}]{} confirms that the effect is not sizeable. The numerical estimates of ref. [@Fadin:1993dz] also suggest that this effect is small, negligible for LEP2. On the other hand, in the future high-statistics experiments it is worth to examine the above effects for the photons with $E_\gamma \sim \Gamma_W$. It was proposed in ref. [@Fadin:1993dz] that it may even provide an independent relatively precise measurement of $\Gamma_W$.
Summarising, the presence of the extra unphysical soft photons with $E_\gamma < \Gamma$ in EEX (and its version implemented in [[YFSWW]{}]{}3) due to setting to zero all QED interference effects between the production and decay processes is not harmful at the precision level of . For the higher-precision requirements, like that in FCC-ee, one should go back to CEEX$_R$, from which EEX$_{R}$ is derived, and get back for $E_\gamma \sim \Gamma$ fully exclusive realistic photon distributions.
Summary and outlook
===================
In the present paper we have proposed a solution to the long-standing problem of the systematic treatment of the soft and hard photon emission from the unstable charged particles and the interferences between production and decay parts of the process, at any perturbative order. This is of practical importance for high-precision measurements of $W^+W^-$-pair production at electron–positron colliders, such as FCC-ee/ILC/CLIC, and for single-$W$ production at hadron colliders, such as LHC/FCC-hh, as well as in many other processes with production and decay of charged unstable particles of any spin. So far there is no practical implementation of the full-scale calculation in the proposed scheme. However, it has been outlined how to accomplish it in the framework of some existing Monte Carlo (MC) event generators.
Our study has been focused on the process $e^+e^- \to W^+W^- \to 4f$ which is to be used e.g. for the high-precision $W$-boson mass and width measurements in the planned electron–positron colliders, particularly FCC-ee. We have argued that the most economical (and perhaps the only feasible) way to achieve the required accuracy of theoretical prediction for this process it to apply the so-called pole expansion (POE) to the general process of $e^+e^- \to 4f$, and then to calculate the electroweak (EW) radiative corrections separately for each term of such an expansion to an appropriate order in the coupling constant $\alpha$. More specifically, for the leading term in POE, i.e. the so-called double-pole contribution which comprises two resonant $W$-bosons, one would need to include the fixed-order EW corrections up to [${\cal O}(\alpha^2)$]{} for the on-shell-like $W$-pair production and $W$ decay processes, while for the non-leading terms, i.e. the single-pole and non-pole contributions, the EW corrections at [${\cal O}(\alpha^1)$]{} would be sufficient. The calculations of the [${\cal O}(\alpha^1)$]{} EW corrections are already available for the whole $e^+e^- \to 4f$ process, while the [${\cal O}(\alpha^2)$]{} ones for the double-resonant contribution do not exist yet, however they are feasible, in our opinion, by the time of the planned FCC-ee physics run.
In addition to the above fixed-order radiative corrections, in order to reach the requisite theoretical precision for the above process, one needs to include higher-order QED corrections corresponding to multiphoton emission from the initial- and final-state leptons as well as from the intermediate $W$-bosons. We have argued that the best framework in which all this can be accomplished is the so-called coherent exclusive exponentiation (CEEX) scheme. Its main advantage over the traditional YFS exclusive exponentiation (EEX) method is that it operates directly at the level of spin amplitudes. Because of that, all multiphoton effects related to radiation from the resonant $W$-bosons and to non-factorisable interferences can be accounted for in a straightforward way. So far, the CEEX methodology was applied to $e^+e^- \to 2f$ in the [[KKMC]{}]{} event generator and proved to be crucial in providing precision theoretical predictions for this process necessary for the LEP experiments.
We have provided the respective general cross-section formulae for the double-pole, single-pole and non-pole contributions to the charged-current $e^+e^- \to 4f$ process which can be a basis for an appropriate MC implementation. An important ingredient in that is resummation of real-photon emissions including radiation from the intermediate $W$-bosons and derivation of the corresponding virtual-photon form-factor, done explicitly in Appendices A, B and C. Our approach exploited the similarity between the virtual- and real-emission QED form-factors guaranteed by the infra-red cancellations.
We have also discussed the relation of the above CEEX realisation to the existing EEX implementation in terms of the hybrid MC program called [[KandY]{}]{}, being the combination of two MC event generators: [[KORALW]{}]{} and [[YFSWW]{}]{}3. In this implementation, the [${\cal O}(\alpha^3)$]{} YFS exponentiation for initial-state radiation in the process $e^+e^- \to 4f$ was combined with the fixed-order [${\cal O}(\alpha^1)$]{} EW corrections in the $W$-pair production and multiphoton radiation in the $W$-decays generated by the [[PHOTOS]{}]{} program, while all the non-factorisable interferences were neglected. Such a solution proved to be good enough for the LEP2 accuracy, but for the expected precision of the FCC-ee experiments it will not suffice. However, it can constitute a good starting point and a MC platform for development and implementation of the CEEX scheme described in this paper. In parallel, one can also develop an EEX approximation of the full-scale CEEX solution which will be important for its numerical cross-checks. For this, the implementation of EEX for the $W$-boson decays in the [[WINHAC]{}]{} program can be used to replace the corresponding [[PHOTOS]{}]{} radiation in [[KandY]{}]{}.
Acknowledgments {#acknowledgments .unnumbered}
===============
Useful discussions with B.F.L Ward and Z. Was are acknowledged.
[**Appendices**]{}
Factoring photon-emission from $W$ {#appendixA}
==================================
The following considerations are valid for a charged unstable particle of any spin, eg. $W^\pm$, $\tau^\pm$ or $t$-quark. Let us start with a simple identity for two propagators related to single photon emission from an internal charged particle line $$\frac{1}{(Q_0^2-M^2)(Q_1^2-M^2)}
= \frac{1}{(Q_0^2-Q_1^2)(Q_1^2-M^2)}-\frac{1}{(Q_0^2-Q_1^2)(Q_0^2-M^2)}$$ where $M^2=M_W^2+iM_W\Gamma_W$.
The kinematics is depicted in fig \[fig:Wgam1\].
![Single emission from the internal W line.[]{data-label="fig:Wgam1"}](Wgam1.pdf){height="15mm" width="60mm"}
Noticing that $Q_0^2-Q_1^2 =2k_1Q_0-k_1^2=2k_1Q_1+k_1^2$, we may rewrite the above as follows: $$\label{eq:ident1phot}
\frac{1}{(Q_0^2-M^2)(Q_1^2-M^2)}
= \frac{1}{(2k_1Q_1+k_1^2)(Q_1^2-M^2)}+\frac{1}{(-2k_1Q_0+k_1^2)(Q_0^2-M^2)}.$$ The reader will recognise the first term as representing a photon (eikonal) emission factor in the [*production*]{} part of the process times a resonance propagator (with the reduced four momentum $Q_1=Q_0-k_1$) and the second term as the analogous emission factor in the [*decay*]{} process times the resonance propagator (with the four-momentum $Q_0=Q_1+k_1$). Each of the two terms look IR-divergent, however the two IR divergences cancel – the difference is finite. In the original expression it was the resonance width $\Gamma_W$ which was providing an infrared regulator for a photon with the momentum $k_1=Q_1-Q_2$.
![Multiple emission from the internal W line.[]{data-label="fig:WgamM"}](WgamN.pdf){height="15mm" width="140mm"}
Let us now consider the general case of the $n$-photon emissions from the internal charged particle line, depicted in fig. \[fig:WgamM\], in the soft-photon approximation. The reorganisation of the product of the propagators starts with the following identity: $$\label{eq:intermediate}
\begin{aligned}
&\frac{1}{(Q_0^2-M^2)(Q_1^2-M^2)\dots(Q_n^2-M^2) }\\
& \qquad=\sum_{j=0}^{n}
\frac{1}{ \prod_{i=0}^{j-1} (Q_i^2-Q_j^2) \; \;(Q_j^2-M^2) \prod_{i=1}^{n-j}(Q_{j+i}^2-Q_j^2) }.
\end{aligned}
$$ It can be proven using the mathematical induction method. Assuming that the identity is true for $n$, let us prove it for $n+1$. Using a short-hand notation $y_i= Q_i^2-M^2$, one obtains[^23] $$\begin{split}
&\prod\limits_{i=0}^{n+1} \frac{1}{y_i}
= \frac{1}{ y_{n+1}}\; \sum_{j=0}^{n} \frac{1}{y_j }\;
\prod\limits_{i\neq j}^n \frac{1}{(y_i-y_j)}\;
=\sum_{j=0}^{n}\; \frac{1}{(y_{n+1}-y_j) }
\Big( \frac{1}{y_j } - \frac{1}{y_{n+1} }\Big)
\prod\limits_{i\neq j}^n \frac{1}{(y_i-y_j)}
\\&
=\sum_{j=0}^{n}\; \frac{1}{y_j }
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
- \frac{1}{y_{n+1}} \sum_{j=0}^{n}\;
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
\\&
=\sum_{j=0}^{n+1}\; \frac{1}{y_j }
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
-\frac{1}{y_{n+1} }
\prod\limits_{i\neq n+1}^{n+1} \frac{1}{(y_i-y_{n+1})}
- \frac{1}{y_{n+1}} \sum_{j=0}^{n}\;
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
\\&
=\sum_{j=0}^{n+1}\; \frac{1}{y_j }
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
- \frac{1}{y_{n+1}} \sum_{j=0}^{n+1}\;
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}
=\sum_{j=0}^{n+1}\; \frac{1}{y_j }
\prod\limits_{i\neq j}^{n+1} \frac{1}{(y_i-y_j)}.
\end{split}$$ Alternatively, one can prove it with the help of partial fractioning with respect to $M^2$: $$\label{eq:partfract}
\begin{aligned}
&\frac{1}{(Q_0^2-M^2)(Q_1^2-M^2)\dots(Q_n^2-M^2) }
=\sum_{j=0}^{n} \frac{A_j}{Q_j^2-M^2}.
\end{aligned}
$$ Multiplying eq. (\[eq:partfract\]) in a standard way by $Q_j^2-M^2$ and substituting $Q_j^2=M^2$ we obtain $$\label{eq:partfract2}
\begin{aligned}
A_j & =
\frac{1}{ \prod_{i=0}^{j-1} (Q_i^2-Q_j^2)
\prod_{i=1}^{n-j}(Q_{j+i}^2-Q_j^2) }.
\end{aligned}
$$
Let us now examine the soft-photon limit in eq.\[eq:intermediate\]. Taking the $j$-th term, we may identify $$\label{eq:appr1}
\begin{aligned}
Q_0^2-Q_j^2 &\simeq(2k_jQ_j+k_j^2)+\dots+(2k_2Q_j+k_2^2) +(2k_1Q_j+k_1^2),\\
Q_1^2-Q_j^2 &\simeq(2k_jQ_j+k_j^2)+\dots+(2k_2Q_j+k_2^2), \\
Q_{j-1}^2-Q_j^2 &= (2k_jQ_j+k_j^2).
\end{aligned}$$ and $$\label{eq:appr2}
\begin{aligned}
Q_{j+1}^2-Q_j^2 & = (-2k_{j+1}Q_j+k_{j+1}^2)\\
Q_{n-1}^2-Q_j^2 & \simeq(-2k_{j+1}Q_j+k_{j+1}^2)+\dots+(-2k_{n-1}Q_j+k_{n-1}^2) \\
Q_{n}^2 -Q_j^2 & \simeq(-2k_{j+1}Q_j+k_{j+1}^2)+\dots+(-2k_{n-1}Q_j+k_{n-1}^2) +(-2k_{n}Q_j+k_{n}^2)
\end{aligned}$$ In the above equations we have neglected the subleading products $k_ik_j$. This is allowed in the soft-photon approximation. On the other hand, terms $k_i^2$ could also be omitted in the soft-photon approximation, but they are kept because they render virtual photon integrals UV-finite.
In the next step we perform the usual sum over permutation over all photons. This will lead to a “Poissonian” emission formula, separately for the resonance production and decay stages of the entire process, with the explicit sum over the assignments of photons to the production, denoted by the index $P$, and to the decay, denoted by the index $D$. We start from eq. (\[eq:intermediate\]) switching to a more compact notation: $$\begin{aligned}
R(Q^2_i)&= Q_i^2-M^2,\\
N^{P}_i &= 2k_{i}Q_j+k_{i}^2=(Q_j+k_{i})^2-Q_j^2,\\
N^{D}_i &=-2k_{i}Q_j+k_{i}^2=(Q_j-k_{i})^2-Q_j^2.
\end{aligned}$$ Inserting the relations of eqs. (\[eq:appr1\]) and (\[eq:appr2\]) into eq. (\[eq:intermediate\]) and summing over permutations we obtain $$\label{eq:identity}
\begin{aligned}
&\sum_{permut.} \frac{1}{R(Q_0^2)R(Q_1^2)\dots R(Q_n^2) }\\
&=\sum_{permut.}
\sum_{j=0}^{n} \Bigl[
\frac{1}{N^{P}_1+N^{P}_2+N^{P}_3+\dots N^{P}_j}\;
\frac{1}{N^{P}_2+N^{P}_3+\dots N^{P}_j} \dots \frac{1}{N^{P}_j}\Bigr]\\
&\qquad\qquad\qquad
\times
\frac{1}{R(Q^2_j)} \times \Bigl[
\frac{1}{N^{D}_{j+1}}\;
\frac{1}{N^{D}_{j+1}+N^{D}_{j+2}} \dots
\frac{1}{N^{D}_{j+1}+N^{D}_{j+2}\dots N^{D}_{n}} \Bigr],\\
\end{aligned}$$ where for $j=0$ and $j=n$, respectively, the term in the first/second square-bracket pair should read as $1$. Next, for each $j$-th term we split the sum over all permutations of $(1,2,3,\dots,n)$ into two separate sums: one over permutations of $(1,2,3,\dots,j)$ and another over permutations of $(j+1,j+2,\dots,n)$. These two sums are performed[^24]. The sum over $\big( {n \atop j} \big)$ assignments of photons to production and decay remains. Alternatively, the entire remaining sum can be represented as a sum over $\sum_j \big( {n \atop j} \big) = (1+1)^n= 2^n$ terms (photon assignments) as follows $$\label{eq:principal-ident}
\begin{aligned}
&\sum_{permut.} \frac{1}{R(Q_0^2)R(Q_1^2)\dots R(Q_n^2) }\\
&\qquad\qquad =
\sum_{\wp=(P,D)^n}
\prod_{\wp_i=P} \frac{1}{(Q_\wp +k_{i})^2 -Q^2_\wp}\times
\frac{1}{R(Q^2_\wp)}\times
\prod_{\wp_i=D} \frac{1}{(Q_\wp -k_{i})^2 -Q^2_\wp},
\end{aligned}$$ where $$Q\wp = Q_0 - \sum_{\wp_i=P} k_i= Q_n + \sum_{\wp_i=D} k_i.$$ The vectors $\wp=(\wp_1,\wp_2, \dots \wp_n)$ of the photon assignments whose components have values equal to $P$ or $D$, while $\sum_{\wp_i=P}$ ($\prod_{\wp_i=P}$) denotes the sum over (product of) all $i$ for which $\wp_i=P$, i.e. all photon which belong to the production stage of the process.
Main features of eq. (\[eq:principal-ident\]), the principal result of this Appendix, are the following:
- Its left-hand side represents “raw” Feynman diagrams for multiple-photon emission from the charged-particle internal line.
- Its right-hand side includes two photon emission factors: one for the production part of the process (resonance formation) and the second one for the decay part of the process (resonance decay).
- It includes the single-resonance propagator of the standard form, with the complex mass $M$, and the four-momentum $Q_\wp$, which comprises momenta of all photons assigned to the resonance decay.
- It is rather striking that all photon-emission factors look as if photons were emitted by the charged particle of the mass $Q_\wp^2$! This is, of course, intuitively well justified and quite appealing.
- The fact that the [*coherent*]{} sum is performed over all the photon assignments to the production and the decay reflects the QED gauge invariance and the Bose–Einstein statistics.
- It holds both for the virtual and real photons (this is why we have kept $k_i^2$).
Resummation of real-photon emissions {#app:b}
====================================
In this Appendix we show how to do the resummation of the amplitude of the multiple-real-photon emission. We expect that because of IR cancellations the basic algebraic structure of our derivation holds for the integrands of multiloop corrections.
Let us begin with a short summary of the YFS method performed in a combinatorial way. The process under consideration is $$e(p_a)\bar\nu_e(p_b)\longrightarrow W \longrightarrow \mu(p_c)\bar\nu_\mu(p_d).$$ At first, we consider the standard YFS scheme without radiation nor recoil from $W$. As proven by YFS [@Yennie:1961ad], the IR radiation comes entirely from the charged external legs ($e$ and $\mu$) and has a form of soft currents. The sum of graphs with $N$ real emissions is the following: $$\label{eq:multi-real1}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\\ \simeq &
\sum_{l=0}^N \sum_{\pi}^{N!}
\biggl(\frac{2p_a^{\mu_1}}{2p_a k_{\pi_1}}
\frac{2p_a^{\mu_2}}{2p_a k_{\pi_1}+2p_a k_{\pi_2}}
\dots
\frac{2p_a^{\mu_l}}{2p_a k_{\pi_1}+2p_a k_{\pi_2}+\dots+2p_a k_{\pi_l}}
\biggr)
\\ & \times
\biggl(\frac{-2p_c^{\mu_{l+1}}}{2p_c k_{\pi_{l+1}}}
\frac{-2p_c^{\mu_{l+2}}}{2p_c k_{\pi_{l+1}}+2p_c k_{\pi_{l+2}}}
\dots
\frac{-2p_c^{\mu_{N}}}{2p_c k_{\pi_{l+1}} +2p_c k_{\pi_{l+2}}
+\dots +2p_ck_{\pi_N}}
\biggr)
\\ & \times
\frac{1}{p_{ab}^2-M^2}.
\end{aligned}$$ We execute now the sum over permutations of photons within the $a$ and $c$ sub-groups according to the formula of footnote \[foot:23\]. This turns complicated sums into simple products: $$\label{eq:multi-real2}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\frac{1}{p_{ab}^2-M^2}
\sum_{l=0}^N \sum_{\pi/\pi_l/\pi_{N-l}}^{N!/l!/(n-l)!}
\biggl(\prod_{i=1}^l \frac{2p_a^{\mu_i}}{2p_a k_{\pi_i}}
\biggr)
\biggl(\prod_{i=1}^{N-l} \frac{-2p_c^{\mu_{l+i}}}{2p_c k_{\pi_{l+i}}}
\biggr).
\end{aligned}$$
It takes now a few moments to realise that the combinatorial sum over permutations can be replaced by the sum over partitions (cf. eqs.(\[eq:identity\]) and (\[eq:principal-ident\]))[^25] $$\label{eq:sumpart}
\sum_{l=0}^N \sum_{\pi/\pi_a/\pi_c}^{N!/l!/(N-l)!}
=
\sum_{\wp=(a,c)^N}^{2^N}.$$ Consequently we get $$\label{eq:multi-real3}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\frac{1}{p_{ab}^2-M^2}
\sum_{\wp=(a,c)^N}^{2^N}
\biggl(
\prod_{i=1}^N \frac{2\theta_{\wp_i}p_{\wp_i}^{\mu_i}}{2p_{\wp_i}k_{i}}
\biggr),
\end{aligned}$$ where $\theta$ equals $+1$ for initial state and $-1$ for final state. Finally, we notice that the sum over partitions in eq. (\[eq:multi-real3\]) can be rewritten in a compact form as[^26] $$\label{eq:multi-real4}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\frac{1}{p_{ab}^2-M^2}
\prod_{i=1}^N \biggl(
\frac{2p_{a}^{\mu_i}}{2p_{a} k_{i}}
-\frac{2p_{c}^{\mu_i}}{2p_{c} k_{i}}
\biggr).
\end{aligned}$$
Let us now allow for the radiation from the $W$-boson. We begin by analysing the numerator of the multiple-emission graph of fig. \[fig:WgamM\], i.e. of LHS of eq. (\[eq:intermediate\]). The numerator of the single photon emission with two accompanying $W$ propagators (in the small-photon-momentum limit) looks as follows: $$\label{eq:single-phot}
\begin{aligned}
& \bigl(-g^{\lambda\lambda'}+p^\lambda p^{\lambda'}/M_W^2\bigr)
V(p,k,p-k)_{\lambda'\rho\sigma'}
\bigl(-g^{\sigma'\sigma}+(p-k)^{\sigma'} (p-k)^{\sigma}/M_W^2\bigr)
\\ &
{\operatornamewithlimits{=}}^{k\to 0}
\bigl(-g_{\lambda\sigma}+ p_\lambda p_\sigma/M_W^2\bigr) (-2p_\rho)
+ g_{\lambda\rho} p_\sigma(p^2-M_W^2)
+ g_{\sigma\rho}p_\lambda(p^2-M_W^2),
\end{aligned}$$ where $V(p,k,p-k)_{\lambda'\rho\sigma'}$ is the $W\gamma W$ vertex. Dropping also the terms proportional to $p^2-M_W^2$ (i.e. putting $p$ on-shell) we obtain a self-repeating structure and the numerator of the whole line becomes $$\label{eq:multi-phot}
\bigl(-g_{\lambda\sigma}+ Q_{j\lambda} Q_{j\sigma}/M_W^2\bigr)
\prod_{i=1}^n (-2p_{\mu_i}).$$ Now inclusion of the radiation from $W$ into eq. (\[eq:multi-real1\]) amounts to the following modification: $$\label{eq:multi-W-real1}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\\ \simeq &
\sum_{{l_a,l_c,n=0}\atop{l_a+l_c+n=N}}^N \sum_{\pi}^{N!}
\biggl[
\biggl(\frac{2p_a^{\mu_{\pi_1}}}{2p_a k_{\pi_1}}
\frac{2p_a^{\mu_{\pi_2}}}{2p_a k_{\pi_1}+2p_a k_{\pi_2}}
\dots
\frac{2p_a^{\mu_{\pi_{l_a}}}}
{2p_a k_{\pi_1}+2p_a k_{\pi_2}+\dots+2p_a k_{\pi_{l_a}}}
\biggr)
\\ &
\frac{-g^{\lambda\sigma}+Q_{\pi_0}^\lambda Q_{\pi_0}^{\sigma}/M_W^2}
{Q_{\pi_0}^2-M_W^2}
\prod_{i=1}^n
\frac{(-2Q_{\pi_0}^{\mu_{l_a+i}})}{Q_{\pi_{l_a+i}}^2-M_W^2}
\\&
\biggl(\frac{-2p_c^{\mu_{\pi_{l_a+n+1}}}}{2p_c k_{\pi_{l_a+n+1}}}
\frac{-2p_c^{\mu_{\pi_{l_a+n+2}}}}
{ 2p_c k_{\pi_{l_a+n+1}}+2p_c k_{\pi_{l_a+n+2}}}
\dots
\frac{-2p_c^{\mu_{\pi_{l_a+n+l_c}}}}
{2p_c k_{\pi_{l_a+n+1}} +2p_c k_{\pi_{l_a+n+2}} +\dots +2p_c
k_{\pi_{l_a+n+l_c}}}
\biggr)
\biggr],
\end{aligned}$$ where we temporarily choose $Q_0$ as the four-momentum in the numerators. The unmatched indices $\lambda\sigma$ in the $W$ propagator are to be treated “symbolically” as we do not write a complete expression for ${\EuScript{M}}$. At this moment we plug in the formula (\[eq:identity\]) $$\label{eq:multi-W-real2}
\begin{aligned}
&\!\!\!\!\!\!{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N) \simeq
\\ \simeq &
\sum_{{l_a,l_c,n=0}\atop{l_a+l_c+n=N}}^N
\sum_{{l_g,l_h=0}\atop{l_g+l_h=n}}^n \sum_{\pi}^{N!}
\biggl[
\biggl(\frac{2p_a^{\mu_{\pi_1}}}{2p_a k_{\pi_1}}
\frac{2p_a^{\mu_{\pi_2}}}{2p_a k_{\pi_1}+2p_a k_{\pi_2}}
\dots
\frac{2p_a^{\mu_{\pi_{l_a}}}}
{2p_a k_{\pi_1}+2p_a k_{\pi_2}+\dots+2p_a k_{\pi_{l_a}}}
\biggr)
\\&
\biggl(\frac{-2Q_{\pi_0}^{\mu_{\pi_{l_a+1}}}}{2Q_{\pi_{l_g}} k_{\pi_{l_a+1}}}
\frac{-2Q_{\pi_0}^{\mu_{\pi_{l_a+2}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+1}}+2Q_{\pi_{l_g}} k_{\pi_{l_a+2}}}
\dots
\frac{-2Q_{\pi_0}^{\mu_{\pi_{l_a+l_g}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+1}} +2Q_{\pi_{l_g}} k_{\pi_{l_a+2}} +\dots
+2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g}}}
\biggr)
\\ &
\frac{-g^{\lambda\sigma}+Q_{\pi_{l_g}}^\lambda Q_{\pi_{l_g}}^{\sigma}/M_W^2}
{Q_{\pi_{l_g}}^2-M_W^2}
\\&
\biggl(
\frac{2Q_{\pi_0}^{\mu_{\pi_{l_a+l_g+1}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+1}}}
\frac{2Q_{\pi_0}^{\mu_{\pi_{l_a+l_g+2}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+1}}+2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+2}}}
\\&
~~~~~~~~~~~~~~~~~~~ \dots
\frac{2Q_{\pi_0}^{\mu_{\pi_{l_a+l_g+l_h}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+1}} +2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+2}}
+\dots +2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+l_h}}}
\biggr)
\\&
\biggl(\frac{-2p_c^{\mu_{\pi_{l_a+n+1}}}}{2p_c k_{\pi_{l_a+n+1}}}
\frac{-2p_c^{\mu_{\pi_{l_a+n+2}}}}
{2p_c k_{\pi_{l_a+n+1}}+2p_c k_{\pi_{l_a+n+2}}}
\dots
\frac{-2p_c^{\mu_{\pi_{l_a+n+l_c}}}}
{2p_c k_{\pi_{l_a+n+1}} +2p_c k_{\pi_{l_a+n+2}} +\dots +2p_c
k_{\pi_{l_a+n+l_c}}}
\biggr)
\biggr].
\end{aligned}$$ The double sum can be converted into a single one $$\sum_{{l_a,l_c,n=0}\atop{l_a+l_c+n=N}}^N
\sum_{{l_g,l_h=0}\atop{l_g+l_h=n}}^n
=
\sum_{{l_a,l_c,l_g,l_h=0}\atop{l_a+l_c+l_g+l_h=N}}^N$$ and the four groups of permutations can be executed as in eq. (\[eq:multi-real2\]): $$\label{eq:multi-W-real3}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N)
\simeq
\sum_{{l_a,l_c,l_g,l_h=0}\atop{l_a+l_c+l_g+l_h=N}}^N
\sum_{\pi/\pi_{l_a}/\pi_{l_g}/\pi_{l_h}/\pi_{l_c}}
^{N!/l_a!/l_g!/l_h!/l_c!}
\frac{-g^{\lambda\sigma}+Q_{\pi_{l_g}}^\lambda Q_{\pi_{l_g}}^{\sigma}/M_W^2}
{Q_{\pi_{l_g}}^2-M_W^2}
\\ &
\biggl(\prod_{i=1}^{l_a} \frac{2p_a^{\mu_{\pi_i}}}{2p_a k_{\pi_i}}
\biggr)
\biggl(\prod_{i=1}^{l_g}
\frac{-2Q_{\pi_0}^{\mu_{\pi_{l_a+i}}}}{2Q_{\pi_{l_g}} k_{\pi_{l_a+i}}}
\biggr)
\biggl(\prod_{i=1}^{l_h}
\frac{2Q_{\pi_0}^{\mu_{\pi_{l_a+l_g+i}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_a+l_g+i}}}
\biggr)
\biggl(\prod_{i=1}^{l_c}
\frac{-2p_c^{\mu_{\pi_{l_a+l_g+l_h+i}}}}{2p_c k_{\pi_{l_a+l_g+l_h+i}}}
\biggr).
\end{aligned}$$ The first two terms in the brackets describe the emission from the production part (from the lines $a$ and $g$). $Q_{\pi_{l_g}}\equiv Q_g$ is defined there as $Q_g=p_{ab} -K_P,~~K_P=K_a +K_g$ and is the same for all terms of the product for given (fixed) $K_P$, regardless of the choice of $K_a$ and $K_g$. Therefore, these two products can be combined into one as in eqs. (\[eq:multi-real3\]) and (\[eq:multi-real4\]). The same holds for the last two terms which describe emission from the decays with $Q_g=p_{cd} +K_D,~~K_D=K_c +K_h$: $$\label{eq:multi-W-real4}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N)
\simeq
\sum_{{l_P,l_D=0}\atop{l_P+l_D=N}}^N
\sum_{\pi/\pi_{l_P}/\pi_{l_D}}
^{N!/l_P!/l_D!}
\frac{-g^{\lambda\sigma}+Q_{\pi_{l_g}}^\lambda Q_{\pi_{l_g}}^{\sigma}/M_W^2}
{Q_{\pi_{l_g}}^2-M_W^2}
\\ &
\prod_{i=1}^{l_P}\biggl( \frac{2p_a^{\mu_{\pi_i}}}{2p_a k_{\pi_i}}
-\frac{2Q_{\pi_0}^{\mu_{\pi_{i}}}}{2Q_{\pi_{l_g}} k_{\pi_{i}}}
\biggr)
\prod_{i=1}^{l_D} \biggl(
\frac{2Q_{\pi_0}^{\mu_{\pi_{l_P+i}}}}
{2Q_{\pi_{l_g}} k_{\pi_{l_P+i}}}
-\frac{2p_c^{\mu_{\pi_{l_P+i}}}}{2p_c k_{\pi_{l_P+i}}}
\biggr),
\end{aligned}$$ where $l_a+l_g=l_P$ and $l_h+l_c=l_D$. The sum over permutations can be once more replaced by the sum over partitions (cf. eq. (\[eq:sumpart\])): $$\label{eq:multi-W-real5}
\begin{aligned}
&{{\EuScript{M}}^{(0)}_N}^{\mu_1,\dots,\mu_N}(k_1,\dots,k_N)
\simeq
\sum_{\wp=(P,D)^N}^{2^N}
\frac{-g^{\lambda\sigma}+Q_{g}^\lambda Q_{g}^{\sigma}/M_W^2}
{Q_{g}^2-M_W^2}
\prod_{i=1}^{N}
\biggl(\frac{2\theta_{\wp_i}p_{\wp_i}^{\mu_{i}}}{2p_{\wp_i} k_i}
-\frac{2\theta_{\wp_i}Q_{\wp_i}^{\mu_{i}}}
{2Q_{\wp_i} k_i}
\biggr)
\\&
=
\sum_{\wp=(P,D)^N}^{2^N}
\frac{-g^{\lambda\sigma}+Q_{g}^\lambda Q_{g}^{\sigma}/M_W^2}
{Q_{g}^2-M_W^2}
\prod_{i=1}^{N}
j_{\wp_i}^{\mu_i},
~~~j_{P}^{\mu_i}=
\frac{2 p_{a}^{\mu_i}}{2p_{a} k_i}
-\frac{2 Q_{g}^{\mu_{i}}}
{2Q_{g} k_i},
~~~j_{D}^{\mu_i}=
\frac{2 Q_{g}^{\mu_{i}}}
{2Q_{g} k_i}
-\frac{2 p_{c}^{\mu_i}}{2p_{c} k_i}.
\end{aligned}$$ In eq. (\[eq:multi-W-real5\]) we have used a freedom of defining $Q_{\pi_0}$ to replace it with $Q_{\wp_i}\equiv Q_g$. Note that, contrary to $p_X$, the vectors $Q_X$ depend on the choice of partitions, i.e. vary from a partition to partition. This prevents us from collapsing the remaining sum over partitions, quite analogously as in the case of the neutral resonance.
Details of virtual form-factor
==============================
In the following we are going to generalise the YFS [@Yennie:1961ad] virtual form-factor function $\alpha B$ to the general case with charged intermediate resonances.
![ Example of one real and one virtual photon emissions. The electric current is a sum of contributions from all external particles. This is why it is attached to the dashed line which crosses all relevant external lines. The rest of the Feynman diagram is visualised as the internal dark box. []{data-label="fig:current"}](current.pdf){height="80mm" width="80mm"}
In order to introduce the notation, let us first write down explicitly the emission factor for a single-real photon $$j^\mu(k)= ie \sum_{X=a,b,c,d,e,f}\!\!\! Q_X \theta_X\;
\frac{ 2p_X^\mu}{2p_X k},$$ where $\theta_X=+1,-1$ for particles in the initial and final state, respectively, $Q_X$ is the charge of the particle (in the units of the positron charge $e$) and the single-virtual photon current reads $$J^\mu(k)= \sum_{X=a,b,c,d,e,f} \hat{J}_{X}^\mu(k),\qquad
\hat{J}_{X}^\mu(k) \equiv
Q_X \theta_X\; \frac{ 2p_X^\mu\theta_X -k^\mu}
{k^2- 2p_X k\theta_X +i\varepsilon},$$ see fig. \[fig:current\]. For the virtual corrections we always have an even number of the $J^\mu(k)$ currents paired in the so-called virtual $S$-factor $$S(k)= J(k) \circ J(k) =\sum_{ {X=a,b,c,d,e,f}\atop{Y=a,b,c,d,e,f} } J_X(k) \circ J_Y(k),$$ where $$J_X(k) \circ J_Y(k) = J_X(k) \cdot J_Y(-k),\; {\rm for}\; X\neq Y,\quad
J_X(k) \circ J_X(k) = J_X(k) \cdot J_X(k).$$ In fig. \[fig:current\] we illustrate all that in a visual way. The contribution $J_X(k) \cdot J_X(k)$ looks diagrammatically like the self-energy, but in fact it comes from the charge renormalisation, see the discussion in refs. [@Yennie:1961ad; @weinberg95-book].
In the derivation of $\exp(\alpha B)$ of ref. [@Yennie:1961ad] (taking as an example the four-fermion production process) we arrive at a certain stage where the contributions from all the real and virtual photons are factorised.
![ The EEX amplitude for four-fermion production. The electric current is a sum of contributions from all external particles. This is why it is attached to the dashed line which crosses all relevant external lines. We use the collective notation for multiple-photon lines which is defined in the plot. []{data-label="fig:yfs61amp"}](yfs61amp.pdf){height="65mm" width="160mm"}
The corresponding scattering amplitude with $m$ real and any number of virtual photons taken in the soft-photon approximation, visualised in fig. \[fig:yfs61amp\], reads $$\begin{aligned}
&M^{\mu_1\mu_2...\mu_m}(k_1,k_2,...,k_m)=
\\&={\EuScript{M}}\;
\prod_{l=1}^m j^\mu(k_l) \;
\sum_{n=0}^\infty
\frac{1}{n!}
\prod_{i=1}^n
\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_i }{k_i^2-\lambda^2+i\varepsilon}\;
J^\mu(k_i) \circ J_\mu(k_i).
\end{aligned}$$ The sum over virtual photons is done trivially, resulting in the exponential form-factor: $$\begin{aligned}
&M^{\mu_1\mu_2...\mu_m}(k_1,k_2,...,k_m)= {\EuScript{M}}\;
\prod_{l=1}^m j^\mu(k_l) \; e^{\alpha B_{6}},
\\
&B_{6}=
\int \frac{i}{(2\pi)^3}\;
\frac{d^4 k }{k^2-\lambda^2+i\varepsilon}\; J(k) \circ J(k).
\end{aligned}$$ Note that in the residual function ${\EuScript{M}}$ there is no “recoil” dependence on photon momenta, we are therefore limited to very soft photons ($E_\gamma\ll \Gamma_W$) in the process of our interest.
![ The CEEX amplitude for $WW$ production and two decays in the soft-photon approximation. Visualised are all classes of virtual and real photon emissions. []{data-label="fig:ceex"}](ceex.pdf){height="80mm" width="100mm"}
Let us now take into account the double-resonant character of the process, see fig. \[fig:ceex\]. After factorising all the real and virtual soft photons, and introducing a new source of emission from the intermediate resonant $W$’s, we use the identity of eq. (\[eq:principal-ident\]) to arrive at the amplitude depicted in fig. \[fig:ceex\], which can be written explicitly as follows: $$\begin{aligned}
&M^{\mu_{_{1_1}}...\mu_{_{3_{n_3}}}}_{n_1 n_2 n_3}(k_{1_1},...,k_{3_{n_3}})=
{\EuScript{M}}_0\;
\prod_{i_1={1}}^{n_1} j^{\mu_{i_1}}_{P}(k_{i_1}) \;
\prod_{i_2={1}}^{n_2} j^{\mu_{i_2}}_{D_1}(k_{i_2}) \;
\prod_{i_3={1}}^{n_3} j^{\mu_{i_3}}_{D_2}(k_{i_3}) \;
\\&\quad
\sum_{n_4=0}^\infty \frac{1}{n_4!}
\prod_{i_4=1}^{n_4} \alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_4} }{k_{i_4}^2-m_\gamma^2+i\varepsilon}\;
J_P(k_{i_4})\circ J_P(k_{i_4})\;
\\&\quad
\sum_{n_5=0}^\infty \frac{1}{n_5!}
\prod_{i_5=1}^{n_5} \alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_5} }{k_{i_5}^2-m_\gamma^2+i\varepsilon}\;
J_{D_1}(k_{i_5})\circ J_{D_1}(k_{i_5})
\\&\quad
\sum_{n_6=0}^\infty \frac{1}{n_6!}
\prod_{i_6=1}^{n_6} \alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_6} }{k_{i_6}^2-m_\gamma^2+i\varepsilon}\;
J_{D_2}(k_{i_6})\circ J_{D_2}(k_{i_6})
\\&\quad
\sum_{n_7=0}^\infty \frac{1}{n_7!}
\prod_{i_7=1}^{n_7} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_7} }{k_{i_7}^2-m_\gamma^2+i\varepsilon}\;
J_{P}(k_{i_7})\circ J_{D_1}(k_{i_7})
\\&\quad
\sum_{n_8=0}^\infty \frac{1}{n_8!}
\prod_{i_8=1}^{n_8} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_8} }{k_{i_8}^2-m_\gamma^2+i\varepsilon}\;
J_{P}(k_{i_8})\circ J_{D_2}(k_{i_8})
\\&\quad
\sum_{n_9=0}^\infty \frac{1}{n_9!}
\prod_{i_9=1}^{n_9} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_9} }{k_{i_9}^2-m_\gamma^2+i\varepsilon}\;
J_{D_1}(k_{i_9})\circ J_{D_2}(k_{i_9})
\\&\quad
\frac{1}{(p_{cd}+K_2-K_7+K_9)^2-M^2 }\; \frac{1}{(p_{ef}+K_3-K_8-K_9)-M^2 },
\end{aligned}$$ where $K_l=\sum_{i_l=1}^{n_l} k_{i_l}$. We have defined [*nine*]{} groups of photons (labeled by $l$); $l=1,2,3$ corresponds, respectively, to the real emission from the $W$-pair production and their decays, the virtual photons corresponding to the same sources are denoted as the group (4,5,6), while (7,8) corresponds to the virtual photons attached to the production and the decay. Finally, the virtual photons denoted as the group (9) connect the decays of the first and the second $W$-boson.
The most interesting part in the above expression is that the product of two resonance propagators includes all the relevant recoil dependence on the real and virtual photon momenta. This dependence can be read easily from fig. \[fig:ceex\]. We are now not limited by $E_\gamma\ll \Gamma_W$, but rather by $E_\gamma\ll \sqrt{s}$. One important feature is that propagators do not depend on $K_4$, $K_5$ and $K_6$ – this is why the sums over relevant photons can be immediately folded into three standard YFS form-factor $e^{\alpha B}$, for the production and decay processes. This, however, cannot be done for the three virtual-interference contributions because the propagators do depend on $K_7$, $K_8$ and $K_9$. The dependence on the real photons $K_2$ and $K_3$ is not harmful for our task of summing up [*virtual*]{} contributions to the infinite order: $$\begin{aligned}
&M^{\mu_{_{1_1}}...\mu_{_{3_{n_3}}}}_{n_1 n_2 n_3}(k_{1_1},...,k_{3_{n_3}})=
{\EuScript{M}}_0\;
\prod_{i_1={1}}^{n_1} j^{\mu_{i_1}}_{P}(k_{i_1}) \;
\prod_{i_2={1}}^{n_2} j^{\mu_{i_2}}_{D_1}(k_{i_2}) \;
\prod_{i_3={1}}^{n_3} j^{\mu_{i_3}}_{D_2}(k_{i_3}) \;
\\&\quad
e^{\alpha B_P}\; e^{\alpha B_{D_1}}\; e^{\alpha B_{D_2}}\;
\sum_{n_7=0}^\infty \frac{1}{n_7!}
\prod_{i_7=1}^{n_7} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_7} }{k_{i_7}^2-m_\gamma^2}\;
J_P(k_{i_7})\circ J_{D_1}(k_{i_7})\;
\\&\quad
\sum_{n_8=0}^\infty \frac{1}{n_8!}
\prod_{i_8=1}^{n_8} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_8} }{k_{i_8}^2-m_\gamma^2}\;
J_{P}(k_{i_8})\circ J_{D_2}(k_{i_8})
\\&\quad
\sum_{n_9=0}^\infty \frac{1}{n_9!}
\prod_{i_9=1}^{n_9} 2\alpha \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k_{i_9} }{k_{i_9}^2-m_\gamma^2}\;
J_{D_1}(k_{i_9})\circ J_{D_2}(k_{i_9})
\\&\quad
\frac{1}{(U_2-K_7+K_9)^2-M^2 }\; \frac{1}{(V_3-K_8-K_9)-M^2 },
\end{aligned}$$ where $U_2=p_{cd}+K_2$ and $V_3=p_{ef}+K_3$, and $$\alpha B_X = \int \frac{i}{(2\pi)^3}\;
\frac{d^4 k }{k^2-m_\gamma^2 +i\varepsilon}\; J_{X}(k)\circ J_{X}(k),\quad X=P,D_1,D_2.$$ At this point we use the following approximations (valid in the soft-photon limit): $$\begin{aligned}
&\frac{1}{(U_2-K_7+K_9)^2-M^2 }\;
\simeq \frac{1}{ U_2^2-M^2 -2U_2 K_7 +2U_2 K_9 }\;
\\&\quad
= \frac{1}{U_2^2-M^2}\;
\frac{1}{ 1-\sum_{i_7} \frac{2U_2 k_{i_7}}{U_2^2-M^2} +\sum_{i_9} \frac{2U_2 k_{i_9}}{U_2^2-M^2} }\;
\\&\quad
\simeq \frac{1}{U_2^2-M^2}\;
\prod_{i_7} \frac{1}{1 -\frac{2U_2 k_{i_7}}{U_2^2-M^2}}\;
\prod_{i_9} \frac{1}{1 +\frac{2U_2 k_{i_9}}{U_2^2-M^2}}\;
\\&\quad
\simeq \frac{1}{U_2^2-M^2}\;
\prod_{i_7} \frac{U_2^2-M^2}{(U_2-k_{i_7})^2-M^2}\;
\prod_{i_9} \frac{U_2^2-M^2}{(U_2+k_{i_9})^2-M^2}\;,
\end{aligned}$$ which, together with the analogous approximation for the second propagator, allows us to fold-in three remaining sums into exponents: $$\label{eq:ceex-formfac}
\begin{aligned}
&M^{\mu_{_{1_1}}...\mu_{_{3_{n_3}}}}_{n_1 n_2 n_3}(k_{1_1},...,k_{3_{n_3}})=
{\EuScript{M}}_0\;
\prod_{i_1={1_1}}^{n_1} j^{\mu_{i_1}}_{P}(k_{i_1}) \;
\prod_{i_2={1_2}}^{n_2} j^{\mu_{i_2}}_{D_1}(k_{i_2}) \;
\prod_{i_3={1_3}}^{n_3} j^{\mu_{i_3}}_{D_2}(k_{i_3}) \;
\\&\qquad\qquad
e^{ \alpha B_{10}^{\rm CEEX}(p_{cd}+K_2,p_{ef}+K_3 )}\;
\frac{1}{(p_{cd}+K_2)^2-M^2 }\; \frac{1}{(p_{ef}+K_3)-M^2 },
\end{aligned}$$ where $$\begin{aligned}
&\alpha B_{10}^{\rm CEEX}(U,V)= \alpha B_P+\alpha B_{D_1}+\alpha B_{D_2}
+2\alpha B_{P\otimes D_1}(U)
+2\alpha B_{P\otimes D_2}(V)
+2\alpha B_{D_1\otimes D_2}(U,V),\;
\\&
\alpha B_{P\otimes D_1}(U)=
\int \frac{i}{(2\pi)^3}\;
\frac{d^4 k }{k^2-m_\gamma^2+i\varepsilon}\; J_P(k)\circ J_{D_1}(k)
\frac{U^2-M^2}{(U-k)^2-M^2},
\\&
\alpha B_{P\otimes D_2}(V)=
\int \frac{i}{(2\pi)^3}\;
\frac{d^4 k}{k^2-m_\gamma^2+i\varepsilon}\; J_{P}(k)\circ J_{D_2}(k)
\frac{V^2-M^2}{(V-k)^2-M^2},
\\&
\alpha B_{D_1\otimes D_2}(U,V)=
\int \frac{i}{(2\pi)^3}\;
\frac{d^4 k }{k^2-m_\gamma^2+i\varepsilon}\; J_{D_1}(k)\circ J_{D_2}(k)
\frac{U^2-M^2}{(U+k)^2-M^2}\;
\frac{V^2-M^2}{(V-k)^2-M^2}.
\end{aligned}$$ Let us note that in the no-recoil limit $U-k\to U$, $V-k\to V$, i.e. $k\ll
\Gamma_W$, the function $B_{10}^{\rm CEEX}(U,V)$ reduces to $B_6^{\rm CEEX}$, in an analogous way to eq. (\[eq:real-single\]).
In eq. (\[eq:ceex-formfac\]) and in all previous steps the contributions of real photon were taken as just one term (in which we know to which subprocess every real photon belongs) from the grand sum (as defined e.g. in eq. (\[eq:ceex1b-resonanat\])), over all $3^n$ photon assignments $(P,D_1,D_2)^n$, in which we know to which subprocess every real photon belongs[^27]. Let us restore this coherent sum over all photon assignments in the following compact final expression: $$\label{eq:ceex-formfactor}
M^{\mu_1...\mu_n}(k_1,k_2,...,k_n)=\!\!\!\!\!
\sum_{\wp\in(P,D_1,D_2)^n}\!\!\!\!
{\EuScript{M}}_0\; \prod_{i=1}^{n} j^{\mu_i}_{\wp_i}(k_i)\;
e^{ \alpha B_{10}^{\rm CEEX}(U_\wp,V_\wp )}\;
\frac{1}{U_\wp^2-M^2 }\; \frac{1}{V_\wp^2-M^2 },$$ where $U_\wp = p_{cd}+\sum\limits_{\wp_i=D_1} k_i$ and $V_\wp = p_{ef}+\sum\limits_{\wp_i=D_2} k_i$.
Eq. (\[eq:ceex-formfactor\]) is the principal result of this Appendix. The CEEX form-factor of eq. (\[eq:ceex-formfactor\]) is valid for production of a pair of any charged resonances of any spin. The case of a single charged resonance, or more than two charged resonances, can be treated in the same way.
The presented derivation of the virtual form-factor is based to a large extent on the analogy with the real-emission part (see Appendix \[app:b\]) and the cancellation between the real and virtual emissions.
[10]{} url \#1[`#1`]{}urlprefixhref \#1\#2[\#2]{} \#1[\#1]{}
J. Alcaraz, et al., [Precision Electroweak Measurements and Constraints on the Standard Model]{} (2007). [](http://arxiv.org/abs/0712.0929).
S. Schael, et al., [Electroweak Measurements in Electron-Positron Collisions at W-Boson-Pair Energies at LEP]{}, Phys. Rept. 532 (2013) 119–244 (2013). [](http://arxiv.org/abs/1302.3415), [](https://doi.org/10.1016/j.physrep.2013.07.004).
A. Abada, et al., [FCC Physics Opportunities]{}, Eur. Phys. J. C79 (6) (2019) 474 (2019). [](https://doi.org/10.1140/epjc/s10052-019-6904-3).
A. Abada, et al., [FCC-ee: The Lepton Collider]{}, Eur. Phys. J. ST 228 (2) (2019) 261–623 (2019). [](https://doi.org/10.1140/epjst/e2019-900045-4).
A. Blondel, et al., [Standard Model Theory for the FCC-ee: The Tera-Z]{}, in: [Mini Workshop on Precision EW and QCD Calculations for the FCC Studies: Methods and Techniques CERN, Geneva, Switzerland, January 12-13, 2018]{}, 2018 (2018). [](http://arxiv.org/abs/1809.01830).
A. Blondel, A. Freitas, J. Gluza, S. Heinemeyer, S. Jadach, P. Janot, T. Riemann, [Theory Requirements and Possibilities for the FCC-ee and other Future High Energy and Precision Frontier Lepton Colliders]{} (2019). [](http://arxiv.org/abs/1901.02648).
M. W. Grunewald, et al., [Reports of the Working Groups on Precision Calculations for LEP2 Physics: Proceedings. Four fermion production in electron positron collisions]{} (2000). [](http://arxiv.org/abs/hep-ph/0005309).
W. Placzek, [Precision predictions for W pair production at LEP-2]{} (2002) 172–176\[Nucl. Phys. Proc. Suppl.117,172(2003)\] (2002). [](http://arxiv.org/abs/hep-ph/0208138), [](https://doi.org/10.1016/S0920-5632(03)90518-4).
S. Jadach, W. Placzek, M. Skrzypek, B. F. L. Ward, Z. Was, [The Monte Carlo program KoralW version 1.51 and the concurrent Monte Carlo KoralW and YFSWW3 with all background graphs and first order corrections to W pair production]{}, Comput. Phys. Commun. 140 (2001) 475–512 (2001). [](http://arxiv.org/abs/hep-ph/0104049), [](https://doi.org/10.1016/S0010-4655(01)00296-X).
S. Jadach, W. Placzek, M. Skrzypek, B. F. L. Ward, Z. Was, [Precision predictions for (un)stable W+ W- pair production at and beyond LEP-2 energies]{}, Phys. Rev. D65 (2002) 093010 (2002). [](http://arxiv.org/abs/hep-ph/0007012), [](https://doi.org/10.1103/PhysRevD.65.093010).
S. Jadach, W. Placzek, M. Skrzypek, B. F. L. Ward, Z. Was, [Exact O (alpha) gauge invariant YFS exponentiated Monte Carlo for (un)stable $W^+ W^-$ production at and beyond LEP-2 energies]{}, Phys. Lett. B417 (1998) 326–336 (1998). [](http://arxiv.org/abs/hep-ph/9705429), [](https://doi.org/10.1016/S0370-2693(97)01253-7).
S. Jadach, W. Płaczek, M. Skrzypek, B. F. L. Ward, Z. Was, [The Monte Carlo event generator YFSWW3 version 1.16 for W pair production and decay at LEP2/LC energies]{}, Comput. Phys. Commun. 140 (2001) 432–474 (2001). [](http://arxiv.org/abs/hep-ph/0103163), [](https://doi.org/10.1016/S0010-4655(01)00288-0).
J. Fleischer, F. Jegerlehner, M. Zralek, [Radiative Corrections to Helicity Amplitudes for W Pair Production in $e^+ e^-$ Annihilation]{}, Z. Phys. C42 (1989) 409 (1989). [](https://doi.org/10.1007/BF01548446).
K. Kolodziej, M. Zralek, [Helicity amplitudes for spin 0 or spin 1 boson production in $e^+ e^-$ annihilation]{}, Phys. Rev. D43 (1991) 3619–3625 (1991). [](https://doi.org/10.1103/PhysRevD.43.3619).
J. Fleischer, K. Kolodziej, F. Jegerlehner, [W pair production in $e^+ e^-$ annihilation: Radiative corrections including hard bremsstrahlung]{}, Phys. Rev. D47 (1993) 830–836 (1993). [](https://doi.org/10.1103/PhysRevD.47.830).
J. Fleischer, F. Jegerlehner, K. Kolodziej, G. J. van Oldenborgh, [EEWW: A generator for $e^+ e^- \to W^+ W^-$ including one loop and leading photonic two loop corrections]{}, Comput. Phys. Commun. 85 (1995) 29–39 (1995). [](http://arxiv.org/abs/hep-ph/9405380), [](https://doi.org/10.1016/0010-4655(94)00113-G).
S. Jadach, W. Placzek, M. Skrzypek, B. F. L. Ward, Z. Was, [Monte Carlo program KoralW 1.42 for all four-fermion final states in $e^+ e^-$ collisions]{}, Comput. Phys. Commun. 119 (1999) 272–311 (1999). [](http://arxiv.org/abs/hep-ph/9906277), [](https://doi.org/10.1016/S0010-4655(99)00219-2).
D. R. Yennie, S. C. Frautschi, H. Suura, [The infrared divergence phenomena and high-energy processes]{}, Annals Phys. 13 (1961) 379–452 (1961). [](https://doi.org/10.1016/0003-4916(61)90151-8).
S. Jadach, Z. Was, R. Decker, J. H. Kuhn, [The tau decay library TAUOLA: Version 2.4]{}, Comput. Phys. Commun. 76 (1993) 361–380 (1993). [](https://doi.org/10.1016/0010-4655(93)90061-G).
E. Barberio, B. van Eijk, Z. Was, [PHOTOS: A Universal Monte Carlo for QED radiative corrections in decays]{}, Comput. Phys. Commun. 66 (1991) 115–128 (1991). [](https://doi.org/10.1016/0010-4655(91)90012-A).
A. Denner, S. Dittmaier, M. Roth, D. Wackeroth, [Electroweak radiative corrections to e+ e- $\to$ W W $\to$ 4 fermions in double pole approximation: The RACOONWW approach]{}, Nucl. Phys. B587 (2000) 67–117 (2000). [](http://arxiv.org/abs/hep-ph/0006307), [](https://doi.org/10.1016/S0550-3213(00)00511-3).
A. Denner, S. Dittmaier, M. Roth, D. Wackeroth, [RACOONWW1.3: A Monte Carlo program for four fermion production at $e^+ e^-$ colliders]{}, Comput. Phys. Commun. 153 (2003) 462–507 (2003). [](http://arxiv.org/abs/hep-ph/0209330), [](https://doi.org/10.1016/S0010-4655(03)00205-4).
S. Jadach, M. Skrzypek, [QED challenges at FCC-ee precision measurements]{}, Eur. Phys. J. C79 (9) (2019) 756 (2019). [](http://arxiv.org/abs/1903.09895), [](https://doi.org/10.1140/epjc/s10052-019-7255-9).
S. Jadach, B. F. L. Ward, Z. Was, [The Precision Monte Carlo event generator KK for two fermion final states in $e^+ e^-$ collisions]{}, Comput. Phys. Commun. 130 (2000) 260–325 (2000). [](http://arxiv.org/abs/hep-ph/9912214), [](https://doi.org/10.1016/S0010-4655(00)00048-5).
S. Jadach, B. F. L. Ward, Z. Was, [Coherent exclusive exponentiation for precision Monte Carlo calculations]{}, Phys. Rev. D63 (2001) 113009 (2001). [](http://arxiv.org/abs/hep-ph/0006359), [](https://doi.org/10.1103/PhysRevD.63.113009).
S. Jadach, B. F. L. Ward, [YFS2: The Second Order Monte Carlo for Fermion Pair Production at [LEP]{} / [SLC]{} With the Initial State Radiation of Two Hard and Multiple Soft Photons]{}, Comput. Phys. Commun. 56 (1990) 351–384 (1990). [](https://doi.org/10.1016/0010-4655(90)90020-2).
S. Jadach, B. F. L. Ward, [Final state multiple photon effects in fermion pair production at SLC / LEP]{}, Phys. Lett. B274 (1992) 470–472 (1992). [](https://doi.org/10.1016/0370-2693(92)92017-B).
S. Jadach, Z. Was, [Monte Carlo Simulation of the Process e+ e- —> tau+ tau- Including Radiative O(alpha\*\*3) QED Corrections, Mass and Spin]{}, Comput. Phys. Commun. 36 (1985) 191–211 (1985). [](https://doi.org/10.1016/0010-4655(85)90123-7).
S. Jadach, B. F. L. Ward, Z. Was, [The Monte Carlo program KORALZ, version 3.8, for the lepton or quark pair production at LEP / SLC energies]{}, Comput. Phys. Commun. 66 (1991) 276–292 (1991). [](https://doi.org/10.1016/0010-4655(91)90077-X).
W. Placzek, S. Jadach, [Multiphoton radiation in leptonic W boson decays]{}, Eur. Phys. J. C29 (2003) 325–339 (2003). [](http://arxiv.org/abs/hep-ph/0302065), [](https://doi.org/10.1140/epjc/s2003-01223-4).
A. Denner, S. Dittmaier, M. Roth, L. H. Wieders, [Complete electroweak O(alpha) corrections to charged-current $e^+ e^- \to$ 4 fermion processes]{}, Phys. Lett. B612 (2005) 223–232, [\[Erratum: Phys. Lett.B704,667(2011)\]](https://doi.org/10.1016/j.physletb.2011.09.020) (2005). [](http://arxiv.org/abs/hep-ph/0502063), [](https://doi.org/10.1016/j.physletb.2005.03.007).
A. Denner, S. Dittmaier, M. Roth, L. H. Wieders, [Electroweak corrections to charged-current $e^+ e^- \to$ 4 fermion processes: Technical details and further results]{}, Nucl. Phys. B724 (2005) 247–294, [\[Erratum: Nucl. Phys.B854,504(2012)\]](https://doi.org/10.1016/j.nuclphysb.2011.09.001) (2005). [](http://arxiv.org/abs/hep-ph/0505042), [](https://doi.org/10.1016/j.nuclphysb.2005.06.033).
M. Beneke, P. Falgari, C. Schwinn, A. Signer, G. Zanderighi, [Four-fermion production near the W pair production threshold]{}, Nucl. Phys. B792 (2008) 89–135 (2008). [](http://arxiv.org/abs/0707.0773), [](https://doi.org/10.1016/j.nuclphysb.2007.09.030).
S. Actis, M. Beneke, P. Falgari, C. Schwinn, [Dominant NNLO corrections to four-fermion production near the W-pair production threshold]{}, Nucl. Phys. B807 (2009) 1–32 (2009). [](http://arxiv.org/abs/0807.0102), [](https://doi.org/10.1016/j.nuclphysb.2008.08.006).
S. Jadach, W. Płaczek, M. Skrzypek, [Exponentiation in QED and Quasi-Stable Charged Particles]{}, Symmetry 11 (11) (2019) 1389 (2019). [](https://doi.org/10.3390/sym11111389).
R. G. Stuart, [Production cross-sections for unstable particles]{}, Nucl. Phys. B498 (1997) 28–38 (1997). [](http://arxiv.org/abs/hep-ph/9504215), [](https://doi.org/10.1016/S0550-3213(97)00276-9).
S. Jadach, W. Placzek, M. Skrzypek, B. F. L. Ward, Z. Was, [On theoretical uncertainties of the W boson mass measurement at LEP-2]{}, Phys. Lett. B523 (2001) 117–126 (2001). [](http://arxiv.org/abs/hep-ph/0109072), [](https://doi.org/10.1016/S0370-2693(01)01310-7).
W. Beenakker, A. Denner, [Standard model predictions for $W$ pair production in electron - positron collisions]{}, Int. J. Mod. Phys. A9 (1994) 4837–4920 (1994). [](https://doi.org/10.1142/S0217751X94001965).
A. P. Chapovsky, V. A. Khoze, [Screened Coulomb ansatz for the nonfactorizable radiative corrections to the off-shell W+ W- production]{}, Eur. Phys. J. C9 (1999) 449–457 (1999). [](http://arxiv.org/abs/hep-ph/9902343), [](https://doi.org/10.1007/s100529900070).
S. Jadach, B. F. L. Ward, [Exponentiation of Soft Photons in the Monte Carlo: The Case of Bonneau and Martin]{}, Phys. Rev. D38 (1988) 2897, [\[Erratum: Phys. Rev.D39,1471(1989)\]](https://doi.org/10.1103/PhysRevD.38.2897) (1988). [](https://doi.org/10.1103/PhysRevD.39.1471.2).
S. Jadach, B. F. L. Ward, Z. Was, [Coherent exclusive exponentiation CEEX: The Case of the resonant $e^+ e^-$ collision]{}, Phys. Lett. B449 (1999) 97–108 (1999). [](http://arxiv.org/abs/hep-ph/9905453), [](https://doi.org/10.1016/S0370-2693(99)00038-6).
S. Jadach, B. F. L. Ward, [Multi - Photon Monte Carlo for Bhabha Scattering at Low Angles]{}, Phys. Rev. D40 (1989) 3582–3589 (1989). [](https://doi.org/10.1103/PhysRevD.40.3582).
S. Jadach, B. F. L. Ward, Z. Was, [The Monte Carlo program KORALZ, version 4.0, for the lepton or quark pair production at LEP / SLC energies]{}, Comput. Phys. Commun. 79 (1994) 503–522 (1994). [](https://doi.org/10.1016/0010-4655(94)90190-2).
S. Jadach, E. Richter-Was, B. F. L. Ward, Z. Was, [Monte Carlo program BHLUMI-2.01 for Bhabha scattering at low angles with Yennie-Frautschi-Suura exponentiation]{}, Comput. Phys. Commun. 70 (1992) 305–344 (1992). [](https://doi.org/10.1016/0010-4655(92)90196-6).
S. Jadach, W. Placzek, E. Richter-Was, B. F. L. Ward, Z. Was, [Upgrade of the Monte Carlo program BHLUMI for Bhabha scattering at low angles to version 4.04]{}, Comput. Phys. Commun. 102 (1997) 229–251 (1997). [](https://doi.org/10.1016/S0010-4655(96)00156-7).
S. Jadach, B. F. L. Ward, Z. A. Was, S. A. Yost, [Systematic Studies of Exact ${\cal O}(\alpha^2L)$ CEEX EW Corrections in a Hadronic MC for Precision $Z/\gamma^*$ Physics at LHC Energies]{} (2017). [](http://arxiv.org/abs/1707.06502).
S. Jadach, W. Placzek, B. F. L. Ward, [BHWIDE 1.00: O(alpha) YFS exponentiated Monte Carlo for Bhabha scattering at wide angles for LEP-1 / SLC and LEP-2]{}, Phys. Lett. B390 (1997) 298–308, also hep-ph/9608412; The Monte Carlo program BHWIDE is available from [http://cern.ch/placzek]{} (1997). [](http://arxiv.org/abs/hep-ph/9608412), [](https://doi.org/10.1016/S0370-2693(96)01382-2).
M. Greco, G. Pancheri-Srivastava, Y. Srivastava, [Radiative Corrections for Colliding Beam Resonances]{}, Nucl. Phys. B101 (1975) 234–262 (1975). [](https://doi.org/10.1016/0550-3213(75)90304-1).
M. Greco, G. Pancheri-Srivastava, Y. Srivastava, [Radiative Corrections to $e^+
e^- \to \mu^+ \mu^-$ Around the Z0]{}, Nucl. Phys. B171 (1980) 118, [\[Erratum: Nucl. Phys. B197,543(1982)\]](https://doi.org/10.1016/0550-3213(82)90458-8) (1980). [](https://doi.org/10.1016/0550-3213(80)90363-6).
V. S. Fadin, V. A. Khoze, A. D. Martin, [Interference radiative phenomena in the production of heavy unstable particles]{}, Phys. Rev. D49 (1994) 2247–2256 (1994). [](https://doi.org/10.1103/PhysRevD.49.2247).
S. Dittmaier, [A General approach to photon radiation off fermions]{}, Nucl. Phys. B565 (2000) 69–122 (2000). [](http://arxiv.org/abs/hep-ph/9904440), [](https://doi.org/10.1016/S0550-3213(99)00563-5).
S. Weinberg, The Quantum Theory of Fields, The Press Syndicate of the University of Cambridge, 1995 (1995).
[^1]: It is tempting to call them “universal” but, in fact, non-soft subleading collinear perturbative corrections are process-dependent, hence non-universal, while all soft corrections are universal.
[^2]: The genuine EW part of the SM perturbative corrections include non-soft, non-collinear remnants of the QED origin.
[^3]: Including EW [${\cal O}(\alpha^1)$]{} corrections of refs. Refs. [@Fleischer:1988kj; @Kolodziej:1991pk; @Fleischer:1991nw; @Fleischer:1994sq].
[^4]: They depend strongly on experimental cut-offs.
[^5]: This will be done without introducing any parameter in the photon energy distinguishing between soft and hard photons. Minimum photon energy in the Monte Carlo implementation can be set to an arbitrarily low value without any effect on the physical results.
[^6]: The same way as in [[KKMC]{}]{}.
[^7]: See also Appandix \[appendixA\].
[^8]: In fact, we use an approximation for the non-factorisable corrections in terms of the so-called screened Coulomb ansatz of Ref. [@Chapovsky:1999kv].
[^9]: In the case of the sharp $t$-channel exchange singularity in the low-angle Bhabha scattering, the analog of the recoil effect between the electron and positron lines is also worth to take into account in a better way than in EEX$_{B}$.
[^10]: In the case of the low angle Bhabha process neglected are interferences between the electron and positron lines in the Feynman diagram.
[^11]: At LEP energies $\tau$ decays are separated from the production by the giant 2 millimeter distance.
[^12]: Finite $W$ width acts as IR regulator.
[^13]: The gauge invariance is manifest: $j^\mu_P k_{1\mu} = j^\mu_{D_1}k_{1\mu} = j^\mu_{D_2}k_{1\mu} =0.$
[^14]: We hope that the mathematical [*rigour*]{} of this proof will be improved in the future works.
[^15]: The ultimate proof will be provided by someone who will do it in practice.
[^16]: It is typically done using some kinematic manipulations on the four-momenta which are fed into formulae or using Mandelstam variables – they are less sensitive to the presence of spectators.
[^17]: In some four-fermion channels there is no possibility to form a single-resonant $W$.
[^18]: This will be mandatory for the LO corrections.
[^19]: The analogy is however incomplete, because here we take into account photon emission from the intermediate charged $W$ boson, while in ref. [@Jadach:2000ir] neutral resonance $Z$ was considered.
[^20]: Connecting the production and the decays through the spin-density matrix formalism is the logical solution in the EEX case, as for the $\tau$-pair production and decay in [[KORALZ]{}]{}.
[^21]: The baseline distribution has to include all the soft and collinear singularities of the EEX distributions.
[^22]: The transition between these two situations is modeled in our new CEEX in a completely realistic way. It is continuous in the photon energy.
[^23]: The identity $\sum_{j=0}^n \prod_{i=0,i\neq j}^n \frac{1}{x_i-x_j}=0$ is used in the last step.
[^24]: \[foot:23\] Here we use twice the well-known identity $\sum\limits_{perm.}
\frac{1}{a_1(a_1+a_2)(a_1+a_2+a_3)\dots (a_1+a_2+\dots a_n)}
=\frac{1}{a_1a_2\dots a_n}$, where the sum is over all permutations of $(1,2,3,\dots ,n)$.
[^25]: \[foot:13\] Note that the identity (\[eq:sumpart\]) generalises to more than two particles, for example: $$\label{eq:sumpart3}
\sum_{{l_a,l_c,l_e=0}\atop{l_a+l_c+l_e=N}}^N
\sum_{\pi/\pi_a/\pi_c/\pi_e}^{N!/l_a!/l_c!/l_e!}
=
\sum_{\wp=(a,c,e)^N}^{3^N}.$$
[^26]: For instance, for N=2 we have four partitions: $$\frac{2p_{a}^{\mu_1}}{2p_{a} k_{1}}
\frac{2p_{a}^{\mu_2}}{2p_{a} k_{2}}
-
\frac{2p_{a}^{\mu_1}}{2p_{a} k_{1}}
\frac{2p_{c}^{\mu_2}}{2p_{c} k_{2}}
-
\frac{2p_{c}^{\mu_1}}{2p_{c} k_{1}}
\frac{2p_{a}^{\mu_2}}{2p_{a} k_{2}}
+
\frac{2p_{c}^{\mu_1}}{2p_{c} k_{1}}
\frac{2p_{c}^{\mu_2}}{2p_{c} k_{2}}
=
\biggl(
\frac{2p_{a}^{\mu_1}}{2p_{a} k_{1}}
-\frac{2p_{c}^{\mu_1}}{2p_{c} k_{1}}
\biggr)
\biggl(
\frac{2p_{a}^{\mu_2}}{2p_{a} k_{2}}
-\frac{2p_{c}^{\mu_2}}{2p_{c} k_{2}}
\biggr).
\notag$$
[^27]: The final form of the result took shape thanks to the use, at the earlier step, the identity of eq. (\[eq:identity\]).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The detection of the first electromagnetic counterpart to the binary neutron star (BNS) merger remnant GW170817 established the connection between short $\gamma$-ray bursts and BNS mergers. It also confirmed the forging of heavy elements in the ejecta (a so-called [*kilonova*]{}) via the [*r-process nucleosynthesis*]{}. The appearance of non-thermal radio and X-ray emission, as well as the brightening, which lasted more than 100 days, were somewhat unexpected. Current theoretical models attempt to explain this temporal behavior as either originating from a relativistic off-axis jet or a kilonova-like outflow. In either scenario, there is some ambiguity regarding how much energy is transported in the non-thermal electrons versus the magnetic field of the emission region. Combining the VLA (radio) and Chandra (X-ray) measurements with observations in the GeV-TeV domain can help break this ambiguity, almost independently of the assumed origin of the emission. Here we report for the first time on deep H.E.S.S. observations of GW170817 / GRB 170817A between 124 and 272 days after the BNS merger with the full H.E.S.S. array of telescopes, as well as on an updated analysis of the prompt ($<$5 days) observations with the upgraded H.E.S.S. phase-I telescopes. We discuss implications of the H.E.S.S. measurement for the magnetic field in the context of different source scenarios.'
author:
- 'H. Abdalla'
- 'R. Adam'
- 'F. Aharonian'
- 'F. Ait Benkhali'
- 'E.O. Ang[ü]{}ner'
- 'M. Arakawa'
- 'C. Arcaro'
- 'C. Armand'
- 'H. Ashkar'
- 'M. Backes'
- 'V. Barbosa-Martins'
- 'M. Barnard'
- 'Y. Becherini'
- 'D. Berge'
- 'K. Bernl[ö]{}hr'
- 'R. Blackwell'
- 'M. Böttcher'
- 'C. Boisson'
- 'J. Bolmont'
- 'S. Bonnefoy'
- 'J. Bregeon'
- 'M. Breuhaus'
- 'F. Brun'
- 'P. Brun'
- 'M. Bryan'
- 'M. B[ü]{}chele'
- 'T. Bulik'
- 'T. Bylund'
- 'M. Capasso'
- 'S. Caroff'
- 'A. Carosi'
- 'S. Casanova'
- 'M. Cerruti'
- 'T. Chand'
- 'S. Chandra'
- 'A. Chen'
- 'S. Colafrancesco'
- 'M. Cury[ł]{}o'
- 'I.D. Davids'
- 'C. Deil'
- 'J. Devin'
- 'P. deWilt'
- 'L. Dirson'
- 'A. Djannati-Ata[ï]{}'
- 'A. Dmytriiev'
- 'A. Donath'
- 'V. Doroshenko'
- 'L.O’C. Drury'
- 'J. Dyks'
- 'K. Egberts'
- 'G. Emery'
- 'J.-P. Ernenwein'
- 'S. Eschbach'
- 'K. Feijen'
- 'S. Fegan'
- 'A. Fiasson'
- 'G. Fontaine'
- 'S. Funk'
- 'M. F[ü]{}[ß]{}ling'
- 'S. Gabici'
- 'Y.A. Gallant'
- 'F. Gaté'
- 'G. Giavitto'
- 'D. Glawion'
- 'J.F. Glicenstein'
- 'D. Gottschall'
- 'M.-H. Grondin'
- 'J. Hahn'
- 'M. Haupt'
- 'G. Heinzelmann'
- 'G. Henri'
- 'G. Hermann'
- 'J.A. Hinton'
- 'W. Hofmann'
- 'C. Hoischen'
- 'T. L. Holch'
- 'M. Holler'
- 'D. Horns'
- 'D. Huber'
- 'H. Iwasaki'
- 'M. Jamrozy'
- 'D. Jankowsky'
- 'F. Jankowsky'
- 'A. Jardin-Blicq'
- 'I. Jung-Richardt'
- 'M.A. Kastendieck'
- 'K. Katarzy[ń]{}ski'
- 'M. Katsuragawa'
- 'U. Katz'
- 'D. Khangulyan'
- 'B. Kh[é]{}lifi'
- 'J. King'
- 'S. Klepser'
- 'W. Kluźniak'
- 'Nu. Komin'
- 'K. Kosack'
- 'D. Kostunin'
- 'M. Kraus'
- 'G. Lamanna'
- 'J. Lau'
- 'A. Lemière'
- 'M. Lemoine-Goumard'
- 'J.-P. Lenain'
- 'E. Leser'
- 'C. Levy'
- 'T. Lohse'
- 'I. Lypova'
- 'J. Mackey'
- 'J. Majumdar'
- 'D. Malyshev'
- 'V. Marandon'
- 'A. Marcowith'
- 'A. Mares'
- 'C. Mariaud'
- 'G. Martí-Devesa'
- 'R. Marx'
- 'G. Maurin'
- 'P.J. Meintjes'
- 'A.M.W. Mitchell'
- 'R. Moderski'
- 'M. Mohamed'
- 'L. Mohrmann'
- 'J. Muller'
- 'C. Moore'
- 'E. Moulin'
- 'T. Murach'
- 'S. Nakashima'
- 'M. de Naurois'
- 'H. Ndiyavala'
- 'F. Niederwanger'
- 'J. Niemiec'
- 'L. Oakes'
- 'P. O’Brien'
- 'H. Odaka'
- 'S. Ohm'
- 'E. de Ona Wilhelmi'
- 'M. Ostrowski'
- 'I. Oya'
- 'M. Panter'
- 'R.D. Parsons'
- 'C. Perennes'
- 'P.-O. Petrucci'
- 'B. Peyaud'
- 'Q. Piel'
- 'S. Pita'
- 'V. Poireau'
- 'A. Priyana Noel'
- 'D.A. Prokhorov'
- 'H. Prokoph'
- 'G. P[ü]{}hlhofer'
- 'M. Punch'
- 'A. Quirrenbach'
- 'S. Raab'
- 'R. Rauth'
- 'A. Reimer'
- 'O. Reimer'
- 'Q. Remy'
- 'M. Renaud'
- 'F. Rieger'
- 'L. Rinchiuso'
- 'C. Romoli'
- 'G. Rowell'
- 'B. Rudak'
- 'E. Ruiz-Velasco'
- 'V. Sahakian'
- 'S. Saito'
- 'D.A. Sanchez'
- 'A. Santangelo'
- 'M. Sasaki'
- 'R. Schlickeiser'
- 'F. Sch[ü]{}ssler'
- 'A. Schulz'
- 'H. Schutte'
- 'U. Schwanke'
- 'S. Schwemmer'
- 'M. Seglar-Arroyo'
- 'M. Senniappan'
- 'A.S. Seyffert'
- 'N. Shafi'
- 'K. Shiningayamwe'
- 'R. Simoni'
- 'A. Sinha'
- 'H. Sol'
- 'A. Specovius'
- 'M. Spir-Jacob'
- '[Ł]{}. Stawarz'
- 'R. Steenkamp'
- 'C. Stegmann'
- 'C. Steppa'
- 'T. Takahashi'
- 'T. Tavernier'
- 'A.M. Taylor'
- 'R. Terrier'
- 'D. Tiziani'
- 'M. Tluczykont'
- 'C. Trichard'
- 'M. Tsirou'
- 'N. Tsuji'
- 'R. Tuffs'
- 'Y. Uchiyama'
- 'J. van der Walt'
- 'C. van Eldik'
- 'C. van Rensburg'
- 'B. van Soelen'
- 'G. Vasileiadis'
- 'J. Veh'
- 'C. Venter'
- 'P. Vincent'
- 'J. Vink'
- 'F. Voisin'
- 'H.J. V[ö]{}lk'
- 'T. Vuillaume'
- 'Z. Wadiasingh'
- 'S.J. Wagner'
- 'R. White'
- 'A. Wierzcholska'
- 'R. Yang'
- 'H. Yoneda'
- 'M. Zacharias'
- 'R. Zanin'
- 'A.A. Zdziarski'
- 'A. Zech'
- 'A. Ziegler'
- 'J. Zorn'
- 'N. Zywucka'
- 'X. Rodrigues'
bibliography:
- 'EM170817\_HESS.bib'
title: 'Probing the magnetic field in the GW170817 outflow using H.E.S.S. observations'
---
Introduction {#sec:introduction}
============
The Gravitational Wave (GW) event detected on August 17$^{\mathrm{th}}$ 2017 by the advanced LIGO and Virgo detectors resulted from the merger of two neutron stars (NS). The GW signal was followed after $\sim$2 seconds by a short, low-luminosity $\gamma$-ray burst (GRB) and seen by the Fermi-GBM [@GBM2017] and INTEGRAL-SPI [@Integral2017] instruments. Observations in the optical band later associated this GRB (GRB 170817A) as the counterpart of GW170817 and localized it to the host galaxy NGC4993 [@Coulter2017a]. Only flux upper limits to the counterpart of GW170817 have been reported at energies beyond the X-ray domain by the H.E.S.S. Collaboration [@Abdalla2017a]. The fading UV, optical, and infrared radiation was followed by a rising non-thermal radio and X-ray signal after $\sim$9 days. This behavior, as predicted by @Takami2014, is indicative of efficient particle acceleration in the NS merger remnant and subsequent synchrotron emission of accelerated electrons in the ejecta’s magnetic field.
After $\sim$160 days, the synchrotron radiation started to plateau and later fade (Fig. \[fig:obs\]). This is similar to the behavior of a young supernova remnant, and suggests that the ejecta is transitioning from the free-expansion to the Sedov-Taylor phase when the ejected mass of the merger remnant equals the swept-up circumstellar material.
@Rodrigues2019 infer a total kinetic energy in the ejecta of $10^{51}$erg, implying a total number of electrons in the remnant of $10^{55}$. Accelerated electrons should also scatter off the intense radio and X-ray synchrotron radiation field and produce synchrotron self-Compton (SSC) emission. The expected peak energy of the SSC component depends on the maximum accelerated electron energy as probed by the X rays. While the radio to X-ray emission probes the product of energy in electrons, $u_e$, and energy in magnetic fields, $u_B$, the SSC component is proportional to $u_{e}^{2}\cdot u_B$. As shown by @Takami2014 and @Rodrigues2019, observations in the $\gamma$-ray regime can break the ambiguity between $u_B$ and $u_e$ and provide crucial insight into the magnetic field in the ejecta as well as the maximum accelerated particle energy.
In this work, we present deep H.E.S.S. observations of GW170817 / GRB 170817A covering the peak and onset of fading in the X-ray and radio light-curves from 124 days to 272 days after the merger. This measurement is accompanied by an improved analysis of the H.E.S.S. data taken on the early (up to 5 days) kilonova. In the next section we present the H.E.S.S. data set and results, followed by a discussion on the implied magnetic field strength in a non-relativistic kilonova scenario and a relativistic jet scenario. Throughout this work we adopt a distance to the host galaxy NGC4993 of 41.0Mpc, corresponding to a redshift of $z=0.009727$ [@Hjorth2017a][^1].
\[sec:analysis\]Data analysis and results
=========================================
---------- --------------- --------------- ---------- -------------- ---------------------- ---------------------------- -------------- ---------------
Data Set Configuration $T - T_0$ Exposure Energy range $F(>E_\mathrm{th})$ $F(1 - 10 \,\mathrm{TeV})$ Zenith angle Reference
(days) (hours) TeV ergcm$^{-2}$s$^{-1}$ ergcm$^{-2}$s$^{-1}$ $^{\circ}$
CT 5 $0.22 - 5.23$ 3.2 0.27$-$8.55 $<1.5\times10^{-12}$ 58 @Abdalla2017a
CT 1$-$5 $0.22 - 5.23$ 3.2 0.56$-$17.8 $<4.7\times10^{-12}$ $<2.8\times10^{-12}$ 58 this work
CT 1$-$5 $124 - 272$ 53.9 0.13$-$23.7 $<1.6\times10^{-12}$ $<3.2\times10^{-13}$ 24 this work
---------- --------------- --------------- ---------- -------------- ---------------------- ---------------------------- -------------- ---------------
![Shown are the H.E.S.S. observation windows (blue areas), VLA radio data at 3 GHz (blue stars) and 6 GHz (orange circles) [@Hallinan2017; @Alexander2017; @Margutti2018; @Dobie2018; @Alexander2018; @Mooley2018; @Hajela:2019], as well as X-ray data (red crosses) [@Nynka2018; @Hajela:2019]. The H.E.S.S. 1$-$10 TeV energy flux upper limits (green arrows) are derived for the prompt and the long-term follow-up with CT1-5 (c.f. Table \[tab:data\]).[]{data-label="fig:obs"}](MWL_observations_energyflux4.pdf){width="50.00000%"}
The data set was obtained from observations with the H.E.S.S. phase II array, including the upgraded 12m-diameter CT1-4 telescopes [@HESS1U:2020] and the large 28m-diameter CT5 telescope. The analysis presented by @Abdalla2017a used monoscopic data of the 28m telescope obtained between 5.3h and 5.3 days after the BNS merger. Here we extend this analysis to also include data taken with CT1-4. Observations from December 2017 to May 2018 with telescopes pointing 0.5$^{\circ}$ offset from GW170817 were conducted allowing for simultaneous estimation of the background level in the signal region as discussed below. The different data sets are summarized in Table \[tab:data\]. A standard data quality selection is applied to the data [@Aharonian2006a; @Abdallah2017a]. The events have been selected and their direction and energy reconstructed using a Monte-Carlo, template-based, shower model technique [@Parsons2014a], requiring at least two telescope to see the same $\gamma$-ray event. With this method, an energy resolution of $\sim$10% and angular resolution (at 68% containment radius) of 0.08$^{\circ}$ above $\gamma$-ray energies of 200GeV is achieved. The resulting energy threshold of data set is $E_{\mathrm{th}} = 130$GeV. We define a circular region-of-interest centered on the optical position of GW170817 [@Coulter2017a] with a radius of 0.09$^{\circ}$ for data sets and – hereafter referred to as the [*ON region*]{}. 10 to 20 background control regions ([*OFF regions*]{}) are defined radially symmetric with respect to the telescope pointing position for each observation [@Fomin1994]. This technique assures that the $\gamma$-ray signal and background are estimated with the same acceptance and under the same observation conditions. No significant $\gamma$-ray excess above the expected background is detected from the direction of GW170817 in any data set. A second analysis using an independent event calibration and reconstruction [@de-Naurois2009a] confirms the result. A search for significant emission on monthly timescales also does not reveal any signal. For the total data set, 95% confidence level (C. L.) upper limits on the photon flux are derived using the method described by @Rolke2005a and assuming an underlying power-law spectral index of the $\gamma$-ray emission of $\alpha=-2.0$ [c.f. @Rodrigues2019]. Note that systematic errors are subdominant compared to statistical uncertainties when deriving upper limits, and are hence not considered here. All results are summarized in Table \[tab:data\].
![Spectra predicted by the SSC modeling of the remnant of GRB 170817A, 110 days after the merger, for two distinct assumptions on the geometry and expansion speed of the remnant: an isotropic, non-relativistic expansion (blue SSC curves) and a relativistic jet (red SSC curves). For both assumptions, we show the minimum magnetic field strength imposed by the H.E.S.S. upper limits (green arrows). Solid and dashed curves are obtained by considering respectively the minimum and maximum flux values allowed by the X-ray measurements (blue points), while retaining compatibility with the radio data (red points).[]{data-label="fig:bmodel"}](em170817_ssc_3.pdf){width="48.30000%"}
Figure \[fig:obs\] depicts the radio and X-ray flux measurements along with the inferred H.E.S.S. energy flux upper limits in the $1-10$TeV energy range for data sets and . Figure \[fig:bmodel\] shows the inferred energy flux upper limits at the 95% confidence level (C. L.) in the VHE $\gamma$-ray range from GRB 170817A for data set , which has the best sensitivity at TeV energies. In the next section we discuss how the H.E.S.S. results constrain the magnetic field in the GW170817 ejecta in a jet- and in a kilonova scenario.
\[sec:discussion\]Discussion
============================
Recent detections of VHE emission from GRBs over minutes [@MAGIC:2019] and hours [@HESS:2019] motivate the search for very late time emission from GRB-related events, like the remnant of GW170817. The H.E.S.S. differential upper limits can be translated into an integral energy flux limit, for a given assumption on the spectrum of the radiating particles. In turn, this limit provides a constraint on the magnetic field strength under the assumption of one-zone synchrotron emission with corresponding inverse Compton (IC) emission.
Observational evidence suggests that at early times, the kilonova provides the dominant target radiation field in the remnant [@Villar:2017wcc]. However, the decay of this component, whose flux falls steeply with $t^{-2.3}$, results in a late-time dominance of the synchrotron radiation in the source. It is therefore naturally expected that at late times SSC will dominate the remnant’s IC emission.
The measured X-ray flux of the source can be used to infer the X-ray luminosity emitted as synchrotron radiation, $L_X$. In order to consistently model the emission from this electron population, a geometry assumption is necessary. We consider two scenarios: one where the remnant expands isotropically and non-relativistically, and the other where a relativistic jet is launched.
In the isotropic scenario, we assume a volume-filled spherical emitter with radius $R_\mathrm{iso}=\beta c\Delta t$, where $\Delta t=110~\mathrm{days}$ is the time since the merger [^2]. Based on photospheric velocity measurements of the remnant [@Piro:2017ayh], we consider a value of the expansion speed of $\beta=0.2$.
In the relativistic scenario, we consider a jet with speed $\beta=0.94$ (corresponding to a Lorentz factor of $\Gamma=3$) at late times, and a jet opening angle[^3] of $\theta^{'}_{\mathrm{jet}}=5^\circ$, observed at an angle of $\theta_\mathrm{obs}=20^\circ\approx180^\circ/\pi\Gamma$ from the jet axis; the source is therefore observed with a doppler factor $\delta=\Gamma=3$. These parameter values are motivated by radio observations of superluminal motion of the source [@Mooley:2018dlz]. Furthermore, from a purely theoretical standpoint, this value of the Lorentz factor is expected at $\sim100$ day timescales, considering the energy transfer from the shock into a surrounding medium of constant density [@Rees:1992ek]. The emitting region is assumed to be a spherical blob at the front edge of the jet, whose radius is therefore given by $R^\prime_\mathrm{blob} \approx \delta \beta c \Delta t \theta^\prime_\mathrm{jet}$.
The maximum energy of the emitted synchrotron radiation is fixed by X-ray observations, $E_{\mathrm X}\approx 10~\mathrm{keV}$ [@Nynka2018]. This is related to the magnetic field strength in the source, $B^\prime$, and the maximum electron energy of the emitting electrons, $E^\prime_e=\delta^{-1}E_e$, through $E^\prime_{\mathrm X} \propto E^{\prime2}_e B^\prime$. This means that in the relativistic scenario, for a given value of $B^\prime$, the maximum energy of the emitting electrons scales as $E^\prime_e\propto\delta^{-0.5}$. Furthermore, since the jet blob emits isotropically in its own rest frame, the X-ray luminosity, obtained assuming the observed emission is isotropic, $L^{\mathrm{iso}}_X$ deduced from flux measurements relates to the luminosity in the rest frame of the blob $L_X^\prime$ through $L^{\mathrm{iso}}_\mathrm{X}=\delta^{4}L^\prime_{X}$. This luminosity relates to the total number of X-ray emitting electrons, $N^\prime_{e}$, as well as $E_{e}^\prime$ and $B^\prime$, through $L^{\prime}_\mathrm{X}=N_{e}^\prime E_{e}^\prime/\tau^\prime_{\mathrm{syn}}(E^\prime_{e}) \propto N_{e}^\prime E_{e}^{\prime2}B^{\prime2}$, where $\tau^\prime_\mathrm{syn}(E_e^\prime)$ is the synchrotron cooling timescale. Therefore, in the relativistic jet scenario, for a given value of $B^\prime$ and measured X-ray flux, the number of electrons scales as $N^\prime_e\propto\delta^{-3}$. This will affect the results of the SSC model, since for higher values of $\delta$ there will be fewer emitting electrons, and a lower density of synchrotron photons, thus reducing the expected $\gamma$-ray flux.
The $\gamma$-ray luminosity expected from SSC is given in the shock rest frame by $L^\prime_{\mathrm{IC}}= N_{e}^\prime E^\prime_{e}/\tau^\prime_{\mathrm{IC}}(E^\prime_{e})\propto N_{e}^{\prime}(L^\prime_{\mathrm{X}}/(4\pi R^{\prime2} c))E_{e}^{\prime0.5}$, where $\tau^\prime_{\mathrm{IC}}(E^\prime_{e})$ is the IC cooling timescale, and $R^{\prime}$ is the size of the emitting region (either $R^\prime_\mathrm{blob}$ or $R^\prime_\mathrm{iso}$, defined above). The approximate scaling with $E_{e}^{\prime0.5}$ is due to the fact that the IC cooling occurs deep in the Klein-Nishina regime (since the observed synchrotron spectrum peaks at X-ray energies and the Lorentz factor of the ejecta is mild). Putting together the scalings given previously, we obtain that for a given value of $B^\prime$, the IC emission for the relativistic jet scenario scales as $L^\prime_{IC}\propto \theta_\mathrm{jet}^{\prime-2}(\beta\Delta t)^{-2}\delta^{-9.25}$ in the shock rest frame, which corresponds to a luminosity in the observer’s frame of
$$L_{IC}\approx10^{46}\left(\frac{\theta^\prime_\mathrm{jet}}{5^\circ}\right)^{-2}\left(\frac{\beta\Delta t}{0.94\times110~\mathrm{days}}\right)^{-2}\left(\frac{\delta}{3}\right)^{-5.25}~\mathrm{erg/s}.
\label{equ:scaling}$$
Thus, a slower expansion of the emission region would lead to a more compact source and therefore a stronger constraint on the magnetic field strength (both in the jet-like and isotropic scenarios).
In Fig. \[fig:bmodel\] we show the modeled synchrotron emission spectrum (black curves) and the respective SSC emission for both scenarios introduced above. These spectra were obtained with a numerical radiation model, introduced by @Rodrigues2019. In this model, a population of electrons is considered to fill homogeneously the emission region, and to be continuously accelerated to a power-law spectrum, with a possible cut-off at the highest energies. As can be seen in Fig. \[fig:bmodel\], these characteristics can explain radio (red points) and X-ray observations (blue points and blue shaded region). The parameters of the electron population have then been adjusted, and the magnetic field strength minimized, so that the predicted SSC emission does not exceed any of the 95% C.L. upper limits in the VHE $\gamma$-ray range derived from the H.E.S.S. observations (green). The solid vs. dashed curves represent the two extreme cases consistent with the observed X-ray and radio fluxes. In the non-relativistic scenario, the H.E.S.S. limits can constrain the minimum magnetic field strength to $\gtrsim210\,\mu\mathrm{G}$. In contrast, in more highly relativistic scenarios, the lower limit is weakened to the level of $\gtrsim24\,\mu\mathrm{G}$. The reason is that the strong scaling of the SSC flux with the Doppler factor (Eq. \[equ:scaling\]) implies that for a more relativistic outflow, a higher electron number is necessary to reach the flux limits, which implies a lower magnetic field in order to maintain the X-ray flux. As a point of comparison with the lower limit obtained through this analysis, the minimum magnetic field expected at late times downstream of the shock is of the order of $\sim100(\Gamma/3)(B_\mathrm{ISM}/10~\mu\mathrm{G})~\mu\mathrm{G}$ [@Kumar:2012xm], where $B_\mathrm{ISM}$ is the magnetic field strength in the interstellar medium (ISM). Furthermore, observations of the prompt emission of GRB 080916C by Fermi-LAT and 1-day afterglow emission in X rays and optical wavelengths have provided evidence of magnetic fields in the shock that are at the level of the compressed surrounding medium [@Kumar:2009ps], thus suggesting an ISM magnetic field of $\sim10~\mu\mathrm{G}$.
Very high-energy $\gamma$-ray observations can provide a direct probe of the magnetic field in BNS merger remnants. While the radio and X-ray data constrain the synchrotron part of the non-thermal spectrum, a measurement of the IC component in $\gamma$ rays is needed to break the ambiguity between energy in electrons and magnetic fields. Interferometric radio observations and long-term X-ray observations are crucial for inferring the jet properties such as the opening angle, viewing angle, or the Doppler factor. Based on the GW 170817 / GRB 170817A characteristics, one other short GRB may have been seen off-axis (GRB 150101B; [@Troja2018]). Long-term monitoring of BNS mergers in the radio, X-ray and $\gamma$-ray domain is necessary to further constrain the source properties and bridge the gap between the early-time kilonova and non-thermal GRB emission on the one hand, and the long-term behavior and the interaction between the jet and the ISM on the other hand. Current-generation Imaging Atmospheric Cherenkov Telescopes (IACTs) such as H.E.S.S., MAGIC or VERITAS can search for and study $\gamma$-ray counterparts on days-to-month timescales for BNS merger remnants seen under different viewing angles. Furthermore, the future Cherenkov Telescope Array (CTA) will be an order of magnitude more sensitive at around 1TeV, allowing us to constrain the minimum magnetic field in events like GW170817 / GRB 170817A to the mG regime. This will allow CTA to detect VHE $\gamma$-ray emission from events such as GW170817. As noted previously by [@Rodrigues2019], future observations may also be able to better constrain the nature of the non-thermal electrons. One possibility is that the observed radiation is dominated by “fresh” electrons picked up from the surroundings of the merger and accelerated at the shock front, allowing for hard photon spectra into the gamma-ray range. In a different scenario, the emission might originate in “old” electrons, continuously accelerated inside the volume of the ejecta, leading to cooling features at high energies.
Acknowledgments
===============
The support of the Namibian authorities and of the University of Namibia in facilitating the construction and operation of H.E.S.S. is gratefully acknowledged, as is the support by the German Ministry for Education and Research (BMBF), the Max Planck Society, the German Research Foundation (DFG), the Helmholtz Association, the Alexander von Humboldt Foundation, the French Ministry of Higher Education, Research and Innovation, the Centre National de la Recherche Scientifique (CNRS/IN2P3 and CNRS/INSU), the Commissariat à l’É́nergie atomique et aux É́nergies alternatives (CEA), the U.K. Science and Technology Facilities Council (STFC), the Knut and Alice Wallenberg Foundation, the National Science Centre, Poland grant no. 2016/22/M/ST9/00382, the South African Department of Science and Technology and National Research Foundation, the University of Namibia, the National Commission on Research, Science & Technology of Namibia (NCRST), the Austrian Federal Ministry of Education, Science and Research and the Austrian Science Fund (FWF), the Australian Research Council (ARC), the Japan Society for the Promotion of Science and by the University of Amsterdam. We appreciate the excellent work of the technical support staff in Berlin, Zeuthen, Heidelberg, Palaiseau, Paris, Saclay, Tübingen and in Namibia in the construction and operation of the equipment. This work benefitted from services provided by the H.E.S.S. Virtual Organisation, supported by the national resource providers of the EGI Federation. XR was supported by the Initiative and Networking Fund of the Helmholtz Association.
[^1]: At this distance, very-high-energy (VHE; 100GeV$<$E$<$100TeV) photons pair-produce $e^{\pm}$ in interactions with the Extragalactic Background Light (EBL) on their way from GW170817 to Earth. The VHE flux reduction due to the EBL is energy-dependent and varies between 10% and 30% between 1TeV and 10TeV, respectively, assuming the @Franceschini2008a EBL model. Note that the model curves have been derived ignoring the EBL correction
[^2]: We base this discussion on the 110 day timescale, for which there are quasi-simultaneous flux measurements in both radio and X ray bands. As shown in Fig. \[fig:obs\], the flux levels at this timescale are comparable throughout the H.E.S.S. observation window.
[^3]: Throughout this text, primed quantities denote parameter values in the shock rest frame and unprimed ones in the observer’s frame.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A long standing problem in weak lensing is about how to construct cosmic shear estimators from galaxy images. Conventional methods average over a single quantity per galaxy to estimate each shear component. We show that any such shear estimators must reduce to a highly nonlinear form when the galaxy image is described by three parameters (pure ellipse), even in the absence of the point spread function (PSF). In the presence of the PSF, we argue that this class of shear estimators do not likely exist. Alternatively, we propose a new way of measuring the cosmic shear: instead of averaging over a single value from each galaxy, we average over two numbers, and then take the ratio to estimate the shear component. In particular, the two numbers correspond to the numerator and denominators which generate the quadrupole moments of the galaxy image in Fourier space, as proposed in Zhang (2008). This yields a statistically unbiased estimate of the shear component. Consequently, measurements of the n-point spatial correlations of the shear fields should also be modified: one needs to take the ratio of two correlation functions to get the desired, unbiased shear correlation.'
author:
- |
Jun Zhang$^{1, 2}$[^1], Eiichiro Komatsu$^{1}$\
\
$^{1}$Texas Cosmology Center, University of Texas at Austin, Austin, TX 78712, USA\
$^{2}$Department of Astronomy, University of California, Berkeley, CA 94720, USA\
title: Cosmic Shears Should Not Be Measured In Conventional Ways
---
\[firstpage\]
cosmology: gravitational lensing - methods: data analysis - techniques: image processing: large scale structure
Introduction {#intro}
============
Weak gravitational lensing refers to the weak and systematic shape distortions of background source images (galaxies, CMB, etc.) by the foreground inhomogeneous density distributions on cosmological scales. Since this effect only involves gravity, it has been widely used as a direct probe of matter density fluctuations of our Universe (see, , @hj08 for a recent review).
The weak lensing effect can only be probed statistically due to the fact that the intrinsic projected shape of each galaxy is always somewhat random and anisotropic. A central theme in the study of weak lensing is to find unbiased cosmic shear estimators on galaxy images. This is indeed very challenging because the shape distortion due to weak lensing is generally much weaker than the intrinsic variations of the galaxy shapes.
In the early stage of this field, most of the work focused on issues regarding the use of the quadrupole moments of a galaxy image as a shear estimator (@tyson90 [@bonnet95; @kaiser95; @luppino97; @hoekstra98; @rhodes00; @kaiser00]). Ever since then, a number of other shear estimators have been considered in the literature, including moments defined by a certain set of orthogonal functions (@bridle01 [@bernstein02; @refregierbacon03; @massey05; @nakajima07]), the spatial derivatives of the galaxy surface brightness field (@zhang08 [@zhang10a]), etc..
Conventionally, for each shear component, the shear estimator is simply one number derived from a galaxy image, whose statistical mean is supposed to be equal to the true shear value, provided that the intrinsic galaxy image is statistically isotropic. Unfortunately, even in the absence of the PSF, we show that such shear estimators at least do not exist in simple forms, making them hard to use in practice for precise shear measurements (, in the presence of noise). In the presence of the PSF, we argue that such shear estimators do not likely exist. We give reasons for the above statements in §\[not\_exist\]. (The readers who are just interested in our new way of measuring shears may skip this section.)
In §\[alternatives\], we present a new form of shear estimators: instead of having only one number from each galaxy image for each shear component, one can keep [*two*]{} numbers, and use the ratio of their averages over many galaxies to accurately measure the cosmic shear. We find that this new way of measuring the shear can be easily implemented by using the method of Zhang (2008) (Z08 hereafter). The new type of shear estimator requires weak lensing statistics such as the n-point correlation functions of the shear field to be carried out in a slightly unusual way, but with little additional cost. This is discussed in §\[statistics\]. In §\[examples\], we give numerical examples. Finally, we summarize in §\[summary\].
Conventional Shear Estimators {#not_exist}
=============================
Conventional shear estimators are defined as a class of shear estimators, which average over a single quantity per galaxy to estimate each shear component. Most of the existing shear measurement methods belong to this class. For example, the method by Kaiser et al. (1995) and its extensions basically use the quadrupole moments of each Gaussian-Profile-Weighted galaxy image to measure the shear components; in the shapelets method (@refregier03), the value of each shear component is estimated from best-fitting a shapelets model to each observed galaxy image; the method of Bernstein and Jarvis (2002) evaluate the shear components by fitting each galaxy shape (also the PSF) with a series of orthogonal 2D Gaussian-based functions (see, , @massey07 for more examples). A common feature of these methods is that they all generate one quantity per galaxy for each shear component.
In §\[classic\], we start our discussion with the well-known examples of shear estimators consisting of quadrupole moments of galaxy images, and show what the issues are. In §\[no\_PSF\], we show that, even in the absence of the PSF, any conventional shear estimator has to reduce to a [*highly*]{} nonlinear form, making it hard to use in practice. In §\[prove\], we provide arguments as to why we think that conventional shear estimators do not likely exist when a PSF is present.
A Review of the Problem {#classic}
-----------------------
The use of galaxy quadrupole moments as shear estimators has been a central topic in weak lensing for many years. It is therefore easier to start our discussion with the quadrupole moments. To present the issues clearly, let us first consider the case without the PSF or any photon noise. For convenience, we use $(x_1, x_2)$ or $(x, y)$ instead of $(\theta_x, \theta_y)$ for coordinates in 2D in this paper.
Suppose that the surface brightness field of the lensed galaxy image is $f_L(\vtl)$ on the image plane, and that of the original (pre-lensing) galaxy image is $f_S(\vts)$ on the source plane, where $\vtl$ and $\vts$ are the position angles on the image and source planes, respectively. We have the following relations: $$\begin{aligned}
\label{fifstits}
&&f_L(\vtl)=f_S(\vts)\\ \nonumber
&&\vtl=\mathbf{A}\vts\end{aligned}$$ where $\mathbf{A}_{ij}=\delta_{ij}+\Phi_{ij}$, and $\Phi_{ij}=\partial x^L_i/\partial x^S_j-\delta_{ij}$, which are the spatial derivatives of the lensing deflection angle. $\Phi_{ij}$ can also be written as $\partial_{x_i}\partial_{x_j}\Phi$, where $\Phi$ is sometimes called the lensing potential. Matrix $\mathbf{A}$ can be alternatively written in terms of the convergence $\kappa=(\Phi_{11}+\Phi_{22})/2$ and the two shear components $\gamma_1=(\Phi_{11}-\Phi_{22})/2$ and $\gamma_2=\Phi_{12}$.
The quadrupole moments of the lensed galaxy image are defined as follows: $$\label{q_moments}
Q_{ij}=\int d^2\vec{x}x_ix_jf_L(\vec{x})$$ where the origin of the coordinates has been chosen to be the center of the light, , $$\label{origin}
\int d^2\vec{x}\vec{x}f_L(\vec{x})=0$$ Let us also define the ellipticities of the image as: $$\begin{aligned}
\label{epsilon}
\epsilon_1&=&\frac{Q_{11}-Q_{22}}{Q_{11}+Q_{22}}\\ \nonumber
\epsilon_2&=&\frac{2Q_{12}}{Q_{11}+Q_{22}}\\ \nonumber\end{aligned}$$
In the absence of the PSF, the quantities $\epsilon_1$ and $\epsilon_2$ are often thought to be good estimators for $\gamma_1$ and $\gamma_2$ up to the first order in the shear. Let us find out if they are indeed unbiased shear estimators. The observed quadrupole can be rewritten from using eq.(\[fifstits\]) in eq.(\[q\_moments\]): $$\begin{aligned}
\label{q2}
Q_{ij}&=&\int d^2\vec{x}x_ix_jf_S(\mathbf{A}^{-1}\vec{x})\\ \nonumber
&=&\vert \mathrm{det} (\mathbf{A})\vert\int d^2\vec{x}(\mathbf{A}\vec{x})_i(\mathbf{A}\vec{x})_jf_S(\vec{x})\end{aligned}$$ Note that the last step of the above equation is achieved by redefining $\mathbf{A}^{-1}\vec{x}$ as $\vec{x}$. Keeping up to first order in $\kappa$, $\gamma_1$, and $\gamma_2$ in eq.(\[q2\]), we get: $$\begin{aligned}
\label{q3}
Q_{11}-Q_{22}&=&(1+4\kappa)(Q_{11}^S-Q_{22}^S)+2\gamma_1(Q_{11}^S+Q_{22}^S)\\ \nonumber
Q_{12}&=&(1+4\kappa)Q_{12}^S+\gamma_2(Q_{11}^S+Q_{22}^S)\\ \nonumber
Q_{11}+Q_{22}&=&(1+4\kappa)(Q_{11}^S+Q_{22}^S)+2\gamma_1(Q_{11}^S-Q_{22}^S)\\ \nonumber
&+&4\gamma_2Q_{12}^S\end{aligned}$$ where $Q_{ij}^S$ are the quadrupole moments of the original galaxy image defined as: $$\label{q_I_moments}
Q_{ij}^S=\int d^2\vec{x}x_ix_jf_S(\vec{x})$$ Note that the two light centers defined in the image and source planes coincide. Based on eq.(\[q3\]), we find: $$\begin{aligned}
\label{q4}
\epsilon_1&=&\frac{\epsilon_1^S+2\gamma_1}{1+2\gamma_1\epsilon_1^S+2\gamma_2\epsilon_2^S}\\ \nonumber
&=&\epsilon_1^S+2\gamma_1\left[1-\left(\epsilon_1^S\right)^2\right]-2\gamma_2\epsilon_1^S\epsilon_2^S\\ \nonumber
\epsilon_2&=&\frac{\epsilon_2^S+2\gamma_2}{1+2\gamma_1\epsilon_1^S+2\gamma_2\epsilon_2^S}\\ \nonumber
&=&\epsilon_2^S+2\gamma_2\left[1-\left(\epsilon_2^S\right)^2\right]-2\gamma_1\epsilon_1^S\epsilon_2^S\end{aligned}$$ where $$\begin{aligned}
\label{epsilon_S}
\epsilon_1^S&=&\frac{Q_{11}^S-Q_{22}^S}{Q_{11}^S+Q_{22}^S}\\ \nonumber
\epsilon_2^S&=&\frac{2Q_{12}^S}{Q_{11}^S+Q_{22}^S}\end{aligned}$$ Given that the surface brightness distribution of the original galaxy image is statistically isotropic, we have $\langle\epsilon_{1,2}^S\rangle=0$ and $\langle\epsilon_1^S\epsilon_2^S\rangle=0$. Therefore, we find $$\begin{aligned}
\label{q5}
\langle\epsilon_1\rangle&=&2\gamma_1\left[1-\langle\left(\epsilon_1^S\right)^2\rangle\right]\\ \nonumber
\langle\epsilon_2\rangle&=&2\gamma_2\left[1-\langle\left(\epsilon_2^S\right)^2\rangle\right]\end{aligned}$$ This result, Eq.(\[q5\]), clearly shows that $\epsilon_1$ and $\epsilon_2$ are [*not*]{} unbiased shear estimators, as $\langle\left(\epsilon_1^S\right)^2\rangle$ and $\langle\left(\epsilon_2^S\right)^2\rangle$ in the multiplicative factors depend on the galaxy morphology distribution, and cannot be reduced to constant factors. (Also see Eq.(3.29) of Bernstein & Jarvis (2002), or Eq.(9.5.26) of Weinberg (2008).)
One can construct an unbiased estimator of the shear, if one keeps three quantities from each lensed galaxy image: $Q_{11}-Q_{22}$, $2Q_{12}$, and $Q_{11}+Q_{22}$, and use the ratios of their averages. Assuming statistical isotropy of intrinsic galaxy shapes in eq.(\[q3\]), and keeping up to first order in shear/convergence, we have (also see Eq.(9.5.30) of Weinberg (2008)): $$\begin{aligned}
\label{q6}
\frac{1}{2}\frac{\langle Q_{11}-Q_{22}\rangle}{\langle Q_{11}+Q_{22}\rangle}&=&\gamma_1\\ \nonumber
\frac{\langle Q_{12}\rangle}{\langle Q_{11}+Q_{22}\rangle}&=&\gamma_2\end{aligned}$$ This form of shear estimators is not conventional, as one has to keep more than one quantities from each galaxy image for each shear component. It is this class of estimators we shall discuss in this paper in detail.
One may wonder whether unbiased shear estimators in the conventional form ever exist. The answer is yes, at least when the PSF is absent. For example, we find the following unbiased shear estimators: $$\begin{aligned}
\label{q7}
\frac{1}{4}\left\langle \ln\left(\frac{1+\epsilon_1}{1-\epsilon_1}\right)\right\rangle&=&\gamma_1\\ \nonumber
\frac{1}{4}\left\langle \ln\left(\frac{1+\epsilon_2}{1-\epsilon_2}\right)\right\rangle&=&\gamma_2\\ \nonumber\end{aligned}$$ Eq.(\[q7\]) can be checked by applying Taylor expansion of $\ln[(1+\epsilon_i)/(1-\epsilon_i)]$ ($i=1, 2$) to the first order in shear/convergence using eq.(\[q4\]). Eq.(\[q7\]) defines a special type of conventional shear estimators that are accidentally found by us. It is now immediately interesting to ask if there exist other types of unbiased shear estimators in the conventional form. We study this issue specifically in the next two sections. If the readers wish to go directly to the relevant sections on the new estimator, read on from §\[alternatives\].
For notational convenience, we shall abbreviate “conventional shear estimator” as “CSE” in the rest of the paper. Once again, by CSE we mean the shear estimators that are made of just one number measured from a galaxy image for each shear component.
CSE in the Absence of the PSF {#no_PSF}
-----------------------------
In preparation for our main theme of this section, we discuss the spin properties of cosmic shears and their estimators in §\[coor\_spin\]. We then study the forms of CSEs in the absence of the PSF in §\[spin\_2\_esti\].
### The Spin of the CSE {#coor_spin}
To study the forms of the CSEs, it is useful to first consider their properties under coordinate rotations. Suppose we rotate the coordinates $(x, y)$ clockwise by an angle $\theta$. The new coordinates $(x^{\theta}, y^{\theta})$ are related to the old one via the following relation: $$\begin{aligned}
\label{coor_trans}
&&x^{\theta}=x\cos\theta-y\sin\theta \\ \nonumber
&&y^{\theta}=x\sin\theta+y\cos\theta\end{aligned}$$ If we write the position vector as a complex number of the form $x+\ima y$, where $\ima$ is the complex unit, the coordinate transformation under rotation can then be written as: $$\label{coor_trans_complex}
x^{\theta}+\ima y^{\theta} =(x+\ima y)\exp(\ima \theta)$$ For notational brevity, we shall generally use $X^{\theta}$ to denote the value of any quantity $X$ in the new coordinates that are rotated clockwise by an angle $\theta$ with respect to the original coordinates.
Let us now discuss how cosmic shears and their CSEs transform under coordinate rotation. The definitions of the shear components (shown in the beginning of §\[classic\]) involve spatial derivatives; thus, their transformation rules under coordinate rotation can be found from the chain rule: $$\begin{aligned}
\label{chain_rule}
&&\frac{\partial}{\partial x}=\frac{\partial x^{\theta}}{\partial x}\frac{\partial}{\partial x^{\theta}}+\frac{\partial y^{\theta}}{\partial x}\frac{\partial}{\partial y^{\theta}}\\ \nonumber
&&\frac{\partial}{\partial y}=\frac{\partial x^{\theta}}{\partial y}\frac{\partial}{\partial x^{\theta}}+\frac{\partial y^{\theta}}{\partial y}\frac{\partial}{\partial y^{\theta}}\end{aligned}$$ From eq.(\[coor\_trans\]), we get: $$\begin{aligned}
\label{cooefi_trans_derivs}
&&\frac{\partial x^{\theta}}{\partial x}=\frac{\partial y^{\theta}}{\partial y}=\cos\theta\\ \nonumber
&&\frac{\partial y^{\theta}}{\partial x}=-\frac{\partial x^{\theta}}{\partial y}=\sin\theta \end{aligned}$$ Therefore, we have: $$\label{coor_trans_complex_2}
\frac{\partial}{\partial x^{\theta}}+\ima\frac{\partial}{\partial y^{\theta}}=\left(\frac{\partial}{\partial x}+\ima\frac{\partial}{\partial y}\right)\exp(\ima\theta)$$ Taking the square of eq.(\[coor\_trans\_complex\_2\]), we find: $$\begin{aligned}
\label{coor_trans_complex_3}
&&\left(\frac{\partial^2}{\partial (x^{\theta})^2}-\frac{\partial^2}{\partial (y^{\theta})^2}\right)+\ima\left(2\frac{\partial^2}{\partial x^{\theta}\partial y^{\theta}}\right)\\ \nonumber
&=&\left[\left(\frac{\partial^2}{\partial x^2}-\frac{\partial^2}{\partial y^2}\right)+\ima\left(2\frac{\partial^2}{\partial x\partial y}\right)\right]\exp(\ima 2\theta)\end{aligned}$$ Therefore, the shear components, which are 2nd order derivatives of the lensing potential $\Phi$, also transform under coordinate rotation as: $$\label{gamma_rot}
\gamma_1^{\theta}+\ima\gamma_2^{\theta}=\left(\gamma_1+\ima\gamma_2\right)\exp(\ima 2\theta)$$ Because of this property, we usually say that cosmic shears form a spin-2 quantity. In general, a complex quantity, say $\Pi$, is called a spin-$n$ quantity if it transforms as $\Pi^{\theta}=\Pi\exp(\ima n\theta)$ under a clockwise coordinate rotation of angle $\theta$.
Now let us discuss shear estimators. It is straightforward to see that the CSEs defined in eq.(\[q7\]) do not form a spin-2 quantity. More generally, assuming that $\Gamma_1$ and $\Gamma_2$ are the CSEs for $\gamma_1$ and $\gamma_2$, respectively, then, unlike $\gamma_1+\ima\gamma_2$, $\Gamma_1+\ima\Gamma_2$ is not necessarily a spin-2 quantity. However, it turns out that we can regularize any CSEs by turning them into components of a spin-2 quantity. We have the following lemma:
\[turn\_spin\_2\] Based on any pair of CSEs $(\Gamma_1, \Gamma_2)$, one can build a new pair of CSEs $(\Gamma_1', \Gamma_2')$ to form a spin-2 quantity through the following procedure: $$\label{trans_spin_2}
\Gamma_1'+\ima\Gamma_2'=\frac{1}{2\pi}\int_0^{2\pi}d\theta\exp(-\ima 2\theta)\left(\Gamma_1^{\theta}+\ima\Gamma_2^{\theta}\right)$$
Firstly, under a clockwise coordinate rotation by angle $\theta_0$, we have: $$\begin{aligned}
\label{trans_spin_2_rot}
&&\Gamma_1'^{\theta_0}+\ima\Gamma_2'^{\theta_0}\\ \nonumber
&=&\frac{1}{2\pi}\int_0^{2\pi}d\theta\exp(-\ima 2\theta)\left(\Gamma_1^{\theta_0+\theta}+\ima\Gamma_2^{\theta_0+\theta}\right)\\ \nonumber
&=&\frac{1}{2\pi}\int_0^{2\pi}d\theta'\exp[-\ima 2(\theta'-\theta_0)]\left(\Gamma_1^{\theta'}+\ima\Gamma_2^{\theta'}\right)\\ \nonumber
&=&\left(\Gamma_1'+\ima\Gamma_2'\right)\exp(\ima 2\theta_0)\end{aligned}$$ Therefore, $\Gamma_1'+\ima\Gamma_2'$ form a spin-2 quantity. To show that $(\Gamma_1', \Gamma_2')$ are a pair of CSEs, let us take the ensemble average on both sides of eq. (\[trans\_spin\_2\_rot\]): $$\begin{aligned}
\label{trans_ideal_esti}
&&\left\langle\Gamma_1'+\ima\Gamma_2'\right\rangle\\ \nonumber
&=&\frac{1}{2\pi}\int_0^{2\pi}d\theta\exp(-\ima 2\theta)\left\langle\Gamma_1^{\theta}+\ima\Gamma_2^{\theta}\right\rangle\end{aligned}$$ Since by definition, $\Gamma_1^{\theta}$ and $\Gamma_2^{\theta}$ measure the shear values in the rotated coordinates, we have, $$\begin{aligned}
\label{Gamma_theta}
&&\left\langle\Gamma_1^{\theta}\right\rangle=\gamma_1^{\theta}=\gamma_1\cos 2\theta-\gamma_2\sin 2\theta\\ \nonumber
&&\left\langle\Gamma_2^{\theta}\right\rangle=\gamma_2^{\theta}=\gamma_1\sin 2\theta+\gamma_2\cos 2\theta\end{aligned}$$ In a more compact form, we can write eq.(\[Gamma\_theta\]) as: $$\label{Gamma_theta2}
\left\langle\Gamma_1^{\theta}+\ima\Gamma_2^{\theta}\right\rangle=(\gamma_1+\ima\gamma_2)\exp(\ima 2\theta)$$ Using eq. (\[Gamma\_theta2\]) in eq. (\[trans\_ideal\_esti\]), we get: $$\label{trans_ideal_esti2}
\left\langle\Gamma_1'+\ima\Gamma_2'\right\rangle=\gamma_1+\ima\gamma_2$$ which proves that $(\Gamma_1', \Gamma_2')$ are indeed also a pair of CSEs.
Due to the invariance under a coordinate rotation of angle $2\pi$, one can always decompose any shear estimator into components of integer spins. Eq.(\[trans\_spin\_2\]) essentially defines a procedure of isolating the spin-2 components of any CSEs using Fourier transformation. Since the cosmic shears form a spin-2 quantity, only the spin-2 components of any CSEs are the relevant/principle components of the estimators. This point is further supported by the fact that the ensemble averages of spin-$n$ ($n\ne 2$) components of any CSEs are zero, , $$\begin{aligned}
\label{trans_ideal_est3i}
&&\frac{1}{2\pi}\int_0^{2\pi}d\theta\exp(-\ima n\theta)\left\langle\Gamma_1^{\theta}+\ima\Gamma_2^{\theta}\right\rangle\\ \nonumber
&=&\frac{1}{2\pi}\int_0^{2\pi}d\theta\exp(-\ima n\theta)\left(\gamma_1+\ima\gamma_2\right)\exp(\ima 2\theta)\\ \nonumber
&=&0 \mbox{ ( if } n\ne 2\mbox{ )}\end{aligned}$$ Therefore, we only need to focus on spin-2 CSEs from here on.
### Spin-2 CSEs {#spin_2_esti}
In the weak lensing limit, , when the cosmic shear parameters $(\gamma_1, \gamma_2, \kappa)$ are small, any spin-2 CSEs $(\Gamma_1, \Gamma_2)$ can be Taylor expanded to the first order in shear as follows: $$\begin{aligned}
\label{Taylor_weak_lensing}
&&\Gamma_1(\gamma_1,\gamma_2,\kappa)\\ \nonumber
&=&\left(\Gamma_1\right)_0+\gamma_1\left(\frac{\partial\Gamma_1}{\partial\gamma_1}\right)_0+\gamma_2\left(\frac{\partial\Gamma_1}{\partial\gamma_2}\right)_0+\kappa\left(\frac{\partial\Gamma_1}{\partial\kappa}\right)_0\\ \nonumber
&&\Gamma_2(\gamma_1,\gamma_2,\kappa)\\ \nonumber
&=&\left(\Gamma_2\right)_0+\gamma_1\left(\frac{\partial\Gamma_2}{\partial\gamma_1}\right)_0+\gamma_2\left(\frac{\partial\Gamma_2}{\partial\gamma_2}\right)_0+\kappa\left(\frac{\partial\Gamma_2}{\partial\kappa}\right)_0\end{aligned}$$ where $(X)_0$ ($X$ is any quantity) denotes the value of $X$ at $\gamma_1=\gamma_2=\kappa=0$. Since $(\Gamma_1)_0$, $(\Gamma_2)_0$, $(\partial_{\kappa}\Gamma_1)_0$, and $(\partial_{\kappa}\Gamma_2)_0$ are all spin-2 quantities, their ensemble average must vanish. On the other hand, the coefficients associated with $\gamma_1$ and $\gamma_2$ in eq.(\[Taylor\_weak\_lensing\]) can be decomposed into spin-0 and spin-4 components as follows: $$\begin{aligned}
\label{Taylor_weak_lensing2}
&&\left\langle\Gamma_1(\gamma_1,\gamma_2,\kappa)\right\rangle=\gamma_1\left\langle A+B_1\right\rangle+\gamma_2\left\langle C+B_2\right\rangle\\ \nonumber
&&\left\langle\Gamma_2(\gamma_1,\gamma_2,\kappa)\right\rangle=\gamma_1\left\langle B_2-C\right\rangle+\gamma_2\left\langle A-B_1\right\rangle\end{aligned}$$ where $$\begin{aligned}
\label{ABC}
&& A=\frac{1}{2}\left(\partial_{\gamma_1}\Gamma_1+\partial_{\gamma_2}\Gamma_2\right)\\ \nonumber
&& C=\frac{1}{2}\left(\partial_{\gamma_2}\Gamma_1-\partial_{\gamma_1}\Gamma_2\right)\\ \nonumber
&& B_1=\frac{1}{2}\left(\partial_{\gamma_1}\Gamma_1-\partial_{\gamma_2}\Gamma_2\right)\\ \nonumber
&& B_2=\frac{1}{2}\left(\partial_{\gamma_2}\Gamma_1+\partial_{\gamma_1}\Gamma_2\right)\end{aligned}$$ As shown in Appendix A, $A$ is a scalar, $C$ is a pseudo scalar, $B_1+\ima B_2$ is a spin-4 quantity. The ensemble averages of $B_1$ and $B_2$ must vanish. The ensemble average of $C$ vanishes if galaxy images have parity symmetry along any direction in the plane of the sky statistically, which is assumed to be true in this paper. Consequently, for $(\Gamma_1, \Gamma_2)$ to be spin-2 CSEs, we only require $\langle A\rangle=1$. This actually implies that for any individual galaxy, $A=1$. The reason is that $A$ of any single galaxy does not change under coordinate rotation of random angles, and is equal to $\langle A\rangle$ because the galaxies generated by rotations of a single galaxy form a complete set of statistically isotropic samples (, there are no special directions). As a result, any spin-2 CSEs $(\Gamma_1, \Gamma_2)$ must satisfy the following necessary condition: $$\label{requirement}
\frac{\partial\Gamma_1}{\partial\gamma_1}+\frac{\partial\Gamma_2}{\partial\gamma_2}=2$$ This is also a sufficient condition, because $A=1$ directly implies that $\langle A\rangle=1$.
In general, the CSEs are functions of a certain number of shape parameters (, the multipole moments of an image). The functions can be very complicated, and are certainly not fixed by the requirement given by eq.(\[requirement\]). However, for galaxies whose shapes are described by only three parameters (perfect ellipses), any CSE should reduce to a function of just three variables. In this case, we find that any pair of spin-2 CSEs must reduce to a unique form, which is sufficient for us to judge whether CSEs are convenient in practice: namely, if we find that the resulting form is highly non-linear even for such a simple case, then it is reasonable to conclude that CSEs are not so useful for accurate shear measurements from more realistic galaxy shapes as well. This is shown in the rest of this section. For clarity, we refer the readers to Appendix B for the mathematical details/proofs for some of the statements made hereafter in this section.
Let us consider a set of galaxies whose surface brightness profiles can be parametrized as $f_S(R)$ with $R=a(x^2+y^2)+b(x^2-y^2)+2cxy$, where $x$ and $y$ are the coordinates, $f_S(R)$ is a function of a fixed form, and $(a, b, c)$ are the three parameters determining galaxy shapes. For the images to be ellipses, we require the following three things: 1. $f_S(R)$ decays sufficiently fast when $R$ becomes large; 2. $a+b>0$; 3. $a^2-b^2>c^2$. For example, if $f_S(R)=H(R_c-R)$ ($H$ is the step function) and $(a, b, c)$ satisfy the above conditions, the galaxy surface brightness is then distributed evenly inside the ellipse defined by $a(x^2+y^2)+b(x^2-y^2)+2cxy\le R_c$. When such images are weakly lensed, the three conditions are not violated, and $(a, b, c)$ becomes $(a', b', c')$ without changing the form of $f_S$. In other words, weak lensing does not introduce additional degrees of freedom to the galaxy shapes. Note that otherwise, one has to consider using more than 3 parameters to construct shear estimators. Among the three parameters, there are indeed only two degrees of freedom useful for shear measurement: the ratios of the parameters. This is because the overall amplitudes of $(a, b, c)$ only change the galaxy size, not its shape. As shown in Appendix B, the ellipticities $(\epsilon_1, \epsilon_2)$ defined in §\[classic\] are directly equal to $(-b/a, -c/a)$, therefore, we can write the shear estimators as functions of only $\epsilon_1$ and $\epsilon_2$.
We can further show that $(\Gamma_1, \Gamma_2)$ must take the following form: $$\label{G_1_G_2_form}
\Gamma_1+\ima\Gamma_2=(\epsilon_1+\ima\epsilon_2)g(u)$$ where $u=\epsilon_1^2+\epsilon_2^2$, and $g$ is a one-variable complex function, whose form is to be determined later in this section. To see why eq.(\[G\_1\_G\_2\_form\]) is true, one can use the Taylor expansion to write $\Gamma_1$ and $\Gamma_2$ as power series of $\epsilon_1$ and $\epsilon_2$ [^2]. Consequently, one can write $\Gamma_1+\ima\Gamma_2$ as power series of $\epsilon_1+\ima\epsilon_2$ and $\epsilon_1-\ima\epsilon_2$, whose spins are $2$ and $-2$ respectively. Since $\Gamma_1+\ima\Gamma_2$ is a spin-2 quantity, each term in the power series must also be a spin-2 quantity. Therefore, in each term of the power series, the power on $\epsilon_1+\ima\epsilon_2$ must be larger than that on $\epsilon_1-\ima\epsilon_2$ by exactly one, , each term must take the form of $\alpha(\epsilon_1+\ima\epsilon_2)(\epsilon_1^2+\epsilon_2^2)^n$, where $n$ is any non-negative integer, and $\alpha$ is a coefficient which can be any complex number at this point. As a result, the shear estimators must have the form defined in eq.(\[G\_1\_G\_2\_form\]). To find out what $g(u)$ is, let us write it as $g_1(u)+\ima g_2(u)$. Eq. (\[G\_1\_G\_2\_form\]) then becomes: $$\begin{aligned}
\label{G_1_G_2_real}
&&\Gamma_1=\epsilon_1g_1(u)-\epsilon_2g_2(u)\\ \nonumber
&&\Gamma_2=\epsilon_1g_2(u)+\epsilon_2g_1(u)\end{aligned}$$ Using the constraint in eq.(\[requirement\]), we find: $$\begin{aligned}
\label{constraint}
2&=&\frac{\partial\Gamma_1}{\partial\gamma_1}+\frac{\partial\Gamma_2}{\partial\gamma_2}\\ \nonumber
&=&\frac{\partial\Gamma_1}{\partial\epsilon_1}\frac{\partial\epsilon_1}{\partial\gamma_1}+\frac{\partial\Gamma_1}{\partial\epsilon_2}\frac{\partial\epsilon_2}{\partial\gamma_1}+\frac{\partial\Gamma_2}{\partial\epsilon_1}\frac{\partial\epsilon_1}{\partial\gamma_2}+\frac{\partial\Gamma_2}{\partial\epsilon_2}\frac{\partial\epsilon_2}{\partial\gamma_2}\\ \nonumber
&=&2(2-u)g_1(u)+4u(1-u)\frac{dg_1}{du}\end{aligned}$$ It is interesting to note that eq.(\[constraint\]) does not place any constraints on $g_2(u)$, , it can be any real function. This is because $g_2(u)$ simply adds unnecessary even-parity terms into odd parity ones, and vice versa, without affecting the ensemble averages and the spin of the shear estimators. For convenience, we set $g_2(u)=0$ hereafter.
Eq.(\[constraint\]) is a typical first-order ordinary differential equation. It can be solved by introducing an integrating factor $k(u)$ which satisfies: $$\label{ku}
k(u)(2-u)=\frac{d}{du}\left[2k(u)u(1-u)\right]$$ Multiplying both sides of eq.(\[constraint\]) with $k(u)$, we get: $$\label{g_1_u}
\frac{d}{du}\left[2k(u)u(1-u)g_1(u)\right]=k(u)$$ It is now straightforward to solve both eq.(\[ku\]) and eq.(\[g\_1\_u\]). The results are: $$\begin{aligned}
\label{results_g1}
&&k(u)\propto(1-u)^{-3/2}\\ \nonumber
&&g_1(u)=\frac{1}{u}\left(1+C\sqrt{1-u}\right) \end{aligned}$$ where $C$ is a real number constant. To guarantee that $\Gamma_1$ and $\Gamma_2$ do not diverge when $\epsilon_1$ and $\epsilon_2$ approach zero, we need $C=-1$. Finally, we find the unique form for the spin-2 CSEs: $$\label{G_1_G_2_form_final}
\Gamma_1+\ima\Gamma_2=(\epsilon_1+\ima\epsilon_2)\frac{1-\sqrt{1-\epsilon_1^2-\epsilon_2^2}}{\epsilon_1^2+\epsilon_2^2}$$ Regarding the uniqueness, it is useful to note that if we transform the CSEs defined in eq.(\[q7\]) into spin-2 shear estimators using the procedure given in eq.(\[trans\_spin\_2\]), we achieve the same shear estimators as those shown in eq.(\[G\_1\_G\_2\_form\_final\]).
We have shown that the principle components (spin-2) of any pair of CSEs have to take specific and [*highly*]{} nonlinear forms for galaxies of elliptical shapes. This feature makes CSEs not convenient in practice (, in the presence of noise).
CSEs in the Presence of the PSF {#prove}
-------------------------------
Any CSEs which correct for the PSF effect also have to reduce to the forms given in eq.(\[G\_1\_G\_2\_form\_final\]) in the limit of zero PSF size when the galaxy images have pure elliptical shapes. For this reason, the conclusion in the previous section is already sufficient to argue against the usefulness of CSEs in practice. For academic interests, we provide the following arguments for why CSEs may not even exist in the presence of the PSF:
In the presence of the point spread function, structural details of galaxy images on scales smaller than the size of the PSF are smeared out. This implies that there are only a finite number of shape parameters (, multipole moments up to some order) available for constructing shear estimators. On the other hand, the derivatives of the lower order shape parameters (, lower order multipole moments) with respect to the cosmic shears depend on the higher order shape parameters because of the PSF, suggesting the requirement for an infinite number of shape parameters to form the shear estimators. Combining the above two reasons, we find it unlikely to form CSEs when a PSF is present. The mathematical details of the above statements are given in Appendix C.
A New Way of Estimating Shears {#alternatives}
==============================
As searching for optimal shear estimators is actively ongoing nowadays (@heymans06 [@massey07; @bridle09; @bridle10]), it is important to realize that CSEs (“conventional shear estimators,” by which we mean the shear estimators that are made of just one number from a galaxy image for each shear component) are hard to use in practice due to their unavoidable complex forms even in the absence of the PSF (simpler forms, such as the quadrupole moments, are biased estimators, as shown in § 2.1).
Therefore, existing shear estimators of the conventional type must quantify the bias factor when estimating the shear, which can be achieved numerically (see, , Erben et al. 2001, Bacon et al. 2001, or most recently, @heymans06 [@massey07; @bridle10]) or estimated analytically (, shear susceptibility in KSB \[Kaiser et al. 1995\] and derived methods, or responsivity factor in Bernstein & Jarvis 2002 and similar methods), although most people have been mainly focusing on the systematic errors caused by the photon noise and the PSF. However, to achieve percent or even sub-percent level accuracy in cosmic shear measurements, it does not seem enough to completely rely on numerical tests using computer-generated galaxies of limited morphology richness, or approximate analytical methods. Unfortunately, in the presence of PSF, most of the existing shear measurement methods are too complicated or too model-dependent (@vb10 [@bernstein10]) to allow for an accurate analytic analysis of the systematic errors in their shear estimators.
The method of Z08 (see also @zhang10a for the treatment of photon noise and the pixelation effect) is easily amenable to the corrections described in eq.(\[q5\]), and can also account for the PSF correction. Not only is it simple, but also well defined regardless of the morphologies of galaxies and the PSF. We show here how to properly use this method (instead of using it as CSEs) to recover the cosmic shear in an unbiased way.
The Idea
--------
The basic idea of Z08 is to use the spatial derivatives of the galaxy surface brightness field to measure the cosmic shears. It relies on the fact that gravitational lensing does not only distort the overall shape of the object, but also locally modifies the anisotropy of the gradient field of the surface brightness. As it allows for using the shape information from galaxy substructures, the method of Z08 can potentially improve on the signal-to-noise ratio of the shear measurements.
It is shown in Z08 that the shear measurement should be carried out in the Fourier space, in which any PSF can be transformed into the desired isotropic Gaussian form through multiplications, and the spatial derivatives of the surface brightness field can be easily measured. The cosmic shear can be estimated using the following relations: $$\begin{aligned}
\label{shear12PSF}
&&\frac{1}{2}\frac{\langle (\partial_1f_O)^2-(\partial_2f_O)^2\rangle}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle}=-\gamma_1 \\ \nonumber
&&\frac{\langle\partial_1f_O\partial_2f_O\rangle}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle}=-\gamma_2\end{aligned}$$ where $$\label{Delta}
\Delta=\frac{\beta^2}{2}\vec{\nabla}f_O\cdot\vec{\nabla}(\nabla^2f_O)$$ $\beta$ is the scale radius of the isotropic Gaussian PSF $W_{\beta}$, which is defined as: $$W_{\beta}(\vec{\theta})=\frac{1}{2\pi\beta^2}\exp\left(-\frac{\vert\vec{\theta}\vert^2}{2\beta^2}\right)$$ $f_O$ is the surface brightness field. $\partial_i$ denotes $\partial /\partial x_i$. As shown in Appendix D, the method of Z08 effectively utilizes the quadrupole moments in the Fourier space to measure the cosmic shears.
A New Unbiased Estimator {#m_zhang08}
------------------------
Now, here is an important point: in order to implement this method, we must make it clear what we mean by the angular brackets in eq.(\[shear12PSF\]). First, we need to measure the derivatives of the surface brightness and average them within a single galaxy. Let us denote this averaging by $\langle\rangle_g$, and write: $$\begin{aligned}
&&\frac{1}{2}\frac{\langle (\partial_1f_O)^2-(\partial_2f_O)^2\rangle_g}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g} \\
&&\frac{\langle\partial_1f_O\partial_2f_O\rangle_g}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g}\end{aligned}$$ Of course, these are still extremely noisy as they use only one galaxy. The question is then, “how do we average these quantities over many galaxies to obtain an unbiased estimator of the shears?”
If one uses these quantities as if they were the CSEs, then one would simply average them over many galaxies. However, this will produce a biased estimator: [^3] $$\begin{aligned}
\label{shear12PSF_wrong}
&&\left\langle\frac{1}{2}\frac{\langle (\partial_1f_O)^2-(\partial_2f_O)^2\rangle_g}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g}\right\rangle_{en}=-\gamma_1(1-\delta_1) \\ \nonumber
&&\left\langle\frac{\langle\partial_1f_O\partial_2f_O\rangle_g}{\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g}\right\rangle_{en}=-\gamma_2(1-\delta_2)\end{aligned}$$ where $\delta_1$ and $\delta_2$ are the ensemble averages of functions of multipole moments of the galaxy images in Fourier space, and $\langle\rangle_{en}$ denotes the ensemble average over many galaxies. The derivation of the forms of $\delta_1$ and $\delta_2$ is given in Appendix D. $\delta_1$ and $\delta_2$ are generally nonzero and dependent on the galaxy morphology.
Instead, we need to take the ensemble averages of the numerator and the denominator separately first, and then divide them to obtain an unbiased estimator: $$\begin{aligned}
\label{shear12PSF2}
&&\frac{1}{2}\frac{\langle\langle (\partial_1f_O)^2-(\partial_2f_O)^2\rangle_g\rangle_{en}}{\langle\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g\rangle_{en}}=-\gamma_1 \\ \nonumber
&&\frac{\langle\langle\partial_1f_O\partial_2f_O\rangle_g\rangle_{en}}{\langle\langle (\partial_1f_O)^2+(\partial_2f_O)^2+\Delta\rangle_g\rangle_{en}}=-\gamma_2\end{aligned}$$ This is the main result of this paper, and the form of the unbiased estimator that we propose to use for the actual analysis of the weak lensing data.
Of course, one could divide the left hand sides of eq.(\[shear12PSF\_wrong\]) by $1-\delta_1$ and $1-\delta_2$ to obtain an unbiased estimator. This is similar to correcting the measured shear for a multiplicative bias that is evaluated from the same ensemble of galaxies. In this sense, eq.(\[shear12PSF\_wrong\]) provides the exact definitions for the multiplicative biases for $\gamma_1$ and $\gamma_2$. However, since $\delta_1$ and $\delta_2$ in eq.(\[shear12PSF\_wrong\]) involve many high order Fourier-space multipole moments of the surface brightness field, evaluation of these terms from simulations (which are incomplete anyway) can be highly uncertain. Even worse, the multiplicative bias mentioned here is not even a constant, but depends on the morphological distribution of the galaxies. This makes the conventional way of measuring shear correlation functions even more challenging, as one must take into account the [*correlations*]{} of the multiplicative biases, as will be shown in §\[statistics\].
In summary, according to eq.(\[shear12PSF2\]), for each shear component, two quantities from each galaxy should be kept, and the ratios of their ensemble averages yield unbiased estimates for the corresponding shear components. Finally, it is important to note that, to efficiently use eq.(\[shear12PSF2\]), the surface brightness of each participating galaxy should be normalized to have roughly the same maximum value, so that faint galaxies are not much less weighted than their brighter counterparts. The details regarding the optimal weighting scheme as a function of the galaxy luminosity should also take into account the photon noise. This is a separate topic, and will be studied in a future work.
Comments on Errors due to Finite Number of Galaxies {#error_budget}
---------------------------------------------------
Strictly speaking, Eq.(\[shear12PSF2\]) holds when we average over an infinite number of galaxies; however, as we shall show in this section, the error that we make by having a finite number of galaxies for averaging is much smaller than the statistical errors, and thus the estimator remains unbiased for practical applications.
For simplicity, we use eq.(\[q6\]) rather than our main equation \[eq.(\[shear12PSF2\])\] in the following discussion, but the conclusion will be the same for eq.(\[shear12PSF2\]).
Let us use $\langle X\rangle_N$ to denote the average of the quantity $X$ over $N$ galaxies. From eq.(\[q3\]), we get: $$\begin{aligned}
\label{q3_error}
\left\langle Q_{11}-Q_{22}\right\rangle_N&=&(1+4\kappa)\left\langle Q_{11}^S-Q_{22}^S\right\rangle_N\\ \nonumber
&+&2\gamma_1\left\langle Q_{11}^S+Q_{22}^S\right\rangle_N\\ \nonumber
\left\langle Q_{12}\right\rangle_N&=&(1+4\kappa)\left\langle Q_{12}^S\right\rangle_N+\gamma_2\left\langle Q_{11}^S+Q_{22}^S\right\rangle_N\\ \nonumber
\left\langle Q_{11}+Q_{22}\right\rangle_N&=&(1+4\kappa)\left\langle Q_{11}^S+Q_{22}^S\right\rangle_N\\ \nonumber
&+&2\gamma_1\left\langle Q_{11}^S-Q_{22}^S\right\rangle_N+4\gamma_2\left\langle Q_{12}^S\right\rangle_N\end{aligned}$$ Consequently, we have: $$\begin{aligned}
\label{q4_error}
\frac{1}{2}\frac{\left\langle Q_{11}-Q_{22}\right\rangle_N}{\left\langle Q_{11}+Q_{22}\right\rangle_N}&=&\frac{1}{2}\Delta_1+\gamma_1(1-\Delta_1^2)-\gamma_2\Delta_1\Delta_2\\ \nonumber
\frac{\left\langle Q_{12}\right\rangle_N}{\left\langle Q_{11}+Q_{22}\right\rangle_N}&=&\frac{1}{2}\Delta_2+\gamma_2(1-\Delta_2^2)-\gamma_1\Delta_1\Delta_2\end{aligned}$$ where $$\begin{aligned}
\label{def_error}
\Delta_1&=&\frac{\left\langle Q_{11}^S-Q_{22}^S\right\rangle_N}{\left\langle Q_{11}^S+Q_{22}^S\right\rangle_N}\\ \nonumber
\Delta_2&=&\frac{2\left\langle Q_{12}^S\right\rangle_N}{\left\langle Q_{11}^S+Q_{22}^S\right\rangle_N}\end{aligned}$$ Here, the terms $\Delta_1$ and $\Delta_1\Delta_2$ contribute to random errors because their ensemble averages vanish, whereas the terms $\Delta_1^2$ and $\Delta_2^2$ lead to systematic biases because their ensemble averages do not vanish. Fortunately, as $\Delta_{1, 2}$ scales as $1/\sqrt{N}$ and $\gamma_{1,
2}\ll 1$, the amplitudes of such systematic biases are always much smaller than the sizes of the statistical errors. Therefore, the results from this new type of shear estimators may be regarded as unbiased for practical applications. Numerical verifications will be given in §\[examples\].
Shear Statistics - n-Point Correlations {#statistics}
=======================================
The cosmic shear field can only be probed statistically. This is mainly due to the intrinsic variations of the galaxy shapes and the spatial fluctuations of the shear components. As a result, the shear statistics is usually studied in the form of n-point spatial correlation functions of the shear field. The previous discussions and measurements in the literature are based on “conventional” shear estimators (CSEs), , one often assumes that the following is true: $$\label{assume_wr}
\langle\Gamma\rangle_{en}=\gamma$$ where $\gamma$ can be either $\gamma_1$ or $\gamma_2$, and $\Gamma$ is a CSE for $\gamma$. For individual galaxies, eq.(\[assume\_wr\]) implies: $$\label{assume_wr2}
\Gamma=\gamma+\Psi$$ where $\Psi$ satisfies $\langle\Psi\rangle_{en}=0$[^4]. It is usually assumed that $\Psi$’s of different galaxies do not correlate with each other[^5]. Therefore, the n-point correlation functions of the shear field can be directly measured by the correlations of $\Gamma$’s.
However, in §\[no\_PSF\], we have shown that such a $\Gamma$ at least does not exist in a convenient form. Instead, as proposed in §\[m\_zhang08\], we can use the new form of shear estimators defined in eq.(\[shear12PSF2\]) to probe the cosmic shear in an unbiased way. Let us now find out how to measure the n-point shear correlation functions with the new form of shear estimators. Numerical examples are given in §\[examples\].
For notational convenience, the type of shear measurement in eq.(\[shear12PSF2\]) can be symbolized as follows: $$\label{symbol}
\frac{\langle A\rangle_{en}}{\langle B\rangle_{en}}=\gamma$$ where $\gamma$ can be either $\gamma_1$ or $\gamma_2$, and $A$ and $B$ are properties of a galaxy, such as those defined in eq.(\[shear12PSF2\]). Similar to eq.(\[assume\_wr2\]), eq.(\[symbol\]) implies the following: $$\label{symbol2}
A=\gamma B+C$$ where $C$ satisfies $\langle C\rangle_{en}=0$. If we assume that the $C$ of any galaxy does not correlate with the $B$’s and $C$’s of other galaxies[^6], the n-point correlation functions of the shear field can be probed using the following relation: $$\label{symbol3}
\left\langle\gamma(\vec{x}_1)\gamma(\vec{x}_2)\ldots\gamma(\vec{x}_n)\right\rangle_{en}=\frac{\left\langle A(\vec{x}_1)A(\vec{x}_2)\ldots A(\vec{x}_n)\right\rangle_{en}}{\left\langle B(\vec{x}_1)B(\vec{x}_2)\ldots B(\vec{x}_n)\right\rangle_{en}}$$ The ensemble averages are taken over a large number of galaxies whose relative positions $\vec{x}_i-\vec{x}_j$ ($i,j=1,2,\ldots,n$) are fixed. In practice, the n-point shear correlation functions can be measured using: $$\label{symbol33}
\frac{\sum A(\vec{x}_{i_1})A(\vec{x}_{i_2})\ldots A(\vec{x}_{i_n})}{\sum B(\vec{x}_{i_1})B(\vec{x}_{i_2})\ldots B(\vec{x}_{i_n})}$$ where the sum is taken over all the galaxy groups that satisfy the positional constraints. Note that [*the ratio is taken after the summations*]{}. The standard deviation ($\sigma$) of the correlation function in such a measurement can be calculated as follows: $$\begin{aligned}
\label{symbol34}
\sigma^2&\doteq& \left\langle\left[\frac{\sum A(\vec{x}_{i_1})A(\vec{x}_{i_2})\ldots A(\vec{x}_{i_n})}{\sum B(\vec{x}_{i_1})B(\vec{x}_{i_2})\ldots B(\vec{x}_{i_n})}\right]^2\right\rangle_{en}\\ \nonumber
&\doteq&\left\langle\frac{\sum A^2(\vec{x}_{i_1})A^2(\vec{x}_{i_2})\ldots A^2(\vec{x}_{i_n})}{\left[\sum B(\vec{x}_{i_1})B(\vec{x}_{i_2})\ldots B(\vec{x}_{i_n})\right]^2}\right\rangle_{en}\\ \nonumber
&=&\frac{N\left\langle A^2(\vec{x}_1)A^2(\vec{x}_2)\ldots A^2(\vec{x}_n)\right\rangle_{en}}{N^2\left\langle B(\vec{x}_1)B(\vec{x}_2)\ldots B(\vec{x}_n)\right\rangle_{en}^2}\\ \nonumber
&=&\frac{1}{N}\left(\frac{\langle A^2\rangle_{en}}{\langle B\rangle_{en}^2}\right)^n\end{aligned}$$ where $N$ is the total number of galaxy groups (, the number of galaxy pairs for 2-point correlations) used.
To summarize, in the new type of shear measurement, the shear correlation function should be measured using the ratio of two ensemble averages, as shown in eq.(\[symbol3\]). If $B$ in eq.(\[symbol\]) is viewed as a multiplicative bias, we need to measure the [*correlations*]{} of these multiplicative biases as well in order to get the correct shear correlation functions.
Numerical Tests {#examples}
===============
In this section, we show how accurately one can recover the cosmic shears and their 2-point correlation functions with the method proposed in §\[alternatives\] and §\[statistics\]. For a comparison, we also show the results using the method of Z08 in the “conventional” (but wrong) way, , eq.(\[shear12PSF\_wrong\]), but without taking into account the biases, $\delta_1$ and $\delta_2$.[^7] Since our focus is to demonstrate principles, we do not include photon noise or the pixelation effect in this paper, and we approximate the PSF as an isotropic Gaussian. Including these effects is straightforward. (Note also that the conventional estimators yield biased results even in this idealized case.)
More comprehensive tests of the method of Z08 have been presented in [@zhang10b], which further improves the accuracy of shear measurement by including the second order shear/convergence terms. As shown in that paper, the second order correction is proportional to the convergence $\kappa$. We simply set $\kappa=0$ in the numerical simulations here. The residual systematic error on the recovered shear ($\gamma_1$ or $\gamma_2$) therefore should be expected to have a magnitude comparable to the third order terms in shear (, $\gamma_1^3$, $\gamma_1^2\gamma_2$, $\gamma_1\gamma_2^2$, $\gamma_2^3$). The errors shown in the rest of this section are all at the $1\sigma$ confidence level.
Image Generation {#image}
----------------
The mock galaxy images we use in our numerical tests are generated by the algorithms introduced in Z08 and Zhang (2010a), , each galaxy is generated as a collection of point sources. The reason is simple: one can accurately and easily mimic the lensing effect by displacing the points. It also allows us to generate galaxies of complex morphologies. There are two types of galaxies we use in this paper: 1. randomly oriented regular galaxies, each of which contains an exponential disk in the galactic plane (no bulge); 2. irregular galaxies being made of points generated by the trajectories of 2D random walks. For simplicity, the PSF is always an isotropic Gaussian function, whose scale radius is four times the grid size to avoid the pixelation problem. All the lengths in our simulations are in units of the grid size in the rest of this section. The dimension of the grid is $64\times 64$.
1-Point Statistics {#one_point}
------------------
As our first example, we study how accurately a single input cosmic shear can be recovered by a large number of mock galaxies, , the 1-point statistics. We use the regular type mock galaxies as introduced in §\[image\]. Each disk galaxy is composed of ten point sources which are randomly distributed within a radius of 7. The intensity of a point is an exponentially decaying function of its distance to the center of the disk with a decay length equal to 7. The galactic disk is then projected onto the source plane in a random direction. For each input shear value, we use $10^7$ mock galaxies to recover the shear.
To quantify the accuracy of shear recovery, we adopt the standard technique in the weak lensing community by using the “multiplicative bias” $m_i$ and the “additive bias” $c_i$, which are defined as: $$\begin{aligned}
\label{mc1}
&&\gamma_1^{measured}=(1+m_1)\gamma_1^{input}+c_1\\ \nonumber
&&\gamma_2^{measured}=(1+m_2)\gamma_2^{input}+c_2\end{aligned}$$
Our simulations use six sets of input shear values ($\gamma_1$, $\gamma_2$). They are: ($0.05$, $-0.05$), ($0.03$, $-0.03$), ($0.01$, $-0.01$), ($-0.01$, $0.01$), ($-0.03$, $0.03$), ($-0.05$, $0.05$). The recovered shear values as well as the linear fitting results for $\gamma_1$ and $\gamma_2$ are shown in Table \[one\_p\_gamma1\] and Table \[one\_p\_gamma2\] respectively. Note that we also list the values of $\chi^2$’s and $Q$’s for the goodness of linear fitting (see @press92 for details). For a comparison, we show in the last column of each table the quality of the shear recoveries using the “conventional” (but wrong) way, given by eq.(\[shear12PSF\_wrong\]).
The tables show that the shear recovery can be very accurate if we use the method of Z08 in the proper way, , eq.(\[shear12PSF2\]). On the other hand, if Z08 is used in the “conventional” way, , eq.(\[shear12PSF\_wrong\]), we find nonzero multiplicative biases ($m_1$, $m_2$) for both $\gamma_1$ and $\gamma_2$. The additive biases $c_1$ and $c_2$ are consistent with being zero. These results agree with our conclusions in §\[m\_zhang08\].
[c|c|c]{}
Input $\gamma_1$ & $\gamma_1$ measured in & $\gamma_1$ measured in\
& the proper way & the “conventional” way\
0.05 & 0.05004$\pm$0.00008 & 0.05217$\pm$0.00009\
0.03 & 0.03004$\pm$0.00008 & 0.03132$\pm$0.00009\
0.01 & 0.01006$\pm$0.00008 & 0.01049$\pm$0.00009\
$-0.01$ & $-0.01002\pm$0.00008 & $-0.01046\pm$0.00009\
$-0.03$ & $-0.02997\pm$0.00008& $-0.03126\pm$0.00009\
$-0.05$ & $-0.05005\pm$0.00008 & $-0.05218\pm$0.00009\
& &\
& &\
Linear Fitting & Using $\gamma_1$’s from & Using $\gamma_1$’s from\
Results & the proper way & the “conventional” way\
$m_1$ & $(0.7\pm1.0)\times10^{-3}$ & $(43.4\pm1.1)\times10^{-3}$\
$c_1$ & $(1.6\pm3.4)\times 10^{-5}$ & $(1.2\pm3.7)\times10^{-5}$\
($\chi^2_1$, $Q_1$) & (0.58, 0.96) & (0.58, 0.96)\
[c|c|c]{}
Input $\gamma_2$ & $\gamma_2$ measured in & $\gamma_2$ measured in\
& the proper way & the “conventional” way\
0.05 & 0.05011$\pm$0.00008 & 0.05226$\pm$0.00009\
0.03 & 0.02996$\pm$0.00008 & 0.03126$\pm$0.00009\
0.01 & 0.00989$\pm$0.00008 & 0.01033$\pm$0.00009\
$-0.01$ & $-0.01010\pm$0.00008 & $-0.01055\pm$0.00009\
$-0.03$ & $-0.03000\pm$0.00008 & $-0.03126\pm$0.00009\
$-0.05$ & $-0.05011\pm$0.00008 & $-0.05225\pm$0.00009\
& &\
& &\
Linear Fitting & Using $\gamma_2$’s from & Using $\gamma_2$’s from\
Results & the proper way & the “conventional” way\
$m_2$ & $(1.4\pm1.0)\times10^{-3}$ & $(44.2\pm1.1)\times10^{-3}$\
$c_2$ & $(-4.1\pm3.4)\times 10^{-5}$ & $(-3.5\pm3.7)\times10^{-5}$\
($\chi^2_2$, $Q_2$) & (3.3, 0.51) & (3.4, 0.50)\
2-Point Correlations {#two_point}
--------------------
Let us now test the recovery of the 2-point shear correlations using the method of Z08 in the way that is proposed in §\[statistics\]. In each test, we use two large groups of galaxies that are given ($\gamma_1$, $\gamma_2$) and ($\gamma_1'$, $\gamma_2'$), respectively. Each group of each test contains $10^7$ galaxies. The values of $\gamma_1$, $\gamma_2$, $\gamma_1'$, and $\gamma_2'$ vary from galaxy to galaxy. They are assumed to be normally distributed with the following covariance matrix: $\langle\gamma_1^2\rangle=\langle\gamma_1'^2\rangle=\langle\gamma_2^2\rangle=\langle\gamma_2'^2\rangle=0.04^2$, $\langle\gamma_1\gamma_2\rangle=\langle\gamma_1\gamma_2'\rangle=\langle\gamma_1'\gamma_2\rangle=\langle\gamma_1'\gamma_2'\rangle=0$, and $\langle\gamma_1\gamma_1'\rangle$ and $\langle\gamma_2\gamma_2'\rangle$ are to be specified in each test. The purpose of the test is to find out how accurately the 2-point correlations $\langle\gamma_1\gamma_1'\rangle$ and $\langle\gamma_2\gamma_2'\rangle$ can be recovered. For the tests presented here, we use the irregular type of mock galaxies, each of which is made of ten point sources generated by the 2D random walks. Each step size of the random walks is a random number between 0 and 2. The radius of each galaxy is limited to be less than 7.
In table \[two\_p\_gamma1\] and \[two\_p\_gamma2\], we report the results of six tests with six different sets of ($\langle\gamma_1\gamma_1'\rangle$, $\langle\gamma_2\gamma_2'\rangle$): (0.001,0.001), (0.0006, 0.0006), (0.0002, 0.0002), ($-0.0002$, $-0.0002$), ($-0.0006$, $-0.0006$), ($-0.001$, $-0.001$). As in §\[one\_point\], to characterize the accuracy of the method, we again use the multiplicative biases ($m_{11}$, $m_{22}$) and additive biases ($c_{11}$, $c_{22}$) that are defined as follows: $$\begin{aligned}
\label{mc2}
&&\langle\gamma_1\gamma_1'\rangle^{measured}=(1+m_{11})\langle\gamma_1\gamma_1'\rangle^{input}+c_{11}\\ \nonumber
&&\langle\gamma_2\gamma_2'\rangle^{measured}=(1+m_{22})\langle\gamma_2\gamma_2'\rangle^{input}+c_{22}\end{aligned}$$ We also show the results from the “conventional” way of using Z08 for a comparison.
Our results again show no systematic errors for the proper way of using Z08. In contrast, the “conventional” way tends to underestimate the amplitudes of the shear correlations (negative multiplicative bias). The signs of the multiplicative biases in the “conventional” cases are opposite to those found in §\[one\_point\]. This is because we have used two different types of mock galaxies.
[c|c|c]{}
Input $\langle\gamma_1\gamma_1'\rangle$ & $\langle\gamma_1\gamma_1'\rangle$ measured in & $\langle\gamma_1\gamma_1'\rangle$ measured in\
& the proper way & the “conventional” way\
$10^{-3}$ & $(0.99\pm0.03)\times10^{-3}$ & $(0.71\pm0.02)\times10^{-3}$\
$6\times10^{-4}$ & $(6.3\pm0.3)\times10^{-4}$ & $(4.5\pm0.2)\times10^{-4}$\
$2\times10^{-4}$ & $(1.7\pm0.3)\times10^{-4}$ & $(1.2\pm0.2)\times10^{-4}$\
$-2\times10^{-4}$ & $(-2.1\pm0.3)\times10^{-4}$ & $(-1.4\pm0.2)\times10^{-4}$\
$-6\times10^{-4}$ & $(-6.1\pm0.3)\times10^{-4}$ & $(-4.5\pm0.2)\times10^{-4}$\
$-10^{-3}$ & $(-0.99\pm0.03)\times10^{-3}$ & $(-0.72\pm0.02)\times10^{-3}$\
& &\
& &\
Linear Fitting & Using $\langle\gamma_1\gamma_1'\rangle$’s from & Using $\langle\gamma_1\gamma_1'\rangle$’s from\
Results & the proper way & the “conventional” way\
$m_{11}$ & -0.001$\pm$0.018 & -0.281$\pm$0.012\
$c_{11}$ & $(-0.18\pm1.2)\times 10^{-5}$ & $(-4.1\pm7.9)\times10^{-6}$\
($\chi^2_{11}$, $Q_{11}$) & (2.5, 0.64) & (2.3, 0.68)\
[c|c|c]{}
Input $\langle\gamma_2\gamma_2'\rangle$ & $\langle\gamma_2\gamma_2'\rangle$ measured in & $\langle\gamma_2\gamma_2'\rangle$ measured in\
& the proper way & the “conventional” way\
$10^{-3}$ & $(0.99\pm0.03)\times10^{-3}$ & $(0.71\pm0.02)\times10^{-3}$\
$6\times10^{-4}$ & $(5.8\pm0.3)\times10^{-4}$ & $(4.2\pm0.2)\times10^{-4}$\
$2\times10^{-4}$ & $(2.0\pm0.3)\times10^{-4}$ & $(1.3\pm0.2)\times10^{-4}$\
$-2\times10^{-4}$ & $(-1.7\pm0.3)\times10^{-4}$ & $(-1.3\pm0.2)\times10^{-4}$\
$-6\times10^{-4}$ & $(-6.0\pm0.3)\times10^{-4}$ & $(-4.4\pm0.2)\times10^{-4}$\
$-10^{-3}$ & $(-1.03\pm0.03)\times10^{-3}$ & $(-0.73\pm0.02)\times10^{-3}$\
& &\
& &\
Linear Fitting & Using $\langle\gamma_2\gamma_2'\rangle$’s from & Using $\langle\gamma_2\gamma_2'\rangle$’s from\
Results & the proper way & the “conventional” way\
$m_{22}$ & 0.000$\pm$0.018 & -0.284$\pm$0.012\
$c_{22}$ & $(-0.61\pm1.2)\times 10^{-5}$ & $(-6.2\pm7.9)\times10^{-6}$\
($\chi^2_{22}$, $Q_{22}$) & (1.8, 0.77) & (1.6, 0.82)\
Summary
=======
Conventionally, in the studies of weak lensing, for each shear component ($\gamma_1$ or $\gamma_2$), one hopes to construct a single quantity from each background galaxy image, whose ensemble average is equal to the true value of a component of the true shear field. We have shown that such conventional shear estimators (CSEs) [*do not exist in convenient forms*]{} even in the absence of the PSF.
Based on the method of Zhang (2008), we have proposed to measure the cosmic shear in a new way: using the ratio of the ensemble averages of two galaxy properties to estimate each shear component. (Also see § 9.2 of Weinberg (2008) for a similar study.) We have shown that, using both analytic analyses and numerical examples, the new way of estimating cosmic shears is unbiased, and does not contain systematic errors to the first order in shear at least. The new type of shear measurement demands shear statistics such as n-point correlation functions to be measured in an unconventional way as well, but with little additional cost.
Acknowledgements {#acknowledgements .unnumbered}
================
The author acknowledges Gary Bernstein, Toshifumi Futamase, Yi Mao, Pengjie Zhang for helpful discussions, and Shanghai Astronomical Observatory (SHAO), National Astronomical Observatories of China (NAOC) for their hospitality. In particular, JZ would like to thank Gary Bernstein for pointing out a mistake in the proof for the nonexistence of CSEs in the presence of PSF in the previous version of this manuscript. JZ is currently supported by the TCC Fellowship of Texas Cosmology Center of the University of Texas at Austin, and was previously supported by the TAC Fellowship of the Theoretical Astrophysics Center of UC Berkeley, where part of this work was done. EK is supported in part by NSF grants AST-0807649 and PHY-0758153 and NASA grant NNX08AL43G.
Bacon D., Refregier A., Clowe D., Ellis R., 2001, MNRAS, 325, 1065
Bernstein G., 2010, MNRAS, 406, 2793
Bernstein G. & Jarvis M., 2002, AJ, 123, 583
Bonnet H. & Mellier Y., 1995, A&A, 303, 331
Bridle S., Gull S., Bardeau S., Kneib J., 2001, in Scientific N. W., ed., Proceedings of the Yale Cosmology Workshop
Bridle S. et al., 2009, Annals of Applied Statistics, Vol.3, No.1, 6-37, arXiv: 0802.1214
Bridle S. et al., 2010, MNRAS, 405, 2044
Erben T., Van Waerbeke L., Bertin E., Mellier Y., Schneider P., 2001, A&A, 366, 717
Heymans C. et al., 2006, MNRAS, 368, 1323
Hoekstra H., Franx M., Kuijken K., Squires G., 1998, ApJ, 504, 636
Hoekstra H. & Jain B., 2008, Annual Review of Nuclear and Particle Science, 58, 99, arXiv:0805.0139
Kaiser N., 2000, ApJ, 537, 555
Kaiser N., Squires G. & Broadhurst T., 1995, ApJ, 449, 460
Luppino G. & Kaiser N., 1997, ApJ, 475, 20
Massey R. et al., 2007, MNRAS, 376, 13
Massey R. & Refregier A., 2005, MNRAS, 363, 197
Nakajima R. & Bernstein G., 2007, AJ, 133, 1763
Press W., Flannery B., Teukolsky S., Vetterling W., 1992, [*Numerical Recipes*]{}, Cambridge Univ. Press, 2nd ed.
Refregier A., 2003, MNRAS, 338, 35
Refregier A. & Bacon D., 2003, MNRAS, 338, 48
Rhodes J., Refregier A. & Groth E., 2000, ApJ, 536, 79
Tyson J., Wenk R. & Valdes F., 1990, ApJL, 349, L1
Voigt L. & Bridle S., 2010, MNRAS, 404, 458
Weinberg S., 2008, [*Cosmology*]{}, Oxford University Press, 1st ed.
Zhang J., 2008, MNRAS, 383, 113
Zhang J., 2010a, MNRAS, 403, 673
Zhang J., 2010b, submitted to MNRAS, arXiv:1002.3614
Appendix A – The Derivatives of the Spin-2 CSEs {#appendixA .unnumbered}
===============================================
Under a clockwise coordinate rotation of angle $\theta$, the cosmic shear components transform according to the following rule: $$\label{shear_rotation}
\gamma_1^{\theta}+\ima\gamma_2^{\theta}=\left(\gamma_1+\ima\gamma_2\right)\exp(\ima 2\theta)$$ Using the chain rule, we then find: $$\label{shear_deriv_rotation}
\frac{\partial}{\partial\gamma_1^{\theta}}+\ima\frac{\partial}{\partial\gamma_2^{\theta}}=\left(\frac{\partial}{\partial\gamma_1}+\ima\frac{\partial}{\partial\gamma_2}\right)\exp(\ima 2\theta)$$ On the other hand, we know that the spin-2 shear estimators $(\Gamma_1, \Gamma_2)$ transform as: $$\label{Estimator_transform}
\Gamma_1^{\theta}+\ima\Gamma_2^{\theta}=\left(\Gamma_1+\ima\Gamma_2\right)\exp(\ima 2\theta)$$ Apply eq.(\[shear\_deriv\_rotation\]) onto eq. (\[Estimator\_transform\]), we get: $$\begin{aligned}
\label{spin-4}
&&\left(\frac{\partial\Gamma_1^{\theta}}{\partial\gamma_1^{\theta}}-\frac{\partial\Gamma_2^{\theta}}{\partial\gamma_2^{\theta}}\right)+\ima\left(\frac{\partial\Gamma_1^{\theta}}{\partial\gamma_2^{\theta}}+\frac{\partial\Gamma_2^{\theta}}{\partial\gamma_1^{\theta}}\right)\\ \nonumber
&=&\left[\left(\frac{\partial\Gamma_1}{\partial\gamma_1}-\frac{\partial\Gamma_2}{\partial\gamma_2}\right)+\ima\left(\frac{\partial\Gamma_1}{\partial\gamma_2}+\frac{\partial\Gamma_2}{\partial\gamma_1}\right)\right]\exp(\ima 4\theta)\end{aligned}$$ Eq.(\[spin-4\]) clearly shows that the derivatives $(\partial_{\gamma_1}\Gamma_1-\partial_{\gamma_2}\Gamma_2)+\ima(\partial_{\gamma_2}\Gamma_1+\partial_{\gamma_1}\Gamma_2)$ form a spin-4 quantity. We can also apply eq.(\[shear\_deriv\_rotation\]) onto the complex conjugate of eq. (\[Estimator\_transform\]), and get: $$\begin{aligned}
\label{spin-0}
&&\left(\frac{\partial\Gamma_1^{\theta}}{\partial\gamma_1^{\theta}}+\frac{\partial\Gamma_2^{\theta}}{\partial\gamma_2^{\theta}}\right)+\ima\left(\frac{\partial\Gamma_1^{\theta}}{\partial\gamma_2^{\theta}}-\frac{\partial\Gamma_2^{\theta}}{\partial\gamma_1^{\theta}}\right)\\ \nonumber
&=&\left(\frac{\partial\Gamma_1}{\partial\gamma_1}+\frac{\partial\Gamma_2}{\partial\gamma_2}\right)+\ima\left(\frac{\partial\Gamma_1}{\partial\gamma_2}-\frac{\partial\Gamma_2}{\partial\gamma_1}\right)\end{aligned}$$ Therefore, $\partial_{\gamma_1}\Gamma_1+\partial_{\gamma_2}\Gamma_2$ is a scalar, and $\partial_{\gamma_2}\Gamma_1-\partial_{\gamma_1}\Gamma_2$ is a pseudo-scalar (it has odd parity).
Appendix B – Some Mathematical Details Regarding Galaxies of Elliptical Shapes {#appendixB .unnumbered}
==============================================================================
The surface brightness profiles of the galaxies used in §\[spin\_2\_esti\] are parametrized as $f_S(R)$, where $R=a(x^2+y^2)+b(x^2-y^2)+2cxy$, and $x$ and $y$ are the coordinate variables. Curves of constant values of $R$ can be ellipses, hyperbolas, parabolas, or lines depending on the values of $a$, $b$, and $c$. For our purposes, we only need ellipses. This requires $a+b>0$ and $a^2-b^2>c^2$ according to the matrix theory. Since linear coordinate transformations do not spoil these relations, weakly lensed images of these galaxies are still ellipses. For the same reason, one can also easily show that the lensed images are still parametrized by the same function $f_S(R)$ with $R=a'(x^2+y^2)+b'(x^2-y^2)+2c'xy$, where $a'=a(1-2\kappa)-2\gamma_1b-2\gamma_2c$, $b'=b(1-2\kappa)-2\gamma_1a$, and $c'=c(1-2\kappa)-2\gamma_2a$.
Let us now show that the ellipticity parameters $(\epsilon_1, \epsilon_2)$ for galaxies of the form $f_S\left[a(x^2+y^2)+b(x^2-y^2)+2cxy\right]$ are equal to $(-b/a, -c/a)$. According to the definitions in §\[classic\], we have: $$\label{e1e2}
Q_{ij}=\int d^2\vec{x}x_ix_jf_S\left[a(x^2+y^2)+b(x^2-y^2)+2cxy\right]$$ This integration can be carried out using the following linear coordinate transformation: $$\begin{aligned}
\label{ct_e1e2}
&&x=\frac{1}{\sqrt{a+b}}(\beta x'+\alpha y')\\ \nonumber
&&y=\frac{1}{\sqrt{a-b}}(\alpha x'+\beta y')\end{aligned}$$ where $$\begin{aligned}
\label{ab_ct}
\alpha&=&\frac{1}{2}\left(\sqrt{1-\frac{c}{\sqrt{a^2-b^2}}}-\sqrt{1+\frac{c}{\sqrt{a^2-b^2}}}\right)\\ \nonumber
&\times&\sqrt{\frac{a^2-b^2}{a^2-b^2-c^2}}\\ \nonumber
\beta&=&\frac{1}{2}\left(\sqrt{1-\frac{c}{\sqrt{a^2-b^2}}}+\sqrt{1+\frac{c}{\sqrt{a^2-b^2}}}\right)\\ \nonumber
&\times&\sqrt{\frac{a^2-b^2}{a^2-b^2-c^2}}\end{aligned}$$ Using $(x', y')$ defined in eq.(\[ct\_e1e2\]) to replace $(x, y)$ in eq.(\[e1e2\]), we get: $$\begin{aligned}
\label{QQs}
&&Q_{11}-Q_{22}\\ \nonumber
&=&\frac{\beta^2-\alpha^2}{\sqrt{a^2-b^2}}\int d^2\vec{x'}\left[\frac{a(\beta^2-\alpha^2)}{a^2-b^2}(x'^2-y'^2)\right.\\ \nonumber
&-&\left.\frac{b(\alpha^2+\beta^2)}{a^2-b^2}(x'^2+y'^2)-\frac{4b\alpha\beta}{a^2-b^2}x'y'\right]f_S(x'^2+y'^2)\\ \nonumber
&&Q_{11}+Q_{22}\\ \nonumber
&=&\frac{\beta^2-\alpha^2}{\sqrt{a^2-b^2}}\int d^2\vec{x'}\left[-\frac{b(\beta^2-\alpha^2)}{a^2-b^2}(x'^2-y'^2)\right.\\ \nonumber
&+&\left.\frac{a(\alpha^2+\beta^2)}{a^2-b^2}(x'^2+y'^2)+\frac{4a\alpha\beta}{a^2-b^2}x'y'\right]f_S(x'^2+y'^2)\\ \nonumber
&&2Q_{12}\\ \nonumber
&=&\frac{\beta^2-\alpha^2}{\sqrt{a^2-b^2}}\int d^2\vec{x'}\left[\frac{2\alpha\beta}{\sqrt{a^2-b^2}}(x'^2+y'^2)\right.\\ \nonumber
&+&\left.\frac{2(\alpha^2+\beta^2)}{\sqrt{a^2-b^2}}x'y'\right]f_S(x'^2+y'^2)\end{aligned}$$ Given that the function $f_S$ only depends on $x'^2+y'^2$, we have: $$\begin{aligned}
\label{facts}
&&\int d^2\vec{x'}(x'^2-y'^2)f_S(x'^2+y'^2)\\ \nonumber
&=&\int d^2\vec{x'}x'y'f_S(x'^2+y'^2)=0\end{aligned}$$ It is now straightforward to calculate $(\epsilon_1, \epsilon_2)$: $$\begin{aligned}
\label{e_12s}
&&\epsilon_1=\frac{Q_{11}-Q_{22}}{Q_{11}+Q_{22}}=-\frac{b}{a}\\ \nonumber
&&\epsilon_2=\frac{2Q_{12}}{Q_{11}+Q_{22}}=-\frac{c}{a}\end{aligned}$$
Appendix C – Why CSEs Do Not Likely Exist In the Presence of PSF {#appendixC .unnumbered}
=================================================================
For technical convenience, let us work in Fourier space. According to the notations in §\[classic\], we use $\widetilde{f_L}(\vec{k}^L)$ and $\widetilde{f_S}(\vec{k}^S)$ to denote the Fourier transformations of the lensed galaxy image $f_L(\vtl)$ and the original galaxy image $f_S(\vts)$ respectively. Their relations are given by the following equations: $$\begin{aligned}
\label{Fourier}
&&\widetilde{f_L}(\vec{k}^L)=\int d^2\vtl e^{\ima\vec{k}^L\cdot\vtl}f_L(\vtl)\\ \nonumber
&&\widetilde{f_S}(\vec{k}^S)=\int d^2\vts e^{\ima\vec{k}^S\cdot\vts}f_S(\vts)\end{aligned}$$ Using the relations defined in eq.(\[fifstits\]), we get: $$\begin{aligned}
\label{Fourier2}
\widetilde{f_L}(\vec{k}^L)&=&\int d^2\vts \left\vert \mathrm{det} \left(\frac{\partial \vtl}{\partial \vts}\right)\right\vert e^{\ima\vec{k}^L\cdot(\mathbf{A}\vts)}f_S(\vts)\\ \nonumber
&=&\vert \mathrm{det} (\mathbf{A})\vert\int d^2\vts e^{\ima(\mathbf{A}\vec{k}^L)\cdot\vts}f_S(\vts)\\ \nonumber
&=&\vert \mathrm{det} (\mathbf{A})\vert\widetilde{f_S}(\mathbf{A}\vec{k}^L)\end{aligned}$$ Eq. (\[Fourier2\]) simply means that under lensing, the Fourier transformation of the galaxy image is changed from $\widetilde{f_S}(\vec{k})$ to $\vert \mathrm{det} (\mathbf{A})\vert\widetilde{f_S}(\mathbf{A}\vec{k})$.
Due to the presence of the PSF, the Fourier transformation of the observed image $\widetilde{f_O}$ is related to that of the lensed image $\widetilde{f_L}$ via: $$\label{f_O}
\widetilde{f_O}(\vec{k})=\widetilde{W}(\vec{k})\widetilde{f_L}(\vec{k})$$ where $\widetilde{W}(\vec{k})$ is the Fourier transformation of the PSF. Without loss of generality, in the rest of our discussion, we use the isotropic Gaussian PSF, , $\widetilde{W}(\vec{k})=\widetilde{W}_{\beta}(\vec{k})=\exp(-\beta^2\left\vert\vec{k}\right\vert^2/2)$. The advantage of working in Fourier space is that the PSF is included as a multiplicative factor, rather than a convolution as in real space. Combining eq.(\[Fourier2\]) and eq.(\[f\_O\]), we get: $$\label{f_O2}
\widetilde{f_O}(\vec{k})=\widetilde{W}_{\beta}(\vec{k})\vert \mathrm{det} (\mathbf{A})\vert\widetilde{f_S}(\mathbf{A}\vec{k})$$ Since the PSF profile in Fourier space typically falls off quickly when the wave number exceeds the inverse of the size of the PSF, it strongly suppresses the power of the observed images on small scales. Therefore, only a finite number of Fourier modes are available for providing shape information. Recovering information on arbitrarily small scales is never feasible in practice due to noise and numerical problems.
To form CSEs, let us use the multipole moments of galaxy images to represent the shape information, which are defined as: $$\label{defineM}
M_{ij}=\int d^2\vec{k}k_1^ik_2^j\widetilde{f_O}(\vec{k})$$ where $i$ and $j$ are non-negative integers. Note that due to the finite degrees of freedom of the shape information, one can equivalently choose other basis (, shapelets) to study the same issue without affecting the conclusion. For simplicity but without loss of generality, let us only consider galaxies that are invariant under the parity transformation $\vec{x}\to -\vec{x}$. Note that weak lensing does not change this property. For this type of galaxies, the imaginary part of $\widetilde{f_O}(\vec{k})$ is always zero, and $M_{ij}$’s are real. Furthermore, $M_{ij}$ is zero when $i+j$ is an odd number. In this case, the shear estimators $\Gamma_1$ and $\Gamma_2$ can be written as functions of the $M_{ij}$ with $i+j$ being even numbers only. For some of the lowest order $M_{ij}$’s, we can find out how they transform under lensing using eq.(\[f\_O2\]): $$\begin{aligned}
\label{Mldl}
M_{ij}&=&\int d^2\vec{k}k_1^ik_2^j\widetilde{W}_{\beta}(\vec{k})\vert \mathrm{det} (\mathbf{A})\vert\widetilde{f_S}(\mathbf{A}\vec{k})\\ \nonumber
&=&\int d^2\vec{k}(\mathbf{A}^{-1}\vec{k})_1^i(\mathbf{A}^{-1}\vec{k})_2^j\widetilde{W}_{\beta}(\mathbf{A}^{-1}\vec{k})\widetilde{f_S}(\vec{k})\end{aligned}$$ The last step in the above equation is achieved by redefining $\mathbf{A}\vec{k}$ as $\vec{k}$. By keeping the terms in matrix $\mathbf{A}$ up to the first order in shear, one can straightforwardly show the following: $$\begin{aligned}
\label{Mij_trans}
&&M_{20}-M_{02}\\ \nonumber
&=&(1-2\kappa)(M_{20}^S-M_{02}^S)+\kappa\beta^2\left(M_{40}^S-M_{04}^S\right)\\ \nonumber
&+&2\gamma_2\beta^2\left(M_{31}^S-M_{13}^S\right)\\ \nonumber
&-&\gamma_1\left[2(M_{20}^S+M_{02}^S)-\beta^2\left(M_{40}^S+M_{04}^S-2M_{22}^S\right)\right]\\ \nonumber
\\ \nonumber
&&M_{11}\\ \nonumber
&=&(1-2\kappa)M_{11}^S-\gamma_2\left(M_{20}^S+M_{02}^S-2\beta^2 M_{22}^S\right)\\ \nonumber
&+&\kappa\beta^2\left(M_{31}^S+M_{13}^S\right)+\gamma_1\beta^2\left(M_{31}^S-M_{13}^S\right)\\ \nonumber
\\ \nonumber
&&M_{20}+M_{02}\\ \nonumber
&=&(1-2\kappa)(M_{20}^S+M_{02}^S)+\kappa\beta^2\left(M_{40}^S+2M_{22}^S+M_{04}^S\right)\\ \nonumber
&-&\gamma_1\left[2(M_{20}^S-M_{02}^S)-\beta^2\left(M_{40}^S-M_{04}^S\right)\right]\\ \nonumber
&-&\gamma_2\left[4M_{11}^S-2\beta^2\left(M_{31}^S+M_{13}^S\right)\right]\end{aligned}$$ where $$\label{MS}
M_{ij}^S=\int d^2\vec{k}k_1^ik_2^j\widetilde{W}_{\beta}(\vec{k})\widetilde{f_S}(\vec{k})$$
The conventional cosmic shear estimators ($\Gamma_1, \Gamma_2$) are functions of $M_{ij}$. According to our discussion in §\[spin\_2\_esti\], the two functions have to satisfy the following relation: $$\label{requirement_again}
\frac{\partial\Gamma_1}{\partial\gamma_1}+\frac{\partial\Gamma_2}{\partial\gamma_2}=2$$ In the presence of PSF, we are unable to find out whether one can find CSEs that satisfy the most general requirement given in eq.(\[requirement\_again\]). However, we can show that there do not exist CSEs satisfying a slightly stronger condition: $$\label{requirement_stronger}
\frac{\partial\Gamma_1}{\partial\gamma_1}=\frac{\partial\Gamma_2}{\partial\gamma_2}=1$$ In addition to eq. (\[requirement\_again\]), eq.(\[requirement\_stronger\]) simply imposes another requirement that the spin-4 parts of the derivatives of the shear estimators with respect to the shears are zero. We can rewrite eq.(\[requirement\_stronger\]) for, , $\Gamma_1$, using the chain rule as: $$\label{Gamma_1_chain_rule}
1=\sum_{ij}\frac{\partial\Gamma_1}{\partial M_{ij}}\frac{\partial M_{ij}}{\partial \gamma_1}$$ Since there are only a finite number of multipole moments available for constructing the shear estimators, we assume the maximum value of $i+j$ is $N$ (N is an even integer), , we have: $$\label{Gamma_1_chain_rule2}
1=\sum_{i+j\le N}\frac{\partial\Gamma_1}{\partial M_{ij}}\frac{\partial M_{ij}}{\partial \gamma_1}$$ Because the right side of eq.(\[Gamma\_1\_chain\_rule2\]) is evaluated at $\gamma_1=\gamma_2=\kappa=0$, both $\partial\Gamma_1/\partial M_{ij}$ and $\partial M_{ij}/\partial \gamma_1$ are functions of only $M_{ij}^S$’s, , the multipole moments in the absence of lensing.
As in eq.(\[Mij\_trans\]), one can show in general that $\partial M_{ij}/\partial \gamma_1$ involves higher order multipole moments, , $M_{i'j'}^S$’s with $i'+j'>i+j$. In particular, when $i+j=N$, $\partial M_{ij}/\partial \gamma_1$ depends linearly on $M_{i'j'}^S$’s with $i'+j'>N$. To satisfy the constraint in eq.(\[Gamma\_1\_chain\_rule2\]), the coefficients in front of the terms proportional to $M_{i'j'}^S$’s ($i'+j'>N$) must vanish because they are independent of the multipole moments with $i+j\le N$. As a result, we find that $\partial\Gamma_1/\partial M_{ij}$ has to vanish when $i+j=N$. In other words, we have: $$\label{Gamma_1_chain_rule3}
1=\sum_{i+j\le N-2}\frac{\partial\Gamma_1}{\partial M_{ij}}\frac{\partial M_{ij}}{\partial \gamma_1}$$ We can now recursively use the above reasoning to show that there does not exist an $N$ that can satisfy eq.(\[Gamma\_1\_chain\_rule2\]). Therefore, we can never find CSEs of the type defined in eq.(\[requirement\_stronger\]). Though, one may still expect to find CSEs based on the most general requirement defined in eq.(\[requirement\_again\]). However, even in this case, CSEs [*must*]{} reduce to highly nonlinear forms for galaxies of pure elliptical shapes when the PSF effect is small, as we have shown in §\[spin\_2\_esti\]. This feature is already sufficient for arguing against their usefulness in practice.
Appendix D – Derivation of the Multiplicative Biases Resulting From a Misuse of Z08 {#appendixD .unnumbered}
===================================================================================
Let us now calculate the terms $\delta_1$ and $\delta_2$ defined in eq.(\[shear12PSF\_wrong\]). The averages of the spatial derivatives of the surface brightness field of a single galaxy can be related to the Fourier modes of the image. The Fourier transformation has been defined in eq.(\[Fourier\]) in Appendix C. Following the notations of Appendix C, we find: $$\label{RF_derive}
\partial_if_O(\vec{x})=\int\frac{d^2\vec{k}}{(2\pi)^2}(-\ima k_i)e^{-\ima\vec{k}\cdot\vec{x}}\widetilde{f_O}(\vec{k})$$ and $$\begin{aligned}
\label{RF_derivative_aves}
&&\left\langle\partial_if_O(\vec{x})\partial_jf_O(\vec{x})\right\rangle_g\\ \nonumber
&&=\frac{\int_Sd^2\vec{x}}{S}\partial_if_O(\vec{x})\partial_jf_O(\vec{x})\\ \nonumber
&&=\frac{\int_Sd^2\vec{x}}{S}\frac{\int d^2\vec{k}\int d^2\vec{k'}}{(2\pi)^4}(-k_ik'_j)e^{-\ima(\vec{k}+\vec{k'})\cdot\vec{x}}\widetilde{f_O}(\vec{k})\widetilde{f_O}(\vec{k'})\\ \nonumber
&&=\frac{1}{S}\frac{\int d^2\vec{k}\int d^2\vec{k'}}{(2\pi)^4}(-k_ik'_j)(2\pi)^2\delta_D^2(\vec{k}+\vec{k'})\widetilde{f_O}(\vec{k})\widetilde{f_O}(\vec{k'})\\ \nonumber
&&=\frac{1}{S}\frac{\int d^2\vec{k}}{(2\pi)^2}k_ik_j\left\vert\widetilde{f_O}(\vec{k})\right\vert^2\end{aligned}$$ where $S$ is the total area of the map containing the galaxy. Similarly, one can derive the following relation: $$\begin{aligned}
\label{RF_derive2}
\left\langle\vec{\nabla}f\cdot\vec{\nabla}(\nabla^2f_O)\right\rangle_g=-\frac{1}{S}\frac{\int d^2\vec{k}}{(2\pi)^2}\left\vert\vec{k}\right\vert^4\left\vert\widetilde{f_O}(\vec{k})\right\vert^2\end{aligned}$$ Eq.(\[RF\_derivative\_aves\]) and eq.(\[RF\_derive2\]) allow us to transform eq.(\[shear12PSF\_wrong\]) into its version in Fourier space: $$\begin{aligned}
\label{shearFourier}
&&\frac{1}{2}\left\langle\frac{P_{20}-P_{02}}{P_{20}+P_{02}-\beta^2D_4/2}\right\rangle_{en}=-\gamma_1(1-\delta_1)\\ \nonumber
&&\left\langle\frac{P_{11}}{P_{20}+P_{02}-\beta^2D_4/2}\right\rangle_{en}=-\gamma_2(1-\delta_2)\end{aligned}$$ where $$\begin{aligned}
\label{defineP}
&&P_{ij}=\int d^2\vec{k}k_1^ik_2^j\left\vert\widetilde{f_O}(\vec{k})\right\vert^2\\ \nonumber
&&D_n=\int d^2\vec{k}\left\vert\vec{k}\right\vert^n\left\vert\widetilde{f_O}(\vec{k})\right\vert^2\end{aligned}$$ Note that $D_4=P_{40}+2P_{22}+P_{04}$. It is now clear that the method of Z08 basically utilizes the quadrupole moments in the Fourier space to measure the cosmic shear.
Using eq.(\[f\_O2\]), we can find out how $P_{ij}$ transform under lensing: $$\begin{aligned}
\label{pldl}
P_{ij}&=&\int d^2\vec{k}k_1^ik_2^j\left\vert\widetilde{W}_{\beta}(\vec{k})\vert \mathrm{det} (\mathbf{A})\vert\widetilde{f_S}(\mathbf{A}\vec{k})\right\vert^2\\ \nonumber
&=&\vert\mathrm{det} (\mathbf{A})\vert\int d^2\vec{k}(\mathbf{A}^{-1}\vec{k})_1^i(\mathbf{A}^{-1}\vec{k})_2^j\\ \nonumber
&\times&\left\vert\widetilde{W}_{\beta}(\mathbf{A}^{-1}\vec{k})\widetilde{f_S}(\vec{k})\right\vert^2\end{aligned}$$ The last step in the above equation is achieved by redefining $\mathbf{A}\vec{k}$ as $\vec{k}$. For the isotropic Gaussian PSF, $\widetilde{W}_{\beta}(\vec{k})=\exp(-\beta^2\left\vert\vec{k}\right\vert^2/2)$. By keeping the terms in matrix $\mathbf{A}$ up to the first order in shear, one can straightforwardly show the following: $$\begin{aligned}
\label{Pij_trans}
&&P_{20}-P_{02}\\ \nonumber
&=&P_{20}^S-P_{02}^S+2\kappa\beta^2\left(P_{40}^S-P_{04}^S\right)+4\gamma_2\beta^2\left(P_{31}^S-P_{13}^S\right)\\ \nonumber
&-&2\gamma_1\left[P_{20}^S+P_{02}^S-\beta^2\left(P_{40}^S+P_{04}^S-2P_{22}^S\right)\right]\\ \nonumber
\\ \nonumber
&&P_{11}\\ \nonumber
&=&P_{11}^S-\gamma_2\left(P_{20}^S+P_{02}^S-4\beta^2 P_{22}^S\right)+2\kappa\beta^2\left(P_{31}^S+P_{13}^S\right)\\ \nonumber
&+&2\gamma_1\beta^2\left(P_{31}^S-P_{13}^S\right)\\ \nonumber
\\ \nonumber
&&P_{20}+P_{02}\\ \nonumber
&=&P_{20}^S+P_{02}^S-2\gamma_1\left[P_{20}^S-P_{02}^S-\beta^2\left(P_{40}^S-P_{04}^S\right)\right]\\ \nonumber
&-&4\gamma_2\left[P_{11}^S-\beta^2\left(P_{31}^S+P_{13}^S\right)\right]+2\kappa\beta^2D_4^S\\ \nonumber
\\ \nonumber
&&D_4\\ \nonumber
&=&D_4^S-2\kappa\left(D_4^S-\beta^2D_6^S\right)\\ \nonumber
&-&2\gamma_1\left[2\left(P_{40}^S-P_{04}^S\right)-\beta^2\left(P_{60}^S+P_{42}^S-P_{24}^S-P_{06}^S\right)\right]\\ \nonumber
&-&4\gamma_2\left[2\left(P_{31}^S+P_{13}^S\right)-\beta^2\left(P_{51}^S+2P_{33}^S+P_{15}^S\right)\right]\\ \nonumber\end{aligned}$$ where $$\begin{aligned}
\label{PS}
&&P_{ij}^S=\int d^2\vec{k}k_1^ik_2^j\left\vert\widetilde{W}_{\beta}(\vec{k})\widetilde{f_S}(\vec{k})\right\vert^2\\ \nonumber
&&D_n^S=\int d^2\vec{k}\left\vert\vec{k}\right\vert^n\left\vert\widetilde{W}_{\beta}(\vec{k})\widetilde{f_S}(\vec{k})\right\vert^2\end{aligned}$$
By keeping the terms up to the first order in shear, and using the fact that the ensemble averages are taken over statistically isotropic galaxy samples, one can now directly find expressions for $\delta_1$ and $\delta_2$ of eq.(\[shear12PSF\_wrong\]): $$\begin{aligned}
\label{deltas}
\delta_1&=&\left\langle\frac{\left( P_{20}^S-P_{02}^S\right)^2}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\\ \nonumber
&-&2\beta^2\left\langle\frac{\left( P_{20}^S-P_{02}^S\right)\left( P_{40}^S-P_{04}^S\right)}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\\ \nonumber
&+&\frac{\beta^4}{2}\left\langle\frac{\left( P_{20}^S-P_{02}^S\right)\left( P_{60}^S+P_{42}^S-P_{24}^S-P_{06}^S\right)}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\\ \nonumber
\\ \nonumber
\delta_2&=&\left\langle\frac{4\left( P_{11}^S\right)^2}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\\ \nonumber
&-&8\beta^2\left\langle\frac{P_{11}^S\left( P_{31}^S+P_{13}^S\right)}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\\ \nonumber
&+&2\beta^4\left\langle\frac{P_{11}^S\left( P_{51}^S+2P_{33}^S+P_{15}^S\right)}{\left( P_{20}^S+P_{02}^S-\beta^2D_4^S/2\right)^2}\right\rangle_{en}\end{aligned}$$
\[lastpage\]
[^1]: E-mail:jzhang@astro.as.utexas.edu
[^2]: We do not consider shear estimators that cannot be Taylor expanded as power series of the galaxy shape parameters in this paper. Such shear estimators likely suffer numerical instabilities in practice.
[^3]: This bias was not detected in the numerical calculations presented in Z08, as the number of galaxies ($10^4$) used for the average was still too small.
[^4]: Note that for our purpose, it is not necessary to know the form of $\Psi$.
[^5]: This is at least true if the relevant galaxies are separated by a large physical distance. Detailed discussions about the correlations of $\Psi$’s belong to the topic of “Galaxy Intrinsic Alignment”, which is beyond the scope of this paper.
[^6]: Such correlations again belong to the topic of “Galaxy Intrinsic Alignment,” which is not considered in this paper.
[^7]: Note that Z08 did not detect these biases because it did not use enough galaxies ($10^4$) in the numerical tests. Here, we use $10^7$ galaxies (or pairs of galaxies for the 2-point correlation function) for each test.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate the action of time dependent detunings upon the excitation inversion of a Cooper pair box interacting with a nanomechanical resonator. The method employs the Jaynes-Cummings model with damping, assuming different decay rates of the Cooper pair box and various fixed and t-dependent detunings. It is shown that while the presence of damping plus constant detunings destroy the collapse/revival effects, convenient choices of time dependent detunings allow one to reconstruct such events in a perfect way. It is also shown that the mean excitation of the nanomechanical resonator is more robust against damping of the Cooper pair box for convenient values of t-dependent detunings.'
address:
- '$^{1}$Universidade Estadual de Goiás, Rod. BR 153, 3105, 75132-903 Anápolis, GO, Brazil.'
- '$^{\dagger}$Instituto de Física, Universidade Federal de Goiás, 74001-970 Goiânia, GO, Brazil.'
- '$^{\ddagger}$Universidade Paulista, Rod. BR 153, km 7, 74845-090 Goiânia, GO, Brazil.'
author:
- 'C. Valverde$^{1,\dagger,\ddagger}$, A.T. Avelar$^{\dagger}$ and B. Baseia$^{\dagger}$'
title: Controlling Excitations Inversion of a Cooper Pair Box Interacting with a Nanomechanical Resonator
---
Introduction
============
A popular and exactly soluble model in quantum optics is the Jaynes-Cumming model (JCM). It describes the interaction of a two-level atom with a single-mode of the electromagnetic field [@1; @2; @2.1; @2.2; @2.3; @2.4; @2.5]. Over the last two decades various extensions of the ordinary JCM have been used in various directions, e.g., as adapted to: (i) the study of interaction of a three-level atom with a two-mode squeezed vacuum [@10]; (ii) the study of atom-field interaction in the presence of a cavity damping [@11]; (iii) the same as in (i), including an additional (nonlinear) Kerr medium [@12]; (iv) the two-level atoms inside a cavity acted upon by an external field control [@13]; (v) study of the nonlinear dynamical evolution of a driven two-photon Jaynes-Cummings model [@14]; (vi) the study of a generalized Jaynes-Cummings models, including dissipation [15,16,16a]{} and multiphoton interactions [@17; @bb1]; etc. In all these cases, with interest either on the field or on atomic properties, the theoretical approach traditionally assumes the atom-field coupling as a constant parameter. Comparatively, the number of works in the literature is very small when one considers such coupling and the atomic frequency as time dependent parameters [@law; @l1; @l22; @l2; @l3; @13.1], including time dependent amplitudes [@abdalla]. However, this scenario is also relevant; for example, the state of two qubits (qubits stand for quantum bits) with a desired degree of entanglement can be generated via a time dependent atom-field coupling [@olaya]. Actually, such coupling can modify the dynamical properties of the atom and the field, with transitions that involve a large number of photons [@yang]. In general, these studies are simplified by neglecting the atomic decay from an excited level. Theoretical treatments taking into account this complication of the real world may employ a modified JCM. In these case, as expected, the state describing the system decoheres, since the presence of dissipation destroys the state of a system as time flows.
In the present work we extend what we have learned from the JCM applied to the atom-field interaction to investigate a more advantageous system, from the experimental viewpoint (faster response, better controllability, and useful scalability for quantum computation [@y1] ) by considering a nanomechanical resonator (NR) interacting with a Cooper pair box (CPB) [17,16,18]{}. Such nanodevice has been explored extensively in the literature, e.g., to investigate: (i) quantum nondemolition measurements [@ak1; @e1], (ii) decoherence of nonclassical states, as Fock states and superposition states describing mesoscopic systems [@e2; @c2], etc. The fast advance in the technique of fabrication in nanotechnology has implied great interest in the study of the NR system in view of its potential applications, as a sensor - to be used in biology, astronomy, quantum computation, and in quantum information [@ak2; @bb2; @bb3], to implement the quantum qubit [ak3]{} and in the production of nonclassical states, as Fock state [akk,bb4]{}, Schrödinger’s cat state [@akk2], squeezed states [a34]{}, clusters states [@ak], etc. In particular, when accompanied by superconducting charge qubits, the NR has been used to prepare entangled states [@akk1]. Zhou et al.[@a34] have proposed a scheme to prepare squeezed states using a NR coupled to a CPB qubit; in this proposal the NR-CPB coupling is under an external control while the connection between these two subsystems plays an important role in quantum computation. Such a control is achieved via convenient change of the system parameters, which can set “on" and “off" the interaction between the NR and the CPB, on demand.
In this report we will investigate the CPB excitation inversion, its control, and the average photon number in the NR. We will consider dissipation in the CPB due to a decay rate from excited to ground states. We will also verify in which way the time dependence of the CPB-NR coupling modifies these two properties. To this end we must solve the time evolution of the whole CPB-NR system, via the approach presented in the following Section.
Model Hamiltonian for the CPB-NR system
=======================================
A Josephson charge qubit system has been used to couple with a NR. Here we study a modified model where a CPB is coupled to a NR, as shown in Fig. [cooper]{} below. The scheme is inspired by the works of Jie-Qiao Liao et al. [@ak3] and Zhou et al. [@a34] where we have substituted each Josephson junction by two of them. This creates a new configuration that includes a third loop. A superconducting CPB charge qubit is adjusted via a voltage $V_{1}$ at the system input and a capacitance $C_{g}$. We want that the scheme attains an efficient tunneling effect for the Josephson energy. In Fig. \[cooper\] we observe three loops: one great loop between two small ones. This makes it easier controlling the external parameters of the system since the control mechanism includes the input voltage $V_{1}$ plus three external fluxes $\Phi _{L},$ $\Phi _{r}$ and $\Phi _{e}(t)$. In this way one can induce small neighboring loops. The great loop contains a NR which is modeled as a harmonic oscillator with a high-Q mode of frequency $\Omega $ and its effective area in the center of the apparatus changes as the NR oscillates, which creates an external flux $\Phi _{e}(t)$ that provides the CPB-NR coupling.
In pursuing the quantum behavior of a macro scale object the nano scale mechanical resonator plays an important role. At sufficiently low temperature the zero-point fluctuation of the NR will be comparable to its thermal Brownian motion. The detection of zero-point fluctuations of the NR can give a direct test of Heisenberg’s uncertainty principle. With a sensitivity up to 10 times the amplitude of the zero-point fluctuation, LaHaye et al [@lahaye] have experimentally detected the vibrations of a 20 MHz mechanical beam of tens of micrometres size. For a 20MHz mechanical resonator its temperature must be cooled below 1mK to suppress the thermal fluctuation. For a GHz mechanical resonator a temperature of 50mK is sufficient to effectively freeze out its thermal fluctuation and let it enter the quantum regime. This temperature is already attainable in dilution refrigerators.
![*Model for the CPB-NMR coupling.*[]{data-label="cooper"}](cpb.ps){width="10cm" height="12cm"}
In this work we will assume the four Josephson junctions being identical, with the same Josephson energy $E_{J}^{0}$, the same being assumed for the external fluxes $\Phi _{L}$ and $\Phi _{r}$, i.e., with same magnitude but opposite sign: $\Phi _{L}=-\Phi _{r}=\Phi _{x}$. This interaction actually couples the two subsystems. Together with the free Hamiltonian of flux qubit and NR, the Hamiltonian of the whole system reads
$$\hat{H}=\Omega \hat{a}^{\dagger }\hat{a}+4E_{c}\left( N_{g}-\frac{1}{2}\right) \hat{\sigma}_{z}-4E_{J}^{0}\cos \left( \frac{\pi \Phi _{x}}{\Phi _{0}}\right) \cos \left( \frac{\pi \Phi _{e}}{\Phi _{0}}\right) \hat{\sigma}_{x},
\label{a1}$$
where $\hat{a}^{\dagger }(\hat{a})$ is the creation (annihilation) operator for the NR excitation, with frequency $\Omega $ and mass $m$; $E_{J}^{0}$ and $E_{c}$ are respectively the energy of each Josephson junction and the charge energy of a single electron; $C_{g}$ and $C_{J}^{0}$ are the input capacitance and the capacitance of each Josephson tunnel, respectively. $\Phi _{0}=h/2e$ is the quantum flux and $N_{g}=C_{g}V_{1}/2e$ is the charge number in the input with the input voltage $V_{1}$. We have used the Pauli matrices to describe our system operators, where the states $\left\vert
0\right\rangle $ and $\left\vert 1\right\rangle $ represent the number of extra Cooper pairs in the superconducting island. We have: $\hat{\sigma}_{z}=\left\vert 0\right\rangle \left\langle 0\right\vert -\left\vert
1\right\rangle \left\langle 1\right\vert $, $\hat{\sigma}_{x}=\left\vert
0\right\rangle \left\langle 1\right\vert -\left\vert 1\right\rangle
\left\langle 0\right\vert $ and $E_{C}=e^{2}/\left( C_{g}+4C_{J}^{0}\right)
. $
The magnetic flux can be written as the sum of two terms, $$\Phi _{e}=\Phi _{1}+B\ell \hat{x}, \label{a4}$$where the first term $\Phi _{1}$ is the induced flux, corresponding to the equilibrium position of the NR and the second term describes the contribution due to the vibration of the NR; $B$ represents the magnetic field created in the loop. We have assumed the displacement $\hat{x}$ described as $\hat{x}=x_{0}(\hat{a}^{\dagger }+\hat{a})$, where $x_{0}=\sqrt{m\Omega /2}$ is the amplitude of the oscillation. Substituting the Eq.([a4]{}) in Eq.(\[a1\]) and controlling the flux $\Phi _{1}$ we can adjust $\cos \left( \frac{\pi \Phi _{1}}{\Phi _{0}}\right) =0$, and making the approximation $\pi B\ell x/\Phi _{0}<<1$ the above Hamiltonian results as (in rotating wave approximation), $$\hat{H}=\Omega \hat{a}^{\dagger }\hat{a}+\frac{1}{2}\omega _{0}\hat{\sigma}_{z}+\lambda _{0}(\hat{\sigma}_{+}\hat{a}+\hat{a}^{\dagger }\hat{\sigma}_{-}),$$where the constant coupling $\lambda _{0}=-4E_{J}^{0}\cos \left( \frac{\pi
\Phi _{x}}{\Phi _{0}}\right) \left( \frac{\pi B\ell x_{0}}{\Phi _{0}}\right)
$ and the effective energy $\omega _{0}=8E_{c}\left( N_{g}-\frac{1}{2}\right) .$ An important advantage of this coupling mechanism is its easy and convenient controllability.
Next, we will extend the previous approach to a more general scenario by substituting $\Omega \rightarrow \omega (t)=\Omega +f\left( t\right) $ and $\lambda _{0}\rightarrow \lambda (t)=\lambda _{0}\left[ 1+f\left( t\right)
/\Omega \right] $ [@yang; @c1; @jf]; in addition we assume the presence of a constant decay rate $\gamma $ in the CPB; $\omega _{0}$ is the transition frequency of the CPB and $\lambda _{0}$ stands for the CPB-NR coupling. $\hat{\sigma}_{\pm }$ and $\hat{\sigma}_{z}$ are the CPB transition and excitation inversion operators, respectively; they act on the Hilbert space of atomic states and satisfy the commutation relations $\left[ \hat{\sigma}_{+},\hat{\sigma}_{-}\right] =\hat{\sigma}_{z}$ and $\left[ \hat{\sigma}_{z},\hat{\sigma}_{\pm }\right] =\pm \hat{\sigma}_{\pm }$. As well known, the coupling parameter $\lambda (t)$ is proportional to $\sqrt{\omega
(t)/V\left( t\right) }$, where the time dependent quantization volume $V\left( t\right) $ takes the form $V\left( t\right) =V_{0}/\left[ 1+f\left(
t\right) /\Omega \right] $ [@scully; @l2; @jf]. Accordingly, we obtain the new (*non hermitian*) Hamiltonian
$$\hat{H}=\omega (t)\hat{a}^{\dagger }\hat{a}+\frac{1}{2}\omega _{0}\hat{\sigma}_{z}+\lambda (t)(\hat{\sigma}_{+}\hat{a}+\hat{a}^{\dagger }\hat{\sigma}_{-})-i\frac{\gamma }{2}\left\vert 1\right\rangle \left\langle 1\right\vert .
\label{b1}$$
It is worth remembering that non Hermitian Hamiltonians (NHH) are largely used in the literature. To give some few examples we mention: Ref. [@nh5], where the authors use a NHH and an algorithm to generalize the conventional theory; Ref. [@nh1], using a NHH to get information about entrance and exit channels; Ref. [@nh6], using non Hermitian techniques to study canonical transformations in quantum mechanics; Ref. [@nh7], solving quantum master equations in terms of NHH; Ref. [@nh3], using a new approach for NHH to study the spectral density of weak H-bonds involving damping; Ref. [@nh8], studying NHH with real eigenvalues; Ref. [@nh4], using a canonical formulation to study dissipative mechanics exhibing complex eigenvalues; Ref. [@nh9], studying NHH in non commutative space, and more recently: Ref. [@nh10], studying the optical realization of relativistic NHH; Ref. [@l2], studying the evolution of entropy of atom-field interaction; Ref. [@l22], using a damping JC-Model to study entanglement between two atoms, each of them lying inside different cavities
Solving the CPB-NR system
=========================
The state that describes this time dependent system can be written in the form
$$\left\vert \Psi \left( t\right) \right\rangle =\sum\nolimits_{n=0}^{\infty
}(C_{0,n}\left( t\right) \left\vert 0,n\right\rangle +C_{1,n}\left( t\right)
\left\vert 1,n\right\rangle ), \label{b2}$$
where $\left\vert 0,n\right\rangle $ $\left( \left\vert 1,n\right\rangle
\right) $ represents the CPB in its excited state $\left\vert 1\right\rangle
$ (ground state $\left\vert 0\right\rangle $). Taking the CPB initially prepared in its excited state $\left\vert 1\right\rangle $ and the NR in a coherent states $\left\vert \alpha \right\rangle $, and expanding coherent state component in the Fock’s basis, i.e., $\left\vert \alpha \right\rangle
=exp(-|\alpha |^{2}/2)\sum_{n=o}^{\infty }(\alpha ^{n}/\sqrt{n!})|n\rangle $, we have $\left\vert \alpha \right\rangle =\sum\nolimits_{n=0}^{\infty
}F_{n}\left\vert n\right\rangle $. Assuming the NR and CPB decoupled at $t=0 $ and the initial conditions $C_{0,n}\left( 0\right) =0$ and $\sum\nolimits_{n=0}^{\infty }\left\vert C_{1,n}\left( 0\right) \right\vert
^{2}=1$ we may write the Eq. (\[b2\]) as $\left\vert \Psi \left( 0\right)
\right\rangle =\sum\nolimits_{n=0}^{\infty }F_{n}\left\vert 1,n\right\rangle
.$
The time dependent Schrödinger equation for this system is
$$i\frac{d\left\vert \Psi \left( t\right) \right\rangle }{dt}=\hat{H}\left\vert \Psi \left( t\right) \right\rangle , \label{b5}$$
with the Hamiltonian $\hat{H}$ given in Eq. (\[b1\]).* *Substituting Eq.(\[b1\]) in Eq.(\[b5\]) we get the (coupled) equations of motion for the probabilitity amplitudes $C_{1,n}(t)$ and $C_{0,n+1}(t)$:$$\begin{aligned}
\frac{\partial C_{1,n}(t)}{\partial t} &=& -in\omega (t)C_{1,n}(t)-\frac{i}{2}\omega _{0}C_{1,n}(t)-i\lambda (t)\sqrt{n+1}C_{0,n+1}(t)-\frac{\gamma }{2}C_{1,n}(t),\phantom {aaaaaa} \label{b8} \\
\frac{\partial C_{0,n+1}(t)}{\partial t}&=&-i(n+1)\omega (t)C_{0,n+1}(t)+\frac{i}{2}\omega _{0}C_{0,n+1}(t)-i\lambda (t)\sqrt{n+1}C_{1,n}(t).
\label{b9}\end{aligned}$$The numerical solutions of the coefficients $C_{1,n}(t)$, $C_{0,n+1}(t)$ furnish the quantum dynamical properties of the system, including the CPB-NR entanglement.
As well known, in the presence of decay rate $\gamma $ in the CPB the state of the whole CPB-NR system becomes mixed. In this case its description requires the use of the density operator $\hat{\rho}_{CN}$, which describes the entire system. To obtain the reduced density matrix describing the CPB (NR) subsystem we must trace over variables of the NR (CPB) subsystem. For example, $\hat{\rho}_{NR}=Tr_{CPB}(\hat{\rho}_{CN})$: $$\hat{\rho}_{NR}=\sum\nolimits_{n,n^{\prime }=o}^{\infty }\left[
C_{1,n}(t)C_{1,n^{\prime }}^{\ast }(t)+C_{0,n}(t)C_{0,n^{\prime }}^{\ast }(t)\right] \left\vert n\right\rangle \left\langle n\prime \right\vert .
\label{c2}$$
Excitation Inversion of the CPB
===============================
The CPB excitation inversion, here denoted as $I(t)_{CPB}$, is an important observable of two level systems. It is defined as the difference of probabilities of finding the system in the excited and ground state, as follows$$I(t)_{{CPB}}=\sum\nolimits_{n=0}^{\infty }\left[ \left\vert
C_{1,n}(t)\right\vert ^{2}-\left\vert C_{0,n+1}(t)\right\vert ^{2}\right] .
\label{joao}$$
The Eq. (\[joao\]) allows one to look at the time evolution of the CPB excitation inversion. First, we assume the resonant case $(f(t)=0)$ for different values of the decay rate $\gamma $, with $\alpha =5$ and $\Omega
=\omega _{0}=2000\lambda _{0}$ and assuming the NR initially in a coherent state with the average number of excitations $\left\langle n\right\rangle
=25 $ as shown in Fig. \[inversao1\]. With the exceptions of amplitudes, the plots (a), (b) and (c) of Fig. \[inversao1\] show identical for collapse-revival effects: the higher the decay rate the lower is the amplitude of oscillation. In presence of detuning, where $f(t)=\Delta =const$, where $\Delta \ll \omega _{0}$, $\Omega $, we see that the excitation inversion in Fig. \[inversao2\](a) occurs within the interval $30<\lambda
_{0}t<50$, whereas in Fig. \[inversao2\](b) it occurs in the range $60<\lambda _{0}t<75$ and in Fig. \[inversao2\](c) the excitation inversion goes to zero rapidly. Considering the case of variable detuning, $f(t)=\eta
\sin (\omega \prime t)$, we see in the plots (a), (b) and (c) of Fig. [inversao3]{} that the excitation inversion occurs frequently in Fig. [inversao3]{}(a) and desapears when the parameter $\eta $ increases, as shown in the plots (b) and (c). Now, even considering the worst results obtained in the off-resonant cases, with detuning $\Delta =\eta =60\lambda _{0}$, as shown in Fig. \[inversao2\](c) and Fig. \[inversao3\](c) we see that we can recover the collapse-revival effects via the increase of the parameter $\omega \prime ,$ as shown in the plots (a), (b) and (c) of Fig. [inversao4]{}. So, the parameter $\omega \prime $ plays a fundamental role in the control of collapse and revival effect.
Fig. \[nmf1\] shows plots of the mean value of the NR excitations in the presence of CPB decay rate for various values of amplitude of oscillations (parameter $\eta $). Plots (**d)** and (**h)** are for the resonant case: in (**d)** the decay rate is greater than in (**h)**. Plots (**a)**, (**b)**, and (**c)** are for constant decay rates, with (time independent) detuning that increases from (**c)** $\rightarrow $** (b)** $\rightarrow $** (a)**. Finally, plots (**e)**, (**f)** and (**g)** are for time dependent detuning, with the parameter $\eta $ increasing from (**e)** $\rightarrow $** (f)** $\rightarrow $** (g)**. We note that the three plots for time dependent detunings (**e)**, (**f)** and* (***g**) are better than those for constant detunings (**a)**, (**b)** and (**c**): despite all plots are concerned with the same decay rate, the first group is more robust against decay. For example, comparing the plots (**a**) and (**g)**: although in **(a)** the fixed detuning is $\Delta =60\lambda _{0}$ and in (**g)** maximum detuning is $\Delta _{\max }=\eta =60\lambda _{0}$ we see that in the last case the average value of the NR decays more slowly. One observes in the plot (**c)** of Fig. \[inversao4\] that the interval $\lambda _{0}t$, where the presence of the time dependent detuning recovers the collapse-revival effect, coincides with that in the plot (**g)** of Fig. \[nmf1\] where the mean value of excitation is around 5 times greater than in the case of constant detuning (plot (**a)** of same figure).
![*Time evolution of the Excitation Inversion in the CPB with $\left\langle n\right\rangle =25$, $\Omega =\protect\omega _{0}=2000\protect\lambda _{0}$, for $f(t)=0$ (resonance) and different values of decay rates $\protect\gamma$: (a) $\protect\gamma =0.01\protect\lambda _{0}$, (b) $\protect\gamma =0.05\protect\lambda _{0}$ and (c) $\protect\gamma =0.5\protect\lambda _{0}$.*[]{data-label="inversao1"}](inversao1){width="9cm" height="8cm"}
![*Same as in Fig. \[inversao1\] for $\protect\gamma =0.05\protect\lambda _{0}$ and different values of detunings $($cf. $f(t)=\Delta =const.)$: (a) $\Delta =10\protect\lambda _{0}$ , (b) $\Delta
=20\protect\lambda _{0}$, (c) $\Delta =60\protect\lambda _{0}$.*[]{data-label="inversao2"}](inversao2){width="9cm" height="8cm"}
![Same as in Fig. \[inversao1\] for $\protect\gamma =0.05\protect\lambda _{0}$ and different time-dependent detunings* (cf. $f(t)=\protect\eta \sin (\protect\omega \prime t)$): (a) $\protect\eta=10\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0}$, (b) $\protect\eta=20\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0},$ (c) $\protect\eta=60\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0}.$*[]{data-label="inversao3"}](inversao3){width="9cm" height="8cm"}
![Same as in Fig. \[inversao1\] for $\protect\gamma =0.05\protect\lambda _{0}$ and different time-dependent detunings* (cf. $f(t)=\protect\eta\sin (\protect\omega \prime t)$): (a) $\protect\eta=60\protect\lambda _{0}$ and $\protect\omega \prime =20\protect\lambda _{0}$, (b) $\protect\eta=60\protect\lambda _{0}$ and $\protect\omega \prime =40\protect\lambda _{0},$ (c) $\protect\eta=60\protect\lambda _{0}$ and $\protect\omega \prime =58\protect\lambda _{0}$.*[]{data-label="inversao4"}](inversao4){width="9cm" height="8cm"}
![*Average number of excitations of the NR, versus time, for $\left\langle n(0)\right\rangle =25$, $\Omega =\protect\omega _{0}=2000\protect\lambda _{0}$, for the following cases: (a) off-resonance* $\protect\gamma =0.05\protect\lambda _{0}$, $\Delta =60\protect\lambda _{0},$* (b) off-resonance* $\protect\gamma =0.05\protect\lambda _{0}$, $\Delta =20\protect\lambda _{0},$* (c) off-resonance* $\protect\gamma =0.05\protect\lambda _{0}$, $\Delta =10\protect\lambda _{0},$* (d) on resonance* $\protect\gamma =0.5\protect\lambda _{0},$* (e) off-resonance* $\protect\gamma =0.05\protect\lambda _{0},$ $\protect\eta=10\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0},$* (f) off-resonance* $\protect\gamma =0.05\protect\lambda _{0},$ $\protect\eta=20\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0},$*(g) off-resonance* $\protect\gamma =0.05\protect\lambda _{0},$ $\protect\eta=60\protect\lambda _{0}$ and $\protect\omega \prime =0.05\protect\lambda _{0},$* (h) on resonance* $\protect\gamma =0.05\protect\lambda _{0}$. []{data-label="nmf1"}](nmf1){width="9cm" height="8cm"}
Conclusion
==========
We have considered a Hamiltonian model that describes a CPB-NR interacting system to study the CPB excitation inversion, $I(t)_{{CPB}},$ and the average excitation number of the NR, $\left\langle n(t)\right\rangle _{NR}$. We have also considered the off-resonant case, with various values of the detuning parameter $f\ $($f=0;f=\Delta ;$ and $f=\eta \sin (\omega \prime t)$) and in the presence of CPB decay (about 10 times greater than the (neglected) NR decay). These properties are characteristics of the entangled state that describes this coupled system for various values of the parameters involved. We have assumed the CPB initially in its excited state and the NR initially in a coherent state (see preparation in [@17]). So, the following three scenarios were treated: (*i*) both subsystems in resonance (detuning $f=0$); (*ii*) off-resonance, with a constant detuning $(f=\Delta \neq 0)$, and (*iii*) with a time dependent detuning $(f(t)=\eta \sin (\omega \prime t))$. The results were discussed in the previous Section: in resume, concerning the CPB excitation inversion, an interesting result emerged: although the presence of a constant detuning destroys the collapse and revivals of the excitation inversion, these effects are restituted by the action of convenient time dependent detunings - even in the presence of damping in the CPB; concerning the NR average excitation number, another interesting result appeared: convenient choices of the time dependent detuning $f(t)$ makes the NR subsystem more robust against the decay affecting the CPB subsystem. For constant values of detuning, our numerical results are similar to others in the literature using a master equation (see, e.g., Ref. [@16]).
Finally we emphasize that the change in magnetic flux $\Phi_{e}$ (cf. Fig. \[cooper\]), due to the presence of an external force upon the NR, is the responsible for controlling the parameters $\omega (t)$ and $\lambda (t)$.
Acknowledgements
================
The authors thank the FAPEG (CV) and CNPq (ATA, BB), Brasilian Agencies, for partially supporting this paper.
References {#references .unnumbered}
==========
[99]{} Jaynes E T and Cummings F W 1963 Proc. IEEE 51 89
Joshi A and Puri R R 1987 J. Mod. Opt. 34 1421
Joshi A and Puri R R 1989 J. Mod. Opt. 36 557
Joshi A and Puri R R 1990 Phys. Rev. A 42 4336
Joshi A and Puri R R 1990 Opt. Commun. 75 189
Joshi A and Lawande S V 1989 Opt. Commun. 70 21
Puri R R and Agarwal G S 1987 Phys. Rev. A 35 3433
Lai W K, Buzek V and Knight P L 1991 Phys. Rev. A 44 6043
Quang T, Knight P L and Buzek V 1991 Phys. Rev. A 44 6092
Joshi A and Puri R R 1992 Phys. Rev. A 45 5056
Alsing P, Gao D -S and Carmichael H J 1992 Phys. Rev. A 45 5135
Li F L and Gao S -Y 2000 Phys. Rev. A 62 043809
Joshi A 2000 Phys. Rev. A 62 043812
Budini A A, de Matos Filho R L and Zagury N 2003 Phys. Rev. A 67 033815
Tiwari R P and Stroud D 2008 Phys. Rev. B 77 214520
Hoehne F, Pashkin Yu A, Astafiev O, Faoro L, Ioffe L B, Nakamura Y and Tsai J S 2010 Phys. Rev. B 81 184112
LaHaye M D, Suh J, Echternach P M, Schwab K C and Roukes M L 2009 Nature 459 960
Chaba A N, Baseia B,Wang C and Vyas R 1996 Phisica A 232 273
Law C K, Zhu S Y and Zubairy M S 1995 Phys. Rev. A 52 4095
Janowicz M 1998 Phys. Rev. A 57 4784
Zhang G F and Xie X C 2010 Eur. Phys. J. D 60 423
Fei J, Xie S Y and Yang Y P 2010 Chin. Phys. Lett. 27 014212
Ateto M S 2010 Int. J. Theor. Phys. 49 276
Abdalla M S,Abdel-Aty M and Obada A S F 2003 Physica A 326 203
Olaya-Castro A, Johnson N F and Quiroga L 2004 Phys. Rev. A 70 020301
Yang Y P, Xu J P, Li G X and Chen H 2004 Phys. Rev. A 69 053406
Astafiev O, Pashkin Y A, Nakamura Y, Yamamoto T and Tsai J S 2004 Phys. Rev. Lett. 93 267007
Xue F, Wang Y D, Sun C P, Okamoto H, Yamaguchi H and Semba K 2007 New J. Phys. 9 35
Ruskov R, Schwab K and Korotkov A N 2005 Phys. Rev. B 71 235407
Irish E K and Schwab K 2003 Phys. Rev. B 68 155311
Siewert J, Brandes T and Falci G 2009 Phys. Rev. B 79 024504
Valverde C and Baseia B 2004 Int. J. Quantum Inf. 2 421
Sun C P, Wei L F, Liu Y and Nori F 2006 Phys. Rev. A 73 022318
Escher B M, Avelar A T, da Rocha T M and Baseia B 2004 Phys. Rev. A 70 025801
Valverde C, Avelar A T, Baseia B and Malbouisson J M C 2003 Phys. lett. A 315 213
Liao J, Wu Q and Kuang L 2008 arXiv:quant-ph/0803.4317v1
Brattke S, Varcoe B T H and Walther H 2001 Phys. Rev. Lett. 86 3534
Maia L P A, Baseia B, Avelar A T and Malbouisson J M C 2004 J. Opt. B: Quantum Semiclass. opt 6 1
Liu Y, Wei L F and Nori F 2005 Phys. Rev. A 71 063820
Zhou X X and Mizel A 2006 Phys. Rev. Lett. 97 267201
Chen G, Chen Z, Yu L and Liang J 2007 Phys. Rev. A 76 024301
Wei L F, Liu Y and Nori F 2006 Phys. Rev. Lett. 96 246803
LaHaye M D, Buu O, Camarota B and Schwab K C 2004 Science 304 74
Valverde C, Avelar A T and Baseia B 2011 arXiv:1104.2106v1
Fei J, Yuan X S and Ping Y Y 2009 Chin. Phys. Soc. 18 3193
Scully M O and Zubairy M S 1997 Quantum Optics (Cambridge: Cambridge University) pp 136-195
Baker H G and Singleton R L 1990 Phys. Rev. A 42 10
Lemm J C, Giraud B G and Weiguny A 1994 Phys. Rev. Lett. 73 420
Lee H and I’yi W S 1995 Phys. Rev. A 51** **982
Visser P M and Nienhuis G 1995 Phys. Rev. A 52 4727
Belharaya K, Blaise P and Rousseau O H 2003 Chem. Phys. 293 9
Faria C F M and Fring A 2007 Laser Phys. 17 424
Rajeev S G 2007 Annals of Physics 322 1541
Giri P R and Roy P 2009 Eur. Phys. J. C 60** **157
Longhi S 2010 Phys. Rev. Lett. 105 013903
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
A [*$\mathcal{K}^n_2$-set*]{} is a set of zero Lebesgue measure containing a translate of every plane in an $(n-2)$–dimensional manifold in $\mathrm{Gr}(n,2)$, where the manifold fulfills a curvature condition. We show that this is a natural class of sets with respect to the Kakeya problem and prove that $\dim_H(E)\ge
7/2$ for all $\mathcal{K}^4_2$-sets $E$. When the underlying field is replaced by ${\mathbb{C}}$, we get $\dim_H(E)\ge 7$ for all $\mathcal{K}^4_2$-sets over ${\mathbb{C}}$, and we construct an example to show that this is sharp. Thus $\mathcal{K}^4_2$-sets over ${\mathbb{C}}$ do not necessarily have full Hausdorff dimension.
address: 'Department of Mathematics, Universidad Autónoma de Madrid, 28049 Madrid, Spain'
author:
- 'Keith M. Rogers'
title: On a planar variant of the Kakeya problem
---
Introduction {#intro}
============
A [*Besicovitch set (with lines)*]{} in ${\mathbb{R}}^n$ is a set of zero Lebesgue measure that contains a translate of every member of $\mathrm{Gr}(n,1)$, where $\mathrm{Gr}(n,1)$ is the Grassmanian manifold of 1–dimensional linear subspaces of ${\mathbb{R}}^n.$
A construction of A. Besicovitch [@be] led to the surprising fact that such sets exist. The Kakeya conjecture asserts that Besicovitch sets must have full Hausdorff dimension.
We note that $\mathrm{Gr}(n,1)$ has dimension $n-1$ and a line has dimension 1, and the union of these lines will fill the space and have dimension $n$. Informally, the Kakeya conjecture asserts that under translation of the lines, the dimensions continue to add up, or that the intersection remains negligible.
The conjecture has been solved in the affirmative in the plane, but is open for $n\ge3$. T. Wolff [@wo] proved that $\dim_H(E) \ge(n+2)/2$ for all Besicovitch sets $E$ in ${\mathbb{R}}^n,$ where $\dim_H$ denotes Hausdorff dimension, and there has been much recent progress for higher dimensions (see [@bo2], [@ka1], [@ka2], [@la], [@ta]).
N. H. Katz, I. Łaba and T. Tao [@ka] have shown that the Minkowski dimension of a Kakeya set in ${\mathbb{R}}^3$ is strictly greater than $5/2$. They also show that if a Kakeya set in ${\mathbb{R}}^3$ has dimension close to $5/2$, then it must exhibit a certain structural property that they call ‘planiness’. Roughly speaking, most of the lines that pass through a point in the Besicovitch set, lie in the union of a small number of planes. It seems reasonable then, to consider the variant of the problem where lines are replaced by planes.
An [*$(n,2)$-set*]{} is a set that contains a translate of every member of $\mathrm{Gr}(n,2)$, where $\mathrm{Gr}(n,2)$ is the Grassmanian manifold of 2–dimensional linear subspaces of ${\mathbb{R}}^n.$ J. Bourgain [@bo] proved that $(4,2)$-sets have strictly positive Lebesgue measure, so that there are no, so called, Besicovitch $(4,2)$-sets. Similarly, there are no Besicovitch $(3,2)$-sets. T. Mitsis [@mi] recently claimed that $(n,2)$-sets have full Hausdorff dimension when $n\ge5$, although unfortunately the argument is incomplete.
As $\mathrm{Gr}(n,2)$ has dimension $2(n-2)$ and a plane has dimension 2, these results do not have the same informal interpretation as that of the Kakeya problem. Indeed the planes of an $(n,2)$-set will inevitably have, in some sense, nontrivial intersection.
Initially, it appears reasonable to ask whether sets containing translates of an $(n-2)$–dimensional manifold in $\mathrm{Gr}(n,2)$ necessarily have full Hausdorff dimension. Without further restriction on the manifold however, this will fail. For instance, a 2–dimensional manifold in $\mathrm{Gr}(4,2)$ could have all of its planes contained in a 3–space, so that the set consisting of the union of these planes would have Hausdorff dimension 3. We outlaw this by adding a curvature condition which is essentially a version of the Wolff axiom ([@wo], [@ka], [@mo]) for planes.
Let $\pi,\pi'\in\mathrm{Gr}(n,2),$ and define the [*major angle between $\pi$ and $\pi'$*]{} by $$\theta(\pi,\pi')=\|\mathrm{proj}_\pi-\mathrm{proj}_{\pi'}\|,$$ where $\mathrm{proj}_\pi:{\mathbb{R}}^n\to\pi$ is the orthogonal projection onto $\pi$. We note that $\theta$ takes values from zero to one, so this is not a standard angle. As $\theta$ is a metric, we can induce a measure $\mu$ on the manifold $\mathcal{M}\subset\mathrm{Gr}(n,2)$, which we normalise to have total mass one. We denote by $B_\epsilon(\pi)$ the ball $$\{\pi'\in\mathrm{Gr}(n,2): \theta(\pi,\pi')<\epsilon\},$$ and denote by $\mathcal{M}^{\epsilon}_{{\Lambda}_m}$ the set $$\{\pi'\in\mathcal{M}:\min_{\pi\subset{\Lambda}_m}\theta(\pi,\pi')<\epsilon\},$$ where ${\Lambda}_m\subset{\mathbb{R}}^n$ is an $m$–space with $3\le m\le n-1$.
An $(n-2)$–dimensional manifold $\mathcal{M} \subset \mathrm{Gr}(n,2)$ is said to be [*curved*]{} if there exists a constant $C$ such that $$\label{pope}
\frac{\mu(\mathcal{M}^{\epsilon_1}_{{\Lambda}_m}\cap
B_{\epsilon_2}(\pi))}{\epsilon_1^{m-2}\epsilon_2^{n-m}}<C$$ for all $\epsilon_1,\epsilon_2>0,$ $m$–spaces ${\Lambda}_m\subset{\mathbb{R}}^n$ and $\pi\in\mathcal{M}.$
The condition forces the dimension of $\{\pi\in\mathcal{M}:\pi\subset\Lambda_m\}$ to be less than or equal to $m-2$ for all $m$–spaces ${\Lambda}_m,$ and the manifold to be more evenly distributed in $\mathrm{Gr}(n,2)$. The reason that this is known as an axiom in the Kakeya problem, is that the corresponding condition is automatically fulfilled by $\mathrm{Gr}(n,1)$. This is not the case for planes, so we are obliged to include it in our definition.
A [*$\mathcal{K}^n_2$-set*]{} is a set of zero Lebesgue measure that contains a translate of every member of an $(n-2)$–dimensional curved manifold $\mathcal{M}\subset\mathrm{Gr}(n,2).$
As $\{ l\times {\mathbb{R}}: l\in \mathrm{Gr}(n-1,1)\}$ is an $(n-2)$–dimensional curved manifold, a Besicovitch set that is constructed by taking the cross product of a Besicovitch set in ${\mathbb{R}}^{n-1}$ with a copy of ${\mathbb{R}}$ is also a $\mathcal{K}^n_2$-set. Thus we have the existence of such sets.
In Section \[two\] we will prove the following results.
\[dimension0\] Let $E$ be a $\mathcal{K}^3_2$-set. Then $\dim_H(E)=3.$
Thus $\mathcal{K}^3_2$-sets have full Hausdorff dimension. Our main concern will be the proof of the following result.
\[dimension\] Let $E$ be a $\mathcal{K}^4_2$-set. Then $\dim_H(E)\ge
7/2$.
If we replace the underlying field by ${\mathbb{C}}$ and multiply the exponents of $\epsilon_1$ and $\epsilon_2$ in (\[pope\]) by two, then the proof of Theorem \[dimension\] can be modified to obtain the following bound. The modification involves little more than changing the relevant exponents.
\[cork\] Let $E$ be a $\mathcal{K}^4_2$-set over ${\mathbb{C}}$. Then $\dim_H(E)\ge 7$.
In Section \[three\], we will construct an example to show that Theorem \[cork\] is sharp. Thus $\mathcal{K}^4_2$-sets over ${\mathbb{C}}$ do not necessarily have full Hausdorff dimension.
Geometric preliminaries {#one}
=======================
Throughout ${\delta}$ will be a real parameter such that $0<{\delta}\ll 1$, and $\pi$ will denote a plane, and never the usual number. We say that a set $\Pi\subset\mathrm{Gr}(n,2)$ is [*$\delta$–separated*]{} if $\theta(\pi,\pi')> \delta$ for all $\pi,\pi'\in\Pi$.
We use $A{\lesssim}B$ to denote the estimate $A\le C_\epsilon{\delta}^{-\epsilon} B$ for all $\epsilon>0$, where $C_\epsilon$ is a constant depending only on $\epsilon$ and the manifold $\mathcal{M}$. This notation will be convenient, as factors of $\log(1/{\delta})$ will simply disappear. We use $A\sim B$ to denote $B/2< A\le B.$
We will require another notion of an angle between two planes. The need for two notions is created by the fact that in ${\mathbb{R}}^n$, where $n\ge4$, the intersection of two planes can be a point, and not necessarily a line. Let $l, l'\in \mathrm{Gr}(n,1),$ and define the [*minor angle between $\pi$ and $\pi'$*]{} by $$\underline{\theta}(\pi,\pi')=\min_{l\subset \pi,l'\subset
\pi'}\|\mathrm{proj}_l-\mathrm{proj}_{l'}\|.$$ Informally, $\underline{\theta}$ can be considered to be the smallest angle between two planes. If $\pi,\pi'\in\mathrm{Gr}(n,2)$ intersect in a line, then $\underline{\theta}(\pi,\pi')=0$, so that $\underline{\theta}$ is not a metric.
Define a [*plate*]{} $P_\pi$ to be the image of $[0,1]\times[0,1]\times\underbrace{[0,{\delta}]\times\cdots\times[0,{\delta}]}_{n-2}$ under a rotation and translation, such that its face of area one is parallel to $\pi\in\mathrm{Gr}(n,2)$. Define $S_\pi$ to be the central unit square of the plate $P_\pi$. When $P_\pi\cap P_{\pi'}\neq \emptyset$ and $\theta(\pi,\pi')=\phi$, we say the plates [*intersect at major angle $\phi$.*]{} Similarly, when $P_\pi\cap P_{\pi'}\neq \emptyset$ and ${\underline{\theta}}(\pi,\pi')=\phi$, we say the plates [*intersect at minor angle $\phi$.*]{}
We will require the following lemmas. The first is due to Mitsis [@mi] and is a natural extension of an observation of A. Córdoba [@co]. We include the proof for convenience. The third is a natural extension of an observation of Wolff [@wo].
\[miti\] Let $\pi_1,
{\pi_2}\in\mathrm{Gr}(n,2)$. Then $$|P_{\pi_1}\cap P_{\pi_2}|{\lesssim}\frac{{\delta}^{n-1}}{\theta(\pi_1,\pi_2)}.$$
We have that $P_{\pi_1}\cap P_{\pi_2}$ is contained in $(R_1\cap R_2)\times R_3$, where $R_1$ and $R_2$ are 2–dimensional rectangles of dimension $1\times{\delta}$, and $R_3$ is an $(n-2)$–dimensional cuboid of dimension $1\times\underbrace{{\delta}\times\cdots\times{\delta}}_{n-3}.$ The rectangles $R_1$ and $ R_2$ intersect at an angle ${\gtrsim}\theta(\pi_1,\pi_2)$, so by elementary geometry, $$\mathcal{L}^2(R_1\cap R_2){\lesssim}\frac{{\delta}^2}{\theta(\pi_1,\pi_2)},$$ and the lemma follows.
\[miti2\] Let $\pi_1, {\pi_2}\in\mathrm{Gr}(n,2)$. Then $$|P_{\pi_1}\cap P_{\pi_2}|{\lesssim}\frac{{\delta}^{n}}{\theta(\pi_1,\pi_2)\underline{\theta}(\pi_1,\pi_2)}.$$
By translation we can suppose that the origin is contained in $P_{\pi_1}\cap P_{\pi_2}$. Let $x\in P_{\pi_1}\cap P_{\pi_2}$ and define $l\in\mathrm{Gr}(n,1)$ to be the line that passes through the origin and $x$. Define $l_1, l_2\in\mathrm{Gr}(n,1)$ to be the orthogonal projections of $l$ onto $\pi_1$ and $\pi_2$ respectively. Now by elementary geometry, $$\|\mathrm{proj}_{l}-\mathrm{proj}_{l_1}\|{\lesssim}{\delta}/|x|,$$ and $$\|\mathrm{proj}_{l}-\mathrm{proj}_{l_2}\|{\lesssim}{\delta}/|x|.$$ By the triangle inequality, $$\underline{\theta}(\pi_1,\pi_2)\le \|\mathrm{proj}_{l_1}-\mathrm{proj}_{l_2}\|{\lesssim}{\delta}/|x|,$$ so that $$|x|{\lesssim}{\delta}/\underline{\theta}(\pi_1,\pi_2).$$ If we denote the diameter of $P_{\pi_1}\cap P_{\pi_2}$ by $\alpha,$ then we see that $\alpha{\lesssim}{\delta}/\underline{\theta}(\pi_1,\pi_2)$.
As in the previous proof, $P_{\pi_1}\cap P_{\pi_2}$ is contained in $(R_1\cap R_2)\times R_3$, where $R_1$ and $R_2$ are 2–dimensional rectangles of dimension $1\times{\delta}$, and $R_3$ is an $(n-2)$–dimensional cuboid of dimension $\alpha\times\underbrace{{\delta}\times\cdots\times{\delta}}_{n-3}.$ We are able to reduce the length of the long side of $R_3$ as we have a bound on the diameter of $P_{\pi_1}\cap P_{\pi_2}$.
The rectangles $R_1$ and $R_2$ intersect at an angle ${\gtrsim}\theta(\pi,\pi')$, so by elementary geometry, $$\mathcal{L}^2(R_1\cap R_2){\lesssim}\frac{{\delta}^2}{\theta(\pi,\pi')},$$ and the lemma follows.
Finally we prove a quantitative version of the fact that three planes intersecting in distinct lines are contained in a 3–plane.
\[trouble\] Let $\pi_0,\pi_1,\pi_2\in \mathrm{Gr}(n,2),$ and define $\Sigma=\{x\in{\mathbb{R}}^n : d(x, S_{\pi_0})> \nu\},$ where ${\delta}<\nu<1$. Suppose that $P_{\pi_1},
P_{\pi_2}$ intersect $P_{\pi_0}$ at major angles $\sim 1$ and minor angles $< \phi$, where ${\delta}\le \phi \le 1,$ and suppose that $$P_{\pi_1}\cap P_{\pi_2}\cap \Sigma\neq\emptyset.$$ Then there is a 3–space ${\Lambda},$ chosen independently of $\pi_2$, such that $$\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_1)<\phi
\,\,\,\,\,\,\,\,
\mathrm{and}\,\,\,\,\,\,\,\,
\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_2){\lesssim}\phi/\nu.$$
By translation we can suppose that the origin is contained in the set $$\{ x : d(x,P_{\pi_0}\cap P_{\pi_1})<{\delta}\}\cap S_{\pi_0},$$ where $d(x,A)=\inf_{y\in A} |x-y|$. Now as $\underline{\theta}(\pi_0,\pi_1)<\phi$, there are lines $l_0,l_1\in \mathrm{Gr}(n,1)$ that are contained in $\pi_0$ and $\pi_1$ respectively, such that $$\|\mathrm{proj}_{l_0}-\mathrm{proj}_{l_1}\|<\phi.$$ Let $l_1'\in\mathrm{Gr}(n,1)$ be the line contained in $\pi_1$ that is orthogonal to $l_1$, and define ${\Lambda}$ to be the 3–space spanned by $\pi_0$ and $l_1'$.
Define $\pi_1'\in\mathrm{Gr}(n,2)$ to be the plane spanned by $l_0$ and $l_1'$, so that $\pi_1'$ is contained in ${\Lambda}$. Now by elementary geometry, we have $\theta(\pi_1,\pi_1')<\phi$ and $$\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_1)<\phi.$$
Define $T:{\mathbb{R}}^n\to{\mathbb{R}}^n$ to be a translation that maps a point in $$\{ x : d(x,P_{\pi_0}\cap P_{\pi_2})<{\delta}\}\cap S_{\pi_0}$$ to the origin. We note that ${\Lambda}$ and $\Sigma$ are essentially invariant under the action of $T,$ so that if $\zeta\in P_{\pi_1}\cap P_{\pi_2}\cap \Sigma$, then there exists a $\zeta'\in {\Lambda}\cap\Sigma$ such that $|T(\zeta)-\zeta'|{\lesssim}\phi$.
Now as $\underline{\theta}(\pi_0,\pi_2)<\phi$, there are lines $l_0',l_2\in \mathrm{Gr}(n,1)$ that are contained in $\pi_0$ and $\pi_2$ respectively, such that $$\|\mathrm{proj}_{l_0'}-\mathrm{proj}_{l_2}\|<\phi.$$ Define $\pi_2'\in\mathrm{Gr}(n,2)$ to be the plane that contains $l_0'$ and $\zeta',$ so that $\pi_2'$ is contained in ${\Lambda}$. Now by elementary geometry, we have $\theta(\pi_2,\pi_2'){\lesssim}\phi/\nu$ and $$\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_2){\lesssim}\phi/\nu,$$ as required.
The main argument {#oneandahalf}
=================
The [*(concave) triangle inequality*]{} states that when $p\le1$, $$\|\sum_{k=0}^N f_k\|_p^p \le\sum_{k=0}^{N} \|f_k\|_p^p,$$ and this will frequently take the following form: If $\|f_k\|_p^p{\lesssim}C$ for all $k$, then $$\|\sum_{k=0}^{\lfloor\log_2(1/{\delta})\rfloor} f_k\|_p^p{\lesssim}C.$$ Similarly, the [*pigeonhole principle*]{} will often take the form: If $$\sum_{k=0}^{\lfloor\log_2(1/{\delta})\rfloor} \|f_k\|{\gtrsim}C,$$ then for some $k,$ we have $\|f_{k}\| {\gtrsim}C.$
The following lemma will be key to the proof of Theorem \[dimension0\].
\[main0\] Let $\Pi$ be a ${\delta}$–separated subset of a 1–dimensional manifold ${\mathcal{M}}$ in $\mathrm{Gr}(3,2)$. Then $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{2}{\lesssim}1.$$
We note that as $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{2}^2=\|\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_
{P_\pi}\chi_{P_{\pi'}}\|_{1},$$ it will suffice to show $$\|\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}{\lesssim}1.$$ Now $$\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_{P_\pi}\chi_{P_{\pi'}}\le
\sum_{k=0}^{\lfloor\log_2(1/{\delta})\rfloor}\sum_{\pi,\pi':\theta(\pi,\pi')\sim2^{-k
}}\chi_{P_\pi}\chi_{P_{\pi'}}+\sum_{\pi\in\Pi}\chi_{P_\pi},$$ as $\Pi$ is ${\delta}$–separated. Thus, by the triangle inequality, it will suffice to show $$\label{dis0}
\|\sum_{\pi,\pi':\theta(\pi,\pi')\sim2^{-k}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}{\lesssim}1$$ for all $k$ less than or equal to $\lfloor\log_2(1/{\delta})\rfloor$, and $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{1}{\lesssim}1.$$ Again, by the triangle inequality, $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{1}\le \sum_{\pi\in\Pi}|P_\pi|{\lesssim}1,$$ as $\#\Pi{\lesssim}{\delta}^{-1}$ and $|P_\pi|={\delta}.$ Thus it remains to show (\[dis0\]) for each $k$, which we now consider to be fixed.
Using the metric $\theta$ on the 1–dimensional manifold $\mathcal{M}$, we can cover $\Pi$ by a constant multiple of $2^k$ balls $\{B_j\}$, with radius a constant multiple of $2^{-k}.$ We can also choose the cover so that if $\theta(\pi,\pi')\sim2^{-k}$, then $\pi$ and $\pi'$ are both contained in some $B_j$. Hence, by the triangle inequality, it will suffice to prove $$\label{roar0} \|\sum_{\substack{\pi,\pi'\in
B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}{\lesssim}2^{-k}$$ for all $k$ less than or equal to $\lfloor\log_2(1/{\delta})\rfloor$, and each ball $B_j$.
Without loss of generality, we can suppose that $B_j$ is centered on the $x_1x_2$–plane. Define the dilation $L:{\mathbb{R}}^3\to {\mathbb{R}}^3$ by $$L(x_1,x_2,x_3)=(x_1,x_2,2^kx_3).$$ We scale our geometric configuration by $L$, so that $$\begin{aligned}
\label{meek0}
\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}&=2^{-k}
\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{L(P_\pi)}\chi_{L(P_{\pi'})}\|_{1}.
\end{aligned}$$ Now as we have essentially changed ${\delta}$ to $2^k{\delta}$, and $\theta(L(\pi),L(\pi'))\sim 1,$ if we can prove (\[roar0\]) when $k=0$, then $$\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{L(P_\pi)}\chi_{L(P_{\pi'})}\|_{1}
\le
C_\epsilon(2^k{\delta})^{-\epsilon}< C_\epsilon{\delta}^{-\epsilon}{\lesssim}1,$$ so that by (\[meek0\]), $$\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}{\lesssim}2^{-k}.$$ Thus it will suffice to prove (\[roar0\]) when $k=0$.
Now $$\|\sum_{ \theta(\pi,\pi')\sim 1}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{1}\le
\sum_{ \theta(\pi,\pi')\sim 1}|P_\pi\cap P_{\pi'}|{\lesssim}(\#\Pi)^2{\delta}^{2}{\lesssim}1,$$ by Lemma \[miti\], and we are done.
The following lemma will be key to the proof of Theorem \[dimension\].
\[main\] Let $\Pi$ be a ${\delta}$–separated subset of a 2–dimensional curved manifold ${\mathcal{M}}$ in $\mathrm{Gr}(4,2)$. Then $$\label{mainy}
\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{5/3}{\lesssim}{\delta}^{-1/5}.$$
The proof is based on the ideas of Wolff [@wo]. The key geometric fact used there, is that three lines intersecting in distinct points are contained in a plane. The corresponding fact here, is that three planes intersecting in distinct lines are contained in a 3–plane. Unfortunately, the intersection between two planes can be a point as well as a line.
In order to deal with the different types of intersection, we use the bilinear reduction of T. Tao, A. Vargas and L. Vega [@ta2], which can also be found in [@ta1]. This enables us to quantify, using the minor angle, how near the planes are to intersecting in lines. When the planes are not intersecting in lines, we are able to use Lemma \[miti2\], in place of Lemma \[miti\], in compensation.
We make the bilinear reduction. Essentially this means we will begin by attempting to copy the proof of Lemma \[main0\]. We note that as $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{5/3}^{5/3}=\|\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_
{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6},$$ it will suffice to show $$\|\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}{\lesssim}{\delta}^{-1/3}.$$ Now $$\sum_{\pi\in\Pi}\sum_{\pi'\in\Pi}\chi_{P_\pi}\chi_{P_{\pi'}}\le
\sum_{k=0}^{\lfloor\log_2(1/{\delta})\rfloor}\sum_{\pi,\pi':\theta(\pi,\pi')\sim2^{-k
}}\chi_{P_\pi}\chi_{P_{\pi'}}+\sum_{\pi\in\Pi}\chi_{P_\pi},$$ as $\Pi$ is ${\delta}$–separated. Thus, by the triangle inequality, it will suffice to show $$\label{dis}
\|\sum_{\pi,\pi':\theta(\pi,\pi')\sim2^{-k}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^
{5/6}{\lesssim}{\delta}^{-1/3}$$ for all $k$ less than or equal to $\lfloor\log_2(1/{\delta})\rfloor$, and $$\label{dis2}
\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{5/6}^{5/6}{\lesssim}{\delta}^{-1/3}.$$ Again by the triangle inequality, $$\|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{5/6}^{5/6}\le \|\sum_{\pi\in\Pi}\chi_{P_\pi}\|_{1}\le \sum_{\pi\in\Pi}|P_\pi|{\lesssim}1<{\delta}^{-1/3},$$ as $\#\Pi{\lesssim}{\delta}^{-2}$ and $|P_\pi|={\delta}^{2}$, so it remains to show (\[dis\]) for each $k$, which we now consider to be fixed.
Using the metric $\theta$ on the 2–dimensional manifold $\mathcal{M}$, we can cover $\Pi$ by a constant multiple of $2^{2k}$ balls $\{B_j\}$, with radius a constant multiple of $2^{-k}.$ We can also choose the cover so that if $\theta(\pi,\pi')\le2^{-k}$, then $\pi$ and $\pi'$ are both contained in some $B_j$. Hence, by the triangle inequality, it will suffice to prove $$\label{roar} \|\sum_{\substack{\pi,\pi'\in
B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}{\lesssim}2^{-2k}{\delta}^{-1/3}$$ for all $k$ less than or equal to $\lfloor\log_2(1/{\delta})\rfloor$, and each ball $B_j$.
Without loss of generality, we can suppose that $B_j$ is centered on the $x_1x_2$–plane. Define the dilation $L:{\mathbb{R}}^4\to {\mathbb{R}}^4$ by $$L(x_1,x_2,x_3,x_4)=(x_1,x_2,2^kx_3,2^kx_4).$$ As $\det L=2^{2k}$, if we scale our geometric configuration by $L$, then $$\begin{aligned}
\label{belle}
\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}&=2^{-2k}
\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{L(P_\pi)}\chi_{L(P_{\pi'})}\|_{5/6}^{5/6}.
\end{aligned}$$
Essentially we have changed ${\delta}$ to $2^k{\delta},$ and $\theta(L(\pi),L(\pi'))\sim 1,$ so that if we can prove (\[roar\]) when $k=0$, then $$\label{sebastian}
\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{L(P_\pi)}\chi_{L(P_{\pi'})}\|_{5/6}^{5/6}
\le
C_\epsilon(2^k{\delta})^{-1/3-\epsilon}< C_\epsilon{\delta}^{-1/3-\epsilon}{\lesssim}{\delta}^{-1/3}.$$ Now by combining (\[belle\]) and (\[sebastian\]), $$\|\sum_{\substack{\pi,\pi'\in B_j\cap\Pi\\
\theta(\pi,\pi')\sim2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}{\lesssim}2^{-2k}{\delta}^{-1/3},$$ so it will suffice to prove (\[roar\]) when $k=0$.
Now as $$\sum_{\pi,\pi'}\chi_{P_\pi}\chi_{P_{\pi'}}\le
\sum_{k=0}^{\lfloor\log_2(1/{\delta})\rfloor}\sum_{\pi,\pi':{\underline{\theta}}(\pi,\pi')\sim2^{-k}}
\chi_{P_\pi}\chi_{P_{\pi'}}+\sum_{\pi,\pi':{\underline{\theta}}(\pi,\pi')<
{\delta}}\chi_{P_\pi}\chi_{P_{\pi'}},$$ by the triangle inequality, it will suffice to show $$\label{points} \|\sum_{\substack{\theta(\pi,\pi')\sim1\\ {\underline{\theta}}(\pi,\pi')\sim
2^{-k}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}{\lesssim}{\delta}^{-1/3}$$ for all $k$ less than or equal to $\lfloor\log_2(1/{\delta})\rfloor,$ and $$\label{lines}
\|\sum_{\substack{\theta(\pi,\pi')\sim 1\\ {\underline{\theta}}(\pi,\pi')< {\delta}}}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}{\lesssim}{\delta}^{-1/3}.$$ In (\[points\]) the planes are only intersecting in points, and in (\[lines\]) the planes are almost intersecting in lines.
To prove (\[points\]), we fix $k$ and define $F$ by $$F=\{x\in {\mathbb{R}}^4:\sum_{\pi,\pi'\in\Pi:(*)}\chi_{P_\pi}\chi_{P_{\pi'}}(x)\sim \lambda\},$$ where $(*)$ denotes the conditions $\theta(\pi,\pi')\sim 1$ and ${\underline{\theta}}(\pi,\pi')\sim 2^{-k}$. We have $$\|\sum_{\pi,\pi'\in\Pi:(*)}\chi_{P_\pi}\chi_{P_{\pi'}}\|_{5/6}^{5/6}
{\lesssim}\sum_{\lambda} \lambda^{5/6}|F|,$$ where $\lambda$ ranges dyadically up to a constant multiple of ${\delta}^{-4}$. Thus, by the triangle inequality, it will suffice to show the weak type inequality $$\label{res}
\lambda^{5/6}|F|{\lesssim}{\delta}^{-1/3}.$$ We can assume that $|F|$ is greater than ${\delta}^3,$ as otherwise we are done.
Define $\Pi_{\nu}$ by $$\Pi_{\nu}=\{\pi\in\Pi: |P_\pi\cap F |\sim {\nu}| P_\pi|\}.$$ We will use the pigeonhole principle to find a single plate that intersects many other plates that have density ${\nu}{\gtrsim}\sqrt{\lambda} |F|.$
By definition, $$\int_{F}\sum_{\pi, \pi'\in\Pi:(*)}\chi_{P_\pi}\chi_{P_{\pi'}} {\gtrsim}\lambda|F|,$$ so that $$\sum_{{\nu}, {\nu}'}\sum_{\pi\in\Pi_{\nu}}\sum_{ \pi'\in\Pi_{{\nu}'}:(*)} |P_\pi\cap P_{\pi'}\cap
F|=\int_{F}\sum_{\pi,\pi\in\Pi:(*)}\chi_{P_\pi}\chi_{P_{\pi'}}{\gtrsim}\lambda|F|{\gtrsim}{\delta}^3,$$ where the sums over ${\nu}$ and ${\nu}'$ range dyadically from zero to one. The summands where ${\nu}$ or ${\nu}'$ is less than a large power of ${\delta}$ can be absorbed by the larger summands. Thus, by the pigeonhole principle, we can find ${\nu}$ and ${\nu}'$, which we now fix, such that $$\label{pig}
\int_{F}\sum_{\pi\in\Pi_{\nu}}\sum_{\pi'\in\Pi_{{\nu}'}:(*)}
\chi_{P_\pi}\chi_{P_{\pi'}} {\gtrsim}\lambda|F|.$$ Let $$F'=\{x\in F:\sum_{\pi\in\Pi_{\nu}}\sum_{\pi'\in\Pi_{{\nu}'}:(*)}
\chi_{P_\pi}\chi_{P_{\pi'}}{\gtrsim}\lambda\},$$ so that by (\[pig\]) and the pigeonhole principle, $$\int_{F'}\sum_{\pi\in\Pi_{\nu}}\sum_{\pi'\in\Pi_{{\nu}'}:(*)}
\chi_{P_\pi}\chi_{P_{\pi'}} {\gtrsim}\lambda|F|.$$ By definition, $$\sum_{\pi\in\Pi_{{\nu}}}\sum_{\pi'\in\Pi_{{\nu}'}:(*)}\chi_{P_\pi}\chi_{P_{\pi'}}\le\lambda$$ on $F$, so that $|F'|{\gtrsim}|F|.$ Now as $$\sum_{\pi\in\Pi_{\nu}}\sum_{\pi'\in\Pi_{{\nu}'}:(*)}
\chi_{P_\pi}\chi_{P_{\pi'}}{\gtrsim}\lambda$$ on $F'$, we can suppose that $$\sum_{\pi\in\Pi_{\nu}}\chi_{P_\pi}{\gtrsim}\sqrt{\lambda}$$ on some $F''\subset F'$, where $|F''|{\gtrsim}|F'|.$ Thus $$\int_{F}\sum_{\pi\in\Pi_{\nu}}\chi_{P_\pi}\ge \int_{F''}\sum_{\pi\in\Pi_{\nu}}\chi_{P_\pi}
{\gtrsim}\sqrt{\lambda}|F''|{\gtrsim}\sqrt{\lambda}|F|.$$ By definition, ${\nu}{\delta}^2{\gtrsim}|P_\pi\cap F|$ for all $\pi\in\Pi_{{\nu}},$ so that $$\sum_{\pi\in\Pi_{\nu}}{\nu}{\delta}^2{\gtrsim}\sum_{\pi\in\Pi_{\nu}}|P_\pi\cap F|=\int_{F}\sum_{\pi\in\Pi_{\nu}}\chi_{P_\pi}{\gtrsim}\sqrt{\lambda}|F|.$$ Now as $\#\Pi_{\nu}{\lesssim}{\delta}^{-2}$, we have ${\nu}{\gtrsim}\sqrt{\lambda}|F|.$
From (\[pig\]), we also have $$\sum_{\pi'\in\Pi_{{\nu}'}}\int_{P_{\pi'}}\sum_{\pi\in\Pi_{{\nu}}:(*)} \chi_{P_\pi} {\gtrsim}\lambda|F|,$$ so that, there exists a $\pi_0\in\Pi_{{\nu}'}$, such that $$\label{pigs}
\int_{P_{\pi_0}}\sum_{\pi\in\Pi_{{\nu}}:(*)} \chi_{P_\pi} {\gtrsim}{\delta}^{2}\lambda|F|,$$ where $(*)$ denotes the conditions $\theta(\pi,\pi_0)\sim 1$ and ${\underline{\theta}}(\pi,\pi_0)\sim 2^{-k}$. By Lemma \[miti2\], $$\label{pigss}
\int_{P_{\pi_0}} \chi_{P_\pi}=|P_\pi\cap P_{\pi_0}|{\lesssim}2^{k}{\delta}^{4},$$ so that if we define $\mathcal{P}$ by $$\mathcal{P}=\{\pi\in\Pi_{\nu}: \theta(\pi,\pi_0)\sim1,\,\,{\underline{\theta}}(\pi,\pi_0)\sim 2^{-k},\,\,P_\pi\cap P_{\pi_0}\neq
\emptyset \},$$ then by (\[pigs\]), we have $$\label{argh2}
\#\mathcal{P}{\gtrsim}2^{-k}{\delta}^{-2}\lambda|F|.$$ Thus we have a large set of planes, with density ${\nu}{\gtrsim}\sqrt{\lambda} |F|,$ that intersect $\pi_0$.
As $\theta(\pi,\pi_0)\sim 1$ for all $\pi\in{\mathcal{P}}$, we can define $\Sigma=\{x\in{\mathbb{R}}^4 : d(x,S_{\pi_0})>{\nu}/C\}$ for some sufficiently large constant $C$, so that $$\int_{F}\chi_{P_\pi\cap\Sigma}{\gtrsim}{\nu}{\delta}^{2}.$$ Summing over ${\mathcal{P}}$, we have $$\int_{F}\sum_{\pi\in{\mathcal{P}}}\chi_{P_\pi\cap\Sigma}{\gtrsim}{\nu}{\delta}^{2}\#{\mathcal{P}},$$ so that by the Cauchy–Schwarz inequality, $$\label{argh}
\|\sum_{\pi\in{\mathcal{P}}}\chi_{P_\pi\cap\Sigma}\|_2{\gtrsim}\frac{{\nu}{\delta}^{2}\#{\mathcal{P}}}{|F|^{1/2}}.$$
We will use the geometry of the construction to bound the left hand side of (\[argh\]) from above, in order to obtain (\[res\]). If we fix a $\pi_1\in{\mathcal{P}}$, then by Lemma \[trouble\], there exists a 3–space ${\Lambda}$, such that $$\label{mere}
\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_1)<2^{-k},$$ and if $\pi_2\in \mathcal{P}$ and $P_{\pi_1}\cap P_{\pi_2}\cap \Sigma\neq\emptyset,$ then $$\label{mere2}
\pi_2\in \mathcal{M}^{{\nu}^{-1}2^{-k}}_\Lambda,$$ where $\mathcal{M}^{{\nu}^{-1}2^{-k}}_\Lambda=\{\pi'\in\mathcal{M}:\min_{\pi\subset{\Lambda}}\theta(\pi,\pi')<{\nu}^{-1}2^{-k}\}.$ Now, as $\mathcal{M}$ is curved, there exists a constant $C$ such that $$\label{pope2}
\frac{\mu(\mathcal{M}^{\epsilon_1}_{{\Lambda}}\cap
B_{\epsilon_2}(\pi))}{\epsilon_1\epsilon_2}<C$$ for all $\epsilon_1,\epsilon_2>0$, 3–spaces $\Lambda$, and $\pi\in\mathcal{M}.$ Thus, as ${\mathcal{P}}$ is ${\delta}$–separated, we see that $$\#\{ \pi_2\in{\mathcal{P}}:\theta(\pi_1,\pi_2)\le 2^{-j},\,\,\, P_{\pi_1}\cap P_{\pi_2}\cap
\Sigma\neq\emptyset\}{\lesssim}{\nu}^{-1}2^{-k}2^{-j}{\delta}^{-2},$$ where we take $j$ to be less than or equal to $\lfloor\log_2(1/{\delta})\rfloor$. By Lemma \[miti\], $$|P_{\pi_1}\cap P_{\pi_2}\cap\Sigma|{\lesssim}2^j{\delta}^{3}$$ when $\theta(\pi_1,\pi_2)\sim 2^{-j}$, so that $$\sum_{\pi_2\in{\mathcal{P}}: \theta(\pi_1,\pi_2)\sim 2^{-j}}|P_{\pi_1}\cap P_{\pi_2}\cap \Sigma|{\lesssim}{\nu}^{-1}2^{-k}{\delta}.$$ Hence, by the triangle inequality, $$\sum_{\pi_2\in{\mathcal{P}}}|P_{\pi_1}\cap P_{\pi_2}\cap \Sigma|{\lesssim}{\nu}^{-1}2^{-k}{\delta},$$ so that by summing over ${\mathcal{P}}$, we have $$\sum_{\pi_1\in{\mathcal{P}}}\sum_{\pi_2\in{\mathcal{P}}}|P_{\pi_1}\cap P_{\pi_2}\cap \Sigma|{\lesssim}{\nu}^{-1}2^{-k}{\delta}\#{\mathcal{P}}.$$ Now $$\| \sum_{\pi\in{\mathcal{P}}}\chi_{P_{\pi}\cap \Sigma}\|^2_2=\sum_{\pi_1\in{\mathcal{P}}}\sum_{\pi_2\in{\mathcal{P}}}|P_{\pi_1}\cap P_{\pi_2}\cap \Sigma|,$$ so that $$\label{second}
\| \sum_{\pi\in{\mathcal{P}}}\chi_{P_{\pi}\cap \Sigma}\|_2{\lesssim}({\nu}^{-1}2^{-k}{\delta}\#{\mathcal{P}})^{1/2}.$$
Combining the equations (\[argh\]) and (\[second\]), and the fact that $\#\mathcal{P}{\gtrsim}2^{-k}{\delta}^{-2}\lambda|F|$, we obtain $${\nu}^3\lambda {\lesssim}{\delta}^{-1}.$$ Using the fact that $\nu{\gtrsim}\sqrt{\lambda}|F|$, we have $$\lambda^{5/2}|F|^{3}{\lesssim}{\delta}^{-1},$$ and we take the third root to obtain (\[res\]) as required.
To prove (\[lines\]), we argue in the same way. We let $(*)$ denote the conditions $\theta(\pi,\pi')\sim 1$ and ${\underline{\theta}}(\pi,\pi')<{\delta}$. We apply Lemma \[miti\] in place of Lemma \[miti2\], so that the estimate (\[pigss\]) becomes $$\int_{P_{\pi_0}} \chi_{P_\pi}=|P_\pi\cap P_{\pi_0}|{\lesssim}{\delta}^{3},$$ and $\#\mathcal{P}{\gtrsim}{\delta}^{-1}\lambda|F|.$ The expressiones in (\[mere\]) and (\[mere2\]) are changed to $$\min_{\pi\subset{\Lambda}} \theta(\pi,\pi_1)<{\delta}\,\,\,\,\,\mathrm{ and }\,\,\,\,\,
\pi_2\in \mathcal{M}^{{\nu}^{-1}{\delta}}_\Lambda,$$ so that $$\| \sum_{\pi\in{\mathcal{P}}}\chi_{P_{\pi}\cap \Sigma}\|_2{\lesssim}({\nu}^{-1}{\delta}^{2}\#{\mathcal{P}})^{1/2}.$$ As before, we combine these equations with (\[argh\]), which is unchanged, so that $$\lambda^{5/2}|F|^{3}{\lesssim}{\delta}^{-1},$$ and we are done.
Proof of Theorems \[dimension0\] and \[dimension\] {#two}
==================================================
The following argument is well known, and can be found in [@bo]. Let $\{B(x_j,r_j)\}$ be a covering of a $\mathcal{K}^n_2$-set $E$, where $r_j\le1/4$. We are required to show $$\sum_j r_j^{\frac{n+3}{2}-\epsilon}\ge C_\epsilon>0$$ for all $\epsilon>0$.
Define $E_k$ by $$E_k=E\cap\bigcup_{r_j\sim2^{-k}}B(x_j,r_j)$$ for $k\ge 2$. As $E$ is a $\mathcal{K}^n_2$-set, for all $\pi\in\mathcal{M}$ there is a square $S_\pi\subset E$ with a corresponding plate $P_\pi$. By the pigeonhole principle, for all $\pi\in\mathcal{M}$ there is a $k_\pi$ such that $$\mathcal{L}^2(S_\pi\cap E_{k_\pi})\ge \frac{1}{k_{\pi}^2}{\gtrsim}1.$$ Thus $\mathcal{M}=\bigcup_{k\ge 2}\mathcal{M}_k,$ where $$\mathcal{M}_k=\{\pi\in\mathcal{M}:\mathcal{L}^2(S_\pi\cap E_k)\ge \frac{1}{k^2}\}.$$ By the pigeonhole principle again, there exists a $k_0$ such that $$\mu(\mathcal{M}_{k_0})\ge \frac{1}{k_0^2}{\gtrsim}1,$$ where $\mu$ is the induced measure on $\mathcal{M}$, normalised to have total mass 1.
Let $\delta=2^{-k_0},$ and let $\Pi$ be a maximal ${\delta}$–separated subset of $\mathcal{M}_{k_0},$ so that $\#\Pi{\gtrsim}{\delta}^{2-n}$. Define $
J=\{ j : r_j\sim {\delta}\}
$ and $E_{\delta}=\bigcup_{j\in J}B(x_j,2{\delta}),$ so that $$\int_{E_{\delta}}\chi_{P_\pi}{\gtrsim}{\delta}^{n-2}\mathcal{L}^2(S_\pi\cap E_{k_0})\ge\frac{{\delta}^{n-2}}{k_0^2}{\gtrsim}{\delta}^{n-2}.$$ Now as $\#\Pi{\gtrsim}{\delta}^{2-n}$, we have $$\sum_{\pi\in \Pi} \int_{E_{\delta}}\chi_{P_\pi}{\gtrsim}1,$$ so that by Hölder’s inequality, $$|E_{\delta}|^{\frac{2}{n+1}}\|\sum_{\pi\in \Pi}\chi_{P_\pi}\|_{\frac{n+1}{n-1}}{\gtrsim}1.$$ By Lemmas \[main0\] and \[main\], we have $$\|\sum_{\pi\in \Pi}\chi_{P_\pi}\|_{\frac{n+1}{n-1}}{\lesssim}{\delta}^{\frac{3-n}{n+1}},$$ where $n=3$ or $4$, so that $$|E_{\delta}|{\gtrsim}{\delta}^{\frac{n-3}{2}}.$$
On the other hand, we have $\#J{\delta}^n{\gtrsim}|E_{\delta}|$, so that $$\#J\ge C_\epsilon{\delta}^{-\frac{n+3}{2}+\epsilon}$$ for all $\epsilon>0$. Hence, when $n=3$ or $4$, $$\sum_j r_j^{\frac{n+3}{2}-\epsilon}\ge \#J \left(\frac{{\delta}}{2}\right)^{\frac{n+3}{2}-\epsilon}\ge C_\epsilon'>0$$ for all $\epsilon>0$, and we are done. $\square$
Sharpness in the complex case {#three}
=============================
We construct an example, inspired by the Heisenberg group example in [@ka], to show that Theorem \[cork\] is sharp. Define $E\subset{\mathbb{C}}^4$ by $$E=\{(z_1,z_2,z_3,z_4)\in{\mathbb{C}}^4 :
\mathrm{Im}(z_1\overline{z_2})=\mathrm{Im}(z_3\overline{z_4})\},$$ so that $\dim_H(E)=7$. Define the planes $\pi_{u,v}$ by $$\pi_{u,v}=
\{(1,-|v|\mathrm{Im}(u)i,u,|v|)z+(1,|v|\mathrm{Im}(u)i,\frac{\overline{u}v}
{|v|},v)z':z,z'\in{\mathbb{C}}\},$$ and the manifold $\mathcal{M}$ by $$\mathcal{M}=\{\pi_{u,v}: \,u,v\in{\mathbb{C}}, \,\,\mathrm{Im}(u)\neq0,\,\,\mathrm{Im}(v)\neq0\}.$$ It is not hard to calculate that the planes are contained in $E$, and it is clear that $\mathcal{M}$ is a 2–dimensional manifold as a subset of $\mathrm{Gr}(4,2)$ over ${\mathbb{C}}$.
It remains to show that $\mathcal{M}$ is curved. It will suffice to show that $\{\pi\in\mathcal{M} :
\pi\subset\Lambda\}$ is no more than 1–dimensional for all 3–spaces $\Lambda$. Now if a plane in $\mathcal{M}$ is contained in a 3–space $\Lambda,$ then $$\begin{aligned}
(1,-|v|\mathrm{Im}(u)i,u,|v|)\cdot(a_1,a_2,a_3,a_4)=a_1-a_2|v|
\mathrm{Im}(u)i+a_3u+a_4|v|=0 \end{aligned}$$ and $$(1,|v|\mathrm{Im}(u)i,\frac{\overline{u}v}{|v|},v)\cdot(a_1,a_2,a_3,a_4)=a_1+a_2
|v|\mathrm{Im}(u)i+a_3\frac{\overline{u}v}{|v|}+a_4v=0$$ for some normal $(a_1,a_2,a_3,a_4)\in{\mathbb{C}}^4$. We can multiply the second equation by $|v|/v$ and subtract it from the first to solve for $\mathrm{Im}(u)$ in terms of $v$. Substituting back into the first equation, we fix $u$ in terms of $v$, so that the set of planes contained in $\Lambda$ is parametrized by a single variable. Thus the restriction of $\mathcal{M}$ to a 3–space is no more than 1–dimensional, and $E$ is a $\mathcal{K}^4_2$-set.
Thus, in the complex case, the curvature condition is not sufficient to guarantee nontrivial intersection, even before translation. This example does not extend to the reals (or the finite fields, as there is no square root), and the curvature condition is stronger over ${\mathbb{R}}$ than over ${\mathbb{C}}$. It seems possible that the real and complex cases are qualitatively different. Thus the problem of sharp lower bounds for the Hausdorff dimension of real $\mathcal{K}^n_2$-sets is open and interesting for $n\ge 4.$
Final remarks
=============
We could define the $\mathcal{K}^n_{2}$-sets so that they only contain a unit square parallel to each direction plane, and not necessarily the whole plane.
Theorems \[dimension0\] and \[dimension\] would extend in a natural way to $\mathcal{K}^n_{k}$-sets, where $k=n-1$ or $n-2$, and the planes are replaced by $k$–planes. We note that the $\mathcal{K}^4_{2}$-set over ${\mathbb{C}}$ of the previous section may not be a $\mathcal{K}^8_{4}$-set over ${\mathbb{R}}$, as the curvature condition is stronger over ${\mathbb{R}}$.
It also seems likely that we could adapt the proofs to obtain the corresponding maximal function estimates. We neglect these potential generalizations mainly for expository purposes.
[^1]
[99]{}
A. Besicovitch, On Kakeya’s problem and a similar one, [*Mat. Zeit.*]{} [**27**]{} (1928), 312–320. J. Bourgain, Besicovitch type maximal operators and applications to Fourier analysis, [*Geom. Funct. Anal.*]{} [**1**]{} (1991), no. 2, 147–187. J. Bourgain, On the dimension of Kakeya sets and related maximal inequalities, [*Geom. Funct. Anal.*]{} [**9**]{} (1999), no. 2, 256–282. A. Córdoba, The Kakeya maximal function and the spherical summation multipliers, [*Amer. J. Math.*]{} [ **99**]{} (1977), no. 1, 1–22. N. H. Katz and T. Tao, Bounds on arithmetic projections, and applications to the Kakeya conjecture, [*Math. Res. Lett.*]{} [**6**]{} (1999), no. 5-6, 625–630. N. H. Katz, I. Łaba and T. Tao, An improved bound on the Minkowski dimension of Besicovitch sets in ${\bf R}\sp 3$, [*Ann. of Math.(2)*]{} [**152**]{} (2000), no. 2, 383–446. N. H. Katz and T. Tao, New bounds for Kakeya problems, [*J. Anal. Math.*]{} [**87**]{} (2002), 231–263. I. Łaba, T. Tao, An improved bound for the Minkowski dimension of Besicovitch sets in medium dimension, to appear. T. Mitsis, $(n,2)$-sets have full Hausdorff dimension, [*Rev. Mat. Iberoamericana*]{}, [**20**]{} (2004), no. 2, 381-393. G. Mockenhaupt and T. Tao, Restriction and Kakeya phenomena for finite fields, [*Duke Math. J.*]{} [**121**]{} (2004), no. 1, 35–74. T. Tao, A new bound for finite field Besicovitch sets in four dimensions, to appear in [*Pacific J. Math.*]{} T. Tao, Restriction theorems and applications, available at [ www.math.ucla.edu/ tao]{}. T. Tao, A. Vargas and L. Vega, A bilinear approach to the restriction and Kakeya conjectures, [*J. Amer. Math. Soc.*]{} [**11**]{} (1998), no. 4, 967–1000. T. Wolff, An improved bound for Kakeya type maximal functions, [*Rev. Mat. Iberoamericana*]{} [**11**]{} (1995), no. 3, 651–674.
[^1]: Many thanks to Gerd Mockenhaupt for suggesting a similar planar variant problem which led to the consideration of this one. Thanks to Peter Sjögren, Thomas Duyckaerts, Fulvio Ricci, Ana Vargas and Toby Bailey for helpful conversations, and thanks to an anonymous referee for pointing out a mistake in an earlier version.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let $X$ be a smooth complex projective curve of genus $g\geq 2$. In this article, we prove that a stable tame parabolic vector bundle ${\mathcal{E}}$ on $X$ is very stable, i.e. ${\mathcal{E}}$ has no non-zero nilpotent Higgs field preserving the parabolic structure, if and only if the restriction of the Hitchin map to the space of parabolic Higgs fields with nilpotent residue is a proper map. The same result holds in the setup of strongly parabolic Higgs bundles. Both results follow from [@Hacen Lemma 1.3], once the image of the Hitchin map restricted to suitable leaves has been proven to be affine of the right dimension. This characterisation of wobbliness proves the equivalence between this notion and shakiness. We use this and the techniques in [@PP] to prove that shaky bundles are the image of the exceptional divisor of a suitable resolution of a certain rational map when the moduli space is smooth. As corollaries to the aforementioned results, we obtain equivalent ones for the moduli space of vector bundles.'
address: |
Ana Peón-Nieto\
Laboratoire de Mathématiques J.A. Dieudonné\
UMR 7351 CNRS\
Université de Nice Sophia-Antipolis\
06108 Nice Cedex 02, France
author:
- 'Ana Peón-Nieto'
title: |
Wobbly and shaky bundles\
and resolutions of rational maps
---
Introduction
============
Let $X$ be a smooth complex projective curve of genus $g\geq 2$ and let $K$ be its canonical bundle. Consider the moduli spaces ${\mathrm{N}}(n,d)$ of vector bundles of rank $n$ and degree $d$, and ${\mathrm{M}}(n,d)$ of Higgs bundles with the same invariants. These are schemes parametrising vector bundles (resp. Higgs bundles, namely pairs $(E,\varphi)$ where $E$ is a vector bundle and $\varphi\in H^0(X,{\mathrm{End}}(E)\otimes K)$ is the Higgs field) of fixed rank and degree. This article is concerned with two loci in ${\mathrm{N}}(n,d)$: the wobbly and the shaky loci. Wobbly bundles are stable bundles in ${\mathrm{N}}(n,d)$ admitting non-zero nilpotent Higgs fields. The name wobbly first appeared in [@DPLectures], but the interest in these objects (or their complement, very stable bundles) dates back to the works of Drinfeld and Laumon from the 80’s [@Drinfeld; @L]. Their study has recently experienced a new impulse, partially motivated by their role in the programme by Donagi–Pantev to prove the geometric Langlands correspondence from the abelianisation of Higgs bundles [@DPLectures; @DPHecke]. Examples of this are [@PP; @Hacen], where a criterion for very stability is proven, and [@PaP], that proves pure codimensionality of the wobbly locus in rank two, as conjectured by Drinfeld [@L Remark 3.6.ii]. The definition of a shaky bundle is more involved. To understand it, note that $T^*{\mathrm{N}}^s(n,d)\subset{\mathrm{M}}(n,d)$, where ${\mathrm{N}}^s(n,d)$ denotes the stable locus, is a dense open set. This defines a surjective rational map $\xymatrix{r:{\mathrm{M}}(n,d)\ar@{-->}[r]&{\mathrm{N}}(n,d),}$ defined beyond the locus of semistable Higgs bundles whose underlying bundle is unstable. Shaky bundles are stable bundles in ${\mathrm{N}}(n,d)$ over which the Hitchin map is not proper. The equality of the shaky and wobbly loci was conjectured in [@DPLectures], where they announce the rank two case, as well as the possibility to understand shaky loci in terms of the exceptional divisor of a suitable resolution of the rational map $r$. Donagi–Pantev observe in [*loc. cit.*]{} that the restriction of $r$ to generic Hitchin fibers can be resolved, and the image of the exceptional divisor (a priori dependant on the fiber and the resolution) should be the shaky locus. In [@DPHecke], they work out the case of tame nilpotent rank two parabolic Higgs bundles on the projective line minus five points, where equality indeed holds.
In [@PP], Pauly and the author proved the equality of the wobbly and shaky loci in the case of vector bundles. Zelaci [@Hacen] generalised it to principal bundles in an elegant way via a general theorem in commutative algebra. The advantage of the latter approach is that is allows for a more straightforward extension to other contexts (parabolic bundles, bundles with level structures, etc). The former in turn, gives a geometric construction that offers a strategy to see shaky bundles in terms of exceptional divisors of a resolution of $r$.
In this article we extend [@PP; @Hacen] to the parabolic bundle setup. Using the results in both these articles, and most crucially [@Hacen Lemma 1.3], we prove the equality of the wobbly and the shaky locus. It next follows from the techniques developed in [@PP] that in the smooth moduli space case, the shaky locus can in fact be characterised in terms of a resolution of $r$.
Structure of the paper and summary of results {#structure-of-the-paper-and-summary-of-results .unnumbered}
---------------------------------------------
Let $D\subset X$ be a reduced divisor, let $\mathcal{P}:=\{P_x:x\in D\}$ be a set of conjugacy classes of parabolic subgroups of ${\mathrm{GL}}(n,{\mathrm{C}})$ indexed by $D$, identified with partitions $\{{\overline{n}}_x=(n_{x,1},\dots, n_{x,r_x}),x\in D\}$ of the rank. A quasi-parabolic vector bundle of rank $n$ and flag type ${\mathcal{P}}$ is a pair ${\mathcal{E}}=(E,\sigma)$ where $E$ is a rank $n$ vector bundle and $\sigma:E|_D\longrightarrow \prod_{x\in D} {\mathrm{GL}}(n,{\mathbb{C}})/P_x$ is a reduction of the structure group to the given parabolic subgroup at each point of $D$. Equivalently, we may identify $\sigma$ with a set of flags $\{E_{x,0}=0\subseteq E_{x,1}\subsetneq\dots\subsetneq E_{x,r_x}=E_x\}_{x\in D}$. Let $\alpha_x=(\alpha_{x,1},\dots,\alpha_{x,r_x})$ be an increasing $r_x$-uple of negative real numbers. The assignment of the weights $\alpha_{x,i}$ to the steps $E_{x,i}$ of the flag defines a parabolic structure on the quasi-parabolic bundle ${\mathcal{E}}$. A parabolic Higgs bundle is a pair $({\mathcal{E}},\varphi)$ where ${\mathcal{E}}$ is a parabolic vector bundle and $\varphi\in H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))$, with ${\mathrm{ParEnd}}({\mathcal{E}})\subset{\mathrm{End}}(E)$ the subsheaf of endomorphism of $E$ preserving the flags at the prescribed points.
We may define a stability notion for each $\alpha$ as above, yielding moduli spaces of parabolic bundles ${\mathrm{N}}({\mathcal{P}},\alpha)$ and parabolic Higgs bundles ${\mathrm{M}}({\mathcal{P}},\alpha)$. The latter is Poisson with symplectic leaves given by Higgs fields whose associated graded endomorphism has fixed orbits at the residues. Particularly important are the strongly parabolic leaf (corresponding to the zero orbit), and the regular leaves (which are generic). Strongly parabolic Higgs bundles have Higgs fields belonging to the subsheaf ${\mathrm{SParEnd}}({\mathcal{E}})\subset{\mathrm{ParEnd}}({\mathcal{E}})$ of strongly parabolic endomorphisms.
Consider the Hitchin fibration $$h_{{\mathcal{P}},\alpha} : {\mathrm{M}}({\mathcal{P}},\alpha) \longrightarrow {\mathrm{H}}_D = \bigoplus_{i=1}^n {\mathrm{H}}^0(X, K^i(iD)),$$ defined by associating to a parabolic Higgs bundle $({\mathcal{E}}, \varphi)$ the coefficients of the characteristic polynomial of $\varphi$. If ${\mathcal{E}}$ is a stable parabolic vector bundle of type $({\mathcal{P}},\alpha)$, there is an embedding of ${\mathrm{H}}^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))$ into the moduli space ${\mathrm{M}}({\mathcal{P}},\alpha)$. Motivated by [@PP; @Hacen], we consider the restriction $h_{{\mathcal{E}},st}$ (respectively $h_{{\mathcal{E}},nilp}$) of the Hitchin map to the vector space $V_{{\mathcal{E}}, st}={\mathrm{H}}^0(X,{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D))$ of strongly parabolic Higgs fields (resp. $V_{{\mathcal{E}}, nilp}={\mathrm{H}}^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))_{nilp}$, of Higgs fields with nilpotent residue along the flag). If ${\mathcal{E}}$ is stable, this corresponds to the intersection of ${\mathrm{H}}^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))$ with the strongly parabolic symplectic leaf (resp., the closure of the regular nilpotent symplectic leaf). These and other notions are explained in Section \[sec:prel\].
In Section \[sec:criteria\] we define the notion of a strongly very stable parabolic bundle (respectively, very stable), that is, a parabolic bundle admitting no non-zero strongly parabolic Higgs field (resp. no nilpotent parabolic Higgs field). Paralleling the case of vector bundles, we find:
Let ${\mathcal{E}}$ be a stable parabolic bundle, and let $h_{{\mathcal{E}},nilp}$ ($h_{{\mathcal{E}},st}$) be the restriction of the Hitchin map to ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ (resp. ${{\mathrm{V}}_{{\mathcal{E}},st}}$). Then, $$\begin{array}{lcl}
{\mathcal{E}}\textnormal{ \it is (strongly) very stable } &\iff& h_{{\mathcal{E}},nilp} \,(h_{{\mathcal{E}},st})\textnormal{ \it is finite }\\
&\iff& h_{{\mathcal{E}},nilp}\,(h_{{\mathcal{E}},st}) \textnormal{ \it is quasi-finite } \\
&\iff& h_{{\mathcal{E}},nilp} \,(h_{{\mathcal{E}},st})\textnormal{ \it is proper }\\
&\iff& {{\mathrm{V}}_{{\mathcal{E}},nilp}}\,({{\mathrm{V}}_{{\mathcal{E}},st}}){\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha) \textnormal{ \it is proper.}
\end{array}$$
It is worth noting that finiteness follows from very stability as a corollary of [@Hacen Lemma 1.3]. Moreover, in the case of nilpotent Higgs fields, properness of $h_{{\mathcal{E}},nilp}$ is equivalent to properness of $h_{{\mathcal{E}}}:=h_{{\mathcal{P}},\alpha}|_{{{\mathrm{V}}_{{\mathcal{E}}}}}$ (cf. Proposition \[prop:vs\_iff\_proper\]).
An immediate corollary is the fact that if ${\mathcal{E}}$ is strongly very stable, then $V_{{\mathcal{E}},st}$ is a Lagrangian multisection of the restriction $h_{{\mathcal{P}},\alpha, st}$ of the Hitchin map (Corollary \[cor:lagrangian\_multi\]) to the strongly parabolic leaf. This was remarked in the usual setup in [@FGOP Corollary 7.3]. A similar result holds for $h_{{\mathcal{E}},nilp}$, except that in that case it will be Lagrangian for a symplectic leaf possibly of lower dimension than the regular nilpotent (Corollary \[cor:multisection\_nilp\]).
Section \[sec:WandShaky\] explains the relation between wobbliness and shakiness. From Propositions \[prop:st\_vs\_iff\_proper\] and \[prop:vs\_iff\_proper\] we find the following:
There is an equality between the wobbly locus and the shaky locus. The same holds for the strongly wobbly and the strongly shaky loci.
In the vector bundle setup, shakiness of wobbly bundles has been known since [@L; @BNR], and follows from similar arguments to those of [@BNR §4.1]. The proof of this fact can be adapted to the parabolic setup, with a bit of extra work for nilpotent parabolic Higgs bundles given the existence of different leaves and images of the Hitchin map, which is the new contribution of Proposition \[prop:vs\_iff\_proper\].
Section \[sec:W=S\] addresses the identification of shaky bundles with the image of the exceptional divisor under a resolution of $r$ obtained by blowup along the locus ${\mathrm{Un}}$ of stable parabolic Higgs bundles with underlying unstable parabolic bundle when the moduli space is regular. Let $\hat{r}$ be such a resolution, and let ${\mathrm{Ex}}$ be the exceptional divisor, and let ${\mathrm{Sh}}=\hat{r}({\mathrm{Ex}})$. By construction, $\hat{r}$ restricted to ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$ and ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ is again a succession of blowups, and we denote by ${\mathrm{Sh}_{nilp}}$ and ${\mathrm{Sh}}_{st}$ the images of the exceptional divisors. Then:
There is an equality ${\mathrm{Wo}}={\mathrm{Sh}_{nilp}}={\mathrm{Sh}}$, where ${\mathrm{Wo}}\subset{\mathrm{N}}({\mathcal{P}},\alpha)$ denotes the wobbly locus.
Similarly, the strong wobbly locus ${\mathrm{Wo}}_{st}$ satisfies ${\mathrm{Wo}}_{st}={\mathrm{Sh}}_{st}$.
This was proven in [@DPHecke] for parabolic bundles on the projective line minus five points. In fact, these auhtors work on the non smooth case, so shakiness is defined with respect to a specific point of the Hitchin base. They prove that the definition is independent of this choice.
Finally, Section \[sec:classical\] explains how these results yield equivalent ones in the case of non-parabolic vector bundles.
### Acknowledgments {#acknowledgments .unnumbered}
The author would like to thank D. Alfaya, R. Donagi, P.B. Gothen, T. Hausel, J. Martens, T. Pantev, C. Pauly, C. Simpson, A. Zamora and H. Zelaci for useful discussions on these and related questions. Many thanks to J.C. Díaz-Ramos, M. Domínguez-Vázquez and E. García-Río for their hospitality during the author’s stay at Universidade de Santiago de Compostela, where part of this article was written.
Parabolic Higgs bundles {#sec:prel}
=======================
We gather in this section some results that will be useful.
Parabolic bundles and parabolic Higgs bundles
---------------------------------------------
Let $D\subset X$ be a reduced divisor of degree $d$. Let $\mathcal{P}:=\{P_x:x\in D\}$ be a set of parabolic subgroups of ${\mathrm{GL}}(n,{\mathrm{C}})$ indexed by $D$. Denote by $L_x<P_x$ the Levi subgroup, and let $\mathfrak{l}_x:=\mathrm{Lie}(L_x)<\mathfrak{p}_x:=\mathrm{Lie}(P_x)$ be the respective Lie algebras. We will denote by ${\mathcal{O}}=\prod_{x\in D}{\mathcal{O}}_x$ an orbit of $\mathfrak{l}:=\prod_{x\in D}\mathfrak{l}_x$ by conjugation by $L:=\prod_{x\in D} L_x$ (so that ${\mathcal{O}}_x$ is a conjugacy orbit of $\mathfrak{l}_x$). We define $\mathfrak{n}:=\prod_{x\in D}\mathfrak{n}_x<\mathfrak{p}:=\prod_{x\in D}\mathfrak{p}_x$ the nilpotent subalgebra.
A quasi-parabolic bundle of flag type type ${\mathcal{P}}$ (note that the rank of the bundle is determined by ${\mathcal{P}}$, as the subgroups are taken inside a fixed ${\mathrm{GL}}(n,{\mathbb{C}})$) is a pair ${\mathcal{E}}=(E,\sigma)$ where $E$ is a rank $n$ vector bundle and $\sigma:E|_D\longrightarrow \prod_{x\in D} {\mathrm{GL}}(n,{\mathrm{C}})/P_x$ is a reduction of the structure group to the given parabolic at each point of $D$. Without loss of generality, we will assume that $P_x$ is associated to the partition $n=\sum_{i=1}^{r_x}n_{x,i}$, and we will identify $\sigma$ with a collection of flags $\{E_{x,\bullet}:0=E_{x,0}\subset E_{x,1}\subset\dots\subset E_{x,r_x}=E_x\,:\, x\in D\}$. Let $\alpha_x=(\alpha_{x,1},\dots,\alpha_{x,r_x})$ be a increasing $r_x$-uple of negative real numbers, that we may assume to be contained in $[-1,0)$ [@SNonComp]. Moreover, since every moduli space is isomorphic to one with rational weights [@MS §2], we may take $\alpha_{x,i}\in \mathbb{Q}$ for all $x\in D,\,i=1,\dots,r_x$. A parabolic vector bundle of type $({\mathcal{P}},\alpha)$ is a quasi-parabolic vector bundle ${\mathcal{E}}$ of type ${\mathcal{P}}$ together with the assignation of the numbers $\alpha_{x,i}$ to each step of the flag $E_{x,i}$ $i=1,\dots, r_x$. The assignment of these tuples to a quasi-parabolic bundle ${\mathcal{E}}$ defines a parabolic structure on it. The parabolic bundle thus defined can be assigned an invariant called [parabolic degree]{} $${\mathrm{par-}\deg}({\mathcal{E}}):=\deg(E)-\sum_{x\in D}\sum_{i=1}^{r_x}\alpha_{x,i}m_{x,i}$$ where $m_{x,i}=\dim E_{x,i}/E_{x,i-1}$ are called the [multiplicities]{}. The [parabolic slope]{} of ${\mathcal{E}}$ is the invariant $${\mathrm{par-}\mu}({\mathcal{E}}):=\frac{{\mathrm{par-}\deg}({\mathcal{E}})}{{\mathrm{rk}}E}.$$ A parabolic vector bundle ${\mathcal{F}}$ of type $({\mathcal{P}}',\alpha')$ is a parabolic sub-bundle of a parabolic bundle ${\mathcal{E}}$ of type $({\mathcal{P}},\alpha)$ if $F$ is a subbundle of $E$, $P_x'\leq P_x$ and $\alpha_{x,i}'\geq\alpha_{x_i}$. The parabolic bundle ${\mathcal{E}}$ is $\alpha$-semistable if for any proper parabolic sub-bundle ${\mathcal{F}}\subset {\mathcal{E}}$ $${\mathrm{par-}\mu}({\mathcal{F}})\leq {\mathrm{par-}\mu}({\mathcal{E}}).$$ Clearly, to elucidate whether a bundle is semistable, it is enough to consider quasi-parabolic sub-bundles ${\mathcal{F}}$ with minimal compatible weights.
The moduli space ${\mathrm{N}}({\mathcal{P}},\alpha)$ of parabolic bundles of fixed degree and type is a projective variety [@MS Theorem 4.1.] (in particular, it is separated), possibly singular, parametrising S-equivalence classes of semistable parabolic bundles (or equivalently, isomorphism classes of polystable parabolic bundles). Its dimension is [@MS Theorem 5.3] $$\label{eq:dim_N}
\dim {\mathrm{N}}({\mathcal{P}},\alpha)=(g-1)n^2+1+\frac{1}{2}\left(\sum_{x\in D}n^2-\sum_{i=1}^{r_x}m_{x,i}^2\right).$$
A parabolic Higgs bundle of type $({\mathcal{P}},\alpha)$ is a pair $({\mathcal{E}},\varphi)$ where ${\mathcal{E}}$ is a parabolic vector bundle of type $({\mathcal{P}},\alpha)$ and $\varphi\in H^0(X,{\mathrm{End}}(E)\otimes K(D))$ is a Higgs field preserving the quasi-parabolic structure (namely, the flag at the prescribed points). Endomorphisms satisfying this condition are called parabolic, and the corresponding sheaf is denoted by ${\mathrm{ParEnd}}({\mathcal{E}})$. The subsheaf ${\mathrm{SParEnd}}({\mathcal{E}})\subset{\mathrm{ParEnd}}({\mathcal{E}})$ of endomorphisms that induce the zero Higgs field on the graded object is called the sheaf of strongly parabolic endomorphisms. Note that there are exact sequences $$\label{eq:PEnd}
0\longrightarrow{\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow{\mathrm{End}}(E)\longrightarrow\mathfrak{gl}(n,{\mathbb{C}})^{|D|} /\mathfrak{p}\longrightarrow 0,$$ where the rightmost arrow are given by restriction to the divisor $D$, and $$\label{eq:SPEnd}
0\longrightarrow{\mathrm{SParEnd}}({\mathcal{E}})\longrightarrow{\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow\mathfrak{p}/\mathfrak{n}\cong\mathfrak{l}\longrightarrow 0,$$ where the rightmost arrow is given by restriction to the divisor $D$, followed by the surjection $\mathfrak{p}{\twoheadrightarrow}\mathfrak{l}$.
Thus, a parabolic Higgs bundle is a pair $({\mathcal{E}},\varphi)$ where ${\mathcal{E}}$ is a parabolic vector bundle and $\varphi\in H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))$. Semistability is defined similarly to the parabolic case, except that the subsheaves taken into account are those preserved by the Higgs field. The corresponding moduli space ${\mathrm{M}}({\mathcal{P}},\alpha)$ is a quasi-projective variety [@Yoko Theorem 2.4.8] of dimension [@YokoInf Theorem 5.2] $$\dim{\mathrm{M}}({\mathcal{P}},\alpha)=(2g-2+d)n^2+1$$ where $d=\deg D$. See also [@Thaddeus] for strongly parabolic Higgs bundles.
Poisson structure of the moduli space of Higgs bundles {#sec:Poisson}
------------------------------------------------------
The moduli space of Higgs bundles is Poisson with hyperkaehler leaves. Underlying this rich geometry is the nonabelian Hodge correspondence, a diffeomorphism between three moduli spaces: the Dolbeault (or Higgs) moduli space, the De Rham moduli space of meromorphic connections on $(X,D)$ and the Betti moduli space of representations of the punctured surface fundamental group. The nonabelian Hodge correspondence is due to the work of innumerable authors of which we stress in here [@SNonComp; @BB].
The symplectic geometry of the Betti moduli space has mainly been studied by Boalch [@BoalchQuasi]. From the Dolbeault point of view, Bottacin [@Bottacin] and Markman [@Markman] studied the symplectic geometry for meromorphic Higgs bundles (see Section \[sec:classical\] for more details).
Following [@LM §3.2.4], consider the complex $$C_{\bullet}: {\mathrm{ParEnd}}({\mathcal{E}})\stackrel{{\mathrm{ad}}(\varphi)}{\longrightarrow }{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D)$$ whose dual complex reads $$C_{\bullet}^*: {\mathrm{SParEnd}}({\mathcal{E}})\stackrel{{\mathrm{ad}}(\varphi)}{\longrightarrow }{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D)$$ by duality of ${\mathrm{ParEnd}}({\mathcal{E}})$ and ${\mathrm{SParEnd}}({\mathcal{E}})(D)$.
If $({\mathcal{E}},\varphi)$ is stable, then the Zariski tangent space is identified with the hypercohomology group $$T_{({\mathcal{E}},\varphi)}{\mathrm{M}}({\mathcal{P}},\alpha)={\mathbb{H}}^1(C_{\bullet}),\qquad T_{({\mathcal{E}},\varphi)}^*{\mathrm{M}}({\mathcal{P}},\alpha)={\mathbb{H}}^1(C_{\bullet}^*),$$ hence there is a Poisson map $$\label{eq:poisson_map}
\omega_{{({\mathcal{E}},\varphi)}}:T_{({\mathcal{E}},\varphi)}^*{\mathrm{M}}({\mathcal{P}},\alpha)\longrightarrow T_{({\mathcal{E}},\varphi)}{\mathrm{M}}({\mathcal{P}},\alpha)$$ defined by Serre duality.
More precisely, the deformation space ${\mathbb{H}}^1(C_\bullet)$ of any parabolic Higgs bundle sits into an exact sequence $$\label{eq:defs}
0\longrightarrow{\mathbb{H}}^0(C_\bullet)\longrightarrow H^0({\mathrm{ParEnd}}({\mathcal{E}}))\longrightarrow H^0({\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))\longrightarrow$$ $$\phantom{(2.5)0\longrightarrow}{\mathbb{H}}^1(C_\bullet)\longrightarrow H^1({\mathrm{ParEnd}}({\mathcal{E}}))\longrightarrow H^1({\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))\longrightarrow$$ $$\phantom{(2.5) |0\longrightarrow}{\mathbb{H}}^2(C_\bullet)\longrightarrow 0\phantom{ H^1({\mathrm{ParEnd}}({\mathcal{E}}))\longrightarrow H({\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))\longrightarrow}.$$ The term $H^1({\mathrm{ParEnd}}({\mathcal{E}}))$ is the deformation space of the underlying parabolic bundle, which is Serre dual to the space of strongly parabolic Higgs fields. This defines the Poisson map . In particular, if ${\mathcal{E}}$ is stable $$\label{eq:T_N}
T_{{\mathcal{E}}}{\mathrm{N}}({\mathcal{P}},\alpha)=H^1({\mathrm{ParEnd}}({\mathcal{E}})).$$
The symplectic leaves are determined by the orbits of $\mathrm{gr}(\varphi_x)$ for all $x\in D$, namely, the endormorphism induced by the Higgs field on the graded vector space $\mathrm{gr}(E_{x,\bullet}):=\oplus_{i=1}^{r_x}E_{x,i}/E_{x,i-1}$ at each of the punctures $x\in D$. Let $\mathfrak{p}:=\prod_{x\in D}\mathfrak{p}_x$ and let $\mathfrak{l}:=\prod_{x\in D}\mathfrak{l}_x$ be its Levi subalgebra (so that $\mathrm{gr}(\varphi_x)\in\mathfrak{l}_x$). There is a short exact sequence of complexes $$\xymatrix{
0\ar[r]&{\mathrm{SParEnd}}({\mathcal{E}})\ar[r]\ar[d]_{-[\varphi, \cdot]}&{\mathrm{ParEnd}}({\mathcal{E}})\ar[r]\ar[d]_{[\varphi,\cdot ]}&\mathfrak{l}\ar[r]\ar[d]_{[\mathrm{gr}(\varphi), \cdot]}&0\\
0\ar[r]&{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D)\ar[r]&{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D)\ar[r]&\mathfrak{l}\ar[r]&0.
}$$ Note that it particular, the left and central columns are the defomation complexes of the moduli space of strongly parabolic Higgs bundles ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ and ${\mathrm{M}}({\mathcal{P}},\alpha)$ respectively. Thus, taking hypercohomology of the columns, we get a long exact sequence: $$\label{eq:LES_hyper}
0\longrightarrow {\mathbb{C}}\longrightarrow\operatorname{Ker}([\mathrm{gr}(\varphi),\cdot])\longrightarrow T_{({\mathcal{E}},\varphi)}^*{\mathrm{M}}({\mathcal{P}},\alpha)\stackrel{\omega}{\longrightarrow} T_{({\mathcal{E}},\varphi)}{\mathrm{M}}({\mathcal{P}},\alpha)\longrightarrow0.$$ So the rank of $\omega$ is determined by the dimension of the kernel of $\mathrm{gr}(\varphi)$. Let ${\mathcal{O}}$ be a conjugacy orbit in $\mathfrak{l}$. Along ${\mathcal{O}}$ the rank of $\omega$ is constant. We denote by ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}$ the symplectic leaf associated to the orbit ${\mathcal{O}}$. We will use the following special notation for the main leaves under study: the regular nilpotent symplectic leaf ${{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp, reg}$ (corresponding to the regular nilpotent orbit), and the strongly parabolic symplectic leaf ${\mathrm{M}}({\mathcal{P}},\alpha)_{st}$ (corresponding to the zero orbit). Likewise, we define ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}:=\overline{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp, reg}$ to be the closure of the regular nilpotent leaf. Note that ${\mathrm{M}}({\mathcal{P}},\alpha)_{st}\subset {{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$ is the only closed stratum, contained in the closure of all the other nilpotent leaves.
Similarly to the case of vector bundles, denoting by ${\mathrm{N}}({\mathcal{P}},\alpha)^s\subset {\mathrm{N}}({\mathcal{P}},\alpha)$ the locus of stable points, the space $T^*{\mathrm{N}}({\mathcal{P}},\alpha)^s$ is a dense open subset of ${\mathrm{M}}({\mathcal{P}},\alpha)_{st}$ [@YokoInf Remark 5.1]. In particular, for ${\mathcal{E}}\in {\mathrm{N}}({\mathcal{P}},\alpha)$ $H^1(X,{\mathrm{ParEnd}}({\mathcal{E}}))=H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))^*$ is the deformation space of ${\mathcal{E}}$, and matches the Zariski tangent space whenever ${\mathcal{E}}$ is stable.
The next lemma is a straightforward generalisation of [@LM Lemma 2.1], so we omit the proof.
\[lm:no\_nilp\_par\_end\] Let $\mathfrak{s}\subset\mathfrak{l}$ be a nilpotent subalgebra. Let ${\mathcal{E}}$ be an $\alpha$-stable parabolic bundle of flag-type ${\mathcal{P}}$, and ${\mathrm{ParEnd}}({\mathcal{E}})_{\mathfrak{s}}\subset{\mathrm{ParEnd}}({\mathcal{E}})$ be the subsheaf defined by the exat sequence $$0\longrightarrow {\mathrm{ParEnd}}({\mathcal{E}})_{\mathfrak{s}}\longrightarrow{\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow\mathfrak{l}/\mathfrak{s}\longrightarrow 0,$$ where ${\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow\mathfrak{l}/\mathfrak{s}$ is the composition of the restriction ${\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow\mathfrak{p}$ $\varphi\mapsto\varphi|_D$, and the quotients $\mathfrak{p}{\twoheadrightarrow}\mathfrak{l}{\twoheadrightarrow}\mathfrak{l}/\mathfrak{s}$. Then $H^0({\mathrm{ParEnd}}({\mathcal{E}})_{\mathfrak{s}})=0$.
Moreover, $H^0({\mathrm{ParEnd}}({\mathcal{E}}))\cong{\mathbb{C}}$.
The Hitchin map
---------------
The Hitchin map $$\label{eq:Hitchin_map}
h_{{\mathcal{P}},\alpha}\,:\, {\mathrm{M}}({\mathcal{P}},\alpha)\longrightarrow {\mathrm{H}}_{D}:=\bigoplus H^0(X,K^i(iD))$$ assigns to each Higgs bundle the characteristic polynomial of the Higgs field. This is a projective map [@Yoko Corollary 5.12]. In particular, it is proper and of finite type.
Logares–Martens studied the complete integrability of the restriction of to the generic symplectic leaves. These correspond to semisimple regular orbits (see [@LM §3.2.4]). However, many of their arguments extend verbatim to the non-semisimple regular orbits, and we will mention this when needed. Baraglia–Kamgarpour extended the study of the integrable system to strongly parabolic bundles [@BK]. See also [@SWW] for proofs in arbitrary characteristic.
${\mathbb{C}}^\times$-action {#sec:C_action}
----------------------------
By [@SNonComp §8], there is a action of ${\mathbb{C}}^*$ on ${\mathrm{M}}({\mathcal{P}},\alpha)$ given by multiplying the Higgs field by scalars $$\lambda \cdot ({\mathcal{E}}, \varphi) \mapsto ({\mathcal{E}}, \lambda \cdot \varphi).$$ The Hitchin map is ${\mathbb{C}}^\times$-equivariant for this action and a suitable weighted action on ${\mathrm{H}}_D$ (with weights given by the degrees of the generators of the ring of invariants ${\mathbb{C}}[\mathfrak{gl}_n({\mathbb{C}})]^{{\mathrm{GL}}(n,{\mathbb{C}})}$). By properness of the Hitchin map, the limits $$\lim_{\lambda\to 0}\lambda\cdot({\mathcal{E}},\varphi)$$ exist and belong to the nilpotent cone $h_{{\mathcal{P}},\alpha}^{-1}(0)$, as so do the limits $$\lim_{t\to\infty} t\cdot ({\mathcal{E}},\varphi)$$ when $\varphi$ is nilpotent.
Since nilpotency along $D$ is preserved by the ${\mathbb{C}}^\times$-action, ${\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}$ is ${\mathbb{C}}^\times$-invariant. On the other hand, $h_{{\mathcal{P}},\alpha}^{-1}(0)\subset {\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}$, so that studying the nilpotent cone for ${\mathrm{M}}({\mathcal{P}},\alpha)$ amounts to studying it for ${\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}$. Since the divisor $D$ is assumed to be reduced, it follows from [@SNonComp Theorem 8] that fixed points for the ${\mathbb{C}}^\times$-action have underlying parabolic bundle of the form ${\mathcal{E}}=(\bigoplus_{p=1}^s E^p,\sigma)$, with $E^p$ the eigenspace of the automorphism of $E$ swapping $\varphi$ and $t\varphi$ [@SimpsonVariations Lemma 4.1]. In turn, $\sigma$ preserves the structure of a system of Hodge bundles (i.e., letting $n_p={\mathrm{rk}}(E^p)$, ${\overline{n}}=(n_1,\dots,n_s)$ be the corresponding partition and $P_{{\overline{n}}}$ be the associated parabolic, then $P_{n_p}<P_x$, so that $\sigma$ induces a ${\mathrm{GL}}(n_p,{\mathbb{C}})$-equivariant morphism $\sigma: E^p_x\to {\mathrm{GL}}(n_p,{\mathbb{C}})/P_{p,x}$). The fact that $\varphi:E^p\longrightarrow E^{p-1}\otimes K(D)$ follows from the definition of the eigenspaces.
\[rk:Hodge\_bundle\_str\_par\] In particular, if $({\mathcal{E}},\varphi)$ is fixed for the ${\mathbb{C}}^\times$-action, there exists a refinement of the flags such that $({\mathcal{E}},\varphi)$ becomes strongly parabolic with respect to the latter. This is a general fact for nilpotent quasi-parabolic bundles (see the proof of Proposition \[prop:image\_V\_nilp\]), but fixed point have a preferred refinement preserving the structure of a system of Hodge bundles.
Strongly parabolic bundles as bundles on orbicurves {#sec:rational_weights}
---------------------------------------------------
When $\alpha$ is composed entirely by rational elements, the stack of strongly parabolic Higgs bundles is isomorphic to a stack of bundles (respectively Higgs bundles) on orbicurves [@BiswasOrbi; @MichaHilbert] whose orbifold structure is determined by the weights. A good summary is found in [@DPHecke §5.4].
Universal bundles and locally universal bundles {#sec:univ_bundles}
-----------------------------------------------
Yokogawa proved that ${\mathrm{M}}({\mathcal{P}},\alpha)$ is an open subset of a good quotient of a projective scheme by $\mathrm{PGL}(N)$ [@Yoko Theorem 4.6], with stable locus ${\mathrm{M}}({\mathcal{P}},\alpha)^s$ contained in the corresponding geometric quotient, giving the smooth points [@YokoInf Theorem 5.2]. As an open subset ${\mathrm{M}}({\mathcal{P}},\alpha)^s$ containts the moduli space ${\mathrm{M}}({\mathcal{P}},\alpha)^0$ of parabolic Higgs bundles with underlying stable parabolic bundle. The inclusion is strict by [@Boden-Yoko Claim 3.2 (i)] (see also Remark \[rk:Un\]): even for non generic weights there are more stable pairs than those in ${\mathrm{M}}^({\mathcal{P}},\alpha)^0$. By [@YokoInf Remark 5.1], étale locally over ${\mathrm{M}}({\mathcal{P}},\alpha)^0$ there is a universal bundle. The action of the center of ${\mathrm{GL}}(N)$ is non trivial [@Boden-Yoko page 464], and therefore the bundle does in general not descend. It only descends to ${\mathrm{M}}({\mathcal{P}},\alpha)^0$ when the parabolic weights are generic [@Boden-Yoko Proposition 3.2]. However, regardless of the weights, points of ${\mathrm{M}}({\mathcal{P}},\alpha)^{0}$ always have an étale neighbourhood admitting a universal bundle [@Bhosle page 16], [@Boden-Yoko page 464]. The same holds for ${\mathrm{N}}({\mathcal{P}},\alpha)^s$.
Parabolic bundles and very stability {#sec:criteria}
====================================
Criteria for very stability
---------------------------
In this section we prove a criterion for very stability of parabolic bundles via the Hitchin map.
Let us start by recalling [@Hacen Lemma 1.3] which plays a fundamental role in our proof:
[[@Hacen Lemma 1.3]]{}\[lm:Hacen\] Let $ f=(f_1,\dots, f_n):\mathbb{A}^m\longrightarrow\mathbb{A}^n$ be a morphism given by homogeneous polynomials. Then, if $f^{-1}(0)=0$, $f$ is finite.
Given a quasi-parabolic bundle ${\mathcal{E}}$, we denote ${\mathrm{V}}_{\mathcal{E}}:=H^0(X,\mathrm{ParEnd}({\mathcal{E}})\otimes K(D))$. For each conjugacy orbit ${\mathcal{O}}=\prod_{x\in D}{\mathcal{O}}_x$ of $\prod_{x\in D}\mathfrak{l}_x$ (where $\mathfrak{l}_x<\mathfrak{p}_x$ is the Levi subalgebra), denote by ${\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}$ the subset of ${{\mathrm{V}}_{{\mathcal{E}}}}$ corresponding to Higgs fields with ${\mathrm{gr}}(\varphi_x)\in{\mathcal{O}}_x$. When ${\mathcal{O}}=0$, we will denote ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ by ${{\mathrm{V}}_{{\mathcal{E}},st}}=H^0(X,{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D))$, and if ${\mathcal{O}}={\mathcal{O}}_{rn}$ is the regular nilpotent orbit, then the closure of ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ in ${{\mathrm{V}}_{{\mathcal{E}}}}$ is denoted by ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$. This is the set of nilpotent Higgs fields.
\[rk:VO\_as\_intersection\] If ${\mathcal{E}}$ is $\alpha$-stable, then ${{\mathrm{V}}_{{\mathcal{E}}}}\subset{\mathrm{M}}({\mathcal{P}},\alpha)$, ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}={{\mathrm{V}}_{{\mathcal{E}}}}\cap{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}$, and ${{\mathrm{V}}_{{\mathcal{E}},nilp}}={{\mathrm{V}}_{{\mathcal{E}}}}\cap{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$, but the definition of these objects is independent of the stability parameter.
A quasi-parabolic bundle $\mathcal{E}=(E,\{E_{x,\bullet}\,:\,x\in D\})$ is called very stable if it has no non-zero nilpotent parabolic Higgs field $\varphi\in{{\mathrm{V}}_{{\mathcal{E}}}}$. It is called [strongly very stable]{} if and only if it has no non-zero strongly parabolic nilpotent Higgs field.
An $\alpha$-stable bundle that is not (strongly) very stable is called (strongly) wobbly. Let ${\mathrm{Wo}}\subset{\mathrm{N}}({\mathcal{P}},\alpha)$ (respectively ${\mathrm{Wo}}_{st}\subset{\mathrm{N}}({\mathcal{P}},\alpha)$) denote the wobbly locus (resp. the strong wobbly locus).
\[rk:vs\_indep\_weight\]
1. \[it:vs\_depends\_on\_qp\] Although the definition of (strong) very stability only depends on the quasi-parabolic structure, for evey assignment of a parabolic structure $({\mathcal{P}},\alpha)$, very stable bundles are $\alpha$-stable (see Lemma \[lm:vs\_is\_stable\] and Remark \[rk:wobbly\_alpha\]).
2. \[it:vs=nvs\] Note that $\mathcal{E}$ is very stable if and only if there is no Higgs field with nilpotent residue along the flag other than zero.
3. \[it:vs\_is\_svs\] Very stable parabolic bundles are strongly very stable.
When the weights are rational, as observed in Section \[sec:rational\_weights\], the tools developed in [@L] are available to analyse the moduli space of strongly nilpotent Higgs bundles. In fact, one may prove the following.
\[lm:vs\_is\_stable\] Let ${\mathcal{E}}$ be parabolic of type $({\mathcal{P}},\alpha)$. If it is strongly very stable, then it is $\alpha$-stable. In particular, if it is very stable, then it is $\alpha$-stable.
By assigning rational weights $\alpha$, strongly very stable parabolic bundles can be seen as very stable bundles on an orbifold curve, and so they are $\alpha$-stable by [@L Proposition 3.5]. Stability of very stable bundles follows from the above and Remark \[rk:vs\_indep\_weight\] .
\[rk:wobbly\_alpha\] Note that the proof of Lemma \[lm:vs\_is\_stable\] uses the identification of parabolic bundles with bundles on orbicurves, and this requires the assignment of weights. However, since the definition of (strong) very stability is independent of $\alpha$ (cf. Remark \[rk:vs\_indep\_weight\]), from Lemma \[lm:vs\_is\_stable\] we deduce the existence of a common subset to all ${\mathrm{N}}({\mathcal{P}},\alpha)$, namely, the set of very stable bundles ${\mathcal{E}}$. The corresponding wobbly locus, however, depends on the stability parameter $\alpha$.
\[prop:st\_vs\_iff\_proper\] Let ${\mathcal{E}}\in{\mathrm{N}}({\mathcal{P}},\alpha)$ be stable. Let $({\mathcal{E}},\varphi)\in{\mathrm{M}}({\mathcal{P}},\alpha)$, let ${{\mathrm{V}}_{{\mathcal{E}},st}}$ be the space of strongly parabolic Higgs fields on ${\mathcal{E}}$, and let $h_{{\mathcal{E}},st}:=h_{{\mathcal{P}},\alpha}|_{{{\mathrm{V}}_{{\mathcal{E}},st}}}$. Then: $$\begin{array}{lcl}
{\mathcal{E}}\textnormal{ \it is strongly very stable } &\iff& h_{{\mathcal{E}},st} \textnormal{ \it is finite }\\
&\iff& h_{{\mathcal{E}},st} \textnormal{ \it is quasi-finite } \\
&\iff& h_{{\mathcal{E}},st} \textnormal{ \it is proper }\\
&\iff& {{\mathrm{V}}_{{\mathcal{E}},st}}{\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha) \textnormal{ \it is proper.}
\end{array}$$
The fact that very stability implies finiteness follows from affineness of ${{\mathrm{V}}_{{\mathcal{E}},st}}$ and Lemma \[lm:Hacen\].
To see the other implications, we need to work some more. Let ${\mathrm{H}}_{st}\subset {\mathrm{H}}_D$ be the image of ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ under $h_{{\mathcal{P}},\alpha}$. By [@BK Theorem 36], ${\mathrm{H}}_{st}\subset{\mathrm{H}}_D$ is an affine subspace of dimension $\dim {\mathrm{H}}_{st}=\dim{\mathrm{N}}({\mathcal{P}},\alpha)$.
Also by [@MS Theorem 5.3], $\dim{\mathrm{N}}({\mathcal{P}},\alpha)=\dim{H^0(X,{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D))}$ whenever ${\mathcal{E}}$ is stable.
Now, finiteness implies properness and quasi-finiteness. Also, by the above discussion $h_{{\mathcal{E}},st}$ is a map of finite type of affine spaces of the same dimension, hence properness implies quasi-finiteness and hence also finiteness.
Regarding the equivalence of quasi-finiteness and finiteness, quasi-finiteness implies very stability (and thus finiteness) as the existence of a non-zero nilpotent Higgs field $\varphi$ on ${\mathcal{E}}$ would automatically produce a one dimensional subspace in $h^{-1}_{{\mathcal{E}},st}(0)$ (this requires the stability hypothesis on ${\mathcal{E}}$ to make sure none of the Higgs fields in the line ${\mathbb{C}}^\times\cdot \varphi$ are identified).
Finally, the equivalence between properness of ${{\mathrm{V}}_{{\mathcal{E}},st}}{\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha)$ and properness of $h_{{\mathcal{E}},st}$ is a consequence of the valuative criterion for properness (this is exactly the proof of [@PP Proposition 2.2]).
The following result was observed in [@FGOP Corollary 7.3], and extends verbatim to the current context.
\[cor:lagrangian\_multi\]Let ${\mathrm{M}}({\mathcal{P}},\alpha)_{st}$ denote the strongly parabolic symplectic leaf, and let $h_{{\mathcal{P}},\alpha,st}:=h_{{\mathcal{P}},\alpha}|_{{\mathrm{M}}({\mathcal{P}},\alpha)_{st}}$. If ${\mathcal{E}}$ is strongly very stable, then ${\mathrm{V}}_{{\mathcal{E}},st}$ is a Lagrangian multisection of $h_{{\mathcal{P}},\alpha,st}$.
Since the only deformations along ${{\mathrm{V}}_{{\mathcal{E}},st}}$ concern the Higgs field (see Section \[sec:Poisson\]), ${{\mathrm{V}}_{{\mathcal{E}},st}}$ is clearly isotropic.
By Lemma \[lm:vs\_is\_stable\], ${\mathcal{E}}$ is stable, and so from and Serre duality $T_E{\mathrm{N}}({\mathcal{P}},\alpha)=H^0(X,{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D))^*$. Thus, $
\dim{{\mathrm{V}}_{{\mathcal{E}},st}}=\dim{\mathrm{N}}({\mathcal{P}},\alpha)^s\subset {{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$. But $T^*{\mathrm{N}}({\mathcal{P}},\alpha)\subset{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ is dense, so it follows that $\dim{{\mathrm{V}}_{{\mathcal{E}},st}}=\dim{\mathrm{N}}({\mathcal{P}},\alpha)=\frac{1}{2}\dim{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$, and so ${{\mathrm{V}}_{{\mathcal{E}},st}}$ is maximal dimensional.
Finally, ${\mathrm{Im}}(h_{{\mathcal{P}},\alpha,st})$ has the same dimension, and is affine by [@BK Theorem 36]. This, together with properness, yields that $h_{{\mathcal{E}},st}$ is onto, whence the result.
Note that the stability assumptions in Proposition \[prop:st\_vs\_iff\_proper\] and Proposition \[prop:vs\_iff\_proper\] may be dropped in one direction by Lemma \[lm:vs\_is\_stable\].
Let $$\label{eq:h_V}
h_{{\mathcal{E}},nilp} : {{\mathrm{V}}_{{\mathcal{E}},nilp}}\longrightarrow {\mathrm{H}}_D$$ be the restriction of the Hitchin map to the vector space ${{\mathrm{V}}_{{\mathcal{E}},nilp}}:=H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))_{nilp}$ of Higgs fields on ${\mathcal{E}}$ with nilpotent residue along the flag. In order to prove the equivalent to Proposition \[prop:st\_vs\_iff\_proper\] in this setup, let us start by two preliminary results about ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$.
\[lm:oVn\_minus\_Vn\] Let ${\mathcal{E}}$ be a stable parabolic bundle, and let $({\mathcal{F}},\psi)\in{{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{\mathrm{V}}_{{\mathcal{E}},nilp}}$ or $({\mathcal{F}},\psi)\in{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}\setminus{{\mathrm{V}}_{{\mathcal{E}}}}$. Then
1. \[it:F\_uns\] ${\mathcal{F}}$ is unstable.
2. \[it:oVE\_vs\_oVn\] ${{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{\mathrm{V}}_{{\mathcal{E}},nilp}}\neq\emptyset\iff{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}\setminus{{\mathrm{V}}_{{\mathcal{E}}}}\neq\emptyset$.
By [@stacks-project [Tag 0A24](https://stacks.math.columbia.edu/tag/0A24)], given $({\mathcal{F}},\psi)\in{{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{\mathrm{V}}_{{\mathcal{E}},nilp}}$, there exists a discrete valuation ring $R$ and a morphism $$\iota_F:C:={\mathrm{Spec}}(R)\longrightarrow {\mathrm{M}}({\mathcal{P}},\alpha)$$ such that the generic point ${\mathrm{Spec}}(k)$ (where $k$ is the fraction field of $R$) maps to ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$, while the closed point $o$ goes to $({\mathcal{F}},\psi)\in {\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}\setminus{{{\mathrm{V}}_{{\mathcal{E}},nilp}}}$. Since ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}\subset{\mathrm{M}}({\mathcal{P}},\alpha)$ is closed, then ${{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{\mathrm{V}}_{{\mathcal{E}},nilp}}\subset{\mathrm{M}}({\mathcal{P}},\alpha)\setminus{{\mathrm{V}}_{{\mathcal{E}}}}$, namely, it must be ${\mathcal{F}}\neq{\mathcal{E}}$. So it is enough to prove the statement for $({\mathcal{F}},\psi)\in{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}$.
If ${\mathcal{F}}$ were semistable, composing with $r$ would yield $$r\circ \iota_F:C:={\mathrm{Spec}}(R)\longrightarrow {\mathrm{N}}({\mathcal{P}},\alpha)$$ non constant, extending the constant map ${\mathcal{E}}:{\mathrm{Spec}}(k)\longrightarrow {\mathrm{N}}({\mathcal{P}},\alpha)$, thus violating separatedness of ${\mathrm{N}}({\mathcal{P}},\alpha)$ [@MS Theorem 4.1]. This proves .
For , note that ${{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{\mathrm{V}}_{{\mathcal{E}},nilp}}\subset{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}\setminus{{\mathrm{V}}_{{\mathcal{E}}}}$ (by , or simply closedness of ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$). So it is enough to prove the converse.
Let $({\mathcal{F}},\psi)\in{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}\setminus{{\mathrm{V}}_{{\mathcal{E}}}}$. Then, by equivariance of the Hitchin map $({\mathcal{F}}_0,\psi_0):=\lim_{t\to 0}t\cdot ({\mathcal{F}},\psi)\in h^{-1}_{{\mathcal{P}},\alpha}(0)\cap{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}\subset{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}\cap{{\overline{{\mathrm{V}}}}_{{\mathcal{E}}}}={{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}$.
The above proof translates essentially verbatim to the simplest non parabolic vector bundle setup, and corrects a mistake in the proof of [@PP Proposition 2.3]. Indeed, the étale local family considered therein may not exist away from $T^*{\mathrm{N}}^s(n,d)$. Thanks to T. Hausel for pointing this error out to us.
The following adapts [@BK Theorem 36] to general nilpotent Higgs fields.
\[prop:image\_V\_nilp\] Let ${\mathcal{E}}$ be a stable parabolic bundle. Then, the image under the Hitchin map $h_{{\mathcal{P}},\alpha}$ of ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ is contained in an affine space of the same dimension.
First note, that ${{\mathrm{V}}_{{\mathcal{E}},nilp}}={\overline{{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}}}\cap {{\mathrm{V}}_{{\mathcal{E}}}}$ for some nilpotent orbit ${\mathcal{O}}$ such that ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}:={{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}\cap{{\mathrm{V}}_{{\mathcal{E}}}}\neq\emptyset$. Indeed, take a maximal dimensional orbit satisfying ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}\neq\emptyset$ . If ${{\mathrm{V}}_{{\mathcal{E}},nilp}}\neq{\overline{{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}}}\cap {{\mathrm{V}}_{{\mathcal{E}}}}$, then there exists another nilpotent orbit ${\mathcal{O}}'$ with ${\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}'}\neq\emptyset$ and ${\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}'}\not\subset {\overline{{{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}}}$. By affineness of ${{\mathrm{V}}_{{\mathcal{E}}}}$, then ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}+{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}'}\subset{{\mathrm{V}}_{{\mathcal{E}},nilp}}$. Now, using the Jordan block form, we see that either ${\mathcal{O}}'\subset{\overline{{\mathcal{O}}}}$ or there is an element in ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}+{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}'}$ belonging to a higher dimensional orbit, which contradicts our choice of ${\mathcal{O}}$. By irreducibility of ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$, we may conclude that ${{\mathrm{V}}_{{\mathcal{E}},nilp}}={\overline{{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}}}\cap {{\mathrm{V}}_{{\mathcal{E}}}}$.
Since ${{\mathrm{V}}_{{\mathcal{E}},nilp}}\cong\mathbb{A}^d$ is irreducible, the closure of ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ inside ${{\mathrm{V}}_{{\mathcal{E}}}}$ must be all of ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$. Now, for some refinement ${\mathcal{P}}'<{\mathcal{P}}$, the orbit ${\mathcal{O}}$ intersects $\mathfrak{l}:=\prod_{x\in D}\mathfrak{l}'_x$ (where $\mathfrak{l}'_x$ is the Levi subalgebra of $\mathfrak{p}_x'=Lie({\mathcal{P}}_x')$) at $0$. For example, taking ${\mathcal{P}}'$ determined by the iterated kernels of $\varphi_x\in{\mathcal{O}}$, then ${\mathcal{P}}'$ satisfies the property and moreover, it is maximal for it (namely, whenever ${\overline{{\mathcal{P}}}}$ is another refinement with $\varphi$ lifting to a strongly nilpotent field preserving a flag of type ${\overline{{\mathcal{P}}}}$, then ${\overline{{\mathcal{P}}}}<{\mathcal{P}}'$).
Let $\mathfrak{p}_x'=\mathfrak{l}_x'\oplus\mathfrak{n}_x'$, and let $\mathbf{O}:=\prod_{x\in D}\mathfrak{n}_x'\cap\mathfrak{l}_x$. Define ${\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\subset{\mathrm{ParEnd}}({\mathcal{E}})$ via the following exact sequence: $$0\longrightarrow{\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\longrightarrow{\mathrm{ParEnd}}({\mathcal{E}})\longrightarrow\mathfrak{l}/\mathbf{O}\longrightarrow0,$$ where the rightmost arrow is the composition of ${\mathrm{ParEnd}}({\mathcal{E}}){\twoheadrightarrow}\mathfrak{p}{\twoheadrightarrow}\mathfrak{l}{\twoheadrightarrow}\mathfrak{l}/\mathbf{O}$.
Now, let ${\mathcal{E}}'$ denote the quasi-parabolic bundle of type ${\mathcal{P}}'$ induced by the existence of some Higgs field in ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ with residue in ${\mathcal{O}}$. Note that the existence of the exact commutative diagram: $$\xymatrix{
{\mathrm{SParEnd}}({\mathcal{E}}')\ar@{^(->}[r]\ar@{^(->}[d]&{\mathrm{ParEnd}}({\mathcal{E}}')\ar@{^(->}[d]\ar@{->>}[r]&\mathfrak{l}'\ar@{^(->}[d]\\
{\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\ar@{^(->}[r]&{\mathrm{ParEnd}}({\mathcal{E}})\ar@{->>}[d]\ar@{->>}[r]&\mathfrak{l}/\mathbf{O}\ar@{->>}[d]\\
&\mathfrak{p}/\mathfrak{p}'\ar[r]_{\cong}&\mathfrak{l}/\mathfrak{p}'\cap\mathfrak{l}.
}$$ implies that $$\label{eq:Par_O_SPar}
{\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\cong{\mathrm{SParEnd}}({\mathcal{E}}').$$ Note that ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}\subset H^0({\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\otimes K(D))\subset{{\mathrm{V}}_{{\mathcal{E}},nilp}}$, thus, by affiness of the last two subspaces and euality of the dimensions of ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ and ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$, the second inclusion must be an equality and so $\dim H^0({\mathrm{ParEnd}}({\mathcal{E}})_{\mathbf{O}}\otimes K(D))=\dim{{\mathrm{V}}_{{\mathcal{E}},nilp}}$.
Now, by stablity of ${\mathcal{E}}$ and Lemma \[lm:no\_nilp\_par\_end\], we have that $H^0({\mathrm{ParEnd}}({\mathcal{E}})_\mathbf{O})=0=H^0({\mathrm{SParEnd}}({\mathcal{E}}'))$. A simple computation using shows that this implies that $\dim H^0({\mathrm{SParEnd}}({\mathcal{E}}')\otimes K(D))=\dim{\mathrm{N}}({\mathcal{P}}',\alpha')$ (for a generic $\alpha'$). Note that this is independent of the existence of a parameter $\alpha'$ turning ${\mathcal{E}}'$ into a stable parabolic bundle. It then follows form the above and that $$\label{eq:dim_VO}
\dim{{\mathrm{V}}_{{\mathcal{E}},nilp}}=\dim H^0({\mathrm{SParEnd}}({\mathcal{E}}')\otimes K(D))=\dim{\mathrm{N}}({\mathcal{P}}',\alpha')$$ for a suitable $\alpha'$.
Let ${\mathcal{M}}({\mathcal{P}})$ be the moduli stack of Higgs bundles on $X$ of flag type ${\mathcal{P}}$. The Hitchin map $$h_{\mathcal{P}}:{\mathcal{M}}({\mathcal{P}})\longrightarrow{\mathrm{H}}_D$$is defined in the same way. Moreover, there is a morphism ${\mathcal{M}}(\mathcal{P}')\longrightarrow {\mathcal{M}}(\mathcal{P})$ making the following diagram commute: $$\label{eq:Hitchin_maps_neq_pars}
\xymatrix{
{\mathcal{M}}(\mathcal{P}')\ar[d]_{p}\ar[dr]^{h_{\mathcal{P}'}}&\\
{\mathcal{M}}(\mathcal{P})\ar[r]_{h_{\mathcal{P}}}&{\mathrm{H}}_{D}.
}$$ In particular $h_{{\mathcal{P}},\alpha}({{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}})\subset {\mathrm{Im}}(h_{{\mathcal{P}}',st})\subset{\mathrm{Im}}(h_{{\mathcal{P}},st})$ which is affine of dimension equal to $\dim{{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ by [@BK Theorem 36] and .
\[prop:vs\_iff\_proper\] Let ${\mathcal{E}}$ be stable, and let ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ and $h_{{\mathcal{E}},nilp}$ be as in . Then, $$\begin{array}{lcl}
{\mathcal{E}}\textnormal{ \it is very stable } &\iff& h_{{\mathcal{E}},nilp} \textnormal{ \it is finite }\\
&\iff& h_{{\mathcal{E}},nilp} \textnormal{ \it is quasi-finite } \\
&\iff& h_{{\mathcal{E}},nilp} \textnormal{ \it is proper }\\
&\iff& {{\mathrm{V}}_{{\mathcal{E}},nilp}}{\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha) \textnormal{ \it is proper}\\
&\iff& h_{{\mathcal{E}}} \textnormal{ \it is proper }\\
&\iff& {{\mathrm{V}}_{{\mathcal{E}}}}{\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha) \textnormal{ \it is proper.}\\
\end{array}$$
The equivalence between properness of ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ and ${{\mathrm{V}}_{{\mathcal{E}}}}$ follows from Lemma \[lm:oVn\_minus\_Vn\]. The equivalence between properness of $h_{{\mathcal{E}},nilp}$ (resp. $h_{\mathcal{E}}$) and properness of ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ (resp. ${{\mathrm{V}}_{{\mathcal{E}}}}$) is a consequence of the valuative criterion for properness. Thus it is enough to prove all the other equivalences.
The implication that very stable implies finite follows from Lemma \[lm:Hacen\].
To prove the remaining equivalences, we note that it is enough to check the equivalence of finiteness and properness. Indeed, quasi-finiteness automatically implies very stability, as the existence of a nilpotent Higgs field implies the existence of a line of nilpotent Higgs fields, and hence quasi-finiteness is equivalent to finiteness by the first equivalence.
Now, since finiteness implies properness, it is enough to check that properness implies finiteness, for which given that the Hitchin map is of finite type, it is enough to check that fibers are finite dimensional. This follows from Propostion \[prop:image\_V\_nilp\].
The equivalent of Corollary \[cor:lagrangian\_multi\] follows from Propositions \[prop:image\_V\_nilp\] and \[prop:vs\_iff\_proper\]:
\[cor:multisection\_nilp\] Let ${\mathcal{E}}$ be very stable, and let ${\mathcal{O}}$ be a nilpotent orbit such that ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}\neq\emptyset$ is maximal dimensional inside ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$. Then, ${{\mathrm{V}}_{{\mathcal{E}},nilp}}$ is the closure of a Lagrangian multisection of the restriction of $h_{{\mathcal{P}},\alpha}$ to ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}$.
By Propositions \[prop:image\_V\_nilp\] and \[prop:vs\_iff\_proper\], together with isotropicity of ${{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}$ (see Section \[sec:Poisson\]), all we need to check is that: $$\dim{{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}=\frac{1}{2}\dim{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{{\mathcal{O}}}}.$$ By , the rank of $\omega$ at $({\mathcal{E}},\varphi)$ when $\varphi|_D\in{\mathcal{O}}$ is $$\mathrm{{\mathrm{rk}}}(\omega)=\dim{\mathrm{M}}({\mathcal{P}},\alpha)-\dim\mathfrak{z}(\mathrm{gr}(\varphi))+1.$$ Thus, taking $\mathfrak{l}'$ as in the proof of Proposition \[prop:image\_V\_nilp\], we have that $\operatorname{Ker}[\mathrm{gr}(\varphi),\cdot]\cong\mathfrak{l}'$, hence $$\mathrm{{\mathrm{rk}}}(\omega)=2(g-1)n^2+2+\sum_{x\in D} n^2-\dim\mathfrak{l}'_x =2(\dim {{\mathrm{V}}_{{\mathcal{E}},{\mathcal{O}}}}),$$ where the last equality follows from and .
Strongly wobbly bundles and the nilpotent cone
----------------------------------------------
In this section we describe strong wobbliness in terms of the geometry of the nilpotent cone.
Let $C_i\subset h_{{\mathcal{P}},\alpha,st}^{-1}(0)$, $i\in I$ denote the irreducible components of the strongly parabolic nilpotent cone. By [@L Proposition 3.8] one of them coincides with ${\mathrm{N}}({\mathcal{P}},\alpha)$, say $C_0={\mathrm{N}}({\mathcal{P}},\alpha)$, under the map ${\mathcal{E}}\mapsto({\mathcal{E}},0)$ for ${\mathcal{E}}\in{\mathrm{N}}({\mathcal{P}},\alpha)$. Let $C_0^s$ denote the interseccion of $C_0$ and the image of stable locus ${\mathrm{N}}({\mathcal{P}},\alpha)^s\subset {\mathrm{N}}({\mathcal{P}},\alpha)$ under the aforementioned map.
\[lm:wobbly\] The strong wobbly locus of ${\mathrm{N}}({\mathcal{P}},\alpha)$ is the intersection ${\mathrm{Wo}}_{st}:=\bigcup_{i\neq 0}C_i\cap C_0^s$.
The inclusion $\subset$ is clear, as if ${\mathcal{E}}\in {\mathrm{Wo}}$, let $\varphi$ be a strongly parabolic Higgs field on ${\mathcal{E}}$, then $({\mathcal{E}},\varphi)\in C_i$ for some $i\neq 0$. Hence $({\mathcal{E}},0)=\lim_{t\to 0}(E,t\varphi)\in C_i\cap C_0^s$ for some $i\neq 0$.
For the converse, assume $({\mathcal{E}},0)\in C_i\cap C_0^s$. Then $({\mathcal{E}},0)$ is a boundary point of $C_i^s\cap{\mathrm{M}}({\mathcal{P}},\alpha)^0$ inside ${\mathrm{M}}({\mathcal{P}},\alpha)^0$. So by [@PP Lemma 2.4], and stability of ${\mathcal{E}}$, we may find a smooth curve $\psi:Z\longrightarrow C_i^s$ such that $\psi\left(Z\setminus\{z_0\}\right)\subset C_i^s\setminus \partial C_i$ and $\psi(z_0)=({\mathcal{E}},0)$. Now, by stability, one may consider the étale local family $(\mathbf{E},\mathbf{\Phi})$ (whose existence follows from Luna’s slice theorem together with Lemma \[lm:no\_nilp\_par\_end\] and the arguments in Section \[sec:univ\_bundles\]). If ${\mathcal{E}}$ were strongly very stable, the generic point of this family would have to be very stable (by openness of very stability [@L Proposition 3.5]), contradicting the fact that $(\mathbf{E},\mathbf{\Phi})_z\in C_i\setminus C_0$ for $z\neq z_0$.
Wobbly and shaky bundles {#sec:WandShaky}
------------------------
By openness of semistability and irreducibility of ${\mathrm{M}}({\mathcal{P}},\alpha)$, there exists a rational map $$r:\xymatrix{{\mathrm{M}}({\mathcal{P}},\alpha)\ar@{-->}[r]&{\mathrm{N}}({\mathcal{P}},\alpha)}$$ given by forgetting the Higgs field, which is a morphism over the stable locus.
\[defi:shaky\_new\] A stable bundle ${\mathcal{E}}$ such that $h_{{\mathcal{E}}}:r^{-1}({\mathcal{E}})={{\mathrm{V}}_{{\mathcal{E}}}}\longrightarrow {\mathrm{H}}_D$ is not proper is called shaky. It is called strongly shaky if $h_{{\mathcal{E}},st}:r^{-1}({\mathcal{E}})\cap{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}\longrightarrow {\mathrm{H}}_D$ is not proper.
An immediate consequence of Propositions \[prop:st\_vs\_iff\_proper\] and \[prop:vs\_iff\_proper\] is the following.
\[cor:W=S\] There is an equality between the wobbly locus and the shaky locus. The same holds for the strongly wobbly and the strongly shaky loci.
Wobbly loci and resolutions. The regular case {#sec:W=S}
=============================================
In this section we will assume that the parabolic weights are generic. In particular ${\mathrm{M}}({\mathcal{P}},\alpha)$ is smooth, so that all semistable parabolic (Higgs) bundles are stable; moreover, over the open subset ${\mathrm{M}}({\mathcal{P}},\alpha)^0\subset {\mathrm{M}}({\mathcal{P}},\alpha)$ of pairs with underlying stable parabolic bundle there exists a universal bundle (see Section \[sec:univ\_bundles\]).
Recall the rational map from Section \[sec:WandShaky\] $$\label{eq:rat_full}
r:\xymatrix{{\mathrm{M}}({\mathcal{P}},\alpha)\ar@{-->}[r]&{\mathrm{N}}({\mathcal{P}},\alpha)}$$ given by forgetting the Higgs field. Note that restricts to rational maps $$\label{eq:rat_nilp}
r_{n}:\xymatrix{{\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}\ar@{-->}[r]&{\mathrm{N}}({\mathcal{P}},\alpha)},$$ $$\label{eq:rat_st}
r_{s}:\xymatrix{{\mathrm{M}}({\mathcal{P}},\alpha)_{st}\ar@{-->}[r]&{\mathrm{N}}({\mathcal{P}},\alpha)}.$$ All the rational maps above are surjective, as by stability there is an embedding ${\mathrm{N}}({\mathcal{P}},\alpha){\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha)_{st}$ given by ${\mathcal{E}}\mapsto({\mathcal{E}},0)$, which is in fact the composition ${\mathrm{N}}({\mathcal{P}},\alpha){\hookrightarrow}T^*{\mathrm{N}}({\mathcal{P}},\alpha){\hookrightarrow}{\mathrm{M}}({\mathcal{P}},\alpha)_{st}$ of the zero section and the natural embedding.
Let ${\mathrm{Un}}\subset{\mathrm{M}}({\mathcal{P}},\alpha)$ be the subset given by semistable parabolic Higgs bundles with unstable underlying parabolic bundle. We denote by ${\mathrm{Un}}_{nilp}:={\mathrm{Un}}\cap{\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}$ and ${\mathrm{Un}}_{st}:={\mathrm{Un}}\cap{\mathrm{M}}({\mathcal{P}},\alpha)_{st}$.
\[rk:Un\] Non emptyness of ${\mathrm{Un}_{st}}$ (and thus of ${\mathrm{Un}_{nilp}}$ and ${\mathrm{Un}}$) follows, e.g., from the fact that the strongly parabolic nilpotent cone is not irreducible by [@L] (cf. Section \[sec:rational\_weights\]). Indeed, by [@SNonComp §8], given $({\mathcal{E}},\varphi)\in {h^{st}}^{-1}(0)$, then $\lim_{t\to\infty} t\cdot ({\mathcal{E}},\varphi)$ exists and is a fixed point for the ${\mathbb{C}}^\times$-action, hence inside $h^{-1}(0)$. Since taking limits at $0$ and $\infty$ defines Zariski locally trivial affine fibrations with strata of half the dimension, limits at $\infty$ must have unstable underlying bundle (this argument is found in [@Heinloth §1.1]).
In the non parabolic case, letting ${\mathrm{M}}$ denote the moduli space of Higgs bundles and ${\mathrm{N}}$ that of vector bundles, there is an equality ${\mathrm{Un}}:={\mathrm{M}}\setminus T^*{\mathrm{N}}$. This is not the case here, as $T^*{\mathrm{N}}$ consists exclusively of strongly parabolic bundles [@YokoInf Remark 5.1]. When ${\mathcal{P}}$ consists only of Borel subgroups, then ${\mathrm{Un}_{nilp}}:={{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}\setminus T^*{\mathrm{N}}({\mathcal{P}},\alpha)$.
Given the smoothness assumption, Hironaka’s results on the resolution of singularities [@H] ensure that a finite number of blowups resolve the morphism (see the discussion following [@H Question E]), which yields $$\label{eq:res_rat_full}
\xymatrix{
\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}\ar[dr]^{\hat{r}}\ar[d]_{\pi}&\\
{\mathrm{M}}({\mathcal{P}},\alpha)\ar@{-->}^r[r]&{\mathrm{N}}({\mathcal{P}},\alpha)
}$$ Let ${\mathrm{Ex}}=\pi^{-1}({\mathrm{Un}})$ be the exceptional divisor of ${\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}}$. Then, since ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ and ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$ are closed, it follows from [@Ha Corollary II.7.15] that ${\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp}}:=\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}\times_{{\mathrm{M}}({\mathcal{P}},\alpha)}{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$ and ${\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{st}}:=\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}\times_{{\mathrm{M}}({\mathcal{P}},\alpha)}{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{st}}$ are closed subschemes of ${\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}}$ consisting of a finite number of blowups along ${\mathrm{Un}_{nilp}}$ and ${\mathrm{Un}_{st}}$ respectively. Hence we have resolutions $$\label{eq:res_rat_nilp}
\xymatrix{
\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp}\ar[dr]^{\hat{r}_n}\ar[d]_{\pi_n}&\\
{\mathrm{M}}({\mathcal{P}},\alpha)_{nilp}\ar@{-->}_{r_n}[r]&{\mathrm{N}}({\mathcal{P}},\alpha),
}$$ $$\label{eq:res_rat_st}
\xymatrix{
\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{st}\ar[dr]^{\hat{r}_s}\ar[d]_{\pi_s}&\\
{\mathrm{M}}({\mathcal{P}},\alpha)_{st}\ar@{-->}_{r_s}[r]&{\mathrm{N}}({\mathcal{P}},\alpha).
}$$ In particular, the exceptional divisors ${\mathrm{Ex}}_{nilp}:=\pi_n^{-1}({\mathrm{Un}_{nilp}})$ and ${\mathrm{Ex}}_{st}:=\pi_s^{-1}({\mathrm{Un}_{st}})$ satisfy ${\mathrm{Ex}}_{nilp}={\mathrm{Ex}}\cap{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp}}$ and ${\mathrm{Ex}}_{st}={\mathrm{Ex}}\cap{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{st}}$.
\[defi:shaky\] Denote by ${\mathrm{Sh}}$ (resp. ${\mathrm{Sh}_{nilp}}$ and ${\mathrm{Sh}_{st}}$) the image under $\hat{r}$ (resp. ${\hat{r}_n}$ and ${\hat{r}_s}$) of the exceptional divisor ${\mathrm{Ex}}=\pi^{-1}({\mathrm{Un}})\subset{{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}}}$ (resp. ${\mathrm{Ex}}_{nilp}$ and ${\mathrm{Ex}}_{st}$).
The notation stands, of course, for shaky, as the equality of ${\mathrm{Sh}}$ and the wobbly locus will follow from shakiness of wobbly bundles.
\[thm:W=S\] There is an equality ${\mathrm{Wo}}={\mathrm{Sh}_{nilp}}={\mathrm{Sh}}$, where ${\mathrm{Wo}}\subset{\mathrm{N}}({\mathcal{P}},\alpha)$ denotes the wobbly locus, namely, those stable bundles with a non-zero parabolic nilpotent Higgs field.
Similarly, the strong wobbly locus ${\mathrm{Wo}}_{st}$ satisfies ${\mathrm{Wo}}_{st}={\mathrm{Sh}}_{st}$.
We will prove that ${\mathrm{Wo}}\subset{\mathrm{Sh}_{nilp}}$ and that ${\mathrm{Sh}}\subset{\mathrm{Wo}}$. This proves the statement, as ${\mathrm{Sh}_{nilp}}\subset{\mathrm{Sh}}$.
Let us first check that ${\mathrm{Wo}}\subset{\mathrm{Sh}_{nilp}}$. By Proposition \[prop:vs\_iff\_proper\], if ${\mathcal{E}}\in{\mathrm{Wo}}$, then ${{\mathrm{V}}_{{\mathcal{E}},nilp}}\subset{\mathrm{M}}({\mathcal{P}},\alpha)$ is not proper. So there exists a discrete valuation ring $R$ and a morphism $$\iota_F:C:={\mathrm{Spec}}(R)\longrightarrow {\mathrm{M}}({\mathcal{P}},\alpha)$$ such that the generic point ${\mathrm{Spec}}(k)$ (where $k$ is the fraction field of $R$) maps to ${{\mathrm{V}}_{{\mathcal{E}}}}$ while the closed point $o$ goes to $({\mathcal{F}},\psi)\in {{\overline{{\mathrm{V}}}}_{{\mathcal{E}},nilp}}\setminus{{{\mathrm{V}}_{{\mathcal{E}},nilp}}}$ (cf. [@stacks-project [Tag 0A24](https://stacks.math.columbia.edu/tag/0A24)]). By Lemma \[lm:oVn\_minus\_Vn\], ${\mathcal{F}}$ is unstable.
Now, $\iota_F(C\setminus{\{o\}})=\iota_F({\mathrm{Spec}}(k))$ can be seen as a subset of ${\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp}}$. More precisely, there is a commutative diagram $$\xymatrix{
{\mathrm{Spec}}(k)\ar[d]\ar[r]&{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}_{nilp}}\ar[d]_{\pi_n}\\
{\mathrm{Spec}}(R)\ar[r]_{\iota}\ar@{-->}[ur]^{\hat{\iota}}&{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}.
}$$ Properness of $\pi_n$ implies that $\hat{\iota}$ exists. Clearly, $\hat{\iota}(o)$ belongs to ${\mathrm{Ex}}_{nilp}$, and so $\hat{r}_n(\hat{\iota}(o))\in{\mathrm{Sh}_{nilp}}$. Consider the curve $$\hat{r}_n\circ\hat{\iota}:C\longrightarrow {\mathrm{N}}({\mathcal{P}},\alpha).$$ Since $\iota(C\setminus\{o\})\subset {{\mathrm{V}}_{{\mathcal{E}}}}$, then $\hat{r}_n\circ\hat{\iota}(C\setminus\{o\})=\{{\mathcal{E}}\}$. Separability of ${\mathrm{N}}({\mathcal{P}},\alpha)$ implies that $\hat{r}_n\circ\hat{\iota}(o)={\mathcal{E}}$, and hence ${\mathcal{E}}\in{\mathrm{Sh}_{nilp}}$.
For the converse: let $s\in{\mathrm{Sh}}$. Then, for some $S\in{\mathrm{Ex}}\subset{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}}$, $s=\hat{r}(S)$. By [@PP Lemma 2.4], one may find a smooth curve $Y$ and a morphism $$\xymatrix{j:Y\ar[r]&{\widehat{{\mathrm{M}}({\mathcal{P}},\alpha)}}\\
y\ar@{|->}[r]&({\mathcal{E}}_y,\varphi_y)
}$$ such that $j(Y\setminus\{y_0\})\subset{\mathrm{M}}({\mathcal{P}},\alpha)\setminus{\mathrm{Un}}$, and $j(y_0)=S$. Therefore $$\hat{r}\circ j(Y\setminus\{y_0\})\subset{\mathrm{N}}({\mathcal{P}},\alpha)\setminus {\mathrm{Sh}}, \qquad \hat{r}\circ j(y_0)=s.$$ Consider now the rational map $$\xymatrix{
{\mathbb{C}}\times Y\ar@{-->}[r]^m& {\mathrm{M}}({\mathcal{P}},\alpha)\\
(t,y)\ar@{|->}[r]&t\cdot \pi\circ j(y)
}$$ defined away from $(0,y_0)$. Note that
1. \[it:curves\] $m(t,Y\setminus\{y_0\})\subset{\mathrm{M}}({\mathcal{P}},\alpha)\setminus {\mathrm{Un}}$, as $t\cdot(\pi\circ j(y))=({\mathcal{E}}_y,t\varphi_y)$, for all $y\neq y_0$, where $({\mathcal{E}}_y,\varphi_y):=\pi\circ j(y))$.
2. \[it:curve\_in\_N\_from\_m\] This implies $r(m(t,y))=r(m(0,y))=\hat{r}\circ j (y)$ for all $t\in {\mathbb{C}}$ and all $y\neq y_0$.
Now, by properness of ${\mathrm{N}}({\mathcal{P}},\alpha)$, it follows that ${r}\circ m(y_0)$ exists. By above, separatedness of ${\mathrm{N}}({\mathcal{P}},\alpha)$, and commutativity of , ${r}\circ m(y_0)=s$.
Moreover, by , $m(t,y_0)\in{\mathrm{Un}}$ for all $t\neq 0$. Let $({\mathcal{F}}_0,\psi_0)=\lim_{t\to 0}m(t,y_0)$. Assume that ${\mathcal{F}}_0$ were semistable. Then, so would be the bundle underlying $m(t,y_0)$ for generic $t\in{\mathbb{A}}^1$ for generic $t$ [@Yoko Proposition 1.11]. But $m(t,y_0)=({\mathcal{F}},t\cdot \psi)$ has underlying unstable bundle, which yields a contradiction. Thus, it must be $({\mathcal{F}}_0,\psi_0)\in{\mathrm{Un}}$.
Now, a finite number of blowups allows to resolve the morphism $m$ [@H], yielding $$\hat{m}:\widehat{{\mathbb{C}}\times Y}
\longrightarrow {\mathrm{M}}({\mathcal{P}},\alpha).$$ The exceptional divisor over $(0,y_0)$ is a union of projective lines containing a curve joining $({\mathcal{F}}_0,\psi_0)$ and $({\mathcal{E}},0)$ and contained inside ${{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$ (as it is contained in $h_{{\mathcal{P}},\alpha}^{-1}(0)\subset{{{\mathrm{M}}}({\mathcal{P}},\alpha)_{nilp}}$). An irreducible component $I$ of the exceptional divisor of $\widehat{{\mathbb{C}}\times Y}$ must therefore intersect $\mathrm{N}({\mathcal{P}},\alpha)$ at $({\mathcal{E}},0)$, but also $h^{-1}(0)\setminus{{\mathrm{N}}({\mathcal{P}},\alpha)}$. We claim that $I\cap\mathrm{N}({\mathcal{P}},\alpha)\subset {\mathrm{Wo}}$. Indeed, otherwise, it would intersect the very stable locus (as there are no strictly semistable bundles). These are strongly very stable, which are a dense open set in ${\mathrm{N}}({\mathcal{P}},\alpha)$ by [@L Proposition 3.5]. Moreover, since $I\cap{\mathrm{M}}({\mathcal{P}},\alpha)^0$ is non empty, then it is dense, so one may conclude that bundles in $I$ with underlying strongly very stable bundle are dense. Thus, by irreducibility of $I$, it would be $I\subset{\mathrm{N}}({\mathcal{P}},\alpha)$, contradicting the fact that $I$ intersects the complement of ${\mathrm{N}}({\mathcal{P}},\alpha)$. In particular, $s\in I\cap\mathrm{N}({\mathcal{P}},\alpha)\subset {\mathrm{Wo}}$.
To see that ${\mathrm{Wo}}_{st}={\mathrm{Sh}_{st}}$, we note that the arguments adapt verbatim if we work in the strong nilpotent leaf.
When the moduli space ${\mathrm{M}}({\mathcal{P}},\alpha)$ is not smooth, Hironaka does not apply to the full moduli space. It however does to the reduced schemes underlying all Hitchin fibers. So working the equivalent to Theorem \[thm:W=S\] out requires a finer study of the Hitchin fibration, and the dependence of shakiness on the characteristic. Moreover, in order to compare all loci involved, strictly semitable points need to be tracked and discarded.
Non-parabolic Higgs pairs {#sec:classical}
=========================
This section shows that the results in the parabolic setup imply equivalent ones in the usual (non parabolic) setup.
Let $D\subset X$ be a reduced divisor, possibly zero. A $D$-twisted Higgs pair is $(E,\varphi)$ where $E$ is a vector bundle and $\varphi\in H^0(X,{\mathrm{End}}(E)\otimes K(D))$.
The moduli space ${\mathrm{M}}_D(n,d)$ of $D$-twisted Higgs pairs of rank $n$ and degree $d$ is a quasi-projective variety [@Nitin]. It is Poisson, with symplectic leaves given by the orbits of the residue of the Higgs field along the divisor [@Markman Corollary 8.10], [@Bottacin Theorem 4.7.5 ]. Let ${\mathrm{N}}(n,d)$ denote the moduli space of vector bundles of rank $n$ and degree $d$.
A bundle $E$ is called $D$-very stable if it has no $D$-twisted nilpotent Higgs field. If $D=0$, we just say that $E$ is very stable. The complement of the $D$-very stable locus ${\mathrm{N}}^{D-vs}(n,d)$ inside the stable locus ${\mathrm{N}}^s(n,d)\subset{\mathrm{N}}(n,d)$ is called the $D$-wobbly locus and denoted by $\mathbf{W}_D$.
From Lemma \[lm:vs\_is\_stable\] $D$-very stable bundles are stable. As a corollary to Proposition \[prop:st\_vs\_iff\_proper\] we obtain:
\[cor:usual\_criterion\] Let $E$ be a stable vector bundle. Let $D\subset X$ be a reduced divisor, possibly zero, and let ${\mathrm{V}}_{E,D}:=H^0(X,{\mathrm{End}}(E)\otimes K(D))$, ${\mathrm{V}}_{E,D, n}:=H^0(X,{\mathrm{End}}(E)\otimes K(D))_{nilp}$. Consider the restrictions of the Hitchin map $h_{E,D}:=h|_{V_{E,D}}$ and $h_{E,D,n}:=h|_{V_{E,D,n}}$. Then: $$\begin{array}{lcl}
E \textnormal{ \it is $D$-very stable } &\iff& h_{E,D,n} \textnormal{ \it is finite }\\ &\iff& h_{E,D,n} \textnormal{ \it is quasi-finite }\\ &\iff& h_{E,D,n} \textnormal{ \it is proper }\\ &\iff& h_{E,D} \textnormal{ \it is proper .}
\end{array}$$
Take ${\mathcal{P}}=\{{\mathrm{GL}}(n,{\mathrm{C}})\,:\, x\in D\}$, so that all flags are trivial. In that case strong very stability of a quasi-parabolic bundle ${\mathcal{E}}=(E,\sigma)$ of type ${\mathcal{P}}$ is equivalent to $E$ being $0$-very stable, while very stability of ${\mathcal{E}}$ is $D$-very stability of $E$. Indeed, this follows easily from the fact that $H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))_{nilp}=H^0(X,{\mathrm{End}}(E)\otimes K(D))_{nilp}\supset H^0(X,{\mathrm{End}}(E)\otimes K)=H^0(X,{\mathrm{SParEnd}}({\mathcal{E}})\otimes K(D))$ and preservation of nilpotency under inclusion given the existence of a commutative diagram $$\xymatrix{
H^0(X,{\mathrm{End}}(E)\otimes K)\ar@{^(->}[r]\ar[d]_h&H^0(X,{\mathrm{ParEnd}}({\mathcal{E}})\otimes K(D))_{nilp}\ar[d]^{h_{{\mathcal{E}},nilp}}\\
{\mathrm{H}}_{0}\ar@{^(->}[r]&{\mathrm{H}}_D}.$$ Taking $\alpha=0$, we note that the moduli spaces ${\mathrm{N}}({\mathcal{P}},0)$ and ${\mathrm{M}}({\mathcal{P}},0)_{st}$ are respectively isomorphic to the moduli spaces of vector bundles ${\mathrm{N}}(n,d)$ (where $d$ equals the parabolic degree) and the moduli space of Higgs bundles ${\mathrm{M}}(n,d)$. Likewise ${\mathrm{M}}({\mathcal{P}},0)_{nilp}\cong{\mathrm{M}}_D(n,d)_{nilp}$.
Thus, the statement follows from Proposition \[prop:st\_vs\_iff\_proper\] for $D=0$ and from Proposition \[prop:vs\_iff\_proper\] for $D\neq 0$.
When $D=0$, Corollary \[cor:usual\_criterion\] is [@PP Theorem 1.1].
Denote by ${\mathrm{N}}(n,d)^{D-vs}$ the $D$-very stable locus, and let $\mathrm{W}_D:={\mathrm{N}}(n,d)^s\setminus {\mathrm{N}}(n,d)^{D-vs}$ be the $D$-wobbly locus. From Lemma \[lm:wobbly\], making $\alpha=0$, one gets
Let $C_i\, i\in I$ be the irreducible components of the nilpotent cone of ${\mathrm{M}}(n,d)$, with $C_0={\mathrm{N}}(n,d)$. Then, $\mathbf{W}_0=\bigcup_{i\in I\setminus 0}C_i\cap C_0^s$.
Likewise, ${\hat{r}_s}$ and ${\hat{r}_n}$ in and provide resolutions (when $(n,d)=1$) of the rational maps $$\label{eq:res_usual}
\xymatrix{r_0:{\mathrm{M}}(n,d)\ar@{-->}[r]&{\mathrm{N}}(n,d)}\qquad \xymatrix{r_D:{\mathrm{M}}_D(n,d)_{nilp}\ar@{-->}[r]&{\mathrm{N}}(n,d)}.$$ Then, ${\mathrm{Sh}_{st}}$ and ${\mathrm{Sh}_{nilp}}$ from Definition \[defi:shaky\] are identified with certain loci ${\mathrm{N}}(n, d)$ that we will denote by $\mathbf{S}_0$ and $\mathbf{S}_D$, image of the exceptional divisors over the loci $\mathbf{Un}_0\subset{\mathrm{M}}(n,d)$ and $\mathbf{Un}_D\subset{\mathrm{M}}_D(n,d)_{nilp}$ of Higgs bundles with underlying unstable bundle.
Finally, we also obtain the equivalent of Theorem \[thm:W=S\]. Assume that $(n,d)=1$ (equivalently, the weight $0$ is generic).
\[thm:usual\_W=S\] There are equalities $\mathbf{W}_D=\mathbf{S}_D$.
Strictly speaking, from Theorem \[thm:W=S\], one can deduce that for any resolution of the rational map and obtained by restricting a resolution of , the corresponding strongly shaky locus equals the strongly wobbly locus. We note that since we are only concerned with the images of the exceptional divisors, the same result will hold if instead of a restriction one considers arbitrary resolutions of and by successive blowups along $\mathbf{Un}_D$.
[DDDD]{}
D. Baraglia and M. Kamgarpour, [*On the image of the parabolic Hitchin map*]{}. Quarterly Journal of Mathematics, 69 2: 681–708. doi:10.1093/qmath/hax055.
A. Beauville, M.S. Narasimhan, and S. Ramanan, [*Spectral curves and the generalised theta divisor*]{}, J. Reine Angew. Math. 398 (1989), 169–179.
U. N. Bhosle, [*Parabolic vector bundles on curves*]{}. Ark. Mat. 27 (1989), no. 1, 15–22.
O. Biquard and P. Boalch, [*Wild nonabelian Hodge theory on curves*]{}, Compos. Math. 140 (2004) no. 1, 179–204.
I. Biswas, [*Parabolic bundles as orbifold bundles*]{}, Duke Math. J. Volume 88, Number 2 (1997), 305–325.
P. Boaclh, [*Quasi-Hamiltonian geometry of meromorphic connections*]{}, Duke Math. J. 139 (2007) no. 2, 369–405.
H.U. Boden and K. Yokogawa, [*Rationality of moduli spaces of parabolic bundles*]{}, J. London Math. Soc. (2) 59 (1999) 461–478.
F. Bottacin, [*Symplectic geometry on moduli spaces of stable pairs*]{}, Ann. Scient. Ec. Norm. Sup., 4ème série, t. 28, (1995) 391–433.
W. Chuang, D. Diaconescu, R. Donagi, T. Pantev, [*Parabolic refined invariants and Macdonald polynomials*]{} Preprint <https://arxiv.org/abs/1311.3624>.
R. Donagi and T. Pantev, [*Geometric Langlands and non-abelian Hodge theory*]{}, Surveys in differential geometry, Vol. XIII. Geometry, analysis, and algebraic geometry: forty years of the Journal of Differential Geometry, 85–116, International Press, Somerville, MA, 2009.
R. Donagi and T. Pantev, [*Parabolic Hecke eigensheaves*]{}. Preprint, <https://arxiv.org/abs/1910.02357>.
E.Franco, P.B. Gothen, A.G. Oliveira and A.Peón-Nieto, [*Unramified covers and branes on the Hitchin system*]{}, Preprint, <https://arxiv.org/pdf/1802.05237.pdf>.
V.G. Drinfeld, Letter to P.Deligne, 22 june 1981.
M. Groechenig, [*Hilbert schemes as moduli of Higgs bundles and local systems*]{}, Int. Math. Res. Not. IMRN 2014, no. 23, 6523-6575.
R. Hartshorne, [*Algebraic Geometry*]{}, GTM 52 (1977), Springer-Verlag.
J. Heinloth, [*The intersection form on moduli spaces of twisted $PGL_n$ Higgs bundles vanishes*]{}, Math. Ann. 365 (2016) no. 3–4, 1499–1526.
H. Hironaka, [*Resolution of singularities of an algebraic variety over a field of characteristic zero: I*]{}, Ann. Math, 79 (1964), 109–203.
G. Laumon, [*Un analogue global du cône nilpotent*]{}, Duke Math. Jour. 57 (1988), 647–671.
M. Logares and J. Martens. [*Moduli of parabolic Higgs bundles and Atiyah algebroids*]{} Journal für die reine und angewandte Mathematik (2010): 89–116.
E. Markman, [*Spectral curves and integrable systems*]{}, Compositio Mathematica, tome 93, n. 3 (1994), 255–290.
M. Maruyama and K. Yokogawa, [*Moduli of parabolic stable sheaves*]{}, Math. Ann. 293 (1992), 77–99.
Mehta and Seshadri, [*Moduli of vector bundles on curves with parabolic structures*]{}, Math. Ann. 248 (1980), no. 3, 205–239.
N. Nitsure, Moduli space of semistable pairs on a curve, Proc. London Math. Soc. (3) 62 (1991), no. 2, 275–300.
S. Pal and C. Pauly, [*The wobbly divisors of the moduli space of rank two vector bundles*]{}. Preprint, ArXiv:1803.11315.
C. Pauly and A. Peón-Nieto, [*Very stability and properness of the Hitchin map*]{}, A. Geom Dedicata (2018). <https://doi.org/10.1007/s10711-018-0333-6>.
C. Simpson, [*Harmonic bundles on non compact curves*]{}, J. Amer. Math. Soc., Vol. 3, No. 3 (Jul., 1990), pp. 713–770.
C. Simpson, [*Higgs bundles and local systems*]{}, Publ. Math.I.H.É.S., tome 75 (1992), 5–95.
, <https://stacks.math.columbia.edu/>.
X. Su, B. Wang, X. Wen, [*Parabolic Hitchin Maps and Their Generic Fibers*]{}. Preprint: arXiv:1906.04475 \[math.AG\].
M. Thaddeus, [*Variation of moduli of parabolic Higgs bundles*]{}, J. reine angew. Math. 547 (2002), 1–14.
K. Yokogawa, [*Compactification of moduli of parabolic sheaves and moduli of parabolic Higgs sheaves*]{}, J. Math. Kyoto Univ. Vol. 33, n. 2 (1993), 451–504.
K. Yokogawa, [*Infinitesimal deformation of parabolic Higgs sheaves*]{}, Internat. J. Math. 6, no. 1 (1995),125–148.
H. Zelaci, [*On very stability of principal $G$-bundles*]{}m Geom. Dedicata (2020) 204: 165. <https://doi.org/10.1007/s10711-019-00447-z>.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Bacteria typically have a few thousand different proteins. The number of proteins with a given charge is a roughly Gaussian function of charge — centered near zero, and with a width around ten (in units of the charge on the proton). We have used the charges on E. coli’s proteins to estimate the changes in the second virial coefficients of all its proteins as the concentration of a 1:1 salt is increased. The second virial coefficient has dimensions of volume and we find that on average it decreases by about twice the average volume of a protein when the salt concentration is increased from $0.2$ to 1 Molar. The standard deviation of the decrease is of the same order. The consequences of this for the complex mixture of proteins inside an E. coli cell, are briefly discussed.'
author:
- 'Richard P. Sear'
title: 'The effects of added salt on the second virial coefficients of the complete proteome of [*E. coli*]{}'
---
Introduction
============
The genomes of a number of organisms are already known and more are being completed at a rate of perhaps one a month. Once a genome [@nomen] has been sequenced the amino-acid sequences of [*all*]{} its proteins are known. The complete set of proteins of an organism is generally referred to as its proteome. Here we use genome data for [*E. coli*]{} to systematically estimate the change in the interactions of all the proteins of this bacterium when the salt concentration is varied. [*E. coli*]{} can grow in environments with a very wide range of salt concentrations [@cayley91], and so its proteins must function [*in vivo*]{} over a wide range of salt concentrations. The potassium ion concentration inside the cell can vary from approximately $150$ to $300$ mM [@cayley91]; potassium is the predominant cation in living cells. Clearly, the proteins must remain soluble over this range, and they should bind to the other proteins which they are required to bind to in order to function, but they should not interact strongly with other proteins. The study of a proteome is often called proteomics. Here, as the physical properties of proteins are studied (as opposed to their chemical properties such as catalytic function) we are doing what may be called physical proteomics.
There has been extensive theoretical work on the salt dependence of the interactions in individual proteins, particularly for the protein lysozyme [@poon00; @warren02; @muschol97]. See Refs. for corresponding experimental work. However, as far as the author is aware, this is the first attempt to characterise the interactions of [*all*]{} the proteins of an organism. We will consider the proteins separately, i.e., as single component solutions. Of course inside a bacterium the proteins exist as a mixture of thousands of components. Future work will consider multi-component mixtures of the proteins. We have chosen [*E. coli*]{} as it is a bacterium, and therefore a relatively simple organism, and as it has been extensively studied. However, the distribution of charges on the proteins of almost all organisms is very similar and so our results apply to almost all organisms, including [*H. sapiens*]{}. The only exceptions are some extremophiles [@runcong01].
In the next section we use genome data to estimate the charges on the proteins of [*E. coli*]{}. This data is used in the third section where we calculate the variation in their second virial coefficients as the salt concentration is varied. The last section is a conclusion.
The charges on proteins of E. coli
==================================
[*E. coli K-12*]{} has a proteome of 4358 proteins. The amino acid sequences of all of them are known from the sequencing of its genome [@coligen; @ebi]. [*K-12*]{} is the name of a strain of [*E. coli*]{}. Runcong and Mitaku have analysed the charge distributions of a number of organisms using a simple approximate method of estimating the charge on a protein at neutral pH from its amino-acid sequence [@runcong01]. We will follow their analysis but use a slightly different approximation for the charge on a protein with a given amino-acid sequence.
Of the 20 amino acids, 5 have pK values such that they should be at least partially charged at neutral pH [@thecell; @stryer]. These are two highly acidic amino acids, aspartic acid and glutamic acid, two highly basic amino acids, lysine and arginine and one somewhat basic amino acid, histidine. Aspartic and glutamic acids have pK’s far below 7 and lysine and arginine have pK’s far above 7 and so we assume that all 4 of these amino acids are fully charged at neutral pH. Aspartic and glutamic acids then each contribute $-1$ to the charge on a protein, and lysine and arginine each contribute $+1$. Histidine has a pK of around 6-6.5 (this will depend on the environment of the amino acid). The equation for the fraction $f$ of a basic group such as histidine that is charged at a given pH is $$f=1/\left(1+10^{pH-pK}\right),$$ where pK is the pK value for the basic group. This equation is just the Henderson-Hasselbalch equation [@stryer] rearranged. Taking pK$=6.5$ [@stryer] and at pH=7 we have that the fraction of histidines charged is $f=0.24$. As this is small we assume for simplicity that all the histidine amino acids are uncharged.
Thus, with these assumptions for the charges on these 5 amino acids, our estimate for the net charge $Q$ on a protein is simply given by $$Q=n_K+n_R-n_D-n_E
\label{qdef}$$ where $n_K$, $n_R$, $n_D$ and $n_E$ are the the protein’s total numbers of lysines, arginines, aspartic acids and glutamic acids, respectively. The subscripts $K$, $R$ etc. correspond to the standard single letter codes for the amino acids [@stryer; @thecell]. The charge $Q$ is in units of $e$ where $e$ is the elementary charge.
Note that Runcong and Mitaku [@runcong01] assume that the histidine amino acids contribute $+1$ to the charge, that is the only difference between our analysis and that of Runcong and Mitaku. As the histidine amino acid is quite a rare amino acid, approximately 1 in 50 amino acids is a histidine, the difference between the results we obtain and those of Runcong and Mitaku [@runcong01] is not large but our charges are shifted to more negative values. Using Runcong and Mitaku’s approximation the mean charge on a protein is $7.11$ units more positive than the mean charge we find here. As a check on our algorithm, we can compare the prediction of equation (\[qdef\]) for chicken lysozyme to that of a titration experiment to determine the charge. Equation (\[qdef\]) predicts that chicken lysozyme [@aalys] has a net charge of 8 at neutral pH. Titration experiments on lysozyme give a titratable charge of close to $8.5$ at pH=7 [@lysoq].
Using equation (\[qdef\]) we can obtain estimates for the charges of all 4358 proteins of [*E. coli*]{} [@coligen; @ebi; @gerstein98; @mitaku99]. The results are shown in Fig. \[qcoli\], where we have plotted the number of proteins $n$ as a function of net charge $Q$. The distribution is centered almost at a net charge $Q=0$, and for not-too-large $|Q|$ the distribution is roughly symmetric and Gaussian. The mean charge is $-3.15$. Given the approximate nature of our equation for the charge on a protein, equation (\[qdef\]), the data is probably consistent with a mean charge of 0. The approximation scheme of Mitaku and Runcong [@runcong01] yields a mean charge of $+3.96$. Also, although when $|Q|$ is not too large the distribution can be seen to be reasonably symmetric, [*E. coli*]{} has 12 proteins with charges $<-50$ but none with charges $>+50$. Excluding proteins with very large charges, $|Q|>30$, the root mean square charge equals $9.1$.
A number of other organisms, both other bacteria and eukaryotes such as yeast, have had the charge distribution on their proteomes determined by Runcong and Mitaku [@runcong01] and by the author [@unpub]. Almost all of them have a roughly Gaussian distribution centered approximately at zero, like the distribution in Fig. \[qcoli\]. The exceptions are some extremophiles. Extremophiles are organisms that live in extreme environments, for example [*Halobacterium sp.*]{} lives in environments with very high levels of salt [@microbio]. The cytosol of [*Halobacterium sp.*]{} contains much higher levels of potassium ions than do other organisms so perhaps it is not a surprise that the distribution of charges on its proteins is different.
We have fitted the Gaussian function $$n(Q)=\frac{1739}{\sigma}\exp(-(Q-{\overline Q})^2/2\sigma^2).
\label{pq}$$ to the data for the number of proteins as a function of their charge. It is drawn as the solid curve in Fig. \[qcoli\]. The fit parameters are mean charge ${\overline Q}=-2.16$ and standard deviation $\sigma=8.32$. 1739 is $4358/(2\pi)^{1/2}$ and so the distribution is normalised so that its integral gives the total number of proteins. Within a couple of standard deviations of the mean the Gaussian function fits the data well but it underestimates the numbers of proteins with charges such that $|Q-{\overline Q}|$ is several times the standard deviation.
We also note that there is a correlation between the net charge $Q$ on a protein and its size, measured by the number of amino acids $M$. Figure \[scat\_coli\] is a scatter plot of charge $Q$ and number of amino acids $M$ for the proteins of [*E. coli*]{}. Although at any particular size $M$ there is a wide distribution of charges, on average the more highly charged proteins are larger than average. We expect the volume of a protein to scale with $M$.
Salt dependence of the second virial coefficients
=================================================
Consider a dilute solution of a single one of the proteins of [*E. coli*]{}. Apart from water, the only other constituents are a 1:1 salt at a concentration $c_s$ and a buffer which controls the pH while making a negligible contribution to the ionic strength. Here we will always assume the pH=7 but other pH’s can be considered if the net charges on the proteins can be calculated. Also, the counterions of the protein are assumed the same as either the anions or cations of the salt, depending on the sign of $Q$. The interactions between the protein molecules in the salt solution can be characterised by means of the protein’s second virial coefficient $B_2$: a function of temperature, pH and salt concentration.
Proteins are complex molecules and we are unable to calculate from first principles the absolute value of $B_2$ for any of the 4358 proteins possessed by [*E. coli*]{}. However, predicting the change in the second virial coefficient when the salt concentration varies is a much easier problem, [*if*]{} we assume that changing the salt concentration changes only the direct electrostatic interaction between the net charges of a protein. This is a strong assumption but studies of the simple protein lysozyme have shown that the variation of its second virial coefficient can be described using a simple model which only includes its net charge [@poon00; @warren02]. Here we will follow Warren [@warren02] and apply his analysis of lysozyme to the complete set of proteins of [*E. coli*]{}. We will discuss which proteins are likely to be less well described by this theory than is lysozyme.
A protein molecule of charge $Q$ is surrounded by its counterions and as the concentration of the protein is increased so is the counterion density. This increase in the counterion density decreases the translational entropy of the counterions and this contributes a positive amount to the second virial coefficient. See Warren [@warren02] and references therein for details. $B_2$ has the form [@warren02] $$B_2=B_2^{(ne)}+\frac{Q^2}{4c_s},
\label{donnan}$$ where $B_2^{(ne)}$ is an assumed constant term due to excluded volume interactions and other interactions which are insensitive to salt concentration. The second term is from the counterions and the salt. It is quadratic in the charge and so of course is zero for uncharged proteins and is independent of the sign of the net charge on a protein.
As stated above we are unable to calculate the absolute value of $B_2$; the values of $B_2^{(ne)}$ of the proteins are unknown. However, we can calculate the difference in $B_2$ when the salt concentration is changed from $c_s=c_1$ to $c_s=c_2$. It is $$\Delta B_2=\frac{Q^2}{4}\left(\frac{1}{c_2}-\frac{1}{c_1}\right).
\label{dond}$$ This is easy to calculate for any protein and in Fig. \[db2\] we have plotted the number of proteins $n$ as a function of the change in their second virial coefficient, $\Delta B_2$, when the salt concentration is decreased from 1M to $0.2$M. The results are given in units of nm$^3$. For comparison the volume of a typical bacterial protein is about 60nm$^3$ and so if a protein were to interact solely via a hard repulsion it would have a second virial coefficient of about 4 times its volume or about 240nm$^3$.
Results for proteins with $-30\le Q\le30$ are shown. Proteins with larger titratable charges are likely to have an effective charge lower than $Q$, see Refs. and references therein. From the linear Poisson-Boltzmann equation, the potential (divided by $e$) at the surface of a spherical particle with charge $Q$ and radius $a$ is $Q\lambda_BkT/((1+\kappa a)a)$. $\lambda_B=e^2/(4\pi\epsilon kT)$ is the Bjerrum length, and and $\kappa^{-1}$ is the Debye screening length, given by $\kappa^2=8\pi\lambda_Bc_s$. For the dielectric constant of water 80 times that in vacuum and at room temperature, $\lambda_B=0.7$nm. Globular proteins are approximately spherical and typically have radii around 2 to 4nm. Taking a protein with a radius of 3nm, in salt at a concentration $c_s=0.1$M, we have that for $Q=30$, the potential at the surface is about $2kT$. Larger charges correspond to larger surface potentials and these large potentials bind oppositely charged ions to the surface reducing the effective charge. On average, this effect will be diminished to a certain extent by the fact that the most highly charged proteins are larger than average. See Fig. \[scat\_coli\], where it is clear that the charge and size of a protein are correlated. Recent simulations by Lobaskin [*et al.*]{} [@lobaskin01] of spheres with radius 2nm and charge $Q=-60$ in the absence of salt found an effective charge of a little under $-20$. Thus we restrict ourselves to proteins with charges of magnitude less than or equal to 30. 4300 of the 4358 proteins, or almost 99%, have charges in this range. The mean change in $B_2$ of these 4300 proteins when the salt concentration is decreased from 1 to $0.2$M is 139nm$^3$ and the standard deviation is 218nm$^3$.
A couple of caveats. The first is that the effect of salt on protein solutions is known to depend not only on whether the salt is a 1:1 salt, a 1:2 salt etc. but also to the nature of ions, whether it is Mg$^{2+}$ or Ca$^{2+}$ for example [@durbin96]. Our generic theory applies only where there are no specific interactions between the salt and the protein. There is good agreement between experiment and theory for lysozyme plus NaCl [@warren02; @poon00] and so we may hope that it applies to NaCl and many proteins but it clearly misses potentially important effects for other salts where there are specific protein-salt interactions. The second is that proteins are not simple charged spheres, for example some have large dipole moments. Dipoles exert net attractions which are screened and hence weakened by added salt. Thus proteins with a small charges but large dipole moments are poorly described by the current theory: if the dipole interactions are dominant then the second virial coefficient may even increase when the salt concentration is increased. Velev [*et al.*]{} [@velev98] discuss this point. Note that although we can estimate the charge on a protein from its amino-acid sequence we cannot estimate its dipole moment without knowing its three-dimensional structure, and so the sequence data from genomics is not adequate to determine dipole moments.
Conclusion
==========
Here we have shown how data from genomics can be used to estimate the charges on the proteins of an organism. We then used these charges to estimate the changes in the second virial coefficients of 4300 (99%) of the proteins of [*E. coli*]{} when the salt concentration is changed. Note that [*E. coli*]{} can survive and multiply in external environments with a very wide range of salt concentrations; Cayley [*et al.*]{} [@cayley91] studied the growth of [*E. coli*]{} in environments with salt concentrations ranging from very low to $0.5$ Molar, corresponding to potassium ion concentrations inside the cell of $150$ to $300$ mM [@cayley91]. Thus, studying the change in interactions of proteins with salt concentration is of direct relevance to the [*in vivo*]{} behaviour of proteins. Within molecular biology there is a clear shift of emphasis away from studying the proteins of an organism one or a few at a time, and towards determining the structure and function of large sets of proteins, in particular proteomes. The systematic study of these large sets of proteins is often called proteomics. This work is a first attempt to keep up with this shift by performing a simple theoretical calculation of a solution phase physical property for a complete proteome, rather than for one or a handful of proteins as is usually done. It may be termed physical proteomics.
Charged species $Q$ n Charge density in cytosol (mM)
----------------- ---------------- --------------- --------------------------------
Protein $-21$ to $+15$ $2\times10^6$ $-10$
tRNA $-80$ $2\times10^5$ $-30$
mRNA $-2000$ $1\times10^3$ $-3$
Ribosome $-3000$ $2\times10^4$ $-100$
DNA $-10^{7}$ 1 $-20$
: The charged macromolecular species in the cytosol of [*E. coli*]{}. The data is from Neidhardt [@neidhardt]. $Q$ is the charge on a macromolecule. For proteins the range given is the mean plus and minus twice the standard deviation. $n$ is the total number of molecules of a species per cell. The volume of an [*E. coli*]{} cell is about $10^{-18}$m$^3$ so 1 molecule per cell corresponds to a concentration of about $2\times 10^{-9}$ Molar. A prokaryote ribosome consists of about 4500 bases of RNA plus protein. The ribosomal proteins are mostly quite stongly positively charged and so will decrease the net negative charge. As a rough estimate we settle on a net charge of $-3000$. The charge on a tRNA molecule is around $-80$, $-1$ from each of its bases [@stryer; @thecell]. The charge on a mRNA molecule is on average around $-2000$ [@neidhardt]. The charge on DNA is equal to twice the number of base pairs, 4,639,221 for [*E. coli K-12*]{} [@coligen]. The charge density from the proteins assumes that the proteins have the mean charge of $-3$ that they would have if their density and charge were uncorrelated. \[t2\]
Future work could consider mixtures of proteins, ultimately aiming to understand the cytosol of a living cell, which is a mixture of of order $10^3$ different types of proteins as well as DNA, RNA, ions like ATP and potassium, etc.. This is of course very complex but [*if*]{} in the cytosol the proteins of [*E. coli*]{} are present in amounts which are uncorrelated with their net charge, the mean charge of the proteins will be close to the mean of the distribution of Fig. \[qcoli\]. This is quite small. Neidhardt [@neidhardt] has taken an inventory of the species inside [*E. coli*]{}, and the results for charged macromolecules are shown in Table \[t2\]. The charged macromolecules in a cell are protein, DNA and the various forms of RNA: transfer RNA (tRNA), messenger RNA (mRNA) and the RNA in ribosomes (rRNA). See Refs. for an introduction to the proteins, DNA and RNA. Although for every molecule of tRNA molecule there are 10 of protein, for every ribosome there are 100 molecules of protein, and for every mRNA there are 1000 molecules of protein, the contributions of the tRNA, ribosomes, DNA and proteins, to the overall charge density of the macromolecules are very roughly comparable. The ribosomes contribute the largest amount. The macromolecules are negatively charged and this negative charge is balanced by potassium ions [@thecell]. Thus the cytosol resembles a solution of a negatively charged polyelectrolyte, except that there is not one, relatively simple, macromolecular species, but thousands of rather complex and diverse species of macromolecules.
Acknowledgements {#acknowledgements .unnumbered}
================
It is a pleasure to acknowledge discussions with J. Cuesta, D. Frenkel and P. Warren.
[99]{}
A brief introduction to the biological nomenclature: an organism’s DNA contains many genes, each of which codes for a protein. The complete set of genes is called the organism’s genome and we will refer to the complete set of proteins as its proteome. Some authors use the word proteome somewhat differently, they use it to denote the set of proteins present in the cytosol of an organism at a particular time. See for example [@thecell] for a more detailed definition of a genome.
S. Cayley, B. A. Lewis, H. J. Guttman and M. T. Record Jr, J. Mol. Biol. [**222**]{}, 281 (1991).
W. C. K. Poon, S. U. Egelhaaf, P. A. Beales, A. Salonen and L. Sawyer, J. Phys. Cond.: Matt. [**12**]{}, L569 (2000).
M. Muschol and F. Rosenberger, J. Chem. Phys. [**107**]{}, 1953 (1997).
P. B. Warren, J. Phys.: Cond. Matt. [**14**]{}, 7617 (2002).
O. D. Velev, E. W. Kaler and A. M. Lenhoff, Biophys. J. [**75**]{}, 2682 (1998).
B. Guo, S. Kao, H. McDonald, A. Asanov, L. L. Combs and W. W. Wilson, J. Cryst. Growth [**196**]{}, 424 (1999).
D. F. Rosenbaum, A. Kulkarni, S. Ramakrishnan and C. F. Zukoski, J. Chem. Phys. [**111**]{}, 9822 (1999).
R. Piazza and M. Pierno M, J. Phys.: Cond. Matt. [**12**]{}, A443 (2000).
K. Runcong and S. Mitaku, Genome Informatics [**12**]{}, 364 (2001).
F. R. Blattner [*et al.*]{}, Science [**277**]{}, 1453 (1997).
The complete proteome of [*E. coli K-12*]{}, i.e., the amino-acid sequences of all its proteins, can be downloaded from databases such as that at the European Bioinformatics Institute (http://www.ebi.ac.uk/proteome). [*E. coli K-12*]{} was sequenced by Blattner [*et al.*]{} [@coligen].
B. Alberts, D. Bray, J. Lewis, M. Raff, K. Roberts and J. D. Watson, Molecular Biology Of The Cell (3rd Edition, Garland Publishing, New York, 1994).
L. Stryer, Biochemistry (4th Edition, Freeman, New York, 1995).
The sequence used to determine this charge has a Protein Data Bank ID [@pdb] of 1GWD.
This value is obtained from Fig. 1 of C. Tanford and R. Roxby, Biochemistry [**11**]{}, 2192 (1972).
M. Gerstein and H. Hegyi, FEMS Microbiology Rev. [**22**]{}, 277 (1998).
S. Mitaku, M. Ono, T. Hirokawa, S. Boon-Chieng and M. Sonoyama, Biophysical Chem. [**82**]{}, 165 (1999).
R. P. Sear, unpublished work.
M. T. Madigan, J. M. Martinko and J. Parker, Brock Biology of Microorganisms (Prentice Hall, New Jersey, 1997).
V. Lobaskin, A. Lyubartsev and P. Linse, Phys. Rev. E [**63**]{}, 020401 (2001).
P. B. Warren, J. Chem. Phys. [**112**]{}, 4683 (2000).
S. D. Durbin and G. Feher, Ann. Rev. Phys. Chem. [**47**]{}, 171 (1996).
F. C. Neidhardt, Chemical composition of [*E. coli*]{}, in [*[*E. coli*]{} and [*S. typhimurium*]{}: Cellular and Molecular Biology*]{} edited by F. .C. Neidhardt J. L. Ingraham, K. B. Low, B. Magasanik and H. E. Umbarger (American Society for Microbiology, Washington D. C., 1987).
H. M. Berman, J. Westbrook, Z. Feng, G. Gilliland, T. N. Bhat, H. Weissig, I. N. Shindyalov and P. E. Bourne, Nucleic Acids Res. [**28**]{}, 235 (2000). Web site: http://www.rcsb.org/pdb/.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
This paper investigates the location of the zeros of a sequence of polynomials generated by a rational function with a binomial-type denominator. We show that every member of a two-parameter family consisting of such generating functions gives rise to a sequence of polynomials $\{P_{m}(z)\}_{m=0}^{\infty}$ that is eventually hyperbolic. Moreover, the real zeros of the polynomials $P_{m}(z)$ form a dense subset of an interval $I\subset\mathbb{R}^{+}$, whose length depends on the particular values of the parameters in the generating function.\
author:
- |
Tamás Forgács\
Khang Tran
title: Polynomials with rational generating functions and real zeros
---
Introduction
============
The study of the location of zeros of polynomials is one of the oldest endeavors in mathematics. The prolific mathematical production of the nineteenth century included a number of advances in this endeavor. The unsolvability of the general quintic equation together with the fundamental theorem of algebra led to the consideration that when it comes to extracting information about the zeros of a complex polynomial from its coefficients, one should perhaps strive to determine subsets of $\mathbb{C}$ where the zeros must lie[^1], rather than looking for the exact location of the zeros. The development of the Cauchy theory for analytic functions provided some of the classic machinery suitable for such investigations, including Rouché’s theorem, the argument principle and the Routh-Hurwitz condition for left-half plane stability. We mention these results not only because they are powerful tools, but also because they embody a fundamental dichotomy. Explicit criteria for the location of the zeros of a polynomial in terms of its coefficients may severely restrict the domain to which they apply, whereas ubiquitous applicability of a theorem to various domains may render the result difficult to use.\
The Gauss-Lucas theorem, relating the location of the zeros of $p'(z)$ to those of the polynomial $p(z)$, pioneered a new approach to an old question: instead of studying the zeros of a function, one can study the behavior of the zero set of a function under certain operators. In this light, given the Taylor expansion of a (real) entire function ${\displaystyle {f(x)=\sum_{k=0}^{\infty}\frac{\gamma_{k}}{k!}x^{k}}}$, one can interpret $f(x)$ as the result of the sequence $\left\{ \gamma_{k}\right\} _{k=0}^{\infty}$ acting on the function $e^{x}$ by forming a Hadamard product[^2]. Thus, complex sequences have a dual nature: they are coefficients of ‘polynomials of infinite degree’ (á la Euler), as well as linear operators on $\mathbb{C}[x]$. G. Pólya and J. Schur’s 1914 paper [@ps] was a major mile stone in understanding how sequences (as linear operators) affect the location of the zeros of polynomials. More precisely, Pólya and Schur gave a classification of real sequences that preserve reality of zeros of real polynomials, and initiated a research program on stability preserving linear operators on circular domains, which was recently completed by J. Borcea and P. Brändén ([@BB]).\
Since the work of Pólya and Schur, the study of sequences as linear operators on $\mathbb{R}[x]$ has attracted a great amount of attention. We remark here only that real sequences, when looked at as operators on $\mathbb{R}[x]$, admit a representation as a formal power series $$\left\{ \gamma_{k}\right\} _{k=0}^{\infty}\sim\sum_{k=0}^{\infty}Q_{k}(x)D^{k},$$ where $D$ denotes differentiation, and the $Q_{k}(x)$s are polynomials with degree $k$ or less (see for example [@peetre]). In [@tomandrzej] the first author and A. Piotrowski study the extent to which the ‘generated’ sequence $\left\{ Q_{k}(x)\right\} _{k=0}^{\infty}$ encodes the reality preserving properties of the sequence $\{\gamma_{k}\}_{k=0}^{\infty}$, and find that if this latter sequence is a Hermite-diagonal[^3] reality preserving operator, then all of the $Q_{k}(x)$s must have only real zeros.\
The present paper extends the works of S. Beraha, J. Kahane, and N. J. Weiss ([@bkw]), A. Sokal ([@sokal]) and K. Tran ([@tran], [@tran-1]), by studying a large family of generating functions which give rise to sequences of polynomials with only real zeros. Our main result (see Theorem \[maintheorem\]) concerns a sequence of polynomials, whose generating function is rational with a binomial-type denominator.
\[maintheorem\] Let $n,r\in\mathbb{N}$ such that $\max\{r,n\}>1$, and set $D_{n,r}(t,z):=(1-t)^{n}+zt^{r}$. For all large $m$, the zeros of the polynomial $P_{m}(z)$ generated by the relation $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{D_{n,r}(t,z)}\label{genfunc}$$ lie on the interval $$I=\left\{
\begin{array}{cc} (0,\infty) & \mbox{ if }n,r\ge2\\
(0,n^{n}/(n-1)^{n-1}) & \text{if} r=1\\
((r-1)^{r-1}/r^{r},\infty) & \text{if} n=1
\end{array}\right.$$ Furthermore, if $\mathcal{Z}(P_{m})$ denotes the set of zeros of the polynomial $P_{m}(z)$, then $\bigcup_{m\gg1}\mathcal{Z}(P_{m})$ is dense in $I$.
Although this result is asymptotic in nature, we do believe that in fact all of the generated polynomials have only real zeros. Given that we have no proof of this claim at this time, we pose this stronger statement as an open problem (see Problem \[prob:1\] in Section \[open\]).\
We close this introduction by noting that if $A(z)$ and $B(z)$ are any two (non-zero) polynomials with complex coefficients, then setting $n=1$ and replacing $z$ by $(-1)^{r}A(z)/B(z)^{r}$ and $t$ by $-B(z)t$ in Theorem \[maintheorem\] reproduces the main result in a recent paper by the second author:
Let $P_{m}(z)$ be a sequence of polynomials whose generating function is $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{1+B(z)t+A(z)t^{r}},$$ where $A(z)$ and $B(z)$ are polynomials in $z$ with complex coefficients. There is a constant $C=C(r)$ such that for all $m>C$, the roots of $P_{m}(z)$ which satisfy $A(z)\neq0$ lie on a fixed curve given by $$\Im\left(\frac{B^{r}(z)}{A(z)}\right)=0\qquad\mbox{and}\qquad0\le(-1)^{r}\frac{B^{r}(z)}{A(z)}\le\frac{r^{r}}{(r-1)^{r-1}}$$ and are dense there as $m\to\infty$.
The rest of the paper is organized as follows. In Section \[preliminaries\] we present all the preliminary results needed for the proof of Theorem \[maintheorem\]. Along with a number of technical lemmas, we prove a key proposition (Proposition \[lem:zerosQ\]) concerning estimates for the relative magnitudes of the zeros (in $t$) of $D(t,z)$. Section \[proofmainthm\] is dedicated to the proof of Theorem \[maintheorem\]. Before presenting the proof, we illustrate the techniques to be employed by proving a slightly stronger result for small $n$ and $r$ (Proposition \[illustration\]). The paper concludes with Section \[open\], where we list some open problems related to our investigations.
Preliminaries
=============
We establish Theorem \[maintheorem\] by showing that for large $m$, the polynomials $P_{m}(z)$ have at least as many zeros on $I$ as their degree, and since $I\subset\mathbb{R}$, all zeros of $P_{m}(z)$ must be real when $m\gg1$.Thus, we lose no generality (in retrospect) by assuming that $z\in\mathbb{R}$, which we shall do for the remainder of the paper. We start our investigations with the following
\[degree\] Suppose that the sequence of polynomials ${\displaystyle {\left\{ P_{m}(z)\right\} _{m=0}^{\infty}}}$ is generated by (\[genfunc\]). Then $\deg P_{m}(z) \leq \lfloor m/r\rfloor$ for all $m$.
Rearranging (\[genfunc\]) yields the equation $$((1-t)^{n}+zt^{r})\sum_{m=0}^{\infty}P_{m}(z)t^{m}=1.$$ By equating coefficients we see that the polynomial $P_{m}(z)$ satisfies the recurrence $$((1-\Delta)^{n}+z\Delta^{r})P_{m}(z)=0,\label{eq:recurrence}$$ where the operator $\Delta$ is defined by $\Delta P_{m}:=P_{m-1}$, and $\Delta P_{0}=0$. The claim follows.
Since we are interested in various $z\in\mathbb{R}$ as potential zeros of $P_{m}(z)$, we seek to understand how $z$ and the zeros of $D_{n,r}(t,z)$ (when seen as a polynomial in $t$) correlate. The next result sheds some light on this question.
Let $n,r\in\mathbb{N}$ be such that $\max\{n,r\}>1$. Suppose $z\in\mathbb{R}$, and that $t=|t|e^{-i\theta}$ is a zero of $D_{n,r}(t,z)$. Then the equation $$z=\frac{\sin^{n}\theta}{\sin^{n-r}(\phi-\theta)\sin^{r}\phi}\label{eq:ztheta}$$ holds, where ${\displaystyle {\phi=\frac{(n-1)\pi+r\theta}{n}}}$.
Suppose $z\in\mathbb{R}$. Then $D_{n,r}(t,z)\in\mathbb{R}[t]$, and consequently if $t=|t|e^{-i\theta}$, $\theta\in\mathbb{R}$, is a zero of $D_{n,r}(t,z)$, then so is $|t|e^{i\theta}=te^{2i\theta}$. Rearranging $D_{n,r}(t,z)=D_{n,r}(te^{2i\theta},z)$ yields the equation $$\left(\frac{te^{2i\theta}-1}{t-1}\right)^{n}=e^{2ir\theta},$$ one of whose solutions is $$t=\frac{1-e^{2\pi i(n-1)/n}e^{2ir\theta/n}}{e^{2i\theta}-e^{2\pi i(n-1)/n}e^{2ir\theta/n}}.\label{tell}$$ We set ${\displaystyle {\phi=\frac{(n-1)\pi+r\theta}{n}}}$ and note that $\phi-\theta\neq0$. Hence the substitution ${\displaystyle {e^{2\pi i(n-1)/n}e^{2ir\theta/n}=e^{2i\phi}}}$ in equation (\[tell\]) gives $$t=\frac{1-e^{2i\phi}}{e^{2i\theta}-e^{2i\phi}}=e^{-i\theta}\frac{e^{-i\phi}-e^{i\phi}}{e^{i(\theta-\phi)}-e^{-i(\theta-\phi)}}=\frac{\sin\phi}{\sin(\phi-\theta)}e^{-i\theta},\label{eq:t0form}$$ and consequently $$(1-t)^{n}=\frac{(-\cos\phi\sin\theta+i\sin\phi\sin\theta)^{n}}{\sin^{n}(\phi-\theta)}=(-1)^{n}\frac{\sin^{n}\theta e^{-in\phi}}{\sin^{n}(\phi-\theta)}=-\frac{\sin^{n}\theta e^{-ir\theta}}{\sin^{n}(\phi-\theta)}.$$ Solving $D_{n,r}(t,z)=0$ for $z$ thus gives $$z=-\frac{(1-t)^{n}}{t^{r}}=\frac{\sin^{n}\theta}{\sin^{n-r}(\phi-\theta)\sin^{r}\phi}.\label{eqn:zintheta}$$ The proof is complete.
We note that equation (\[eq:ztheta\]) defines a smooth curve in $(0,\pi/r)\times\mathbb{R}^{+}$ which, as we shall see, contains at least $\lfloor m/r\rfloor$ many points with distinct second coordinates that are all zeros of $P_{m}(z)$ if $m$ is large. For the graphs of $z$ as a function of $\theta$ when $n=3$, $r=1$, and $r=2$, see Figure \[fig:zthetafig\].
![\[fig:zthetafig\]The curve $z(\theta)$ when $n=3,r=1$ (left), and $n=3,r=2$ (right)](n3r1 "fig:")![\[fig:zthetafig\]The curve $z(\theta)$ when $n=3,r=1$ (left), and $n=3,r=2$ (right)](n3r2 "fig:")
We next give the proof of two key properties of the curve defined by (\[eq:ztheta\]).
\[lem:zfunctheta\] Let $n,r\in\mathbb{N}$ be such that $\max\{n,r\}>1$, let ${\displaystyle {\phi=\frac{(n-1)\pi+r\theta}{n}}}$, and let $I$ be the interval as in Theorem \[maintheorem\]. The function ${\displaystyle {z(\theta)=\frac{\sin^{n}\theta}{\sin^{n-r}(\phi-\theta)\sin^{r}(\phi)}}}$ is increasing on $(0,\pi/r)$, and maps this interval onto $I$.
We start by computing three derivatives: $$\begin{aligned}
\frac{d}{d\theta}\left(\frac{\sin\theta}{\sin(\phi-\theta)}\right) & = & \frac{\cos\theta\sin(\phi-\theta)-\sin\theta\cos(\phi-\theta)\left(r/n-1\right)}{\sin^{2}(\phi-\theta)}\nonumber \\
& = & \frac{\sin\phi-r\sin\theta\cos(\phi-\theta)/n}{\sin^{2}(\phi-\theta)},\label{eq:derivfirstfrac}\end{aligned}$$ $$\frac{d}{d\theta}\left(\frac{\sin\theta}{\sin\phi}\right)=\frac{\cos\theta\sin\phi-r\sin\theta\cos\phi/n}{\sin^{2}\phi},\label{eq:derivsecondfrac}$$ and $$\begin{aligned}
\frac{d}{d\theta}\left(\frac{\sin(\phi-\theta)}{\sin\phi}\right) & = & \frac{\sin\phi\cos(\phi-\theta)(r/n-1)-r\sin(\phi-\theta)\cos\phi/n}{\sin^{2}\phi}\nonumber \\
& = & \frac{r\sin\theta/n-\sin\phi\cos(\phi-\theta)}{\sin^{2}\phi}.\label{eq:derivthirdfrac}\end{aligned}$$ Armed with these calculations we now consider three cases, depending on the relative sizes of $n$ and $r$.\
- In this case $z(\theta)$ simplifies to $z(\theta)=\sin\theta/\sin\phi$. This is an increasing function on $(0,\pi/r)$, since its derivative (\[eq:derivsecondfrac\]) is equal to ${\displaystyle {\frac{\sin((n-1)\pi/n)}{\sin((n-1)\pi/n+\theta)}}}$, a quantity strictly bigger than 0 for $\theta\in(0,\pi/r)$. Finally, we observe that $z(\theta)$ is continuous on $(0,\pi/r)$, ${\displaystyle {\lim_{\theta\to0}z(\theta)=0}}$, and that ${\displaystyle {\lim_{\theta\to\pi/r}z(\theta)=+\infty}}$. We conclude that $z(\theta)$ maps $(0,\pi/r)$ to $(0,\infty)$.
- We write $$z(\theta)=\left(\frac{\sin\theta}{\sin(\phi-\theta)}\right)^{n-r}\left(\frac{\sin\theta}{\sin\phi}\right)^{r},$$ and note that the inequalities $$\pi-\frac{\pi}{n}<\phi<\pi\label{eq:phirange}$$ and $$\pi-\frac{\pi}{r}<\phi-\theta<\pi-\frac{\pi}{n}\label{eq:phi-thetarange}$$ hold for $\theta\in(0,\pi/r)$. We now demonstrate that $\sin\theta/\sin(\phi-\theta)$ and $\sin\theta/\sin\phi$ are both increasing functions of $\theta$ on $(0,\pi/r)$. If $2\le r<n$, then $\cos\theta>0$ and $\cos(\phi-\theta)<0$. Consequently (\[eq:derivfirstfrac\]) and (\[eq:derivsecondfrac\]) are positive. By checking the limits as $\theta\rightarrow0$ and $\theta\rightarrow\pi/r$, we see that $z(\theta)$ maps $(0,\pi/r)$ onto $(0,\infty)$. If $r=1$ and $\theta\le\pi/2$, then $\cos\theta\ge0$ and $\cos(\phi-\theta)\le0$ and $z(\theta)$ is an increasing function of $\theta$ for the same reason.\
Finally, we treat the case when $r=1$ and $\pi>\theta>\pi/2$. We write $$\sin\phi=\sin\frac{(n-1)\pi+\theta}{n}=\sin\frac{\pi-\theta}{n},$$ and note that for any angle $0<\alpha<\pi/2$, the inequality $$\sin\frac{\alpha}{n}>\frac{1}{n}\sin\alpha$$ holds, as the two sides agree when $\alpha=0$ and the derivative of the left side is greater than that of the right side when $0<\alpha<\pi/2$. Thus $\sin\phi>\sin\theta/n$ and (\[eq:derivfirstfrac\]) is positive. To show (\[eq:derivsecondfrac\]) is positive, we note that for any angle $0<\alpha<\pi/2$, the inequality $$\frac{1}{n}\sin\alpha\cos\frac{\alpha}{n}>\cos\alpha\sin\frac{\alpha}{n}\label{eq:trigineq}$$ holds since the two sides are equal when $\alpha=0$ and their respective derivatives satisfy $$\frac{1}{n}\cos\alpha\cos\frac{\alpha}{n}-\frac{1}{n^{2}}\sin\alpha\sin\frac{\alpha}{n}>\frac{1}{n}\cos\alpha\cos\frac{\alpha}{n}-\sin\alpha\sin\frac{\alpha}{n}.$$ Applying (\[eq:trigineq\]) with $\alpha=\pi-\theta$ establishes that (\[eq:derivsecondfrac\]) is positive. Thus $z(\theta)$ is an increasing function for all pairs $r,n$ under consideration. We next compute $$\begin{aligned}
\lim_{\theta\rightarrow\pi}\frac{\sin\theta}{\sin(\phi-\theta)} & = & \lim_{\theta\rightarrow\pi}\frac{\sin(\pi-\theta)}{\sin\left((n-1)(\pi-\theta)/n\right)}=\frac{n}{n-1},\label{eq:limitfirstquotient}\\
\lim_{\theta\rightarrow\pi}\frac{\sin\theta}{\sin\phi} & = & \lim_{\theta\rightarrow\pi}\frac{\sin(\pi-\theta)}{\sin((\pi-\theta)/n)}=n.\label{eq:limitsecondquotient}\end{aligned}$$ Consequently, if $1=r<n$, then $z(\theta)$ maps the interval $(0,\pi)$ to $(0,n^{n}/(n-1)^{n-1})$.
- We apply arguments akin to those above to the function $$z(\theta)=\left(\frac{\sin\theta}{\sin\phi}\right)^{n}\left(\frac{\sin(\phi-\theta)}{\sin\phi}\right)^{r-n}$$ and see that $z(\theta)$ is an increasing function of $\theta$ on $(0,\pi/r)$. If $n>1$, then $z$ maps $(0,\pi/r)$ onto $(0,\infty$). If $n=1$, we compute $$\begin{aligned}
\lim_{\theta\rightarrow0}\frac{\sin\theta}{\sin\phi} & = & \frac{1}{r}\qquad\text{and}\label{eq:limitthirdquotient}\\
\lim_{\theta\rightarrow0}\frac{\sin(\phi-\theta)}{\sin\phi} & = & \frac{r-1}{r},\label{eq:limitfourthquotient}\end{aligned}$$ and conclude that $z(\theta)$ maps the interval $(0,\pi/r)$ onto $((r-1)^{r-1}/r^{r},\infty)$. The proof is complete.
We now reformulate the condition $D_{n,r}(t,z)=0$ by rescaling the zeros of $D_{n,r}(t,z)$. Although it may seem insignificant at first, this change in point of view will enable to us derive key magnitude estimates for these zeros (see Proposition \[lem:zerosQ\]). These estimates in turn lay the foundation for the asymptotic analysis, and with that for the proof of Theorem \[maintheorem\], in Section \[proofmainthm\].\
We proceed as follows. Suppose $t=|t|e^{-i\theta}$ is a zero of $D_{n,r}(t,z)$, set $$\begin{aligned}
t_{0} & = & \frac{\sin\phi}{\sin(\phi-\theta)}e^{-i\theta}\qquad\text{(see (\ref{eq:t0form})), and}\\
t_{1} & = & t_{0}e^{2i\theta}=\overline{t_{0}}.\end{aligned}$$ After labeling the remaining $\max\{n,r\}-2$ zeros of $D_{n,r}(t,z)$ as $t_{2},t_{3},\ldots,t_{\max\{n,r\}-1}$, write $$\begin{aligned}
q_{k} & = & \frac{t_{k}}{t_{0}},\qquad\text{and}\\
\zeta_{k} & = & q_{k}e^{-i\theta}\qquad\text{for}\quad0\leq k<\max\{n,r\}.\end{aligned}$$ With this notation we rewrite $(1-t_{k})^{n}+zt_{k}^{r}=0$ as $$-z=\frac{(1-t_{k})^{n}}{t_{k}^{r}}=\frac{(t_{0}^{-1}-q_{k})^{n}}{q_{k}^{r}}t_{0}^{n-r}.$$ Combining (\[eq:t0form\]) with (\[eq:ztheta\]) we see that $D_{n,r}(t_{k},z)=0$ if and only if $$-\frac{\sin^{n}\theta}{\sin^{n-r}(\phi-\theta)\sin^{r}\phi}=\frac{\left(\sin(\phi-\theta)/\sin\phi-\zeta_{k}\right)^{n}}{\zeta_{k}^{r}}\left(\frac{\sin\phi}{\sin(\phi-\theta)}\right)^{n-r},$$ or equivalently $$\left(\frac{\sin(\phi-\theta)}{\sin\phi}-\zeta_{k}\right)^{n}+\zeta_{k}^{r}\left(\frac{\sin\theta}{\sin\phi}\right)^{n}=0.\label{eq:zetatheta}$$ We remark that $k=0,1$ correspond to the two trivial solutions of (\[eq:zetatheta\]), namely $\zeta_{0}=e^{-i\theta}$ and $\zeta_{1}=e^{i\theta}$. The next proposition establishes that these are the solutions of (\[eq:zetatheta\]) of smallest magnitude.
\[lem:zerosQ\] Suppose $n,r\in\mathbb{N}$ are such that $\max\{n,r\}>1$, $\theta\in(0,\pi/r)$ and $\phi=((n-1)\pi+r\theta)/n$. The polynomial $$Q(\zeta)=\left(\frac{\sin(\phi-\theta)}{\sin\theta}-\zeta\frac{\sin\phi}{\sin\theta}\right)^{n}+\zeta^{r}\label{eq:Qzeta}$$ has exactly two zeros on the unit circle. All other zeros of $Q(\zeta)$ lie outside the closed unit disk $|\zeta|\leq1$.
We replace $\zeta$ by $\zeta^{n}$ and consider the zeros[^4] of $$\widetilde{Q(\zeta)}=\left(\frac{\sin(\phi-\theta)}{\sin\theta}-\zeta^{n}\frac{\sin\phi}{\sin\theta}\right)^{n}+\zeta^{rn}.$$ Note that if $\zeta$ is a zero of $\widetilde{Q(\zeta)}$, then $\zeta$ is a zero of $$\begin{aligned}
R(\zeta) & = & \sin(\phi-\theta)-\zeta^{n}\sin\phi-\omega\zeta^{r}\sin\theta\label{eq:Qzetatransf}\\
& = & \sin\theta(-\omega\zeta^{r}-\cos\phi)-\sin\phi(\zeta^{n}-\cos\theta),\nonumber \end{aligned}$$ where $\omega=\exp\{i(\pi+2\pi k)/n\}$ for some $0\leq k<n$. Consequently, it suffices to establish the result for $R(\zeta)$.\
\
![\[fig:Qzeta0\]Generic positioning of the two circles $C_{1}$ and $C_{2}$](Qzeta0)
A priori we don’t know where the points of intersection (if any) of the two circles $C_{1}$ and $C_{2}$ are. A quick calculation using the law of cosines shows however, that in Figure \[fig:Qzeta0\], we must in fact have $x=\sin\phi$, and we obtain the more accurate Figure \[fig:Qzeta\].\
![\[fig:Qzeta\]The two circles $C_{1}$ and $C_{2}$](Qzeta)
**Zeros of $R(\zeta)$ inside the unit disk.** We claim that the loop $\Gamma:=R(e^{i\psi})$, $0\leq\psi<2\pi$, does not intersect the ray $(-\infty,0)$, and hence the winding number of $\Gamma$ about any point on $(-\infty,0)$ is $0$. We demonstrate the claim by treating three different cases, distinguished by whether $r$, $n$, or neither is equal to one.\
If $r\geq2$ and $n\geq2$, then $\cos\theta>0$ and $\cos\phi<0$. Suppose, by way of contradiction, that $\Gamma$ intersects the ray $(-\infty,0)$. Then modulo $2\pi$ the inequalities $$\begin{aligned}
& & -\theta<n\psi<\theta\quad\text{and}\\
& & \phi<r\psi+\alpha<2\pi-\phi\end{aligned}$$ must hold. Whence, the points $p_{1}$ and $p_{2}$ (on $C_{1}$ and $C_{2}$ respectively) corresponding to $\psi$ must lie on the arc of $C_{1}(C_{2}$ resp) that is inside of $C_{2}(C_{1}$ resp), (see Figure \[fig:Qzeta\]). Write $$\begin{aligned}
\theta-n\psi & \equiv & \beta\pmod{2\pi}\quad\text{and}\label{eq:congr1}\\
r\psi+\alpha-\phi & \equiv & \gamma\pmod{2\pi}\label{eq:congr2}\end{aligned}$$ where, given the symmetry about the $x$-axis, we may without loss of generality assume that $0<\beta\le\theta$ and $0<\gamma\le(\pi-r\theta)/n$. We add $r$ times the first congruence to $n$ times the second congruence and obtain $$r\beta+n\gamma\equiv0\mbox{ (mod \ensuremath{2\pi})},$$ which is impossible since $0<r\beta+n\gamma\le\pi$.\
If $n=1$, the fact that the zeros of the function $$R(\zeta)=\sin((r-1)\theta)-\zeta\sin r\theta+\zeta^{r}\sin\theta$$ lie outside the open unit disk is a direct consequence of Lemma 8 in [@tran-1]. If $r=1$, we write $R(\zeta)$ as $$\begin{aligned}
R(\zeta) & = & \sin\theta(-\omega\zeta-\cos\phi)-\sin\phi(\zeta^{n}-\cos\theta)\\
& = & \sin(\pi-\theta)\left(-\omega\zeta+\cos\frac{\pi-\theta}{n}\right)-\sin\frac{\pi-\theta}{n}\left(\zeta^{n}+\cos(\pi-\theta)\right).\end{aligned}$$ Applying the substitutions $\omega\zeta\rightarrow\zeta$, $(\pi-\theta)/n\rightarrow\theta$, and $n\rightarrow r$, reduces this case to the case $n=1$, and completes the proof.
We close this section with a result concerning the multiplicities of the zeros of $Q(\zeta)$, in addition to their location furbished by Proposition \[lem:zerosQ\].
\[lem:distinctzeros\] Let $n,r\in\mathbb{N}$ be such that $\max\{n,r\}>1$, $0<\theta<\pi/r$, and ${\displaystyle {\phi=\frac{(n-1)\pi+r\theta}{n}}}$. The zeros of the polynomial $$Q(\zeta)=\left(\frac{\sin(\phi-\theta)}{\sin\theta}-\zeta\frac{\sin\phi}{\sin\theta}\right)^{n}+\zeta^{r}$$ are distinct, except when
1. $n>r>1$ and $r$ is odd, or
2. $r>n>1$ and $n$ is odd .
If (1) or (2) hold, then there exists a unique $\theta^{*}\in(0,\pi/r)$ so that $Q(\zeta)$ has one double zero, and all of its remaining zeros are distinct.
Suppose that $\zeta_{*}$ is a multiple zero of $Q(\zeta)$, that is, $Q(\zeta_{*})=Q'(\zeta_{*})=0$. If $n=r$, $\zeta_{*}$ would have to satisfy $\zeta_{*}^{n-1}(\sin(\phi-\theta)/\sin\phi)=0$, and hence $\zeta_{*}=0$. Since $Q(0)\neq0$, we conclude that if $n=r$, $Q(\zeta)$ has no multiple zeros.\
If $n\neq r$, the equations $Q(\zeta_{*})=0=Q'(\zeta_{*})$ imply that $$\begin{aligned}
Q'(\zeta_{*}) & = & -n\frac{\sin\phi}{\sin\theta}\left(\frac{\sin(\phi-\theta)}{\sin\theta}-\zeta_{*}\frac{\sin\phi}{\sin\theta}\right)^{n-1}+r\zeta_{*}^{r-1}\nonumber \\
& = & \zeta_{*}^{r-1}\left(n\frac{\zeta_{*}\sin\phi}{\sin(\phi-\theta)-\zeta_{*}\sin\phi}+r\right)\label{eq:Qprime}\\
& = & 0.\nonumber \end{aligned}$$ We note that (\[eq:Qprime\]) has the unique non-zero solution $$\zeta_{*}=-\frac{r}{n-r}\frac{\sin(\phi-\theta)}{\sin\phi}.$$ Substituting this expression into the equation $Q(\zeta^{*})=0$ yields $$\frac{\sin^{n}(\phi-\theta)}{\sin^{n}\theta}\left(\frac{n^{n}}{(n-r)^{n}}+(-1)^{r}\frac{r^{r}}{(n-r)^{r}}\frac{\sin^{n}\theta}{\sin^{r}\phi\sin^{n-r}(\phi-\theta)}\right)=0.\label{eq:Qmultzero}$$ By Lemma \[lem:zfunctheta\], the left hand side of equation (\[eq:Qmultzero\]) does not vanish for $\theta\in(0,\pi/r)$, unless (1) or (2) hold, in which case it vanishes for a unique value $\theta^{*}\in(0,\pi/r)$.
The proof of Theorem \[maintheorem\] {#proofmainthm}
====================================
We now turn our attention to the proof of Theorem \[maintheorem\], which admittedly is technical. As such, that the reader may benefit from seeing the proof of a special case before proceeding to the general case. Recall that the generating function $1/D_{n,r}(t,z)$ of the sequence $\left\{ P_{m}(z)\right\} _{m=0}^{\infty}$ depends on $n,r\in\mathbb{N}$. If $\max\{n,r\}=2$, the result in Theorem 1 is in fact a consequence of Theorem 1 in [@tran]. In the following proposition we treat the case $\max\{n,r\}=3$. As it turns out, in this case the conclusions of Theorem \[maintheorem\] can be strengthened to $P_{m}(z)$ having only real zeros[^5]
\[illustration\] Suppose that $n,r\in\mathbb{N}$ be such that $\max\{n,r\}=3$. Let $D_{n,r}(t,z)=(1-t)^{n}+zt^{r}$, and let $I$ be the real interval in the statement of Theorem \[maintheorem\]. Suppose further that $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{D_{n,r}(t,z)},$$ and write $\mathcal{Z}(P_{m})$ for the set of zeros of $P_{m}(z)$. Then $\mathcal{Z}(P_{m})\subset I$ for all $m\geq0$, and ${\displaystyle {\bigcup_{m=0}^{\infty}\mathcal{Z}(P_{m})}}$ is dense in $I$.
Suppose $\theta\in(0,\pi/r)$ and let $z\in I$ be the point so that $t_{0}=|t_{0}|e^{-i\theta},\,t_{1}=t_{0}e^{2i\theta}$ and $t_{2}\in\mathbb{R}$ are the zeros of $D_{n,r}(t,z)$. Note that by Lemma \[lem:distinctzeros\], $t_{0},t_{1}$ and $t_{2}$ are distinct. Thus by partial fraction decomposition we obtain $$\begin{aligned}
\frac{1}{D_{n,r}(t,z)} & = & -\frac{1}{(t-t_{0})(t_{0}-t_{1})(t_{0}-t_{2})}-\frac{1}{(t-t_{1})(t_{1}-t_{0})(t_{1}-t_{2})}-\frac{1}{(t-t_{2})(t_{2}-t_{0})(t_{2}-t_{1})}\\
& = & \sum_{m=0}^{\infty}\left(\frac{1}{(t_{0}-t_{1})(t_{0}-t_{2})}\frac{1}{t_{0}^{m+1}}+\frac{1}{(t_{1}-t_{0})(t_{1}-t_{2})}\frac{1}{t_{1}^{m+1}}+\frac{1}{(t_{2}-t_{0})(t_{2}-t_{1})}\frac{1}{t_{2}^{m+1}}\right)t^{m},\end{aligned}$$ and consequently, for all $m\geq0$, $$P_{m}(z)=\frac{1}{t_{0}^{m+1}(t_{0}-t_{1})(t_{0}-t_{2})}+\frac{1}{t_{1}^{m+1}(t_{1}-t_{0})(t_{1}-t_{2})}+\frac{1}{t_{2}^{m+1}(t_{2}-t_{0})(t_{2}-t_{1})}.\label{eq:pmcubic}$$ We set $q_{1}=t_{1}/t_{0}$, $q_{2}=t_{2}/t_{0}$, and divide the right hand side of (\[eq:pmcubic\]) by $t_{0}^{m+3}$ to conclude that $z$ is a zero of $P_{m}(z)$ if and only if $$\frac{1}{(1-q_{1})(1-q_{2})}+\frac{1}{q_{1}^{m+1}(q_{1}-1)(q_{1}-q_{2})}+\frac{1}{q_{2}^{m+1}(q_{2}-1)(q_{2}-q_{1})}=0.\label{eq:qexpansion}$$ By multiplying the left hand side of (\[eq:qexpansion\]) by $e^{(m+3)i\theta}$ we obtain the equivalent equation $$-\frac{1}{2i\sin\theta e^{-i(m+1)\theta}(e^{-i\theta}-\zeta_{2})}+\frac{1}{2i\sin\theta e^{i(m+1)\theta}(e^{i\theta}-\zeta_{2})}+\frac{1}{\zeta_{2}^{m+1}(\zeta_{2}^{2}-2\cos\theta\zeta_{2}+1)}=0,$$ where $\zeta_{2}:=q_{2}e^{-i\theta}\in\mathbb{R}$ since $t_{2}\in\mathbb{R}$. Combining the first two summands and factoring the expression ${\displaystyle {\frac{1}{\zeta_{2}^{2}-2\zeta_{2}\cos\theta+1}}}$ establishes that equation (\[eq:qexpansion\]) is equivalent to $$\frac{\sin(m+2)\theta-\zeta_{2}\sin(m+1)\theta}{\sin\theta}+\frac{1}{\zeta_{2}^{m+1}}=0.\label{eq:thetacrit}$$ Finally, using the identity $\sin(m+2)\theta=\sin(m+1)\theta\cos\theta+\cos(m+1)\theta\sin\theta$, we see that equation (\[eq:thetacrit\]) holds if and only if $\theta$ is a zero of $$R_{m}(\theta)=\frac{(\cos\theta-\zeta_{2})\sin(m+1)\theta}{\sin\theta}+\cos(m+1)\theta+\frac{1}{\zeta_{2}^{m+1}}.\label{eq:Pthethacubic}$$ According to Proposition \[lem:zerosQ\], $|\zeta_{2}|>1$, and hence the sign of $R_{m}(\theta)$ alternates at values of $\theta$ for which $\cos(m+1)\theta=\pm1$. Thus, by the Intermediate Value Theorem, there are at least $\left\lfloor m/r\right\rfloor $ zeros $\theta$ of $R_{m}(\theta)$ on $(0,\pi/r)$. Lemma \[lem:zfunctheta\] in turn implies that each of these zeros $\theta$ yields an distinct zero $z(\theta)$ of $P_{m}(z)$ on $I$. Since the degree of $P_{m}(z)$ is $\left\lfloor m/r\right\rfloor $, we see that $\mathcal{Z}(P_{m})\subset I$ for all $m\geq0$. Finally, note that the set of solutions to $\cos(m+1)\theta=\pm1$, $m\geq0$, is dense in $(0,\pi/r)$, which immediately implies that ${\displaystyle {\bigcup_{m=0}^{\infty}\mathcal{Z}(P_{m})}}$ is dense in $I$, and completes the proof.
We continue with some remarks regarding Theorem \[maintheorem\] in the case when $\max\{n,r\}>3$ and $m$ is large. If the the zeros $t_{k}$, $0\le k<\max\{n,r\}$, of $D_{n,r}(t,z)$ are distinct[^6], then the partial fraction decomposition of ${\displaystyle {\frac{1}{D_{n,r}(t,z)}}}$ gives $$\begin{aligned}
\sum_{m=0}^{\infty}P_{m}(z)t^{m} & = & \pm\prod_{k=1}^{\max(n,r)-1}\frac{1}{t-t_{k}}\\
& = & \pm\sum_{k=0}^{\max(n,r)-1}\frac{1}{t-t_{k}}\prod_{l\ne k}\frac{1}{t_{k}-t_{l}}\\
& = & \pm\sum_{m=0}^{\infty}\left(\sum_{k=0}^{\max(n,r)-1}\frac{1}{t_{k}^{m+1}}\prod_{l\ne k}\frac{1}{t_{k}-t_{l}}\right)t^{m}.\end{aligned}$$ Equating coefficients of powers of $t$ in these formal power series shows that for any $m\geq0$, the equation $P_{m}(z)=0$ is equivalent to $$\sum_{k=0}^{\max(n,r)-1}\frac{1}{t_{k}^{m+1}}\prod_{l\ne k}\frac{1}{t_{k}-t_{l}}=0.\label{eq:Pquotients}$$ We multiply equation (\[eq:Pquotients\]) by $t_{0}^{m+\max\{n,r\}}$ and by $e^{(m+\max\{n,r\})i\theta}$ and conclude that $z$ is a zero of $P_{m}(z)$ if and only if $\theta$ is a zero of $$\begin{aligned}
R_{m}(\theta) & := & \sum_{k=0}^{\max(n,r)-1}\frac{1}{\zeta_{k}^{m+1}}\prod_{l\ne k}\frac{1}{\zeta_{k}-\zeta_{l}}\nonumber \\
& = & \sum_{k=0}^{\max\{n,r\}-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})},\label{eq:Ptheta}\end{aligned}$$ where $\zeta_{k}=e^{-i\theta}t_{k}/t_{0}$, $0\le k<\max\{n,r\}-1$, are the roots of $$Q(\zeta)=\left(\frac{\sin(\phi-\theta)}{\sin\theta}-\zeta\frac{\sin\phi}{\sin\theta}\right)^{n}+\zeta^{r}$$ with $\phi=((n-1)\pi+r\theta)/n$. Using symmetric reduction we conclude that $R_{m}(\theta)$ is a real valued continuous function of $\theta$ on $(0,\pi/r)$.\
If $D_{n,r}(t,z)$ has zeros of multiplicity greater than one, then so does $Q(\zeta)$. Thus we must be in either case (1) or (2) in Lemma \[lem:distinctzeros\], and hence there exists a unique $\theta^{*}\in(0,\pi/r)$ corresponding to the double zero $\zeta_{*}$ of $Q(\zeta)$. It is clear that $R_{m}(\theta)$ has a singularity at $\theta^{*}$. We claim that this singularity is removable. Indeed, if $\zeta_{a}=\zeta_{b}=\zeta_{*}$ for some $0\le a<b<\max\{n,r\}-1$, then the sum of the two terms of $R_{m}(\theta)$ involving both $\zeta_{a}$ and $\zeta_{b}$ is equal to $$\frac{m+1}{\zeta_{a}^{m+2}}\prod_{l\neq a,b}\frac{1}{\zeta_{a}-\zeta_{l}}.$$ Thus we may consider $R(\theta)$ to be a real valued continuous function of $\theta$ on $(0,\pi/r)$, regardless of whether or not the zeros of $D_{n,r}(t,z)$ are distinct.\
Recall that the magnitude estimates of Proposition \[lem:zerosQ\] allowed us to dispense with the term in $R_{m}(\theta)$ corresponding to the third zero of $Q(\zeta)$ in Proposition \[illustration\]. In order to make a similar approach work in the general case, we will need the following calculations concerning the dominating terms in $R_{m}(\theta)$.\
Assume that $\theta\in(0,\pi/r)$, set ${\displaystyle {\phi=\frac{(n-1)\pi+r\theta}{n}}}$, and recall that $\zeta_{0,1}=e^{\mp i\theta}$ are the two trivial zeros of $Q(\zeta)$ of minimal magnitude. Using the identity $$\sin(\phi-\theta)-e^{i\theta}\sin\phi=-\sin\theta e^{i\phi},$$ we rewrite the expression for $Q'(\zeta)$ in (\[eq:Qprime\]) as $$\begin{aligned}
Q'(e^{i\theta}) & = & e^{(r-1)i\theta}\left(-n\frac{\sin\phi}{\sin\theta}e^{i(\theta-\phi)}+r\right).\end{aligned}$$ Thus sum of the two terms in (\[eq:Ptheta\]) corresponding $\zeta_{0}=e^{-i\theta}$ and $\zeta_{1}=e^{i\theta}$ can be written as $$2\frac{\Re\left(e^{i(m+1)\theta}Q'(e^{i\theta})\right)}{|Q'(e^{i\theta})|^{2}}=\frac{2}{|Q'(e^{i\theta})|^{2}}\left(A\cos(m+r)\theta-B\sin(m+r)\theta\right),\label{eq:mainterm}$$ where $$\begin{aligned}
A=A(\theta) & = & -n\frac{\sin\phi}{\sin\theta}\cos(\phi-\theta)+r,\label{eq:Atheta}\\
B=B(\theta) & = & n\frac{\sin\phi}{\sin\theta}\sin(\phi-\theta).\nonumber \end{aligned}$$ Just as we did in the proof of Proposition \[illustration\], we will demonstrate that the sign of $R_{m}(\theta)$ changes at points where $\cos(m+r)\theta=\pm1$. At such points $B(\theta)=0$, so in order to be able to track the exact number of sign changes in $R_{m}(\theta)$, it remains to describe the function $A(\theta)$, which we do in the next lemma.
\[lem:Atheta\]Let $n,r\in\mathbb{N}$ be such that $\max\{n,r\}>1$, suppose that $\theta\in(0,\pi/r)$ and set $\phi=((n-1)\pi+r\theta)/n$. The function $$A(\theta)=-n\frac{\sin\phi}{\sin\theta}\cos(\phi-\theta)+r$$ is strictly positive.
If $\cos(\phi-\theta)\le0$ then $A(\theta)>0$ since $\sin\phi>0$ and $\sin\theta>0$ on the interval under consideration. Consequently, if $r\ge2$ and $n\ge2$, then $A(\theta)>0$.\
If $r=1$, then $A(\theta)>0$ if and only if $$\sin\theta-n\sin\phi\cos(\phi-\theta)>0.$$ Note that the left side of the above inequality is $0$ when $\theta=\pi$, and it is a decreasing function of $\theta$ on $(0,\pi)$ because its derivative is negative there: $$\begin{aligned}
& & \frac{d}{d\theta}\left(\sin\theta-n\sin\phi\cos(\phi-\theta)\right)\\
& = & \cos\theta-\cos\phi\cos(\phi-\theta)+n\sin\phi\sin(\phi-\theta)\left(\frac{1}{n}-1\right)\\
& = & (2-n)\sin\phi\sin(\phi-\theta)\\
& < & 0.\end{aligned}$$ It follows that if $r=1$, then $A(\theta)>0$.\
Finally, if $n=1$, then $\phi=r\theta$, and $$\begin{aligned}
A(\theta)\sin\theta & = & r\sin\theta-\sin\phi\cos(\phi-\theta)\\
& = & (r-1)\sin\theta-\cos\phi\sin(\phi-\theta).\end{aligned}$$ Since $A(0)\sin(0)=0$, and $$\frac{d}{d\theta}(A(\theta)\sin(\theta))=(2r-1)\sin(r\theta)\sin((r-1)\theta)>0$$ for $\theta\in(0,\pi/r)$, we conclude that $A(\theta)>0$ in this case as well.
We are now in position to describe the sign changes of $R_{m}(\theta)$ on the interval $(0,\pi/r)$.
\[prop:signchangegeneral\] Suppose that $n,r\in\mathbb{N}$ are such that $\max\{n,r\}>1$, and let $R_{m}(\theta)$ be defined as in equation (\[eq:Ptheta\]) for $m\geq0$. If ${\displaystyle {\theta_{h}=\frac{h\pi}{m+r},(h=1,\ldots,\left\lfloor m/r\right\rfloor ),}}$ denote the values of $\theta$ in $(0,\pi/r)$ which give $\cos(m+r)\theta=\pm1$, then
- $\textrm{sgn}\left(R_{m}(\theta_{h})\right)=(-1)^{h}$, and
- $\textrm{sgn}\left(R_{m}(\pi/r^{-})\right)=(-1)^{\left\lfloor m/r\right\rfloor +1}$.
for all $m$ sufficiently large.
The proof is broken into three cases as dictated by the asymptotic behavior of the expressions $\theta_{h}$ as $m$ goes to infinity. Some of these will stay bounded away from both zero and $\pi/r$, some will approach $0$, and some will tend to $\pi/r$.
Case 1: $\gamma<\theta<\pi/r-\gamma$ for some small fixed $\gamma$ independent of $m$ {#case-1-gammathetapir-gamma-for-some-small-fixed-gamma-independent-of-m .unnumbered}
-------------------------------------------------------------------------------------
Proposition \[lem:zerosQ\] implies that if $2\le k<\max(n,r)$, then $|\zeta_{k}|>1+\epsilon$ for some fixed $\epsilon>0$. Using Lemma \[lem:Atheta\] and equation (\[eq:mainterm\]) we write $$R_{m}(\theta)=\frac{2}{|Q'(e^{i\theta})|^{2}}\left(A\cos(m+r)\theta-B\sin(m+r)\theta\right)+\sum_{k=2}^{\max(n,r)-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})},$$ where the $A(\theta)>0$ and the sum approaches $0$ exponentially fast when $m\rightarrow\infty$. We invoke Lemma \[lem:Atheta\] to conclude that the sign of $R_{m}(\theta_{h})$ is $(-1)^{h}$.
Case 2: $\theta\rightarrow0$ as $m\rightarrow\infty$ {#case-2-thetarightarrow0-as-mrightarrowinfty .unnumbered}
----------------------------------------------------
If $n=1$, the zeros $e^{\pm i\theta}$ of $Q(\zeta)$ converge to $1$, while its remaining zeros converge to those zeros of $T(\zeta)=r-1-r\zeta+\zeta^{r}$ which are greater than one in magnitude[^7], and hence are uniformly separated from the closed unit disk. Thus, following an argument similar to the one in Case 1, we see that the sum in $$R_{m}(\theta)=\frac{2}{|Q'(e^{i\theta})|^{2}}\left(A\cos(m+r)\theta-B\sin(m+r)\theta\right)+\sum_{k=2}^{\max(n,r)-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})},$$ approaches $0$ exponentially fast as $m\rightarrow\infty$. We next calculate $$\begin{aligned}
A(\theta) & = & r-\frac{\sin(r\theta)\cos((r-1)\theta)}{\sin\theta}\\
& = & \left(\frac{r}{3}-r^{2}+\frac{2r^{3}}{3}\right)\theta^{2}+\mathcal{O}(\theta^{4}),\end{aligned}$$ and deduce that $A(\theta)\to0$ at a polynomial rate. Consequently, the sign of $R_{m}(\theta_{h})$ is $(-1)^{h}$ for $m\gg1$.
On the other hand, if $n\ne1$, then $Q(\zeta)$ has $n$ zeros in a neighborhood of $1$, and the possible remaining $\max\{n,r\}-n$ zeros lie outside the closed unit disk. Assume without loss of generality that $\zeta_{k}$ is in a neighborhood of $1$ for $0\le k<n$. For the same reason as in Case 1, the sum $$\sum_{k=n}^{\max(r,n)-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}$$ is either $0$, or it approaches $0$ exponentially fast. We next consider the sum $$\sum_{k=0}^{n-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}.$$ Set $\zeta=1+\epsilon$, $\epsilon\in\mathbb{C}$, and note that $$\begin{aligned}
\sin\left(\phi-\theta\right)-\zeta\sin\phi & = & \sin\frac{\pi+(n-r)\theta}{n}-\zeta\sin\frac{\pi-r\theta}{n}\nonumber \\
& = & \sin\frac{\pi}{n}+\left(\cos\frac{\pi}{n}\right)\left(\frac{(n-r)\theta}{n}\right)-(1+\epsilon)\sin\frac{\pi}{n}+(1+\epsilon)\left(\cos\frac{\pi}{n}\right)\left(\frac{r\theta}{n}\right)+\mathcal{O}\left(\theta^{2}\right)\nonumber \\
& = & \theta\cos\frac{\pi}{n}+\frac{r\theta\epsilon}{n}\cos\frac{\pi}{n}-\epsilon\sin\frac{\pi}{n}+\mathcal{O}\left(\theta^{2}\right).\label{eq:sinediff}\end{aligned}$$ Thus, if $\zeta$ is a zero of $Q(\zeta)$, then $$0=\left(\theta\cos\frac{\pi}{n}+\frac{r\theta\epsilon}{n}\cos\frac{\pi}{n}-\epsilon\sin\frac{\pi}{n}+\mathcal{O}(\theta^{2})\right)^{n}+\theta^{n}(1+\epsilon)^{r}(1+\mathcal{O}(\theta)).$$ We deduce that $$\theta\cos\frac{\pi}{n}+\frac{r\theta\epsilon}{n}\cos\frac{\pi}{n}-\epsilon\sin\frac{\pi}{n}=\omega\theta\left(1+\epsilon\right)^{r/n}\left(1+\mathcal{O}\left(\theta\right)\right),$$ where $\omega$ is an $n$-th root of $-1$. Expanding $(1+\epsilon)^{n/r}$ and solving for $\epsilon$ yields $$\epsilon=\frac{\theta\left(\cos(\pi/n)-\omega\right)}{\sin(\pi/n)}+\mathcal{O}\left(\theta^{2}\right).\label{eq:epsilonform}$$ Substituting (\[eq:sinediff\]) and (\[eq:epsilonform\]) into the expression (\[eq:Qprime\]) for $Q'(\zeta)$ we see that if $\zeta$ is a zero of $Q(\zeta)$, then $$\begin{aligned}
Q'(\zeta) & = & \zeta^{r-1}\left(n\frac{\zeta\sin\phi}{\sin(\phi-\theta)-\zeta\sin\phi}+r\right)\nonumber \\
& = & \zeta^{r-1}\left(\frac{n\zeta\sin(\pi/n)}{\theta\omega}+\mathcal{O}(1)\right).\label{eq:Qprimeasymp}\end{aligned}$$ Thus, $$\sum_{k=0}^{n-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}=\frac{\theta}{n\sin(\pi/n)}\sum_{k=0}^{n-1}\frac{\omega_{k}}{\zeta_{k}^{m+r}}\left(1+\mathcal{O}(\theta)\right),\label{eq:Pasympzero}$$ where $\omega_{k}=e^{(2k-1)\pi i/n}$, $0\le k<n$.
If $\theta\geq\delta/\sqrt{m}$ for some small $\delta$, then equation (\[eq:epsilonform\]) gives $$\begin{aligned}
|\zeta_{k}| & \geq & \left|1+\frac{\theta(\cos(\pi/n)-\omega_{k})}{\sin(\pi/n)}\right|\\
& > & \left|1+\frac{\delta\left(\cos(\pi/n)-\cos\left((2k-1)\pi/n\right)\right)}{2\sqrt{m}\sin(\pi/n)}\right|.\end{aligned}$$ From this inequality we deduce that when $m$ is large, $$\sum_{k=2}^{n-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}$$ approaches $0$ faster than the polynomial decay of $\frac{2}{|Q'(e^{i\theta})|^{2}}\left(A\cos(m+r)\theta-B\sin(m+r)\theta\right)$. Thus the sign of $R_{m}(\theta_{h})$ is again $(-1)^{h}$ for $m\gg1$.\
On the other hand, if $\theta<\delta/\sqrt{m}$ for $\delta\ll1$, then $$\begin{aligned}
\zeta_{k}^{m+r} & = & \left(1+\frac{\theta(\cos(\pi/n)-\omega_{k})}{\sin(\pi/n)}+\mathcal{O}(\theta^{2})\right)^{m+r}\\
& = & \exp\left(\frac{\cos(\pi/n)-\omega_{k}}{\sin(\pi/n)}h\pi\right)\left(1+\mathcal{O}\left(\frac{h^{2}}{m}\right)\right).\end{aligned}$$ We select $\delta$ small enough so that the sign of $R_{m}(\theta_{h})$ is the same as the sign of $$\sum_{k=0}^{n-1}\omega_{k}\exp^{-1}\left(\frac{\cos(\pi/n)-\omega_{k}}{\sin(\pi/n)}h\pi\right),$$ which in turn is determined by the sign of $$\sum_{k=0}^{n-1}\omega_{k}\exp\left(\frac{h\pi\omega_{k}}{\sin(\pi/n)}\right),$$ since ${\displaystyle {\exp^{-1}\left(\cos(\pi/n)/\sin(\pi/n)\right)>0}}$. The next lemma establishes that the sign of the expression in question is $(-1)^{h}$, thereby completing Case 2.
\[aestimates\] Suppose that $n\geq2$. Then for al $h\in\mathbb{N}$, the sign of the expression $$\sum_{k=0}^{n-1}\omega_{k}\exp^{-1}\left(\frac{\cos(\pi/n)-\omega_{k}}{\sin(\pi/n)}h\pi\right),\label{eq:expsum}$$ is $(-1)^{h}$.
Note that the sum of the first two terms of (\[eq:expsum\]) is $$2(-1)^{h}\cos\left(\frac{\pi}{n}\right),$$ whereas the sum of the remaining $n-2$ terms is at most $$\sum_{k=2}^{n-1}\exp^{-1}\left(\frac{\cos(\pi/n)-\cos((2k-1)\pi/n)}{\sin(\pi/n)}h\pi\right),\label{eq:smallsum}$$ a sum which is largest when $h=1$. We use a computer to check that when $n\le85$ and $h=1$, (\[eq:smallsum\]) is less than $|2\cos(\pi/n)|$, hence the sign of (\[eq:expsum\]) is $(-1)^{h}$. For $n>85$ we bound (\[eq:smallsum\]) as follows: $$\begin{aligned}
\sum_{k=2}^{n-1}\exp^{-1}\left(\frac{\cos(\pi/n)-\cos((2k-1)\pi/n)}{\sin(\pi/n)}h\pi\right) & = & \sum_{k=2}^{n-1}\exp^{-1}\left(2\frac{h\pi}{\sin(\pi/n)}\sin\frac{k\pi}{n}\sin\frac{(k-1)\pi}{n}\right)\\
& < & 2\sum_{k=2}^{\left\lfloor n/2\right\rfloor }\exp^{-1}\left(2h\pi\sin\frac{k\pi}{n}\right)\\
& \stackrel{(i)}{<} & 2\sum_{k=2}^{\left\lfloor n/2\right\rfloor }\exp^{-1}\left(\frac{hk\pi^{2}}{n}\right)\\
& \stackrel{(ii)}{<} & \frac{2}{e^{h\pi^{2}/n}(e^{h\pi^{2}/n}-1)},\end{aligned}$$ where inequality $(i)$ follows from the fact that $2\sin(k\pi/n)>k\pi/n$ when $k<n/2$, and inequality $(ii)$ is obtained by replacing the partial geometric sum by the whole series. The reader will note that the inequality $$\frac{2}{e^{h\pi^{2}/n}(e^{h\pi^{2}/n}-1)}<2\cos\frac{\pi}{85}<2\cos\frac{\pi}{n}$$ holds when $e^{h\pi^{2}/n}>1.62$, or when $n<20h$. Consequently, when $n<20h$, $$\sum_{k=2}^{n-1}\omega_{k}\exp^{-1}\left(\frac{\cos(\pi/n)-\omega_{k}}{\sin(\pi/n)}h\pi\right)<|2\cos(\pi/n)|$$ and the sign of (\[eq:expsum\]) is $(-1)^{h}$.\
Consider now the case $h<n/20$. We note that $n>85$ implies that $$\sum_{k=0}^{n-1}\omega_{k}\exp\left(\frac{h\pi\omega_{k}}{\sin(\pi/n)}\right)\approx\sum_{k=0}^{n-1}\omega_{k}e^{hn\omega_{k}}.$$ By expanding the exponential function in a series we obtain $$\begin{aligned}
\sum_{k=0}^{n-1}\omega_{k}e^{hn\omega_{k}} & = & \sum_{k=0}^{n-1}\omega_{k}\sum_{j=0}^{\infty}\frac{1}{j!}\left(hn\omega_{k}\right)^{j}\nonumber \\
& = & \sum_{j=0}^{\infty}\frac{1}{j!}(hn)^{j}\left(\sum_{k=0}^{n-1}\omega_{k}^{j+1}\right).\label{eq:expseries}\end{aligned}$$ The identity $$\sum_{k=0}^{n-1}\omega_{k}^{a}=\left\{ \begin{array}{cc}
0 & \text{if} n \nmid a\\
n(-1)^{a/n} & \text{if }n\mid a
\end{array} \right.,$$ simplifies (\[eq:expseries\]) to $$n\sum_{j=1}^{\infty}(-1)^{j}\frac{1}{(nj-1)!}\left(hn\right)^{jn-1}=:n\sum_{j=1}^{\infty}a_{j}.$$ It is straightforward to show that $|a_{j}|<|a_{j+1}|$ for $j=0,1,\ldots,h-1$, and $|a_{j}|>|a_{j+1}|$ for $j=h,h+1,\ldots$. Since the sum $\sum a_{j}$ is alternating, we conclude that $$\left|\sum_{j=0}^{h-1}a_{j}\right|<|a_{h-1}|,\qquad\text{and}\qquad\left|\sum_{j=h+1}^{\infty}a_{j}\right|<|a_{h+1}|.$$ We now relate the quantities $|a_{h-1}|,|a_{h}|$ and $|a_{h+1}|$. To this end, we compute $$\left|\frac{a_{j+1}}{a_{j}}\right|=\frac{(hn)^{n}}{\prod_{\ell=0}^{n-1}nj+\ell},$$ and consequently, $$\begin{aligned}
\frac{|a_{h-1}|+|a_{h+1}|}{|a_{h}|} & = & \frac{\prod_{\ell=0}^{n-1}(n(h-1)+\ell)}{(nh)^{n}}+\frac{(hn)^{n}}{\prod_{\ell=0}^{n-1}(nh+\ell)}\\
& = & \prod_{\ell=0}^{n-1}\left(1-\frac{n-\ell}{n}\frac{1}{h}\right)+\frac{1}{\prod_{\ell=1}^{n-1}\left(1+\frac{\ell}{nh}\right)}\\
& = & P_{1}+\frac{1}{P_{2}}.\end{aligned}$$ Taking the natural logarithm of $P_{1}$ we obtain $$\begin{aligned}
\ln(P_{1}) & = & \sum_{\ell=0}^{n-1}\ln\left(1-\frac{n-\ell}{n}\frac{1}{h}\right)\\
& \stackrel{\ast}{<} & \sum_{\ell=0}^{n-1}\left(-\frac{n-\ell}{n}\frac{1}{h}\right)\\
& = & -\frac{1}{nh}\frac{n(n+1)}{2}\\
& = & -\frac{n+1}{2h},\end{aligned}$$ where the starred inequality follows from the fact that $\ln x\leq x-1$ for all $x>0$, with equality only when $x=1$. We conclude that $P_{1}<\exp(-\frac{n+1}{2h})$. On the other hand, taking the natural logarithm of $P_{2}$ gives $$\begin{aligned}
\ln(P_{2}) & = & \sum_{\ell=1}^{n-1}\ln\left(1+\frac{\ell}{nh}\right)\\
& \stackrel{\ast}{>} & \sum_{\ell=1}^{n-1}\frac{\ell}{2nh}\\
& = & \frac{1}{2nh}\frac{(n-1)n}{2}\\
& = & \frac{n-1}{4h},\end{aligned}$$ where for the starred inequality we used the fact that $\ln x>\frac{x-1}{2}$ for $1<x<2$. Consequently, ${\displaystyle {P_{2}>\exp\left(\frac{n-1}{4h}\right)}}$. Combining these two results we see that $$P_{1}+\frac{1}{P_{2}}<e^{-\frac{n+1}{2h}}+e^{-\frac{n-1}{4h}}<1,$$ which in turn implies that $|a_{h-1}|+|a_{h+1}|<|a_{h}|$, and hence $$\left|\sum_{j\neq h}a_{j}\right|<|a_{h}|$$ when $h<n/20$ and $n>85$. Thus, the sign of the expression (\[eq:expsum\]) is $(-1)^{h}$ in this case as well.
Case 3: $\theta\rightarrow\pi/r$ as $m\rightarrow\infty$ {#case-3-thetarightarrowpir-as-mrightarrowinfty .unnumbered}
--------------------------------------------------------
When $\theta\rightarrow\pi/r$ and $r=1$, we observe that with $R(\zeta)$ as in (\[eq:Qzetatransf\]), the polynomial $$\frac{R(\zeta)}{\sin\theta}=(-\omega\zeta-\cos\phi)-\frac{\sin\phi}{\sin\theta}(\zeta^{n}-\cos\theta)$$ converges locally uniformly to $$\widetilde{T}(\zeta)=(1-\omega\zeta)-\frac{1}{n}(1+\zeta^{n}).$$ A quick calculation shows that $$\label{eq:ttransform}
n \widetilde{T}(\zeta) \big|_{\omega \zeta \to \zeta}=T(\zeta),$$ the polynomial we treated at the beginning of case 2. Since the transformations used in (\[eq:ttransform\]) preserve the location of the zeros in, on, and outside the unit circle, we conclude that as $\theta\rightarrow\pi/r$, $Q(\zeta)$ converges locally uniformly to a polynomial whose zeros lie outside the closed unit disk besides the double zeros at $\zeta=-1$. As a result, the sign of $R_{m}(\theta_{h})$ is again determined by the sign of $\cos(m+r)\theta_{h}$, and is hence equal to $(-1)^{h}$.\
When $r>1$, the polynomial $Q(\zeta)$ has $r$ zeros approaching the $r$-th roots of $-1$, with the possible remaining $n-r$ zeros (when $n>r)$ tending to $\infty$. We thus consider the sum $$\sum_{k=0}^{r-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}$$ where each $\zeta_{k}$, $0\le k<r$, approaches an $r$-th root of $-1$. We set $e_{k}=e^{(2k-1)\pi i/r}$ and $\zeta_{k}=e_{k}+\epsilon$, for some $\epsilon\in\mathbb{C}$, and write $$\eta=\frac{\pi}{r}-\theta=\frac{\pi}{r}-\frac{h\pi}{m+r}=\frac{m+r-hr}{r(m+r)}\pi.$$ We compute the difference $$\begin{aligned}
\sin\left(\phi-\theta\right)-\zeta\sin\phi & = & \sin\frac{\pi+(n-r)\theta}{n}-\zeta\sin\frac{\pi-r\theta}{n}\\
& = & \sin\frac{\pi}{r}+\frac{(r-n)\eta}{n}\cos\frac{\pi}{r}-\zeta\frac{r\eta}{n}+\mathcal{O}(\eta^{2}),\end{aligned}$$ and use it to rewrite the equation $Q(\zeta_{k})=0$ as $$\begin{aligned}
0 & = & \left(\sin\frac{\pi}{r}+\frac{(r-n)\eta}{n}\cos\frac{\pi}{r}-(e_{k}+\epsilon)\frac{r\eta}{n}\right)^{n}+(\sin\frac{\pi}{r}-\eta\cos\frac{\pi}{r})^{n}(e_{k}+\epsilon)^{r}+\mathcal{O}(\eta^{2})\\
& = & \left(\sin\frac{\pi}{r}+\frac{(r-n)\eta}{n}\cos\frac{\pi}{r}-e_{k}\frac{r\eta}{n}\right)^{n}-(\sin\frac{\pi}{r}-\eta\cos\frac{\pi}{r})^{n}\left(1+\frac{r\epsilon}{e_{k}}\right)+\mathcal{O}(\eta^{2}+\epsilon^{2}).\end{aligned}$$ Taking $n$-th roots yields $$\frac{(r-n)\eta}{n}\cos\frac{\pi}{r}-e_{k}\frac{r\eta}{n}=\frac{r\epsilon}{ne_{k}}\sin\frac{\pi}{r}-\eta\cos\frac{\pi}{r}\left(1+\frac{r\epsilon}{ne_{k}}\right)+\mathcal{O}(\eta^{2}+\epsilon^{2}),$$ and consequently $$\epsilon=\frac{e_{k}(\cos(\pi/r)-e_{k})\eta}{\sin(\pi/r)-\eta\cos(\pi/r)}+\mathcal{O}\left(\eta^{2}\right)=\frac{e_{k}(\cos(\pi/r)-e_{k})\eta}{\sin(\pi/r)}+\mathcal{O}(\eta^{2}).\label{eq:epsilonpi/r}$$ From equation (\[eq:Qprime\]) we also obtain $$\begin{aligned}
Q'(\zeta_{k}) & = & \zeta_{k}^{r-1}\left(n\frac{\zeta_{k}\sin\phi}{\sin(\phi-\theta)-\zeta\sin\phi}+r\right)\\
& = & r\zeta_{k}^{r-1}+\mathcal{O}(\eta)\end{aligned}$$ which we substitute into the sum under consideration to obtain $$\sum_{k=0}^{r-1}\frac{1}{\zeta_{k}^{m+1}Q'(\zeta_{k})}=\frac{1}{r}\sum_{k=0}^{r-1}\frac{1+\mathcal{O}(\eta)}{\zeta_{k}^{m+r}}.\label{eq:Pasymppi/r}$$ For the same reasons as in Case 2, it suffices to consider $\eta<\delta/\sqrt{m}$ for small $\delta$. In this case we have the approximation $$\begin{aligned}
\zeta_{k}^{m+r} & = & e_{k}^{m+r}\left(1+\frac{\cos(\pi/r)-e_{k}}{\sin(\pi/r)}\eta+\mathcal{O}(\eta^{2})\right)^{m+r}\\
& = & e_{k}^{m+r}\exp\left(\frac{\cos(\pi/r)-e_{k}}{\sin(\pi/r)}(m+r)\eta\right)\left(1+\mathcal{O}\left(m\eta^{2}\right)\right).\end{aligned}$$ We choose $\delta$ small so that the sign of $R_{m}(\theta_{h})$ is the same as the sign of $$\sum_{k=0}^{r-1}e_{k}^{-m-r}\exp^{-1}\left(\frac{\cos(\pi/r)-e_{k}}{\sin(\pi/r)}\left(\frac{m}{r}-h+1\right)\pi\right).\label{eq:finalest}$$ The sum of the first two terms in the above sum is $2(-1)^{h}$ and hence, for reasons similar to those in Case 2 (i.e. using the same argument with $m/r-h+1$ in place of $h$), it suffices to consider $r>85$. For such $r$, arguments entirely analogous to those we gave in the proof of Lemma \[aestimates\] establish that the sign of the sum (\[eq:finalest\]) is $(-1)^{h}$. The determine the sign of $R(\pi/r^{-})$, we first note that for $r>85$, the sign of (\[eq:finalest\]) is the same as that of $$\sum_{k=0}^{r-1}e_{k}^{-m-r}e^{(m-rh+r)e_{k}}=\sum_{k=0}^{r-1}e_{k}^{-m-r}\sum_{j=0}^{\infty}\frac{1}{j!}\left((m-rh+r)e_{k}\right)^{j}.\label{eq:doublesum}$$ If we let $m=pr+s$, $0\le s<r$, then the double summation on the right hand side of (\[eq:doublesum\]) can be rewritten as $$r\sum_{j=0}^{\infty}(-1)^{p+1-j}\frac{\left(m+r-rh\right)^{jr+s}}{(rj+s)!}.\label{eq:signpioverr}$$ Now the sign of $R(\pi/r^{-})$ is obtained by replacing $h$ by $(m+r)/r^{-}$ in (\[eq:finalest\]), and thus by setting $j=0$ in (\[eq:signpioverr\]). By doing so we conclude that the sign of $R(\pi/r^{-})$ is $(-1)^{p+1}=(-1)^{\lfloor\frac{m}{r}\rfloor+1}$. The proof is complete.
With Proposition \[prop:signchangegeneral\] at our disposal, we now put the finishing touches on the proof of Theorem \[maintheorem\]. Let $n,r\in\mathbb{N}$ such that $\max\{r,n\}>1$, and suppose that $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{(1-t)^{n}+zt^{r}}=:\frac{1}{D_{n,r}(t,z)}.$$ Given $m\in\mathbb{N}$, every zero of $R_{m}(\theta)$ (see (\[eq:Ptheta\])) corresponds to a distinct zero $z\in I$ of $P_{m}(z)$. Proposition \[prop:signchangegeneral\], together with the Intermediate Value Theorem imply that for $m\gg1$, $R_{m}(\theta)$ has at least $\left\lfloor m/r\right\rfloor $ zeros on $(0,\pi/r)$. The conclusions of Theorem \[maintheorem\] now follow from degree considerations, along with the density of the solutions of the equations $\cos(m+r)\theta=\pm1$, $m=0,1,\ldots$, in the interval $(0,\pi/r)$.
Some open problems {#open}
==================
In light of the conclusions of Theorem $\ref{maintheorem}$ and Proposition $\ref{illustration}$, it is natural to ask whether the zeros of $P_{m}(z)$ lie on $I$ for all $m$.
\[prob:1\] Let $n,r\in\mathbb{N}$ such that $\max\{n,r\}>1$. We consider the sequence of polynomials $P_{m}(z)$ generated by $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{(1-t)^{n}+zt^{r}},$$ and write $\mathcal{Z}(P_{m})$ for the set of zeros of $P_{m}(z)$. Show that $\mathcal{Z}(P_{m})\subset I$ for all $m\ge0$, and ${\displaystyle \bigcup_{m=0}^{\infty}\mathcal{Z}(P_{m})}$ is dense in $I$ where $$I=\left\{ \begin{array}{cc}
(0,\infty) & \text{if} n,r\ge2\\
(0,n^{n}/(n-1)^{n-1}) & \text{if }r=1\\
((r-1)^{r-1}/r^{r},\infty) & \text{ if }n=1
\end{array}\right. .$$
A natural way to extend the problem is to consider, in place of the binomial expression $(1-t)^{n}$,
Let $Q(t)\in\mathbb{R}[t]$ be a real polynomial in $t$ whose zeros are positive real numbers and $Q(0)>0$. Show that for any integer $r\ge0$, the zeros of the polynomial $P_{m}(z)$ generated by $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{1}{Q(t)+zt^{r}}$$ lie on the positive real ray.
As we have seen in the proof of Lemma $\ref{degree}$, the denominator of the generating function gives the recurrence relation for $P_{m}(z)$ whereas the numerator gives rise to the set of initial polynomials. Allowing for more general numerators in the rational generating function is a natural extension of the current work. With some numerical evidence in support, we propose the following
Let $Q(t)\in\mathbb{R}[t]$ be a real polynomial in $t$ whose zeros are positive real numbers and $Q(0)>0$. For any integer $r\ge0$ and any real bivariate polynomial $N(t,z)\in\mathbb{R}[t,z]$ whose degree in $t$ is less than the degree in $t$ of $Q(t)+zt^{r}$, we consider the sequence of polynomials $P_{m}(z)$ generated by $$\sum_{m=0}^{\infty}P_{m}(z)t^{m}=\frac{N(t,z)}{Q(t)+zt^{r}}.$$ Show that there is a fixed constant (depending perhaps on $r$ and $\deg Q$) $C$ such that for any $m$, the number of zeros of $P_{m}(z)$ outside $(0,\infty)$ is less than $C$.
Finally, although in this paper we do not study them explicitly, transformations $T$ with the property that the sequence of polynomials $\left\{ P_{m}^{T}(z)\right\} _{m=0}^{\infty}$ generated by $T[G(t,z)]$ have only real zeros whenever those generated by $G(t,z)$ do are of great interest. We see a natural parallel between these operators, and reality preserving linear operators on $\mathbb{R}[x]$, and believe that understanding them is key to understanding how polynomial sequences with only real zeros might be generated. As such, we pose
Classify operators (bi-linear, or otherwise) $T$ on $\mathbb{R}(t,z)$ with the property that all terms of the sequence of polynomials $\left\{ P_{m}^{T}(z)\right\} _{m=0}^{\infty}$ generated by $T[G(t,z)]$ have only real zeros whenever those generated by $G(t,z)$ do. Any results regarding this problem would be pioneering, even if $G(t,z)$ is restricted to be a rational function of the type discussed in the present paper.
[10]{} S. Beraha, J. Kahane, N. J. Weiss, Limits of zeros of recursively defined polynomials, *Proc. Nat. Acad. Sci. U.S.A.*, **72** (1975), no. 11, 4209.
J. Borcea and P. Brändén, Pólya-Schur master theorems for circular domains and their boundaries,
T. Forgács and A. Piotrowski, Hermite multiplier sequences and their associated operators,
G. H. Hardy, On the zeros of certain class of integral Taylor series II,
I. V. Ostrovskii, Hardy’s generalization of $e^{x}$ and related analogs of cosine and sine,
J. Peetre. Une caractérisation abstraite des opérateurs différentiels. In:
G. Pólya and J. Schur, Über zwei Arten von Faktorenfolgen in der Theorie der algebraischen Gleichungen,
K. Tran, Connections between discriminants and the root distribution of polynomials with rational generating function,
K. Tran, The root distribution of polynomials with a three-term recurrence,
A. Sokal, Chromatic roots are dense in the whole complex plane, *Combin. Probab. Comput.*, **13** (2004), 221-261.
[^1]: Although the publication of
[^2]: We direct the reader to the beautiful works of Hardy [@hardy] and Ostrovskii [@ostr] concerning the zero loci of certain entire functions obtained this way
[^3]: By $\Gamma=\{\gamma_{k}\}_{k=0}^{\infty}$ being a Hermite diagonal operator we simply mean that $\Gamma[H_{n}(x)]=\gamma_{n}H_{n}(x)$ for all $n$, where $H_{n}(x)$ denotes the $n$th Hermite polynomial.
[^4]: The change of variables $\zeta\to\zeta^{n}$ maps the (interior, exterior and the) unit circle to itself (resp). Hence if we prove the proposition for $\widetilde{Q(\zeta)}$, we simultaneously also get the result for $Q(\zeta)$.
[^5]: By convention we consider all constant polynomials to have only real zeros (namely none), including the zero polynomial, which clearly has infinitely many non-real zeros.
[^6]: Recall that the zeros of $D_{n,r}(t,z)$ and those of $Q(\zeta)$ are scaled copies of each other. Hence, if one has distinct zeros, so does the other.
[^7]: $T(\zeta)/(\zeta-1)=\sum_{j=1}^{r-1}\zeta^{j}-(r-1)$ has exactly one zero in the closed unit disk at $\zeta=1$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let $p$ be a prime number. We show in this short note that mod-$p$ maximal compact inductions of a $p$-adic split reductive group do not have irreducible admissible subrepresentations.'
author:
- '**Peng Xu**'
bibliography:
- 'new.bib'
title: '**Mod-$p$ maximal compact inductions do not have irreducible admissible subrepresentations**'
---
Introduction
============
Let $F$ be a non-archimedean local field of residue characteristic $p$, and $G$ be a $p$-adic split reductive group defined over $F$. Let $K$ be a hyperspecial maximal compact subgroup of $G$. Let $\sigma$ be an irreducible smooth $\overline{\mathbf{F}}_p$-representation of $K$. We show in this short note:
\[main in intro\] The compactly induced representation $\textnormal{ind}^G _K \sigma$ does not have irreducible admissible subrepresentations.
Proof of Theorem \[main in intro\]
==================================
Let $\sigma$ be an irreducible smooth $\overline{\mathbf{F}}_p$-representation of $K$.
Assume $\pi$ is an irreducible admissible $\overline{\mathbf{F}}_p$-representation of $G$ contained in $\textnormal{ind}^G _K \sigma$, i.e., we are given a $G$-embedding
$\iota: \pi \hookrightarrow \textnormal{ind}^G _K \sigma$.
Take an irreducible smooth $\overline{\mathbf{F}}_p$-representation $\sigma_1$ of $K$ contained in $\pi$. By Frobenius reciprocity, we get a non-zero $G$-map $\theta$ in the space $\text{Hom}_G (\textnormal{ind}^G _K \sigma_1, \pi)$, which will be denoted by $H(\sigma_1, \pi)$. Since $\pi$ is admissible, the space $H(\sigma_1, \pi)$ is finite dimensional. By composition, the space $H(\sigma_1, \pi)$ is a right module over the spherical Hecke algebra $\mathcal{H} (K, \sigma_1): =\text{End}_G (\textnormal{ind}^G _K \sigma_1)$. As $G$ is split, the algebra $\mathcal{H} (K, \sigma)$ is commutative ([@Her2011a Corollary 1.3]). Therefore we may replace $\theta$ by an eigenvector. That is to say, there is a character $\chi: \mathcal{H} (K, \sigma_1)\rightarrow \overline{\mathbf{F}}_p$ so that $\theta$ is annihilated by the kernel of $\chi$.
Now under our assumption, the composition $\iota\circ \theta$ is a non-zero map in $\text{Hom}_G (\textnormal{ind}^G _K \sigma_1, \textnormal{ind}^G _K \sigma)$. We take a non-zero $T$ in the kernel of $\chi$. We get
$(\iota\circ \theta)\circ T =\iota \circ (\theta\circ T)=0$.
As $T$ and $\iota\circ \theta$ are both non-zero, we get a contradiction by the argument of [@Her2011b Corollary 6.5].
Note that we have assumed $\pi$ is admissible in the theorem. However, in certain cases such an assumption is not necessary, say $G= GL_2/ U(2, 1)$, as in both cases a weaker substitute, i.e., the existence of Hecke eigenvalues ([@B-L94], [@X2018a]), is available. Note also that Daniel Le proved recently that there are non-admissible irreducible mod-$p$ smooth representations of $GL_2 (\mathbf{Q}_{p^3})$ ([@Le2018]).
One interest of Theorem \[main in intro\] is to compare it with the **complex** case: when $G$ is a classical group, a complex maximal compactly induced representation of $G$ might often happen to be irreducible.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We use renormalization group methods to derive equations of motion for large scale variables in fluid dynamics. The large scale variables are averages of the underlying continuum variables over cubic volumes, and naturally live on a lattice. The resulting lattice dynamics represents a perfect discretization of continuum physics, i.e. grid artifacts are completely eliminated. Perfect equations of motion are derived for static, slow flows of incompressible, viscous fluids. For Hagen-Poiseuille flow in a channel with square cross section the equations reduce to a perfect discretization of the Poisson equation for the velocity field with Dirichlet boundary conditions. The perfect large scale Poisson equation is used in a numerical simulation, and is shown to represent the continuum flow exactly. For non-square cross sections we use a numerical iterative procedure to derive flow equations that are approximately perfect.'
author:
- |
E. Katz and U.-J. Wiese\
\
Center for Theoretical Physics,\
Laboratory for Nuclear Science, and Department of Physics\
Massachusetts Institute of Technology (MIT)\
Cambridge, Massachusetts 02139, U.S.A.\
\
MIT Preprint, CTP 2423\
\
title: 'Lattice Fluid Dynamics from Perfect Discretizations of Continuum Flows [^1]'
---
addtoreset[equation]{}[section]{}
Introduction
============
Problems of fluid flow are of great practical importance in such areas as mechanical engineering and meteorology. A full theoretical understanding of fluid flow is very difficult because of the complex dynamics. In particular, in the regime of fully developed turbulence we face a dynamical system with fluctuations on all length scales. The small scale structures influence the dynamics at larger scales in a way that is hard to control analytically. Numerical simulations are therefore an essential tool in fluid dynamics. However, with present day computers a full simulation of realistic problems from first principles is beyond reach. Already simple model problems exhaust the biggest machines available. Hence, in practical applications one makes additional assumptions, especially about the small scale dynamics, which are difficult to justify theoretically, and which involve free parameters. In fact, the primary systematic error in numerical simulations of fluid flow is due to the finite grid size that is necessary to discretize the continuum Navier-Stokes equations.
The main purpose of this paper is to point out that perfect discretizations, i.e. ones that are completely free of grid artifacts, exist and can be constructed explicitly. The existence of perfect discretizations is a direct consequence of Wilson’s renormalization group theory [@Wil74]. The idea of the renormalization group is to deal with the fluctuations at different length scales step by step in scale. In this way the influence of small scale fluctuations on the large scale dynamics can, at least in principle, be controlled exactly. The renormalization group approach has been useful in the study turbulence and has been applied to various problems previously (see [@Ors1], [@Ors2], [@star] for some examples).
Recently, however, in the context of lattice quantum field theories so-called perfect actions have been constructed for various models using the renormalization group. Originally, Hasenfratz and Niedermayer iterated a renormalization group transformation to obtain a perfect discretization for the 2-dimensional $O(3)$ nonlinear $\sigma$-model, and they demonstrated that grid artifacts were eliminated even on coarse lattices [@Has94]. The same method has been applied to lattice fermions [@Wie93; @Bie95], to 4-dimensional pure $SU(3)$ gauge theory [@DeG95] and to full QCD [@Bie95a]. Here we use similar methods in classical physics. In principle, the renormalization group can be used to derive perfect discretizations for any differential equation. Examples of great practical importance are the Laplace and Poisson equations, Maxwell’s equations and the Navier-Stokes equation. A perfect discretization of the Laplace equation was obtained by Bell and Wilson in an early study of scalar lattice field theory [@Bel75]. The perfect action for a classical gauge field — and hence a perfect discretization of Maxwell’s equations — has been constructed in refs.[@DeG95; @Bie95a]. In this work we concentrate on the Navier-Stokes equations. Here we do not treat the problem in its full complexity yet. Specifically, we restrict ourselves to incompressible fluids and we assume static, slow flows.
The first step in our program is to define large scale lattice variables from the underlying continuum fields. In the case of the Navier-Stokes equations these are the velocity and pressure fields. We average the continuum pressure over a cube of side length $a$ to define the large scale pressure variables, which then naturally live on a lattice of spacing $a$ consisting of the cube centers. Similarly, the velocity field is averaged over the interface between two neighboring cubes, and hence the corresponding lattice field lives on the links connecting neighboring lattice points. This construction ensures that the continuity equation maintains a simple form on the lattice. By construction we know exactly how the large scale variables are related to the continuum fields. The goal of the next step is to derive equations of motion for the large scale fields. Solving these equations yields exact results for the averaged continuum quantities. This is in contrast to standard discretization procedures, where the lattice field represents an approximation of the continuum field at the same point. The approximation becomes exact only in the continuum limit $a \rightarrow 0$, because the value of the field at a specific point is influenced by the dynamics at arbitrarily small scales. A perfect discretization, on the other hand, ensures exact results for averaged continuum quantities already at finite $a$. Still, the values of the continuum fields can be reconstructed from the lattice data by using so-called perfect fields.
The second step is to derive exact equations of motion for the lattice fields. One possibility to do that is to iterate a discrete renormalization group transformation. This is done by starting on a very fine lattice (in fact, at the end the lattice spacing of the fine lattice is sent to zero), and then performing an infinite sequence of blocking steps, in which the lattice fields are averaged over blocks, and effective equations of motion for the block variables are derived. Here we proceed differently by blocking directly from the continuum. The result is identical to the discrete, iterative procedure, but the derivation is more transparent. Wilson was the first to discover this technique [@Wil76], which has not been very well known until recently. For example, blocking from the continuum was also used in the derivation of the QCD perfect action [@Bie95a]. In contrast to quantum field theory there is no action for the full Navier-Stokes problem. Still, blocking from the continuum can be directly applied to the equations of motion. The constraint that identifies the large scale lattice variables with the averaged continuum fields is implemented through a Lagrange multiplier. Solving the equations for the continuum variables in terms of the lattice fields and plugging the result back into the constraint equations leads to the equations for the large scale variables.
For numerical applications it is essential that the lattice equations of motion are as local as possible. In this paper we distinguish three types of locality: ultralocal, local and nonlocal. A lattice equation is called ultralocal if it couples the lattice field at a given site to a finite number of neighboring sites only. Consider, for example, the 1-dimensional Laplace equation $$\frac{d^2 f(x)}{d x^2} = 0.$$ Its standard lattice discretization $$\label{standard}
\frac{1}{a^2}[f(x+a) - 2 f(x) + f(x-a)] = 0$$ is ultralocal because it couples the field at the point $x$ to its nearest neighbors at $x+a$ and $x-a$ only. In general, a perfect discretization will not lead to an ultralocal coupling. The couplings to fields at distant sites are non-zero, but exponentially suppressed. In that case we say that the lattice equation is local. By optimizing the parameters of the renormalization group transformation one can control the strength of the exponential decay. It turns out that the parameters can be chosen such that the perfect equations become ultralocal in one dimension. In practice, even in higher dimensions this leads to extremely local perfect lattice equations, whose exponentially suppressed couplings to distant neighbors can safely be neglected. In particular, as it will become clear later, even the ultralocal eq.(\[standard\]) turns out to be perfect. This would not be the case for the common approach to discretization problems, where one interprets lattice derivatives as approximations of continuum derivatives at the same point. Still, even then one can find a perfect discretization. It is a well known result in engineering that a smooth function $f(x)$ can be reconstructed from its values $f(n a)$ on a lattice by $$f(x) = \sum_{n \in {{\sf Z \!\!\! Z}}} f(n a) \frac{\sin \pi(x/a - n)}
{\pi (x/a - n)}.$$ Here we have assumed that $f(x)$ is infinitely differentiable and that its Fourier transform exists. The above equation yields $$\frac{d^2 f(m a)}{dx^2} = - \frac{1}{a^2} \Big[\frac{\pi^2}{3} f(0) +
2 \sum_{n \in {{\sf Z \!\!\! Z}}\backslash \{ m \} } f(n a) \frac{(-1)^{n-m}}{(n-m)^2}\Big].$$ This exactly represents the continuum second derivative of the function at the lattice point $m a$. However, now the lattice derivative comprises terms at arbitrarily large distances, and their contributions are suppressed only power-like. This is what we call a nonlocal coupling. In this case the contributions from large distances cannot be neglected and therefore this equation is not of practical use. In contrast, the advantage of using block averaged fields is that their perfect discretizations are local.
As opposed to previous applications of perfect discretizations, which dealt with infinite volumes or finite systems with periodic boundary conditions, the incorporation of more general boundary conditions is essential in fluid dynamics. While it is straightforward to implement periodic boundary conditions, it is more difficult to treat fixed boundary conditions, which occur in typical fluid dynamics applications. Still, it turns out that fixed boundary conditions can be represented exactly in the perfect lattice equations. Here we restrict ourselves to channels whose cross section is a rectangular polygon. Arbitrarily curved boundaries can in principle be handled using curvelinear coordinates, but we do not elaborate on this issue here.
The paper is organized as follows. In section 2 we introduce the large scale lattice pressure and velocity variables by blocking from the continuum, and we investigate their behavior under discrete renormalization group transformations. Section 3 contains the derivation of perfect equations of motion for Hagen-Poiseuille flow. The parameter in the renormalization group transformation is optimized for ultralocality in two dimensions. We verify explicitly that also in three dimensions continuum physics is exactly reproduced. In section 4 we show analytically that the equations of motion indeed satisfy a factor 2 renormalization group step. Section 5 presents results of a numerical calculation of the lattice fields. In section 6 we construct the perfect velocity field that allows reconstruction of the continuum field from the lattice data. Section 7 deals with static, slow flows in general. We derive the perfect lattice version of the corresponding Navier-Stokes equations. In section 8 we present a perturbative method of introducing the non-linear term to equation of motion. Finally, section 9 contains concluding remarks on how to extend our construction to the full Navier-Stokes equations and to arbitrarily shaped boundaries. We also speculate about applications of perfect discretizations to numerical simulations of turbulent flows.
Large scale lattice variables
=============================
The continuum Navier-Stokes equations are formulated in terms of pressure and velocity variables. Here we consider static flows in $d$ dimensions, and hence pressure $p(y)$ and velocity $v_i(y)$ with $i \in \{1,2,...,d \}$ are functions of the position vector $y$ only. For incompressible fluids the velocity field obeys the continuity equation $$\partial_i v_i(y) = 0.$$ To define the large scale lattice pressure variables $P_x$ the continuum pressure field $p(y)$ is averaged over a $d$-dimensional cube $c_x$ of side length $a$ centered at $x$, $$P_x = \frac{1}{a^d} \int_{c_x} d^dy \ p(y).$$ The variables $P_x$ naturally live on a lattice of spacing $a$ formed by the cube centers $x$. Similarly, for the velocity field we define $$V_{i,x} = \frac{1}{a^{d-1}} \int_{f_{i,x}} d^{d-1}y \ v_i(y),$$ where $f_{i,x}$ is the $(d-1)$-dimensional face separating the cubes $c_{x-a\hat i/2}$ and $c_{x+a\hat i/2}$. Note that $\hat i$ is the unit vector in the $i$-direction. It is natural to associate the velocity variables with the links connecting neighboring lattice sites $x-a\hat i/2$ and $x+a\hat i/2$. Now the $i$-component of $x$ is a half-integer multiple of the lattice spacing. The lattice velocity variables can be interpreted as volume flow rates per cross sectional area of the faces separating neighboring cubes. This definition of the lattice velocity field is natural because the lattice continuity equation then assumes an ultralocal form. It is simply given by $$\begin{aligned}
\delta V_x&=&\sum_i [V_{i,x+a\hat i/2} - V_{x-a\hat i/2}]=
\frac{1}{a^{d-1}} \sum_i \Big[ \int_{f_{i,x+a\hat i/2}} \!\!\!\!\!\!\!\!\!\!
d^{d-1}y \ v_i(y)
- \int_{f_{i,x-a\hat i/2}} \!\!\!\!\!\!\!\!\!\!
d^{d-1}y \ v_i(y) \Big] \nonumber \\
&=&\frac{1}{a^{d-1}} \int_{\partial c_x} d^{d-1}\sigma_i \ v_i(y) =
\frac{1}{a^{d-1}} \int_{c_x} d^dy \ \partial_i v_i(y) = 0.\end{aligned}$$ Here we have used Gauss’ law together with the continuum continuity equation. The definition of lattice variables by blocking from the continuum can be viewed as a renormalization group transformation with an infinite blocking factor. In fact, the same result could be obtained by iterating block factor 2 transformations starting from an arbitrarily fine lattice. To identify the corresponding factor 2 transformation we now block from the continuum to a lattice of twice the lattice spacing $a' = 2 a$. The continuum pressure variables are then integrated over cubes $c'_{x'}$ of side length $a'$ which contain $2^d$ original cubes $c_x$. Hence, the resulting pressure variable $$P'_{x'}=\frac{1}{{a'}^d} \int_{c'_{x'}} d^dy \ p(y) =
\frac{1}{2^d} \sum_{x \in x'} \frac{1}{a^d} \int_{c_x} d^dy \ p(y) =
\frac{1}{2^d} \sum_{x \in x'} P_x$$ on the coarse lattice is a block average of the pressure variables on the fine lattice. We use $x \in x'$ to denote that the cube $c_x$ belongs to the block $c'_{x'}$. The corresponding geometry is illustrated in figure \[blocking\].
\[blocking\]
Similarly, for the velocity field we have $$\begin{aligned}
V'_{i,x'}&=&\frac{1}{{a'}^{d-1}} \int_{f'_{i,x'}} d^{d-1}y \ v_i(y)
\nonumber \\
&=&\frac{1}{2^{d-1}} \sum_{x \in x'} \frac{1}{a^{d-1}} \int_{f_{i,x}} d^{d-1}y
\ v_i(y) = \frac{1}{2^{d-1}} \sum_{x \in x'} V_{i,x}.\end{aligned}$$ Here $x \in x'$ denotes that the sum includes faces $f_{i,x}$ on the fine lattice that belong to the face $f'_{i,x'}$ on the coarse lattice. Of course, by construction the coarse lattice velocity variables also obey the continuity equation. This block factor 2 renormalization group transformation (RGT) will be used later to demonstrate that the structure of the perfect equations of motion reproduces itself under renormalization.
Perfect equations of motion for Hagen-Poiseuille flow
=====================================================
For simplicity we first discuss the derivation of perfect equations of motion in the context of Hagen-Poiseuille flow. The general equations for static, slow flows are derived in section 6. We start from the Navier-Stokes equation $$\partial_t v_i(y) + (v_j(y) \partial_j) v_i(y) =
- \frac{1}{\rho} \partial_i p(y) + \nu \partial_j \partial_j v_i(y).$$ Here $\nu$ is the kinematic viscosity of the fluid, and $\rho$ is its density. We restrict ourselves to incompressible flows, i.e. to a constant density, and thus the continuity equation applies. The appropriate boundary condition for a viscous fluid at a fixed wall is $v_i(y) = 0$ for all $i$. Furthermore, we consider static Hagen-Poiseuille flow, i.e. a constant pressure gradient $\delta p$, in a channel with square cross section $A=L^{d-1}$ along the $d$-direction. Hence, the pressure is given by $p(y) = \delta p \ y_d$. The velocity is along the channel, i.e. only $v_d(y)$ is nonzero and in addition independent of $y_d$. Then the equations reduce to the $(d-1)$-dimensional Poisson equation $$\partial_i \partial_i v_d(y) = \frac{\delta p}{\nu \rho},$$ with Dirichlet boundary conditions $v_d(y) = 0$ at the edge of the square. This equation follows from a variational principle of an action $$S[v_d] = \int_A d^{d-1}y [\frac{1}{2} \partial_i v_d(y) \partial_i v_d(y) +
\frac{\delta p}{\nu \rho} v_d(y)].$$ Note that here $i$ runs from 1 to $(d-1)$ only. The action is useful in our derivation of the perfect equations of motion. Yet we want to stress that the existence of an action is not really necessary. This is crucial because there is no action principle for the full Navier-Stokes problem. In fact, writing the above action was possible only because we have removed the time dependent dissipative element of the full problem and have forced the flow to be static.
What is the perfect discretization of the Poisson equation? To answer that question we consider the large scale lattice variable $$\label{constraint}
V_{d,x} = \frac{1}{a^{d-1}} \int_{f_{d,x}} d^{d-1}y \ v_d(y).$$ Note that the face $f_{d,x}$ is nothing more than a $(d-1)$-dimensional cube, and $V_{d,x}$ behaves practically as a $(d-1)$-dimensional scalar. We want to impose eq.(\[constraint\]) as a constraint for each face $f_{d,x}$. For this purpose we add it to the action using a Lagrange multiplier field $\eta_{d,x}$ that lives on the face $f_{d,x}$ $$\begin{aligned}
S[v_d,V_d,\eta_d]&=&\int_A d^{d-1}y
[\frac{1}{2} \partial_i v_d(y) \partial_i v_d(y) +
\frac{\delta p}{\nu \rho} v_d(y)] \nonumber \\
&+&\sum_x a^{d-1} [\eta_{d,x}
(V_{d,x} - \frac{1}{a^{d-1}} \int_{f_{d,x}} d^{d-1}y \ v_d(y))
- \frac{a^2\alpha}{2} \eta_{d,x}^2].\end{aligned}$$ We have also included a term quadratic in the Lagrange multiplier field. This would not really be necessary, but its coefficient $\alpha$ will allow us to optimize the perfect equation’s locality, and therefore it is of great practical importance. Next we derive equations of motion from the variational principle of the action. Varying the action with respect to $v_d(y)$ yields $$\label{Navierv}
\partial_i \partial_i v_d(y) - \frac{\delta p}{\nu \rho} +
\sum_x \eta_{d,x} \theta_{d,x}(y) = 0,$$ where $\theta_{d,x}(y) = 1$ for $y \in f_{d,x}$ and zero otherwise. Varying the action with respect to $\eta_{d,x}$ one obtains $$\label{Navier}
V_{d,x} - \frac{1}{a^{d-1}} \int_{f_{d,x}} d^{d-1}y \ v_d(y)
- a^2 \alpha \eta_{d,x} = 0.$$ It appears that the constraint of eq.(\[constraint\]) is correctly implemented only for $\alpha = 0$. However, varying the action with respect to $V_{d,x}$ gives $$\eta_{d,x} = 0.$$ To solve the above equations we now go to momentum space. Since the velocity vanishes at the boundary the appropriate Fourier transform takes the form $$v_d(k) = \int_A d^{d-1}y \ v_d(y) \prod_{i=1}^{d-1} \sin(k_i y_i).$$ Here the $$\label{momentum}
k_i = \frac{\pi}{L} m_i,$$ are discrete momenta with $m_i \in {{\sf I \!\! N}}$. Note that $v_d(k)$ extends naturally to negative momentum values because $$v_d(k_1,...,-k_i,...,k_{d-1}) = - v_d(k_1,...,k_i,...,k_{d-1}).$$ Similarly, for the lattice velocity field we have $$V_d(k) = a^{d-1} \sum_x V_{d,x} \prod_{i=1}^{d-1} \sin(k_i y_i).$$ Now the sum is finite and extends over $m_i = 1,2,...,N$ only, where $N = L/a$. The Lagrange multiplier field $\eta_{d,x}$ is transformed analogously. In Fourier space the action takes the form $$\begin{aligned}
&&S[v_d,V_d,\eta_d]=\frac{1}{a^{d-1}}\sum_{k \in (\frac{\pi N}{L})^{d-1}} [\frac{1}{2} k^2 v_d(k)^2
+ \frac{\delta p}{\nu \rho} \delta_L(k) v_d(k)] \nonumber \\
&&+\frac{1}{a^{d-1}}\sum_k [(V_d(k) - \sum_{l \in {{\sf Z \!\!\! Z}}^{d-1}} v_d(k + 2 \pi l/a)
\Pi_d(k + 2 \pi l/a) \prod_{i=1}^{d-1} (-1)^{l_i}) \eta_d(k) \nonumber \\-
&&\frac{a^2 \alpha}{2} \eta_d(k)^2]. \nonumber \\ \\end{aligned}$$ The second sum over $k$ is restricted to $m_i \in \{1,2,...,N\}$. In the above expression we have introduced $$\delta_L(k) = \prod_{i=1}^{d-1} \frac{1 - (-1)^{m_i}}{k_i}$$ as the Fourier transform of the constant 1. The function $$\Pi_d(k) = \prod_{i=1}^{d-1} \frac{2 \sin(k_i a/2)}{k_i a}$$ results from the Fourier transform of $\theta_{d,x}(y)$. In Fourier space eq.(\[Navierv\]) takes the form $$\label{equationofmotion}
- k^2 \ v_d(k) - \frac{\delta p}{\nu \rho} \delta_L(k) + \Pi_d(k) \eta_d(k) = 0.$$ In eq.(\[equationofmotion\]) the momentum $k$ has the form of eq.(\[momentum\]) with $m_i \in N$ being an arbitrary integer. On the other hand, as it stands $\eta_d(k)$ is defined only for momenta with $m_i \leq N$. For other momenta $\eta_d(k)$ naturally extends to $$\begin{aligned}
&&\eta_d(k_1,...,-k_i,...,k_{d-1}) = - \eta_d(k_1,...,k_i,...,k_{d-1}),
\nonumber \\
&&\eta_d(k + 2 \pi l_i \hat i/a)=(-1)^{l_i} \eta_d(k).\end{aligned}$$ Solving eq.(\[equationofmotion\]) yields $$\label{vequation}
v_d(k) = \frac{1}{k^2}[\Pi_d(k) \eta_d(k) - \frac{\delta p}{\nu \rho}
\delta_L(k)].$$ We plug this back into the action and obtain $$S[V_d,\eta_d]=\frac{1}{a^{d-1}}\sum_k [- \frac{1}{2} \eta_d(k) \omega_{dd}(k)^{-1} \eta_d(k)
+ \frac{\delta p}{\nu \rho} \Delta_L(k) \eta_d(k) +
V_d(k) \eta_d(k)],$$ where $$\begin{aligned}
\label{omega}
\omega_{dd}(k)^{-1}&=&\sum_{l \in {{\sf Z \!\!\! Z}}^{d-1}} \frac{1}{(k + 2 \pi l/a)^2}
\Pi(k + 2 \pi l/a)^2 + \alpha a^2 \nonumber \\
\Delta_L(k)&=&\sum_{l \in {{\sf Z \!\!\! Z}}^{d-1}} \delta_L(k + 2 \pi l/a)
\frac{1}{(k + 2 \pi l/a)^2} \Pi_d(k + 2 \pi l/a) (-1)^{\sum_i l_i}.\end{aligned}$$ Minimizing with respect to the auxiliary field $\eta_d(k)$ yields $$\label{etaequation}
\eta_d(k) = \omega_{dd}(k)(V_d(k)
+ \frac{\delta p}{\nu \rho} \Delta_L(k)).$$ Reinserting this into the action finally gives $$S[V_d] = \frac{1}{a^{d-1}}\sum_k [\frac{1}{2} V_d(k) \omega_{dd}(k) V_d(k) +
\omega_{dd}(k) \frac{\delta p}{\nu \rho} \Delta_L(k) V_d(k)].$$ This is a perfect action for the Poisson equation in a finite volume. The resulting perfect equation of motion takes the form $$\label{perfecteq}
\omega_{dd}(k) V_d(k) +
\omega_{dd}(k) \frac{\delta p}{\nu \rho} \Delta_L(k) = 0.$$ By construction it is clear that this equation perfectly represents continuum physics. It was helpful to have an action whose variation produced the perfect equation of motion, but the action was not really necessary. As we will see later, all we really need are the equations of motion. This is important for the full Navier-Stokes problem, for which an action does not exist.
For practical applications eq.(\[perfecteq\]) needs to be transformed back to coordinate space. It should be noted that $\omega_{dd}(k)$ contains the arbitrary parameter $\alpha$. Hence, eq.(\[perfecteq\]) represents a whole family of perfect equations of motion. Now we want to optimize $\alpha$ such that the equations become as local as possible. For this purpose we consider $d=2$. Then the sums in eq.(\[omega\]) can be performed analytically, and one obtains $$\begin{aligned}
\omega_{22}(k_1)^{-1}&=&a^2[\frac{1}{4 \sin^2(a k_1/2)} - \frac{1}{6}
+ \alpha]
\nonumber \\
\Delta_L(k_1)&=& a^3[\frac{1}{8 \sin^3(a k_1/2)} - \frac{1}{6}
\frac{1}{2 \sin(a k_1/2)}](1 - (-1)^{m_1}),\end{aligned}$$ where again $k_1=\pi m_1/L$.
For the choice of $\alpha = 1/6$ this gives $\omega_{22}(k_1) = 4/a^2 \sin^2(a k_1/2)$, which corresponds to the standard ultralocal second derivative in coordinate space. Transforming the whole perfect equation of motion eq.(\[perfecteq\]) to coordinate space one obtains $$\label{perfect2d}
\frac{1}{a^2} [V_{2,x_1+a} - 2 V_{2,x_1} + V_{2,x_1-a}] =
\frac{\delta p}{\nu \rho},$$ for $x_1 = (n - 1/2)a$ with $n \neq 1, N$, i.e. for points away from the boundary. At the left boundary one finds $$\frac{1}{a^2} [V_{2,3a/2} - 3 V_{2,a/2}] =
\frac{2}{3} \frac{\delta p}{\nu \rho},$$ and there is an analogous expression at the right boundary. It is straightforward to verify that the averaged continuum solution does in fact satisfy these equations. In the continuum the Hagen-Poiseuille flow in $d=2$ is given by $$v_2(y_1) = \frac{1}{2} \frac{\delta p}{\nu \rho} y_1(y_1 - L).$$ Consequently, the corresponding large scale lattice variable takes the form $$V_{2,x_1} = \frac{1}{a} \int_{x_1-a/2}^{x_1+a/2} dy_1 \ v_2(y_1)
= \frac{1}{2} \frac{\delta p}{\nu \rho} [x_1(x_1 - L) + \frac{a^2}{12}],$$ which indeed solves eq.(\[perfect2d\]).
In $d=3$ the perfect equations of motion are no longer ultralocal. Still, when we use the optimized $\alpha = 1/6$ from $d=2$ also in 3 dimensions the equations turn out to be extremely local. In coordinate space the equation of motion takes the form $$\label{coordseqn}
\sum_{x'} \omega_{dd}(x,x') V_{d,x'} =
- \frac{\delta p}{\nu \rho} \Delta_L(x),$$ where $\omega_{dd}(x,x')$ and $\Delta_L(x)$ are Fourier transforms of the corresponding quantities in momentum space. Their values for a 3-d system with $N=5$ , $a=1$ are given in tables \[table1\], \[table2\] and \[table3\] for various values of $x$ and plots for the same values are shown in figures 2, \[fig2\], and 4. Also the values for $D_L(x)$ (with $\delta p/\nu \rho$ set to 1) are provided in table \[Tdel\], and displayed in Fig. 5.
$V_{(1/2,1/2)}$ $x=1/2$ $x=3/2$ $x=5/2$ $x=7/2$ $x=9/2$
----------------- ---------- ---------- ---------- ---------- ----------
$y=1/2$ 4.28599 -0.42637 -0.00118 -0.00013 -0.00000
$y=3/2$ -0.42637 -0.18734 -0.00227 -0.00003 -0.00000
$y=5/2$ -0.00118 -0.00227 0.00151 0.00006 0.00000
$y=7/2$ -0.00013 -0.00003 0.00006 -0.00001 0.00000
$y=9/2$ -0.00000 -0.00000 0.00000 0.00000 0.00000
: *Coupling of a corner field*[]{data-label="table1"}
\[fig1\]
$V_{(5/2,5/2)}$ $x=1/2$ $x=3/2$ $x=5/2$ $x=7/2$ $x=9/2$
----------------- ---------- ---------- ---------- ---------- ----------
$y=1/2$ 0.00151 -0.00070 -0.00188 -0.00070 0.00151
$y=3/2$ -00070 -0.19033 -0.61801 -0.19033 -0.00070
$y=5/2$ -0.00188 -0.61801 3.24027 -0.61801 -0.00188
$y=7/2$ -0.00070 -0.19033 -0.61801 -0.19033 -0.00070
$y=9/2$ 0.00151 -0.00070 -0.00188 -0.00070 0.00151
: *Coupling of a center field*[]{data-label="table2"}
\[fig2\]
$V_{(5/2,1/2)}$ $x=1/2$ $x=3/2$ $x=5/2$ $x=7/2$ $x=9/2$
----------------- ---------- ---------- ---------- ---------- ----------
$y=1/2$ -0.00118 -0.42769 3.85830 -0.42769 -0.00118
$y=3/2$ -0.00227 -0.18965 -0.61602 -0.18965 -0.00227
$y=5/2$ -0.00151 -0.00070 -0.00188 -0.00070 -0.00151
$y=7/2$ 0.00006 0.00002 -0.00011 0.00002 0.00006
$y=9/2$ 0.00000 0.00000 -0.00000 0.00000 0.00000
: *Coupling of an edge field*[]{data-label="table3"}
\[fig3\]
\* $x=1/2$ $x=3/2$ $x=5/2$ $x=7/2$ $x=9/2$
--------- --------- --------- --------- --------- ---------
$y=1/2$ 0.45985 0.66567 0.66624 0.66567 0.45985
$y=3/2$ 0.66567 0.99806 0.99902 0.99806 0.66567
$y=5/2$ 0.66624 0.99902 0.99894 0.99902 0.66624
$y=7/2$ 0.66567 0.99806 0.99902 0.99806 0.66567
$y=9/2$ 0.45985 0.66567 0.66624 0.66567 0.45985
: *The values of $D_L(x)$ for $\delta p/\nu\rho$ set to 1*[]{data-label="Tdel"}
\[fdel\]
Invariance of the lattice equations of motion under a factor 2 RGT
==================================================================
We would like to show explicitly that the lattice equations of motions remain unchanged under a factor 2 renormalization group transformation. To do this we must first find the recursion relation, relating the equations on the coarse lattice to the ones on the fine lattice. To find the recursion relation, we add to the action in the fine lattice variables $V_{d,x}$ (with spacing $a$) the constraint involving the coarse lattice variables $V'_{d,x'}$ (with spacing $2a$), $$V'_{d,x'} = \frac{1}{2^{d-1}} \sum_{x \in x'} V_{d,x},$$ using a Lagrange multiplier $\eta_{d,x'}$. Varying the action with respect to the various fields yields equations analogous to the ones relating the continuum fields to the lattice fields $$\begin{aligned}
\label{recursion}
\sum_y V_y \omega_{dd}(y,x) &+& \frac{\delta p}{ \nu \rho}D_L(x) -
\frac{1}{2^{d-1}} \sum_{x} \theta_{d,x'} \eta_{d,x'}(x) = 0 ,\nonumber \\
V'_{d,x'} &-&\frac{1}{2^{d-1}} \sum_{x \in x'} V_{d,x} -
\alpha_2 \eta_{d,x'}=0 , \nonumber \\
\eta_{d,x'}&=&0.\end{aligned}$$ Here $\theta_{d,x'}(x) = 1$ for $x \in x'$ and zero otherwise and $\alpha_2$ is a constant to be adjusted to fit with our earlier choice of optimization parameter $\alpha$. Fourier transforming the above equations gives $$\begin{aligned}
&&V_d(k) \omega_{dd}(k) + \frac{\delta p}{ \nu \rho} \omega_{dd}(k) \Delta_L(k) -
\tilde \Pi(k) \eta_d(k) = 0 , \nonumber \\
&&V'_d(k') - \sum_{m \in \{0,1\}^{d-1}} V_d(k' + \pi m/a) \tilde \Pi(k' + \pi m/a)
\prod_{i=1}^{d-1} (-1)^l_i - \alpha_2 \eta(k') =0 ,\nonumber \\
\eta(k')&=&0 ,\end{aligned}$$ where $k'_i = \pi/L n_i$ with $n_i \in \{1,N/2\}$ only, and $$\tilde \Pi(k) = \prod_{i=1}^{d-1} \cos(k_i a/2) = \frac{\Pi(2k)}{\Pi(k)}$$ is the Fourier transform of $\theta_x'(x)$. We note that in the above equations $\eta_d(k')$ naturally extends to momenta $n_i>N/2$ by $$\eta_d(k'+ \pi m_i \hat i/a) = (-1)^{m_i} \eta_d(k').$$ Solving the first equation for $V_d(k)$ we obtain $$V_d(k) = \frac{1}{2^{d-1}} \omega(k)^{-1} \tilde \Pi(k)
- \frac{\delta p}{ \nu \rho} \Delta_L(k).$$
Inserting this relation into the other two equations and comparing with eq.(\[perfecteq\]) yields recursion formulas for $\omega_{dd}(k)$ and $\Delta_L(k)$: $$\begin{aligned}
\omega'_{dd}(k')^{-1}&=& \sum_{m \in \{0,1\}^{d-1}}
\omega_{dd}(k' + \pi m/a)^{-1} \tilde \Pi(k'+ \pi m/a)^2 + \alpha_2 \nonumber \\
\Delta'_L(k')&=& \sum_{m \in \{0,1\}^{d-1}} \Delta_L(k' + \pi m/a)
\tilde \Pi(k' + \pi m/a) (-1)^{\sum_i m_i}. \end{aligned}$$ We now plug in the explicit form of $\omega_{dd}(k)$ from eq.(\[omega\]) into the first recursion relation $$\begin{aligned}
\omega'_{dd}(k')^{-1}&=& \sum_{m \in \{0,1\}^{d-1}}
\sum_{l \in {{\sf Z \!\!\! Z}}^{d-1}}\frac{1}{(k' + \pi (2l + m)/a)^2}
\Pi(2k' + 2 \pi (2l + m)/a)^2 \nonumber \\
&+&a^2 \alpha \frac{\Pi(2(k'+\pi m))^2}{\Pi(k'+\pi m)^2}
+ \alpha_2 \nonumber \\
&=&\sum_{l' \in {{\sf Z \!\!\! Z}}^{d-1}}\frac{1}{(k' + 2 \pi l'/a')^2} \prod_{i=1}^{d-1}
\frac{4 \sin^2(a'k'_i + 2 \pi l'_i)}{a'k'_i + 2 \pi l'_i} +
a^2 \alpha + \alpha_2 ,\end{aligned}$$ where we have let $l'_i = 2l_i + m_i$ and $a' = 2a$ in the last step. Now, letting $\alpha_2=3a^2\alpha$ we recover the familiar form for $\omega'_{dd}(k')$. Thus, we see that under the factor 2 recursion relation, $\omega_{dd}(k)$ for lattice spacing $a$ transforms into $\omega'_{dd}(k')$ for lattice spacing $2a$. It can be shown that $\Delta_L(k)$ behaves similarly. The fact that the equations transform properly indeed confirms the they are perfect lattice equations of motion.
The recursion relation also serves as a powerful tool for numerical methods because of the locality of the couplings. For example, one cannot solve analytically the lattice equations for a rectangular cross section from which a smaller rectangular slice has been removed. However, eq.(\[recursion\]) is not specific to a given channel geometry and hence it still applies. To find the coupings for inward corner points one need only choose a very fine lattice. On it specify the couplings already known, make ultralocal guesses for the unknown sites, and perform several recursions. Since the couplings are so local, if the lattice is fine enough they serve as a very good approximation to the actual values that define the equations of motion. Thus, the averaging process of the recursive step will tend to converge to fixed values, thereby providing the necessary information for the unknown sites. In fact, we could have derived all the couplings in the above tables in this fashion by making ultralocal naive guesses for the Poissonian and then performing recursions.
In addition, we have numerically calculated the averaged velocity field for $N=6$, $a=1$ and for $N=3$, $a=2$. This was done by first numerically transforming $\omega_{dd}(k)$ and $\omega_{dd}(k)\Delta_L(k)$ back into coordinate space. The values obtained were then used to solve eq.(\[coordseqn\]) for the lattice velocity field. The solution was found iteratively by making an initial guess for the field values and then using the equation of motion to find the field at $x$ in terms of the fields at $x' \neq
x$. With each iterative step the velocity field at $x$ was replaced (if different) by the value of the field as found through the equation of motion. The iteration was stopped once the field converged to stationary values. For the above cases the field values are provided in tables \[3T\], \[6T\] and plotted in figures 6 and 7. We have explicitly checked that a factor 2 RGT on the fine solution gives the coarse lattice fields values. The fact that they satisfy eq.(\[constraint\]) confirms that the equations of motion found are perfect.
\* $x=1/2$ $x=3/2$ $x=5/2$
--------- --------- --------- ---------
$y=1/2$ 0.81972 1.40674 0.81972
$y=3/2$ 1.40674 2.48867 1.40674
$y=5/2$ 0.81972 1.40674 0.81972
: *Velocity field values for $N=3$*[]{data-label="3T"}
\[3flow\]
\* $x=1/2$ $x=3/2$ $x=5/2$ $x=7/2$ $x=9/2$ $x=11/2$
---------- --------- --------- --------- --------- --------- ----------
$y=1/2$ 0.31513 0.68669 0.83510 0.83510 0.68669 0.31513
$y=3/2$ 0.68669 1.58922 1.97725 1.97725 1.58992 0.68669
$y=5/2$ 0.83510 1.97725 2.48758 2.48758 1.97725 0.83510
$y=7/2$ 0.83510 1.97725 2.48758 2.48758 1.97725 0.83510
$y=9/2$ 0.68669 1.58922 1.97725 1.97725 1.58992 0.68669
$y=11/2$ 0.31513 0.68669 0.83510 0.83510 0.68669 0.31513
: *Velocity field for $N=6$*[]{data-label="6T"}
\[6flow\]
The perfect velocity field
==========================
Up to now we have shown that the perfect lattice equations of motion give exact results for the continuum fields averaged over cubes. In many practical cases one is most interested in these averaged quantities. Still, in some cases one would also like to know the values of the continuum fields themselves. In particular, when one works on a very coarse lattice the averaged lattice quantities may not contain enough information, and one would like to extract information for continuum points. Fortunately, this is possible with a perfect discretization. In fact, one can reconstruct the continuum field from the lattice data. For this purpose we insert eq.(\[etaequation\]) into eq.(\[vequation\]) and obtain $$v_d(k) = \frac{1}{k^2}[\Pi_d(k) \omega_{dd}(k)(\frac{1}{a^2} V_d(k) +
\frac{\delta p}{\nu \rho} \Delta_L(k)) - \frac{\delta p}{\nu \rho} \delta_L(k)],$$ which gives the continuum velocity field in terms of the lattice field $V_d$. Transforming back to coordinate space one can reconstruct exactly the continuum velocity field. When one considers the continuum velocity field at lattice points one finds $$v_d(x) = \sum_{x'} \rho_{dd}(x,x') V_{d,x'} +
\frac{\delta p}{\nu\rho} \lambda_d(x).$$ For a 2-d system $\rho_{dd}$ and $\lambda_d$ are ultralocal and $$v_d(x) = \frac{1}{8}(V_{d,x+a} + 6 V_{d,x} + V_{d,x-a}) -
\frac{1}{3} \frac{\delta p}{\nu\rho}$$ for $x \neq a/2,(N-1/2)a$ and for the boundary point $x = a/2$ $$v_d(a/2) = \frac{1}{8}(5 V_{d,a/2} + V_{d,3a/2}) -
\frac{1}{3} \frac{\delta p}{\nu\rho}.$$ For continuum points which are not lattice points one can also obtain the values of $\rho_{dd}$ and $\lambda_d$ by Fourier transformation.
Perfect discretization for static slow flows
============================================
Now we turn to general static flows with not necessarily constant pressure gradients. In this case the Navier-Stokes equation reduces to $$\partial_j \partial_j v_i(y) = \frac{1}{\nu\rho} \partial_i p(y).$$ For reasons of simplicity we restrict ourselves to an infinite volume. The inclusion of boundary conditions is straightforward (though tedious) and can be done in analogy to the Hagen-Poiseuille flow. Although the above equation still follows from an action principle, here we restrict ourselves to working with the equations of motion alone. Eq.(\[Navierv\]) then takes the form $$\partial_i \partial_i v_j(y) - \frac{1}{\nu\rho} \partial_j p +
\sum_x \eta_{j,x} \theta_{j,x}(y) = 0,$$ and we still have $$V_{j,x} - \frac{1}{a^{d-1}} \int_{f_{j,x}} d^{d-1}y \ v_j(y) -
a^2 \alpha \eta_{j,x} = 0, \ \eta_{j,x} = 0.$$ Similarly, the continuity equation leads to $$\frac{1}{\nu\rho}\partial_i v_i(y) + \sum_x \eta_x \theta_x(y) = 0,$$ and we also have $$P_x - \frac{1}{a^d} \int_{c_x} d^dy \ p(y) - a^2 \beta \eta_x= 0, \ \eta_x = 0.$$ Note that we have included a parameter $\beta$ analogous to $\alpha$ to be optimized for locality later. It will turn out that this is not necessary, because the equations of motion can be optimized for locality without it. Since we are now in an infinite volume $$\begin{aligned}
v_j(k)&=&\int d^dy \ v_j(y) \prod_{i=1}^d \exp(-i k_i y_i),\\
p(k)&=&\int d^dy \ p(y) \prod_{i=1}^d \exp(-i k_i y_i),\end{aligned}$$ where $k \in {{\sf I \!\! R}}^d$. Hence, it is convenient to now let $$\begin{aligned}
V_j(k)&=&a^{d-1} \sum_{x} V_{j,x} \prod_{i \neq j} \exp(-i k_i x_i),\\
P(k)&=&a^d \sum_{x} P_x \prod_{i=1}^d \exp(-i k_i x_i),\end{aligned}$$ where $k \in [-\pi/a ,\pi/a]^d$. We note that given the geometry of Fig. 1, if $P_x$ lie at points $x$ whose coordinates are integers, then $V_{j,x}$ will lie at a point $x$ whose $j$-th coordinate is a half integer. Also, as before, it is possible to extend the lattice fields to momenta outside $[-\pi/a, \pi/a]$ by letting $$\begin{aligned}
V_j(k+2 \pi l_i \hat i/a) &=& V_j(k) , i \neq j \nonumber \\
V_j(k+2 \pi l_j \hat j/a) &=& (-1)^{l_j} V_j(k), \nonumber \\
P(k+2 \pi l_i \hat i/a) &=& P(k), \mbox{for all $i$}.\end{aligned}$$ We are thus ready to go to Fourier space where the above equations take the form $$\begin{aligned}
\label{transnavier}
&&-k^2 v_j(k) - \frac{i}{\nu \rho} k_j p(k) +\Pi_j(k) \eta_j(k)=0 \\
\label{aveV}
&&V_j(k) - \sum_{l \in {{\sf Z \!\!\! Z}}^d} v_j(k+2 \pi l/a) \Pi_j(k+2 \pi l/a)(-1)^{l_j}
-a^2 \alpha \eta_j=0 \\
&&\frac{i}{\nu \rho} k_j v_j(k) + \Pi(k) \eta(k)=0 \\
\label{aveP}
&&P(k) - \sum_{l \in {{\sf Z \!\!\! Z}}^d} p(k+2 \pi l/a) \Pi(k+2 \pi l/a) -a^2 \beta \eta(k)=0.\end{aligned}$$ Solving for the continuum fields we obtain $$\begin{aligned}
&&p(k) = -(\nu \rho)^2 \Pi(k) \eta(k) - \frac{i \nu \rho}{k^2} k_j
\Pi_j(k) \eta_j(k) \\
&&v_i(k) = \frac{i \nu \rho}{k^2} k_i \Pi(k) \eta(k) - \frac{k_i k_j
-k^2 \delta_{ij}}{k^4} \Pi_j(k) \eta_j(k).\end{aligned}$$ Reinserting these back into eq.(\[aveV\]) and eq.(\[aveP\]) gives equations of motion for the lattice fields and the Lagrange multiplier fields $$\begin{aligned}
&&V_i(k) - i\nu \rho \omega_i(k) \eta(k) +\omega_{ij}(k)
\eta_{j}(k)=0 \nonumber \\
&&P(k) + (\nu \rho)^2 \omega(k) \eta(k)+ i \nu \rho \omega_j(k)
\eta_j(k)=0.\end{aligned}$$ Here we have introduced the functions $$\begin{aligned}
&& \omega(k)=\sum_{l \in {{\sf Z \!\!\! Z}}^d} \Pi(k+2 \pi l/a)^2 + \beta = 1 + \beta \\
&& \omega_i(k)=\sum_{l \in {{\sf Z \!\!\! Z}}^d} \frac{k_i +2 \pi l_i/a}{(k+2 \pi
l/a)^2} \Pi_i(k + 2 \pi l/a) \Pi(k + 2 \pi l/a)(-1)^{l_i} \\
&& \omega_{ij}(k)^{-1}=\sum_{l \in {{\sf Z \!\!\! Z}}^d} \frac{(k_i +2 \pi l_i/a)(k_j +
2 \pi l_j/a)-(k+2 \pi l/a)^2 \delta_{ij}}{(k+2 \pi
l/a)^4} \nonumber \\
&& \times \Pi_i(k + 2 \pi l/a)\Pi_j(k + 2 \pi l/a)(-1)^{l_i+l_j}+ a^2 \alpha .\end{aligned}$$ We can rewrite the above equations in a simple fashion by letting $V_0=P$ and $\eta_0=\eta$ $$V_{\mu}=\Omega_{\mu \nu}^{-1} \eta_{\nu},$$ where $\Omega_{\mu \nu}$ is defined through the above relations. Finally, we solve for the Lagrange multiplier fields and set them equal to zero to obtain the equations of motion for the lattice fields $$\label{eqofmotionnavier}
\Omega_{ \mu \nu} V_{\nu}=0.$$ We must now optimize for locality. As before, we go to $d=1$ and choose $\alpha$ and $\beta$ so that the equations of motion take an ultralocal form. In one dimension eq.(\[eqofmotionnavier\]) becomes $$\frac{1}{(1+\beta)\alpha-1/(2\sin(k/2))^2}
\left( \begin{array}{cc}
\alpha & i/(2\sin(k/2)) \\
-i/(2\sin(k/2) &
1+\beta
\end{array} \right)
\left( \begin{array}{c} P \\ V \end{array} \right) = 0.$$ Setting $\alpha=0$ and $\beta=0$, we obtain familiar ultralocal equations $$\begin{aligned}
&& 2\sin(k/2) V(k)=0 \\
&& 2 i \sin(k/2) P(k) = (2\sin(k/2))^2 V(k).\end{aligned}$$ Having solved this simplified version of the Navier-Stokes equations we proceed to add the nonlinear term.
Perturbative treatment of the nonlinear term
============================================
Since the nonlinear term of the Navier-Stokes equation is quadratic in the velocity, it is possible to solve the lattice equations of motion only perturbatively. Here we demonstrate how to do this for the first order correction. We begin by introducing the modified version of eq.(\[transnavier\]), now with the new term $$-\frac{i}{\nu} v_j(k) k_j v_i(k) -k^2 v_i(k) -\frac{i}{\nu \rho} k_i v_i(k) +
\Pi_j(k) \eta_j(k) = 0.$$ Because we are only interested in the first order correction we need only keep terms in the nonlinear contribution that are quadratic in the Lagrange multiplier fields. Thus, for the nonlinear term, it is sufficient to express the continuum fields as linear functions of the Lagrange fields. We make the following ansatz for $p^{(2)}$ and $v_i^{(2)}$, the second order contributions: $$\begin{aligned}
&&p^{(2)} = A \eta(k)^2 + \eta(k) B_j(k) \eta_j(k) + \eta_i(k) C_{ij}(k)
\eta_j(k) \\
&&v_i^{(2)} = D_i(k) \eta(k)^2 + \eta(k) E_{ij} \eta_j(k) + F_{ijk}
\eta_j(k) \eta_k(k).\end{aligned}$$ Then $p^{(2)}$ and $v_i^{(2)}$ must obey the following equations, $$\begin{aligned}
&&\frac{i \nu \rho^2}{k^2}k_i \Pi(k)^2 \eta^2-\eta
\frac{\rho (k_i k_j-k^2\delta_{ij})}{k^4}\Pi(k)\Pi_j(k) \eta_j -k^2
v_i^{(2)}-\frac{i}{\nu\rho} k_i p^{(2)}=0. \\
&& k_i v_i^{(2)}(k) = 0.\end{aligned}$$ Solving them, we obtain $$\begin{aligned}
&&p^{(2)}(k) = \nu^2\rho^3 \frac{\Pi(k)^2}{k^2} \eta(k)^2 \\
&&v_i^{(2)}(k) = \rho \eta(k) \frac{k_i k_j - k^2 \delta_{ij}}{k^6} \Pi_j(k)\Pi(k) \eta_j(k).\end{aligned}$$ This procedure can now be repeated for the lattice fields. We make the ansatz, $$\eta_{\lambda}^{(2)}(k)=V_{\mu}(k) A_{\mu\nu}^{\lambda}(k) V_{\nu}(k),$$ where again $V_0=P$. Keeping only the second order terms, yields $$\begin{aligned}
&& -i\nu\rho\omega_i V_{\mu} A_{\mu\nu}^{0} V_{\nu}+\omega_{ij}(k)
V_{\mu} A_{\mu\nu}^{j} V_{\nu}
-\rho V_{\mu} \Omega_{\mu 0}^{T} S_{ij} \Omega_{j\nu} V_{\nu} = 0 \\
&& (\nu\rho)^2 \omega V_{\mu} A_{\mu\nu}^{0} V_{\nu}+i\nu\rho\omega_j V_{\mu} A_{\mu\nu}^{j} V_{\nu}
-\nu^2\rho^3 V_{\mu} \Omega_{\mu 0}^{T} S \Omega_{0 \nu} V_{\nu} = 0.\end{aligned}$$ Here, the following functions have been introduced $$\begin{aligned}
&& S(k)=\sum_{l \in {{\sf Z \!\!\! Z}}^d} \frac{1}{(k+2 \pi
l/a)^2} \Pi_i(k + 2 \pi l/a)^3 \\
&& S_{ij}(k)=\sum_{l \in {{\sf Z \!\!\! Z}}^d} \frac{(k_i +2 \pi l_i/a)(k_j +
2 \pi l_j/a)-(k+2 \pi l/a)^2 \delta_{ij}}{(k+2 \pi
l/a)^6} \nonumber \\
&& \times \Pi_i(k + 2 \pi l/a)\Pi_j(k + 2 \pi l/a)\Pi(k + 2 \pi l/a)(-1)^{l_i+l_j}.\end{aligned}$$ We rewrite the above equations in a more compact form $$\Omega^{-1}_{\sigma\lambda} A_{\mu\nu}^{\lambda}=M_{\mu\nu}^{\sigma}.$$ This, finally gives the first order nonlinear equation of motion $$\Omega_{\mu\nu} V_{\nu} + \Omega_{\mu\nu} V_{\lambda}
M_{\lambda\sigma}^{\nu} V_{\sigma} = 0.$$ In principal this procedure could be repeated to obtain non-linear lattice equations up to any order in the lattice fields.
Conclusions
===========
In this paper we have performed the first steps towards a perfect discretization of the Navier Stokes equations. The essential idea is to define coarse grained variables by averaging the continuum pressure and velocity fields over appropriate spatial regions. The corresponding averaged field variables naturally live on a lattice. In the cases of Hagen-Poisseille flow and slow static flows in general we derived the exact equations of motion for the lattice variables, and hence a perfect discretization of the corresponding continuum equations. The incorporation of boundary conditions is nontrivial, and has been done explicitly for Hagen-Poisseille flow. For practical applications it is essential that the perfectly discretized equations of motion are very local (although in general not ultralocal). In particular, the lattice couplings to distant variables decay exponentially, and one can safely restrict oneself to a few nearby neighbors. In fact, we have performed numerical simulations using the perfectly discretized equations of motion for laminar flow in a channel with quadratic cross section and we have verified explicitly that an arbitrarily coarse lattice indeed gives continuum answers. Our method can be directly applied to more general geometries, although non-cubic lattices require nontrivial modifications. An important aspect of our construction are the so-called perfect fields. Indeed our method allows to interpolate the continuum fields from the lattice data. Hence, especially when we are working on a coarse lattice, we are not limited to lattice points.
Of course, a method like that would be most welcome in numerical simulations of more complicated fluid dynamics systems, especially in the case of turbulence. In these simulations the effects of a finite lattice spacing are the main source of systematic errors. We have made a first step in this direction by deriving the perfect equations of motion in the presence of the nonlinear term that gives rise to turbulence. However, we have included that term only to first order, while a perfect discretization of turbulent flows would require a treatment to all orders. This could be achieved numerically in analogy to [@Has94] by solving a minimization problem on a multigrid. When one wants to simulate turbulence one should not restrict oneself to static flows. Then the definition of the averaged lattice variables should be generalized. In particular, one should also average pressure and velocity in time. In addition, one may want to simulate compressible fluids, which would bring in an averaged density variable. Our techniques are still applicable to these cases, but the actual implementation is nontrivial.
In conclusion, we have proposed a renormalization group treatment of problems in fluid dynamics. The renormalization group implies that perfect discretizations of the continuum Navier-Stokes equations exist. The perfectly discretized equations of motion have been constructed explicitly for slow static flows, and have been optimized for locality. This is essential in practical applications. The renormalization group offers a systematic way of treating small scale effects in fluid dynamics. Realizing this program for the full Navier-Stokes problem requires a lot of work. Still, the first steps that we have taken, have led to promising results.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank W. Evers for very important discussions during the initial stages of this project. We would also like to thank W. Bietenholz for very useful comments and remarks about the paper.
[11]{} K. Wilson and J. Kogut, Phys. Rep. C12 (1974) 75. P. Hasenfratz and F. Niedermayer, Nucl. Phys. B414 (1994) 785. U.-J. Wiese, Phys. Lett. B315 (1993) 417. W. Bietenholz, E. Focht and U.-J. Wiese, Nucl. Phys. B436 (1995) 393. T. DeGrand, A. Hasenfratz, P. Hasenfratz and F. Niedermayer, Nucl. Phys. B454 (1995) 587; 615. W. Bietenholz and U.-J. Wiese, hep-lat/9510026, to appear in Nucl. Phys. B. K. G. Wilson, in New Pathways in High Energy Physics II, ed. A. Perlmutter (Plenum, New York 1976) 243. T. L. Bell and K. Wilson, Phys. Rev. B11 (1975) 3431. Yakhot, V., and Orsag, S. A. (1986) Renormalization Group Analysis of Turbulence. I. Basic Theory J. Sci. Comput. 1, 3-51. Yakhot, A., Orsag, S. A., Yakhot, V., and Israeli, M. (1989) Renormalization Group Formulation of Large-Eddy Simulations. J. Sci. Comput. 4, 139-158. Staroselsky, I.,and Sukoriansky, S.(1993) Renormalization Group Approach to Two-Dimensional Turbulance and $\epsilon$- expansion for Vorticity Euation. In Branover, H., and Unger, Y. (ed.), Advances in Turbulance Studies, Progress in Astron. and Aeron. AIAA, 149, 159-164.
[^1]: This work is supported in part by funds provided by the U.S. Department of Energy (D.O.E.) under cooperative research agreement DE-FC02-94ER40818.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A $Z_{2L}\times Z_{2R}$ generation symmetry in the neutrino sector predicts the atmospheric neutrino mixing to be maximal, and the MNS matrix element $U_{e3}$ to be zero, consistent with observations. Solar neutrino mixing may be maximal but is not required by the symmetry. Neutrino masses of the first two generations are predicted to vanish, providing a first approximation to the oscillation data. The consequence of a smaller $Z_2$ symmetry is also discussed. In that case, deviation from the $Z_{2L}\times Z_{2R}$ result is of the order of the neutrino mass ratio between the first two generations and the third generation.'
address: |
Department of Physics, McGill University\
3600 University St., Montreal, QC, Canada H3A 2T8\
Email: Lam@physics.mcgill.ca
author:
- 'C.S. Lam'
title: 'Consequences of a $Z_2$ Symmetry for Neutrino Oscillations'
---
§[[S]{}]{} $${\left[}
\def$$[\]]{} ${\left(}
\def$[)]{} \#1[\#1]{} \#1[\#1|]{} \#1[|\#1]{} \#1[\[\#1\]]{} \#1[(\[\#1\])]{} ø
ł v\#1[\#1]{} .
Right-handed neutrinos must be present to explain the deficit of solar and atmospheric neutrinos through oscillations. These are unique probes at high energy for they have no standard-model quantum numbers to allow a pollution by standard-model interactions. Recent results from Super-Kamiokande suggest that the mixing for solar neutrinos and the atmospheric neutrinos are both maximal, and the $(\Delta m)^2$ for solar neutrino oscillation is much smaller than the $(\Delta m)^2$ for atmospheric neutrino oscillations [@SK]. The negative result of CHOOZ [@CH] also limits the magnitude of the MNS mixing matrix [@MNS] element $|U_{e3}|^2$ to be smaller than 0.02 to 0.047, depending on the exact value of the atomospheric neutrino mass difference. Is there a simple way to understand these facts? Many models have been proposed [@MO]. We suggest that a $Z_{2L}\times
Z_{2R}$ generation symmetry in the sea-saw scenario will naturally explain many of these observed results. By $Z_2$ we mean a finite group with elements $\{e,g\}$ so that $g^2$ is the identity element $e$. Specifically, under such a symmetry, the atmospheric neutrino mixing is maximal, and the MNS matrix element $U_{13}\equiv
U_{e3}$ is zero. The neutrino masses of the first two generations vanish, which may be a fair approximation to Nature since the mass difference observed in atmospheric neutrino oscillation is much larger than that seen in solar neutrino oscillation. Under such a symmetry, the solar neutrino mixing may be maximal but is not required to be so.
Let $D$ be the Dirac mass matrix for neutrinos in the flavor basis, and $M$ the Majorana mass matrix in the same basis for right-handed neutrinos. We shall assume oscillations to occur only among the active neutrinos, so both are $3 \times 3$ complex matrices, although $M$ is symmetrical. When the right-handed neutrinos are integrated out, the effective mass matrix for the left-handed neutrinos in the flavor basis is [@AK] m’=D\^TM\^[-1]{}D.To diagonalize it, we must use the MNS matrix $U$ to rotate the flavor basis into the energy eigenbasis, then m’=U\^\*mU\^,with $m={\rm diag}(m_1,m_2,m_3)$.
Let $\nu_{aL}$ and $\nu_{aR}$ be respectively the left- and right-handed neutrino field of the $a$th generation. Assume the mass terms in the neutrino Lagrangian to be invariant under the $Z_{2L}\times Z_{2R}$ transformation $\nu_{2A}\to -\nu_{3A},
\ \nu_{3A}\to -\nu_{2A}$, and $\nu_{1A}\to -\nu_{1A}$, where $A$ is either $L$ or $R$. Then the matrices $D$ and $M^{-1}$ must be of the form D&=&m\_D,\
\
M\^[-1]{}&=&m\_M\^[-1]{}. Using , this leads to m’=[2[m\_D]{}\^2(1-\_3)m\_M]{}. Other than the common scale factor which may be complex, the matrix is real and symmetrical, so it can be diagonalized by a real orthogonal transformation. The eigenvalues are $(0,0,1)$, but the matrix $U$ used for diagonalization is not unique because of degeneracy of the first two eigenvalues. For the purpose of later generalization it is useful to deduce these results explicitly from (m)\_[ij]{}&=&U\_[ai]{}(m’)\_[ab]{}U\_[bj]{}\
&=&[2[m\_D]{}\^2(1-\_3)m\_M]{}$U_{2i}-U_{3i}$$U_{2j}-U_{3j}$. In order for $m$ to be diagonal, two of the following three combinations must vanish: $U_{21}-U_{31}, U_{22}-U_{32}$, and $U_{23}-U_{33}$. Let us first consider the case when $U_{23}-U_{33}\not=0$, but the other two combinations vanish. In that case $m_1=m_2=0$ and $m_3=(2{m_D}^2(1-\mu_3)/m_M)
{(U_{23}-U_{33})}^2$, thus giving rise to the extreme limit of a normal hierarchy for the neutrino masses, which is a good approximation to reality because $(\Delta m)^2_{12}\ll
(\Delta m)^2_{23}$. To simplify writing let $a=U_{21}
=U_{31}$, $b=U_{22}=U_{32}$. Then $U$ is of the form U=.We will adopt the usual phase convention so that $c,d,f$ and $g$ are real, and the imaginary part of $a$ and of $b$ are proportional to that of $e$. Normalization of the second and third rows of the matrix $U$ requires $|a|^2+|b|^2+f^2=1=|a|^2+|b|^2+g^2$. Hence $f$ and $g$ have the same magnitude. Orthogonality of the second and the third rows of $U$ then requires $f=-g$ and $f^2=|a|^2+|b|^2=\h$. In other words, atmospheric neutrino mixing is [*maximal*]{}, consistent with the Super-Kamiokande observation. Using the fact that the first row of $U$ must be orthogonal to both the second and the third roles, we conclude that $e=0$, consistent with the CHOOZ observation. In that case there will be no CP violation observable through neutrino oscillations. With the present phase convention both $a$ and $b$ become real, and the magnitudes of $a,b,c,d$ are related by unitarity. So, there is only one free parameter $\theta_{12}$ left in describing $U$. For $c=-d=1/\sqrt{2}$, maximal mixing occurs in solar neutrino mixing, but this is not required to be so by the symmetry. In summary, judging from neutrino oscillations, the $Z_{2L}\times
Z_{2R}$ symmetry appears to be a good approximation to Nature.
The other two solutions of leads to either $m_1\not=0$ or $m_2\not=0$, with the other two diagonal matrix elements of $m$ zero. The former case leads to $U_{11}=0$, and the latter case leads to $U_{12}=0$, neither is consistent with solar neutrino and CHOOZ observations. They will therefore be rejected.
Note also that the symmetry interchanges generations 2 and 3, with a minus sign. If instead it interchanged generations 1 and 2, which at first sight might seem a more reasonable thing to do in view of the mass hierarchy, the result would be the same as above but with the first and third rows of the matrix $U$ permuted. This would not be consistent with experiment.
If the charged lepton and the quark Dirac mass matrices are subject to the same symmetry, then all of them would have a mass spectrum proportional to (0,0,1), not a bad first approximation. In that case the CKM matrix would be $\1$, again a reasonable first approximation given that the Cabibbo angle is small.
There is the question of how a small but non-zero value of $m_{1,2}$ can be obtained. One possibility is to have a small breaking of $Z_{2L}\times Z_{2R}$, into a diagonal $Z_2$ where the left-handed and the right-handed particles are simultaneously transformed. This does not alter the parametrization of $M$, but allows small parameters to occur in $D$ where its present matrix elements are zero. It should be possible to tune these parameters to get finite masses for the first two generations. Since a different set of parameters may be present for the charged leptons, the up-type quarks, and the down-type quarks, it is quite conceivable that different mass spectra can be obtained for these different particles, while obtaining also a realistic CKM matrix.
Let us now explore this scenario more systematically. The $Z_2$ symmetry requires the (symmetric) matrix $m'$ in and to obey $m'_{12}=m'_{13}$ and $m'_{22}=m'_{33}$. Using , these two requirements can be written as \_i\_i\_im\_i&=&0,\
\_i\_i\_im\_i&=&0,where \_i&=&U\^\*\_[1i]{},\
\_i&=&U\^\*\_[2i]{}-U\^\*\_[3i]{},\
\_i&=&U\^\*\_[2i]{}+U\^\*\_[3i]{},thus giving rise to constraints on $U$ controlled by the neutrino mass values. For the extreme hierarchy spectrum $(m_1,m_2,m_3)\propto(0,0,1)\equiv s_1$ discussed above, the constraints are $\alpha_3\beta_3=\gamma_3\beta_3=0$. Being second order equations there are two solutions: (1a) $\alpha_3=\gamma_3=0$, and (1b) $\beta_3=0$. Solution (1a) is the same as the $Z_{2L}\times Z_{2R}$ solution before. To see that, note that the third column of $U$ is fixed by this constraint and unitarity to be $(U_{13},U_{23},U_{33})=(0,f,-f)$, with $f^2=\h$. Orthogonality of the first two columns to the third then fixes $U$ to be of the form , with parameters identical to those in $Z_{2L}\times Z_{2R}$ as per the arguments given below . Solution (1b) gives $U_{23}=-U_{33}$, which is not very restrictive. For a realistic hierarchical spectrum $s_2=(\eta,\epsilon,1)$, with $\epsilon\ll 1$ and $\eta$ either of the same order or much less than $\epsilon$, the constraints become $\alpha_3\beta_3=O(\epsilon)$ and $\gamma_3\beta_3=O(\epsilon)$. The two corresponding solutions are: (2a) $\alpha_3=O(\epsilon)$, $\gamma_3=O(\epsilon)$, and $\beta_3=O(1)$, and (2b) $\beta_3=O(\e)$ and $\alpha_3,\gamma_3=O(1)$. As before, solution (2b) is not very restrictive. For (2a), $|U_{e3}|$ is of order $\e$ and the atmospheric neutrino oscillation is close to maximal, with deviation of the order of $\e=m_2/m_3$. In this way a $Z_2$ symmetry with a realistic mass spectrum as an input gives rise to practically the same predictions as the larger $Z_{2L}\times Z_{2R}$ symmetry.
Whether a more realistic mass spectrum should be obtained by an explicit breaking of $Z_{2L}\times Z_{2R}$ like the one described above, by renormalization-group corrections, or by something else, will be left to future investigations.
This research is supported in part by the Natural Sciences and Engineering Research Council of Canada, and the Fonds pour la formation de Chercheurs et l’Aide à la Recherche of Québec. I am indebted to T.K. Kuo, Greg Mahlon, Gary Shiu, and Tony Zee for stimulating discussions.
[9]{} Y. Suzuki, ‘Solar neutrino results from Super-Kamiokande’, H. Sobel, ‘The study of atmospheric neutrino with Super-Kamiokande’, Neutrino 2000 Conference (Sudbury, Canada; June, 2000) CHOOZE Collaboration, M. Apollonio [*et al*]{}., hep-ex/9907037. Z. Maki, M. Nakagawa, and S. Sakata, [*Prog. Theor. Phys.*]{} [**28**]{} (1962) 870. V. Barger, S. Pakvasa, T.J. Weiler, and K. Whisnant, hep-ph/9806387; A.J. Baltz, A.S. Goldhaber, and M. Goldhaber, hep-ph/9806540; Y. Nomura and T. Yanagida, hep-ph/9807325; G. Altarelli and F. Ferugio, hep-ph/9807353, 9809596, 9812475; S. Raby, hep-ph/9909279; R.N. Mohapatra and S. Nussinov, hep-ph/9807415; S.K. Kang and C.S. Kim, hep-ph/9811379; R. Barbieri, L.J. Hall, G.L. Kane, and G.G. Ross, hep-ph/9901228; E. Ma, hep-ph/9902392, 9909249; R.N. Mohapatra. A. Pérez-Corenzana, and C.A. de S. Pires, hep-ph/9911395; T.K. Kuo, G.-H. Wu, and S.W. Manosour, hep-ph/9912366; D. Chang and A. Zee, hep-ph/9912380; K.R.S. Balaji, A.S. Dighe, R.N. Mohapatra, and M.K. Parida, hep-ph/0001310; A. Aranda, C.D. Carone, and R.F. Lebed, hep-ph/0002044; C.H. Albright and S.M. Barr, hep-ph/0003251; Y. Koide and A. Ghosal, hep-ph/0008129. See, for example, G. Altarelli and F. Feruglio, hep-ph/9905536, and E.Kh. Akhmedov, hep-ph/0001264, for excellent reviews.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Learning high-quality domain word embeddings is important for achieving good performance in many NLP tasks. General-purpose embeddings trained on large-scale corpora are often sub-optimal for domain-specific applications. However, domain-specific tasks often do not have large in-domain corpora for training high-quality domain embeddings. In this paper, we propose a novel *lifelong learning* setting for domain embedding. That is, when performing the new domain embedding, the system has seen many past domains, and it tries to expand the new in-domain corpus by exploiting the corpora from the past domains via meta-learning. The proposed meta-learner characterizes the similarities of the contexts of the same word in many domain corpora, which helps retrieve relevant data from the past domains to expand the new domain corpus. Experimental results show that domain embeddings produced from such a process improve the performance of the downstream tasks.'
author:
- 'Hu Xu[^1^]{.nodecor}, Bing Liu[^1^]{.nodecor}, Lei Shu[^1^]{.nodecor}'
- |
Philip S. Yu[^1,2^]{.nodecor}\
^1^[Department of Computer Science, University of Illinois at Chicago, Chicago, IL, USA]{}\
^2^[Institute for Data Science, Tsinghua University, Beijing, China]{}\
{hxu48, liub, lshu3, psyu}@uic.edu
bibliography:
- 'ijcai18.bib'
title: 'Lifelong Domain Word Embedding via Meta-Learning'
---
Introduction {#intro}
============
Learning word embeddings [@mnih2007three; @turian2010word; @mikolov2013efficient; @mikolov2013distributed; @pennington2014glove] has received a great deal of attention due to its success in numerous NLP applications, e.g., named entity recognition [@sienvcnik2015adapting], sentiment analysis [@maas2011learning] and syntactic parsing [@durrett2015neural]. The key to the success of word embeddings is that a large-scale corpus can be turned into a huge number (e.g., billions) of training examples. Two implicit assumptions are often made about the effectiveness of embeddings to down-stream tasks: 1) the training corpus for embedding is available and much larger than the training data of the down-stream task; 2) the topic (domain) of the embedding corpus is closely aligned with the topic of the down-stream task. However, many real-life applications do not meet both assumptions.
In most cases, the in-domain corpus is of limited size, which is insufficient for training good embeddings. In applications, researchers and practitioners often simply use some general-purpose embeddings trained using a very large general-purpose corpus (which satisfies the first assumption) covering almost all possible topics, e.g., the GloVe embeddings [@pennington2014glove] trained using 840 billion tokens covering almost all topics/domains on the Web. Such embeddings have been shown to work reasonably well in many domain-specific tasks. This is not surprising as the meanings of a word are largely shared across domains and tasks. However, this solution violates the second assumption, which often leads to sub-optimal results for domain-specific tasks, as shown in our experiments. One obvious explanation for this is that the general-purpose embeddings do provide some useful information for many words in the domain task, but their embedding representations may not be ideal for the domain and in some cases they may even conflict with the meanings of the words in the task domain because words often have multiple senses or meanings. For example, we have a task in the programming domain, which has the word “Java”. A large-scale general-purpose corpus, which is very likely to include texts about coffee shops, supermarkets, the Java island of Indonesia, etc., can easily squeeze the room for representing “Java”’ context words like “function”, “variable” or “Python” in the programming domain. This results in a poor representation of the word “Java” for the programming task.
To solve this problem and also the limited in-domain corpus size problem, cross-domain embeddings have been investigated [@bollegala-maehara-kawarabayashi:2015:ACL-IJCNLP; @yang-lu-zheng:2017:EMNLP2017; @bollegala2017think] via transfer learning [@pan2010survey]. These methods allow some in-domain words to leverage the general-purpose embeddings in the hope that the meanings of these words in the general-purpose embeddings do not deviate much from the in-domain meanings of these words. The embeddings of these words can thus be improved. However, these methods cannot improve the embeddings of many other words with domain-specific meanings (e.g., “Java”). Further, some words in the general-purpose embeddings may carry meanings that are different from those in the task domain.
In this paper, we propose a novel direction for domain embedding learning by expanding the in-domain corpus. The problem in this new direction can be stated as follows:
**Problem statement**: We assume that the learning system has seen $n$ domain corpora in the past: $D_{1:n}=\{D_1, \dots, D_n\}$, when a new domain corpus $D_{n+1}$ comes with a certain task, the system automatically generates word embeddings for the $(n+1)$-th domain by leveraging some useful information or knowledge from the past $n$ domains.
This problem definition is in the *lifelong learning* (LL) setting, where the new or $(n+1)$-th task is performed with the help of the knowledge accumulated over the past $n$ tasks [@ChenLiu2016]. Clearly, the problem does not have to be defined this way with the domains corpora coming in a sequential manner. It will still work as long as we have $n$ existing domain corpora and we can use them to help with our target domain embedding learning, i.e., the $(n+1)$-th domain.
The main challenges of this problem are 2-fold: 1) how to automatically identify relevant information from the past $n$ domains with no user help, and 2) how to integrate the relevant information into the $(n+1)$-th domain corpus. We propose a meta-learning based system L-DEM (ifelong omain mbedding via eta-learning) to tackle the challenges.
To deal with the first challenge, for a word in the new domain, L-DEM learns to identify similar contexts of the word in the past domains. Here the context of a word means the surrounding words of that word in a domain corpus. We call such context *domain context* (of a word). For this, we introduce a multi-domain meta-learner that can identify similar (or relevant) domain contexts that can be later used in embedding learning in the new domain. To tackle the second challenge, L-DEM augments the new domain corpus with the relevant domain contexts (knowledge) produced by the meta-learner from the past domain corpora and uses the combined data to train the embeddings in the new domain. For example. for word “Java” in the programming domain (the new domain), the meta-learner will produce similar domain contexts from some previous domains like programming language, software engineering, operating systems, etc. These domain contexts will be combined with the new domain corpus for “Java" to train the new domain embeddings.
The main contributions of this paper are as follows. 1) It proposes a novel direction for domain embedding learning, which is a lifelong or continual learning setting and can benefit down-stream learning tasks that require domain-specific embeddings. 2) It proposes a meta-learning approach to leveraging the past corpora from different domains to help generate better domain embeddings. To the best of our knowledge, this is the first meta-learning based approach to helping domain-specific embedding. 3) It experimentally evaluates the effectiveness of the proposed approach.
Related Works {#rw}
=============
Learning word embeddings has been studied for a long time [@mnih2007three]. Many earlier methods used complex neural networks [@mikolov2013linguistic]. More recently, a simple and effective unsupervised model called skip-gram (or word2vec in general) [@mikolov2013distributed; @mikolov2013linguistic] was proposed to turn a plain text corpus into large-scale training examples without any human annotation. It uses the current word to predict the surrounding words in a context window. The learned weights for each word are the embedding of that word. Although some embeddings trained using large scale corpora are available [@pennington2014glove; @bojanowski2016enriching], they are often sub-optimal for domain-specific tasks [@bollegala-maehara-kawarabayashi:2015:ACL-IJCNLP; @yang-lu-zheng:2017:EMNLP2017; @xu:Short; @Xu2018pro; @xu2017Fun]. However, a single domain corpus is often too small for training high-quality embeddings [@Xu2018pro].
Our problem setting is related to *Lifelong Learning* (LL). Much of the work on LL focused on supervised learning [@Thrun1996learning; @Silver2013; @ChenLiu2016]. In recent years, several LL works have also been done for unsupervised learning, e.g., topic modeling [@chen2014topic], information extraction [@Mitchell2015] and graph labeling [@shu2016lifelong]. However, we are not aware of any existing research on using LL for word embedding. Our method is based on meta-learning, which is very different from existing LL methods. Our work is related to transfer learning and multi-task learning [@pan2010survey]. Transfer learning has been used in cross-domain word embeddings [@bollegala-maehara-kawarabayashi:2015:ACL-IJCNLP; @yang-lu-zheng:2017:EMNLP2017]. However, LL is different from transfer learning or multi-task learning [@ChenLiu2016]. Transfer learning mainly transfers common word embeddings from general-purpose embeddings to a specific domain. We expand the in-domain corpus with similar past domain contexts identified via meta-learning.
To expand the in-domain corpus, a good measure of the similarity of domain contexts of the same word from two different domains is needed. We use meta-learning [@thrun2012learning] to learn such similarities. Recently, meta-learning has been applied to various aspects of machine learning, such as learning an optimizer [@andrychowicz2016learning], and learning initial weights for few-shot learning [@finn2017model]. The way we use meta-learning is about domain independent learning [@JMLR:v17:15-239]. It learns similarities of domain contexts of the same word.
Model Overview
==============
The proposed L-DEM system is depicted in Figure \[fig:fr\]. Given a series of past domain corpora $D_{1:n}=\{D_1, D_2, \dots, D_n\}$, and a new domain corpus $D_{n+1}$, the system learns to generate the new domain embeddings by exploiting the relevant information or knowledge from the past $n$ domains. Firstly, a base meta-learner $M$ is trained from the first $m$ past domains (not shown in the figure) (see Section 4), which is later used to predict the similarities of *domain contexts* of the same words from two different domains. Secondly, assuming the system has seen $n-m$ past domain corpora $D_{m+1:n}$, when a new domain $D_{n+1}$ comes, the system produces the embeddings of the $(n+1)$-th domain as follows (discussed in Section 5): (i) the base meta-learner first is adapted to the $(n+1)$-th domain as $M_{n+1}$ (not shown in the figure) using the $(n+1)$-th domain corpus; (ii) for each word $w_{i}$ in the new domain, the system uses the adapted meta-learner $M_{n+1}$ to identify every past domain $j$ that has the word $w_{i}$ with domain context similar to $w_{i}$’s domain context in the new domain (we simply call such domain context from a past domain *similar domain context*); (iii) all new domain words’ similar domain contexts from all past domain corpora $D_{m+1:n}$ are aggregated. This combined set is called the *relevant past knowledge* and denoted by $\mathcal{A}$; (iv) a modified word2vec model that can take both domain corpus $D_{n+1}$ and the relevant past knowledge of $\mathcal{A}$ is applied to produce the embeddings for the $(n+1)$-th new domain. Clearly, the meta-learner here plays a central role in identifying relevant knowledge from past domains. We propose a pairwise model as the meta-learner.
To enable the above operations, we need a knowledge base (KB), which retains the information or knowledge obtained from the past domains. Once the $(n+1)$-th domain embedding is done, its information is also saved in the KB for future use. We discuss the detailed KB content in Section 5.1.
\[fig:ll\] ![Overview of L-DEM.[]{data-label="fig:fr"}](ll.png "fig:"){width="3.3in"}
Base Meta-Learner
=================
This section describes the base meta-learner, which identifies similar domain contexts. The input to the meta-learner is a pair of word feature vectors (we simply call them *feature vectors*) representing the domain contexts of the same word from two similar / non-similar domains. The output of the meta-learner is a similarity score of the two feature vectors.
Training Examples
-----------------
We assume the number of past domains is large and we hold out the first $m$ domains, where $m \ll n$, as the domains to train and test the base meta-learner. In practice, if $n$ is small, the $m$ domains can be sampled from the $n$ domains. The $m$ domains are split into 3 disjoint sets: training domains, validation domains, and testing domains. To enable the meta-learner to predict the similarity score, we need both positive examples (from similar domains) and negative examples (from dissimilar domains). Since each past domain can be unique (which makes it impossible to have a positive pair from two similar domains), we sub-sample each domain corpus $D_j$ into 2 sub-corpora: $D_{j, k} \sim P(D_i)$, where $1\le j\le m$ and $k=\{1, 2\}$. This sampling process is done by drawing documents (each domain corpus is a set of documents) uniformly at random from $D_j$. The number of documents that a domain sub-corpus can have is determined by a pre-defined sub-corpus (file) size (explained in Section 6). We enforce the same file size across all sub-corpora so feature vectors from different sub-corpora are comparable.
Next, we produce feature vectors from domain sub-corpora. Given a word $w_{i, j, k}$ (instance of the word $w_{i}$ in the domain sub-corpus $D_{j, k}$), we choose its co-occurrence counts on a fixed vocabulary $V_{\textit{wf}}$ within a context window (similar to word2vec) as the word $w_{i, j, k}$’s feature vector $\mathbf{x}_{w_{i, j, k}}$. The fixed vocabulary $V_{\textit{wf}}$ (part of the KB used later, denoted as $\mathcal{K}.V_{\textit{wf}}$) is formed from the top-$f$ frequent words over $m$ domain corpora. This is inspired by the fact that an easy-to-read dictionary (e.g., Longman dictionary) uses only a few thousand words to explain all words of a language. A pair of feature vectors $(\mathbf{x}_{w_{i, j, k}}, \mathbf{x}_{w_{i, j, k'}})$ with $k \neq k'$, forms a postive example; whereas $(\mathbf{x}_{w_{i, j, k}}, \mathbf{x}_{w_{i, j', k}})$ with $j\neq j'$ forms a negative example. Details of settings are in Section 6.
Pairwise Model of the Meta-learner
----------------------------------
We train a small but efficient pairwise model (meta-learner) to learn similarity score. Making the model small but high-throughput is crucial. This is because the meta-learner is required in a high-throughput inference setting, where every word from a new domain needs to have context similarities with the same word from all past domains.
The proposed pairwise model has only four layers. One shared fully-connected layer (with $l_1$-norm) is used to learn two continuous representations from two (discrete) input feature vectors. A matching function is used to compute the representation of distance in a high-dimentional space. Lastly, a fully-connected layer and a sigmoid layer are used to produce the similarity score. The model is parameterized as follows: $$\sigma \big( \bm{W}_2 \cdot \text{abs}\big( ( \bm{W}_1 \cdot \frac{\mathbf{x}_{w_{i, j, k}}}{ |\mathbf{x}_{w_{i, j, k}}|_1 }) - (\bm{W}_1 \cdot \frac{\mathbf{x}_{w_{i, j', k'}} }{ |\mathbf{x}_{w_{i, j', k'}}|_1 } ) \big) + b_2 \big) ,$$ where $|\cdot|_1$ is the $l_1$-norm, $\text{abs}(\cdot)$ computes the absolute value of element-wise subtraction ($-$) as the matching function, $\bm{W}$s and $b$ are weights and $\sigma (\cdot)$ is the sigmoid function. The majority of trainable weights resides in $\bm{W}_1$, which learns continuous features from the set of $f$ context words. These weights can also be interpreted as a general embedding matrix over $V_{\textit{wf}}$. These embeddings (not related to the final domain embeddings in Section \[sec:aet\]) help to learn the representation of domain-specific words. As mentioned earlier, we train the base meta-learner $M$ over a hold-out set of $m$ domains. We further fine-tune the base meta-learner using the new domain corpus for its domain use, as described in the next section.
Embedding Using Past Relevant Knowledge
=======================================
We now describe how to leverage the base meta-learner $M$, the rest $n-m$ past domain corpora, and the new domain corpus $D_{n+1}$ to produce the new domain embeddings.
Identifying Context Words from the Past
---------------------------------------
When it comes to borrowing relevant knowledge from past domains, the first problem is what to borrow. It is well-known that the embedding vector quality for a given word is determined by the quality and richness of that word’s contexts. We call a word in a domain context of a given word a *context word*. So for each word in the new domain corpus, we should borrow all context words from that word’s similar domain contexts. The algorithm for borrowing knowledge is described in Algorithm \[alg:ll\], which finds relevant past knowledge $\mathcal{A}$ (see below) based on the knowledge base (KB) $\mathcal{K}$ and the new domain corpus $D_{n+1}$.
The KB $\mathcal{K}$ has the following pieces of information: (1) the vocabulary of top-$f$ frequent words $\mathcal{K}.V_{\textit{wf}}$ (as discussed in Section 4.1), (2) the base meta-learner $\mathcal{K}.M$ (discussed in Section 4.2), and (3) domain knowledge $\mathcal{K}_{m+1:n}$. The domain knowledge has the following information: (i) the vocabularies $V_{m+1:n}$ of past $n-m$ domains, (ii) the sets of past word domain contexts $C_{m+1:n}$ from the past $n-m$ domains, where each $C_j$ is a set of key-value pairs $(w_{i,j}, \mathcal{C}_{w_{i,j} } )$ and $\mathcal{C}_{w_{i,j} } $ is a list of context words[^1] for word $w_i$ in the $j$-th domain, and (iii) the sets of feature vectors $E_{m+1:n}$ of past $n-m$ domains, where each set $E_{j}=\{ \mathbf{x}_{w_{i, j, k}} | w_i \in V_{j} \text{ and } k=\{1, 2\} \}$.
The relevant past knowledge $\mathcal{A}$ of the new domain is the aggregation of all key-value pairs $(w_t, \mathcal{C}_{w_t})$, where $\mathcal{C}_{w_t}$ contains all similar domain contexts for $w_t$. Algorithm \[alg:ll\] retrieves the past domain knowledge in line 1. Lines 2-4 prepare the new domain knowledge. The BuildFeatureVector function produces a set of feature vectors as $E_{n+1}=\{ \mathbf{x}_{w_{i, n+1, k}} | w_i \in V_{j} \text{ and } k=\{1, 2\}\}$ over two sub-corpora of the new domain corpus $D_{n+1}$. The ScanContextWord function builds a set of key-value pairs, where the key is a word from the new domain $w_{i, n+1}$ and the value $\mathcal{C}_{w_{i,n+1} } $ is a list of context words for the word $w_{i, n+1}$ from the new domain corpus. We use the same size of context window as the word2vec model.
### Adapting Meta-learner
In line 5, AdaptMeta-learner adapts or fine-tunes the base meta-learner $\mathcal{K}.M$ to produce an adapted meta-learner $M_{n+1}$ for the new domain. A positive tuning example is sampled from two sub-corpora of the new domain $(\mathbf{x}_{w_{i, n+1, 1}}, \mathbf{x}_{w_{i, n+1, 2}})$ in the same way as described in Section 4.1. A negative example is exampled as $(\mathbf{x}_{w_{i, n+1, 1}}, \mathbf{x}_{w_{i, j, 2}})$, where $m+1 \le j \le n$. The initial weights of $M_{n+1}$ are set as the trained weights of the base meta-learner $M$.
$(V_{m+1:n}, C_{m+1:n}, E_{m+1:n}) \gets \mathcal{K}_{m+1:n}$ $V_{n+1} \gets \text{BuildVocab}(D_{n+1})$ $C_{n+1} \gets \text{ScanContextWord}(D_{n+1}, V_{n+1})$ $E_{n+1} \gets \text{BuildFeatureVector}(D_{n+1}, \mathcal{K}.V_{\textit{wf}})$ $M_{n+1} \gets \text{AdaptMeta-learner}(\mathcal{K}.M, E_{m+1:n}, E_{n+1})$ $\mathcal{A} \gets \emptyset$ $\mathcal{K}_{n+1} \gets (V_{n+1}, C_{n+1}, E_{n+1}) $
### Retriving Relevant Past Knowledge
Algorithm \[alg:ll\] further produces the relevant past knowledge $\mathcal{A}$ from line 6 through line 16. Line 6 defines the variable that stores the relevant past knowledge. Lines 7-15 produce the relevant past knowledge $\mathcal{A}$ from past domains. The For block handles each past domain sequentially. Line 8 computes the shared vocabulary $O$ between the new domain and the $j$-th past domain. After retrieving the sets of feature vectors from the two domains in line 9, the adapted meta-learner uses its inference function (or model) to compute the similarity scores on pairs of feature vectors representing the same word from two domains (line 10). The inference function can parallelize the computing of similarity scores in a high-throughput setting (e.g., GPU inference) to speed up. Then we only keep the words from past domains with a score higher than a threshold $\delta$ at line 11. Lines 12-14 aggregate the context words for each word in $O$ from past word domain contexts $C_j$. Line 16 simply stores the new domain knowledge for future use. Lastly, all relevant past knowledge $\mathcal{A}$ is returned.
Augmented Embedding Training {#sec:aet}
----------------------------
We now produce the new domain embeddings via a modified version of the skip-gram model [@mikolov2013distributed] that can take both the new domain corpus $D_{n+1}$ and the relevant past knowledge $\mathcal{A}$. Given a new domain corpus $D_{n+1}$ with the vocabulary $V_{n+1}$, the goal of the skip-gram model is to learn a vector representation for each word $w_{i} \in V_{n+1}$ in that domain (we omit the subscript $_{n+1}$ in $w_{i, n+1}$ for simplicity). Assume the domain corpus is represented as a sequence of words $D_{n+1}=(w_1, \dots, w_T)$, the objective of the skip-gram model maximizes the following log-likelihood: $$\label{eq:sg}
\begin{split}
\mathcal{L}_{D_{n+1}} =\sum_{t=1}^{T} \big( \sum_{w_c \in \mathcal{W}_{w_{t}} } \big(\log \sigma (\bm{u}_{w_t}^T\cdot \bm{v}_{w_c}) \\
+ \sum_{w_{c'} \in \mathcal{N}_{w_t} } \log \sigma(-\bm{u}_{w_t}^T\cdot \bm{v}_{w_{c'}} ) \big) \big) ,
\end{split}$$ where $\mathcal{W}_{w_t}$ is the set of words surrounding word $w_t$ in a fixed context window; $\mathcal{N}_t$ is a set of words (negative samples) drawn from the vocabulary $V_{n+1}$ for the $t$-th word; $\bm{u}$ and $\bm{v}$ are word vectors (or embeddings) we are trying to learn. The objective of skip-gram on data of relevant past knowledge $\mathcal{A}$ is as follows: $$\begin{split}
\mathcal{L}_{\mathcal{A}}=\sum_{(w_t, \mathcal{C}_{w_t} ) \in \mathcal{A}} \big( \sum_{w_c \in \mathcal{C}_{w_t}} \big( \log \sigma (\bm{u}_{w_t}^T\cdot \bm{v}_{w_c}) \\
+ \sum_{w_{c'} \in \mathcal{N}_{w_t} } \log \sigma(-\bm{u}_{w_t}^T\cdot \bm{v}_{w_{c'}} ) \big) \big).
\end{split}$$ Finally, we combine the above two objective functions: $$\begin{split}
\mathcal{L}'_{D_{n+1}}=\mathcal{L}_{D_{n+1}} + \mathcal{L}_{\mathcal{A}}.
\end{split}$$ We use the default hyperparameters of skip-gram model [@mikolov2013distributed] to train the domain embeddings.
Experimental Evaluation {#sec:exp}
=======================
Following [@nayak2016evaluating], we use the performances of down-stream tasks to evaluate the proposed method. We do not evaluate the learned embeddings directly as in [@mikolov2013distributed; @pennington2014glove] because domain-specific dictionaries of similar / non-similar words are generally not available. Our down-stream tasks are text classification that usually requires fine-grained domain embeddings.
Datasets
--------
We use the Amazon Review datasets from [@HeMcA16a], which is a collection of multiple-domain corpora. We consider each second-level category (the first level is department) as a domain and aggregate all reviews under each category as one domain corpus. This ends up with a rather diverse domain collection. We first randomly select 56 ($m$) domains as the first $m$ past domains to train and evaluate the base meta-learner. Then from rest domains, we sample three random collections with 50, 100 and 200 ($n-m$) domains corpora, respectively, as three settings of past domains. These collections are used to test the performance of different numbers of past domains. Due to the limited computing resource, we limit each past domain corpus up to 60 MB. We further randomly selected 3 rest domains (*Computer Components* (CC), *Kitchen Storage and Organization* (KSO) and *Cats Supply* (CS)) as new domains for down-stream tasks. These give us three text classification problems, which have 13, 17, and 11 classes respectively. The tasks are topic-based classification rather than sentiment classification. Since the past domains have different sizes (many have much less than 60 MB) and many real-world applications do not have big in-domain corpora, we set the size of the new domain corpora to be 10 MB and 30 MB to test the performance in the two settings.
Evaluation of Meta-Learner
--------------------------
CC KSO CS
------ ------- ------- -------
10MB 0.832 0.841 0.856
30MB 0.847 0.859 0.876
: F1-score of positive predictions of the adapted meta-learner on 3 new domains: Computer Components (CC), Kitchen Storage and Organization (KSO) and Cats Supply (CS).
\[table:fine-tune\]
We select the top $f=5000$ words from all 56 domains’ corpora as word features. Then we split the 56 domains into 39 domains for training, 5 domains for validation and 12 domains for testing. So the validation and testing domain corpora have no overlap with the training domain corpora. We sample 2 sub-corpora for each domain and limit the size of each sub-corpus to 10 MB. We randomly select 2000, 500, 1000 words from each training domain, validation domain, and testing domain, respectively, and ignore words with all-zero feature vectors to obtain pairwise examples. The testing 1000 words are randomly drawn and they have 30 overlapping words with the training 2000 words, but not from the same domains. So in most cases, it’s testing the unseen words in unseen domains. We set the size of a context window to be 5 when building feature vectors. This ends up with 80484 training examples, 6234 validation examples, and 20740 test examples. For comparison, we train a SVM model as a baseline. The F1-score (for positive pairs) of SVM is 0.70, but the F1-score of the proposed base meta-learner model is **0.81**.
To adapt the base meta-learner for each new domain. We sample 3000 words from each new domain, which results in slightly fewer than 6000 examples after ignoring all-zero feature vectors. We select 3500 examples for training, 500 examples for validation and 2000 examples for testing. The F1-scores on the test data is shown in Table 1. Finally, we empirically set $\delta=0.7$ as the threshold on the similarity score in Algorithm 1, which roughly doubled the number of training examples from the new domain corpus. The size of the context window for building domain context is set to 5, which is the same as word2vec.
\[table:pc\]
Baselines and Our System
------------------------
Unless explicitly mentioned, the following embeddings have 300 dimensions, which are the same size as many pre-trained embeddings (GloVec.840B [@pennington2014glove] or fastText English Wiki[@bojanowski2016enriching]).\
**No Embedding (NE)**: This baseline does not have any pre-trained word embeddings. The system randomly initializes the word vectors and train the word embedding layer during the training process of the down-stream task.\
**fastText**: This baseline uses the lower-cased embeddings pre-trained from English Wikipedia using fastText [@bojanowski2016enriching]. We lower the cases of all corpora of down-stream tasks to match the words in this embedding.\
**GoogleNews**: This baseline uses the pre-trained embeddings from word2vec [^2] based on part of the Google News dataset, which contains 100 billion words.\
**GloVe.Twitter.27B**: This embedding set is pre-trained using GloVe[^3] based on Tweets of 27 billion words. This embedding is lower-cased and has 200 dimensions.\
**GloVe.6B**: This is the lower-cased embeddings pre-trained from Wikipedia and Gigaword 5, which has 6 billion tokens.\
**GloVe.840B**: This is the cased embeddings pre-trained from Common Crawl corpus, which has 840 billion tokens. This corpus contains almost all web pages available before 2015. We show that the embeddings produced from this very general corpus are sub-optimal for our domain-specific tasks.\
**New Domain 10M (ND 10M)**: This is a baseline embedding pre-trained only from the new domain 10 MB corpus. We show that the embeddings trained from a small corpus alone are not good enough.\
**New Domain 30M (ND 30M)**: This is a baseline embedding pre-trained only from the new domain 30 MB corpus. We increase the size of the new domain corpus to 30 MB to see the effect of the corpus size.\
**200 Domains + New Domain 30M (200D + ND 30M)**: The embedding set trained by combining the corpora from all past 200 domains and the new domain. We use this baseline to show that using all past domain corpora may reduce the performance of the down-stream tasks.\
**L-DENP 200D + ND 30M**: This is a on-arametric variant of the proposed method. We use TFIDF as the representation for a sentence in past domains and use cosine as a non-parametic function to compute the similarity with the TFIDF vector built from the new domain corpus. We report the results on a similarity threshold of 0.18, which is the best threshold ranging from 0.15 to 0.20.\
**L-DEM Past Domains + New Domain (L-DEM \[P\]D + ND \[X\]M)**: These are different variations of our proposed method L-DEM. For example, “L-DEM 200D + ND 30M” denotes the embeddings trained from a 30MB new domain corpus and the relevant past knowledge from 200 past domains.
Down-stream Tasks and Experiment Results
----------------------------------------
As indicated earlier, we use classification tasks from 3 new domains (“Computer Components”, “Cats Supply” and “Kitchen Storage and Organization”) to evaluate the embeddings produced by our system and compare them with those of baselines. These 3 new domains have 13, 17 and 11 classes (or product types), respectively. For each task, we randomly draw 1500 reviews from each class to make up the experiment data, from which we keep 10000 reviews for testing (to make the result more accurate) and split the rest 7:1 for training and validation, respectively. All tasks are evaluated on accuracy. We train and evaluate each task on each system 10 times (with different initializations) and average the results.
For each task, we use an embedding layer to store the pre-trained embeddings. We freeze the embedding layer during training, so the result is less affected by the rest of the model and the training data. To make the performance of all tasks consistent, we apply the same Bi-LSTM model [@hochreiter1997long] on top of the embedding layer to learn task-specific features from different embeddings. The input size of Bi-LSTM is the same as the embedding layer and the output size is 128. All tasks use many-to-one Bi-LSTMs for classification purposes. In the end, a fully-connected layer and a softmax layer are applied after Bi-LSTM, with the output size specific to the number of classes of each task. We apply dropout rate of 0.5 on all layers except the last one and use Adam [@kingma2014adam] as the optimizer.
Table 2 shows the main results. We observe that the proposed method L-DEM 200D + ND 30M performs the best. The difference in the numbers of past domains indicates more past domains give better results. The GloVe.840B trained on 840 billion tokens does not perform as well as embeddings produced by our method. GloVe.840B’s performance on the CC domain is close to our method indicating mixed-domain embeddings for this domain is not bad and this domain is more general. Combining all past domain corpora together with the new domain corpus (200D + ND 30M) makes the result worse than not using the past domains at all (ND 30M). This is because the diverse 200 domains are not similar to the new domains. The L-DENP 200D + ND 30M performs poorly indicating the proposed parametric meta-learner is useful, except the CC domain which is more general.
Additional Experiments
----------------------
\[table:concat\]
Note that we did not compare with the existing transfer learning methods [@bollegala2017think; @bollegala-maehara-kawarabayashi:2015:ACL-IJCNLP; @yang-lu-zheng:2017:EMNLP2017] as our approaches focus on domain-specific words in a lifelong learning setting, which do not need the user to provide the source domain(s) that are known to be similar to the target domain. One approach to leveraging existing embeddings is to concatenate pre-trained embeddings with domain-specific embeddings[^4]. To demonstrate our method further improves the domain-specific parts of the down-stream tasks, we evaluate two methods: (1) GloVe.840B&ND 30M, which concatenates new domain only embeddings with GloVe.840B; (2) GloVe.840B&L-DEM 200D + ND 30M, which concatenates our proposed embeddings with GloVe.840B. As shown in Table 3, concatenating embeddings improve the performance. Our method boosts the domain-specific parts of the embeddings further.
Conclusions
===========
In this paper, we formulated a domain word embedding learning process. Given many previous domains and a new domain corpus, the proposed method can generate new domain embeddings by leveraging the knowledge in the past domain corpora via a meta-learner. Experimental results show that our method is highly promising.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported in part by NSF through grants IIS-1526499, IIS-1763325, IIS1407927, CNS-1626432 and NSFC 61672313, and a gift from Huawei Technologies.
[^1]: We use list to simplify the explanation. In practice, bag-of-word representation should be used to save space.
[^2]: https://code.google.com/archive/p/word2vec/
[^3]: https://nlp.stanford.edu/projects/glove/
[^4]: Note the ideal LL setting is to perform L-DEM over all domain corpora of the pre-trained embeddings.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We report on a deep, multiwavelength study of the galaxy cluster using X-ray, optical, and 1.4 GHz radio data. This cluster ($z=0.352$) harbors one of the most X-ray luminous cool cores yet discovered, with an equivalent mass cooling rate within the central $50\h70^{-1} \kpc$ is $\sim$700 . Unique features observed in the central core of hint to a wealth of past activity that has greatly disrupted the original cool core. The X-ray and optical data suggest oscillatory motion of the cool core along a roughly north-south direction. We also observe a spiral of relatively cool, dense, X-ray emitting gas connected to the cool core, as well as highly elongated intracluster light (ICL) surrounding the cD galaxy. For a cluster with such a high nominal cooling rate, this cluster is missing the central metallicity peak almost always seen in cool core clusters, which suggests bulk transport of cool gas out to large distances from the center. Extended radio emission is observed surrounding the central AGN, elongated in the east-west direction, spatially coincident with X-ray cavities. The power input required to inflate these ‘bubbles’ is estimated from both the X-ray and radio emission to reside between ${\ensuremath{P_{\mathrm{jet}}}} \sim$4 – 14 $\times 10^{45}$ , putting it among the most powerful jets ever observed. This combination of a powerful AGN outburst and bulk motion of the cool core have resulted in two X-ray bright ridges to form to the north and south of the central AGN at a distance of approximately 25 . The northern ridge has spectral characteristics typical of cool cores: it contains low temperature, high density, metal rich gas and is consistent with being a remnant of the cool core after it was disrupted by the AGN and bulk motions. It is also the site of filaments and young stars. The X-ray spectroscopic cooling rate associated with this ridge is $\sim$165 , which agrees with the estimate of the star formation rate from broad-band optical imaging ($\sim$170 ). appears to harbor one of most profoundly disrupted low entropy cores observed in a cluster, and offers new insights into the survivability of cool cores in the context of hierarchical structure formation.'
bibliography:
- 'MACS1931.bib'
- 'optical.bib'
date: 'Accepted 2010 September 29. Received 2010 August 24; in original form 2010 June 21'
title: ' Extreme AGN Feedback and Cool Core Destruction in the X-ray Luminous Galaxy Cluster MACS J1931.8-2634'
---
X-rays: galaxies: clusters – galaxies: clusters: individual: – galaxies: clusters: general – galaxies: active – galaxies: cooling flows
Introduction {#intro}
============
The relatively cool and dense gas at the centers of many galaxy clusters emits copious X-ray radiation by thermal bremsstrahlung and line emission. In the absence of external sources of heating, this high emission should lead to very rapid cooling ($\tau_{cool} < 1 \ \rm{Gyr}$) and very high rates of mass deposition onto the central cluster galaxy (up to $\sim$1000 ), in turn causing very high star formation rates and strong line emission around 0.5-1.0 [e.g. @Fabian1977; @Cowie1977; @Peterson2006; @McNamara2007]. The lack of such obvious observational signatures [e.g. @Peterson2001; @Peterson2003] provides compelling evidence that some central source of heating must be present. The most plausible source of heating is feedback from the central Active Galactic Nucleus (AGN). Large X-ray cavities filled with radio plasma are clearly seen in many systems, which provide an expected source of heating and turbulence [e.g. @Brueggen2002]. In systems such as the Perseus, Virgo, Centaurus, and Hydra A Clusters [e.g. @Fabian2003; @Forman2005; @Nulsen2005Hydra; @Fabian2006; @Forman2007; @Sanders2007; @Sanders2008; @Simionescu2009 Million et al. submitted], heating of the ICM has been argued to involve sound waves and weak shocks. Many questions remain as to the precise processes by which central AGN activity suppresses cooling flows, and the extent to which these feedback processes require finely tuned parameters. More ambiguous still are the nature and energetics of feedback mechanisms at higher redshifts, as only a few systems with such high classical cooling rates have been studied in detail [@Schindler1997; @Allen2000; @Kitayama2004; @Bohringer2005; @Ogrean2010].
The galaxy cluster is an extreme example of a cluster with a rapidly cooling core, making it an ideal system to test the limits of feedback mechanisms in galaxy clusters. In a short 12 ks observation of taken in October of 2002 [@Allen2004; @Allen2008], X-ray cavities were detected surrounding a bright central AGN. The physical size of these cavities ($\sim$25 $\kpc$) is similar to those observed in the nearby Perseus Cluster [@Fabian2003; @Fabian2006]. Indeed, with its luminous cool core, visible central point source, and very large apparent cooling rate, is in many ways a higher redshift analog to the Perseus Cluster. Deeper observations of were taken with in August of 2008, increasing the total clean exposure to $\sim$100 $\rm{ks}$. The combined X-ray data are presented for the first time here, and are complimented with optical () and radio () observations. Our goals are to acquire a better understanding of the thermodynamic structure around the central AGN and the extent to which feedback from the central AGN counteracts this extreme cooling flow.
The structure of this paper is as follows: Section \[datared\] discusses the reduction of the data. Section \[imaging\] discusses the imaging analysis of the X-ray data, while Section \[deprojanalysis\] presents results on the mass profile and cooling flow of . Section \[spectra\] discusses the spectral analysis. Sections \[Optical\] and \[radiodata\] present the optical and radio data, respectively. Calculations of the energetics associated with the central AGN are discussed in Section \[centralagn\], and the results are discussed and summarized in Section \[discussion\]. Throughout this paper, a [$\Lambda$CDM]{} cosmology is assumed with ${\ensuremath{\Omega_{\Lambda}}}=0.73$, ${{\ensuremath{\Omega_{\mathrm{m}}}}}=0.27$, and $H_{0}=70 \km \s^{-1} \Mpc^{-1}$. At the redshift of ($z=0.352$), 1 corresponds to 4.926 .
Data Reduction and Processing {#datared}
==============================
Two observations of were performed using the Advanced CCD Imaging Spectrometer (ACIS) in October 2002 and August 2008. The standard level-1 event lists produced by the Chandra pipeline processing were reprocessed using the CIAO (version 4.1.2) software package, including the appropriate gain maps and calibration products (CALDB version 4.1.2). Bad pixels were removed and standard grade selections were applied to the event lists. Both observations were taken in VFAINT mode, and the additional information available in this mode was used to improve the rejection of cosmic ray events. The data were cleaned to remove periods of anomalously high background using the standard energy ranges and binning methods recommended by the X-ray Center. The net exposure times after processing are summarized in Table \[Observations\].
[Obs \#]{} [Observation Date]{} Detector Exposure Time (ks)
------------ ---------------------- ---------- --------------------
3282 October 20 2002 ACIS-I 10.0
9382 August 21 2008 ACIS-I 89.5
: \[Observations\]Summary of the two observations of . Exposure times are the net exposure after all cleaning and processing as described in Section \[datared\].
X-ray Imaging Analysis {#imaging}
======================
![\[SurBrightProf\] Combined X-ray surface brightness profile for in the energy range 0.7-2.0 . The red points denote the total surface brightness, while the black points denote the background subtracted surface brightness.](SBProf.ps){width="\columnwidth"}
Surface Brightness Profiles
---------------------------
Flat-fielded images were first created in the energy range from 0.7-2.0 for each of the observations. This energy range was chosen to minimize the impact of astrophysical and instrumental background components. All subsequent imaging analysis was performed in this energy range. Surface brightness profiles were produced from these flat-fielded images centered on the central AGN ($\alpha (2000)=19^{h}31^{m}49^{s}.6, \ \delta (2000)=-26^{h}34^{m}33^{s}.6$). Each surface brightness profile was fit with a King plus constant model in the range of 50 to 400 from the central AGN. The King plus constant model is explicitly given by the form
$$\label{KingModel}
I(r)={\ensuremath{I_{\mathrm{0}}}}\left(1+\left(\frac{r}{{\ensuremath{r_{\mathrm{0}}}}}\right)^{2}\right)^{-\beta}+ {\ensuremath{C_{\mathrm{0}}}}$$
The best-fit constant for each observation was subtracted from each surface brightness profile, and Fig. \[SurBrightProf\] shows the exposure-time weighted average of the two surface brightness profiles. Over the radial range of 10-250 , the profile can be approximately described by a King model with ${\ensuremath{r_{\mathrm{0}}}}=6.54 \pm 0.23 \arcsec$ and $\beta=1.35 \pm 0.01$.
The background level determined from the surface brightness analysis described above was subtracted from each flat-fielded image on a pixel-by-pixel basis. These two images were reprojected into a common aspect solution and combined. This combined image, which is used in all subsequent imaging analysis, is shown in Fig. \[NicerImages\].
Background Subtracted, Flat-Fielded Image {#basicimaging}
-----------------------------------------
On scales larger than $\sim$100 $\kpc$, exhibits a general elliptical symmetry, with the major axis aligned along the approximately north-south direction. There are asymmetries at these scales, however, that can be more readily observed by overlaying the X-ray surface brightness contours, seen in Fig. \[MACSCont\]. The isophotes are clearly not concentric, and the centroid appears to shift north and south of the central AGN with an amplitude as large as $\sim$20-30 $\kpc$. Inside the central 100 , seen in Fig. \[MACSZoom\], the morphology becomes considerably more complex. The bright central point source is surrounded by two bright ridges to the north and south. The northern ridge is brighter and has a relatively sharp boundary to the north, while the southern ridge is trailed by diffuse emission extending further to the south. These features, as well as the varying isophote centroids on large scales are indicative of past oscillations of the core, which currently appears to be moving to the north.
Substructure Analysis
---------------------
### Bandpass Filtering
In order to better resolve small scale structures around the center of the cluster, a high-frequency bandpass filter was applied to the image shown in Fig. \[MACSBasic\]. The functional form of the filter is given as
$$F(k)=\frac{\left(\frac{k}{{\ensuremath{k_{\mathrm{0}}}}}\right)^{2}}{1+\left(\frac{k}{{\ensuremath{k_{\mathrm{0}}}}}\right)^{2}}$$
We have set the scale length ${\ensuremath{k_{\mathrm{0}}}}=10$ . The transformed image is shown in Fig. \[Butterworth\], and shows the central AGN as well as the ridges to the north and south more clearly. Other features more apparent in this image are depressions in the X-ray brightness immediately to the east and west of the central AGN. Nearby systems show clear cavities in the X-ray brightness near the central AGN, similar in shape to these [e.g. @Fabian2003; @Birzan2004]. Unlike the cavities in those systems, however, there is no evidence in Fig. \[Butterworth\] as to where the outer boundary of these cavities might lie.
### Two-Dimensional Surface Brightness Modeling
The X-ray image of Fig. \[MACSBasic\] was also fit with a two-dimensional elliptical beta model of the form $$\centering
I(x,y)= \frac{A}{\left[1+\frac{r(x,y)^{2}}{{\ensuremath{r^{2}_{\mathrm{0}}}}}\right]^{\alpha}},$$
where $A$ is the central normalization, ${\ensuremath{r_{\mathrm{0}}}}$ is the core radius, and $\alpha$ is the power index. The term $r(x,y)$ is given as
$$\begin{aligned}
r(x,y)^{2} &= &\left[(x-{\ensuremath{x_{\mathrm{0}}}})\cos{\theta}+(y-{\ensuremath{y_{\mathrm{0}}}})\sin{\theta}\right]^{2}\\
&+ & \left[\frac{(y-{\ensuremath{y_{\mathrm{0}}}})\cos{\theta}-(x-{\ensuremath{x_{\mathrm{0}}}})\sin{\theta}}{1-\epsilon}\right]^{2},\end{aligned}$$
which gives the distance of any point in the image $(x,y)$ from a fixed center $({\ensuremath{x_{\mathrm{0}}}},{\ensuremath{y_{\mathrm{0}}}})$ with an ellipticity $\epsilon$. The position angle, $\theta$, was also a free parameter in the fit. Since the image has already been background subtracted, no further considerations for the background were included in the fit. The best fit parameters for this model are ${\ensuremath{r_{\mathrm{0}}}}=3.21 \pm 0.04 \arcsec \
$, $\epsilon=0.290 \pm 0.004 $, $\theta=4.11 \pm 0.46 \ $ degrees and $\alpha=1.139 \pm 0.004$.
The fit was then subtracted from the image, and the residuals adaptively smoothed. The smoothed residuals image is shown in Fig. \[EllipBeta\]. With the elliptical model subtracted, a spiral pattern beginning just east of the central AGN emerges. Such a pattern is expected to arise in off-axis mergers [e.g. @Ascasibar2006 see also Section \[discussion\]].
Image Deprojection Analysis {#deprojanalysis}
===========================
An image deprojection analysis of was undertaken following the manner described in @Allen2008 [see also @Schmidt2007]. In brief, the azimuthally averaged X-ray surface brightness profile (centered on the central AGN) and the deprojected temperature profile are combined to simultaneously determine the X-ray emitting gas mass and total mass profiles of the cluster. We assume that the dark+luminous mass distribution follows the model of @Navarro1995 [@Navarro1997] (hereafter NFW profile) $$\rho(r)=\frac{{\ensuremath{\rho_{\mathrm{c}}}}(z){\ensuremath{\delta_{\mathrm{c}}}}}{(r/{\ensuremath{r_{\mathrm{s}}}})(1+r/{\ensuremath{r_{\mathrm{s}}}})^{2}}$$
where $\rho(r)$ is the mass density, ${\ensuremath{\rho_{\mathrm{c}}}}(z)=3H(z)^{2}/8\pi G$ is the critical density for closure at redshift $z$, ${\ensuremath{r_{\mathrm{s}}}}$ is the scale radius, $c$ is the concentration parameter (with $c={\ensuremath{r_{\mathrm{200}}}}/{\ensuremath{r_{\mathrm{s}}}}$) and ${\ensuremath{\delta_{\mathrm{c}}}}=200c^{3}/3\left[\ln{(1+c)}-c/(1+c)\right]$. The length scale [$r_{\mathrm{\Delta}}$]{} is defined such that the enclosed mean mass density is $\Delta$ times the critical density of the Universe at the redshift of the cluster. For a given scale radius and concentration parameter, a model temperature profile can be calculated and compared with the observed deprojected temperature profile. The observed deprojected temperature profile was azimuthally averaged and binned into annular regions with roughly 5,000 counts in each (see Section \[spectra\]). The mass profile parameters were stepped over a range of values to determine best-fit values and uncertainties using a $\chi^{2}$ minimization technique. This analysis assumed hydrostatic equilibrium and spherical symmetry, both of which are obviously not true in the vicinity of the central AGN. To account for these assumptions, the data for the central 50 have been statistically down-weighted by adding 30% systematic uncertainties to the measured temperatures.
Our best fit NFW profile has a scale radius of ${\ensuremath{r_{\mathrm{s}}}}=0.26^{+0.05}_{-0.02} \Mpc$, a concentration parameter of $c=6.25^{+0.40}_{-0.75}$, and an equivalent velocity dispersion [@Allen2002] $\sigma=\sqrt{50}{\ensuremath{r_{\mathrm{s}}}}cH(z)=966^{+60}_{-18} \km\s^{-1}$. The integrated mass profile is shown in Fig. \[gravmass\]. The enclosed mass within ${\ensuremath{r_{\mathrm{2500}}}}=505^{+19}_{-3} \kpc$ is ${\ensuremath{M_{\mathrm{2500}}}}= 2.64^{+0.31}_{-0.06} \times 10^{14} {{\ensuremath{M_{\mathrm{\odot}}}}}$.
The bolometric luminosity, cooling time, and equivalent mass deposition rate profiles have also been determined using the calculations described in detail in @White1997. These are shown in Figs. \[lumin\], \[cooltime\], and \[coolflow\] respectively. These profiles show that there is rapid cooling within the central 50 . Within this region and in the absence of of balancing heat sources, a cooling flow with an equivalent mass deposition rate of $\dot{M} \sim$700 would be expected. The bolometric luminosity within this region is $\sim$1 $\times
10^{45}$, and the cooling time is less than 1 Gyr.
Spectral Analysis {#spectra}
=================
Methods {#spectralmethods}
-------
The observations of are sufficiently deep for high signal-to-noise spectra to be extracted from relatively small regions. This enables us to carry out detailed, spatially resolved measurements of the thermodynamic quantities of the ICM. All spectral analysis was carried out using [XSPEC]{} [@Arnaud2004 version 12.5]. The backgrounds for all spectral analysis were extracted directly from the science observations, specifically from a region roughly the same distance from the center of the detector as the cluster, but on the diagonally opposite chip. The background regions are devoid of point sources and cluster emission.
### Regions of Interest {#spectralbins}
The spectral structure of was measured in both one-dimensional radial profiles and two-dimensional maps. In addition, the spectra of several regions of interest identified from the imaging analysis (specifically the central AGN and bright central northern and southern ridges) were also investigated in detail. To account for contamination by the central AGN, the 2.5 radius region surrounding the central AGN was excluded from the spectral profiles and maps. Region specific response matrices and ancillary response files were created for all spectra.
Spatially resolved spectral maps for the cluster were extracted in regions determined by the contour binning algorithm of @Sanders2006, which creates bins of equivalent signal-to-noise, following contours in surface brightness. Regions were selected to have a signal-to-noise of 30, resulting in approximately $1,000$ counts per bin.
Azimuthally averaged spectral profiles were measured in annular regions with nearly equal numbers of counts. Initially radial profiles were made using annular bins that each contain roughly 3,000 counts. Measuring the metallicity requires data with a higher signal-to-noise, so we also carried out a similar analysis using annular bins with roughly 10,000 counts each. The chosen center for all annuli was the location of the central AGN. The spectral properties of each annular region were measured both in projection and deprojection. Deprojection was implemented using the [PROJCT]{} mixing model in [XSPEC]{}.
### Modeling the Emission
All spectral regions were initially modeled as a single temperature optically thin plasma using the [MEKAL]{} code of @Kaastra1993 incorporating the Fe-L calculations of @Liedhal1995 and the photoelectric absorption models of @McCammon1992. We used the determinations of solar element abundances given by @Anders1989. The abundances of metals ($Z$) were assumed to vary with a common ratio with respect to the Solar values. The single-temperature plasma model has three free parameters: the temperature ($kT$), the metallicity ($Z$), and normalization ($K$).
In each region, the spectral analysis assumed a fixed Galactic absorption column of $8.3 \times 10^{20} \cm^{-2}$ [@Kalberla2005], consistent with the value measured directly from the X-ray spectra. The modified Cash statistic in [XSPEC]{} [@Cash1979; @Arnaud2004] was minimized to determine the best fit model parameters and uncertainties. All uncertainties given are 68% ($\Delta C=1$) confidence intervals, unless otherwise noted. In the analysis of the bright ridges to the north and south of the AGN, the spectroscopic model included an additional [MKCFLOW]{} component, appropriate for a scenario where gas is assumed to cool at constant pressure from an upper temperature down to a lower temperature. The upper temperature and metallicity of the [MKCFLOW]{} component were tied to their corresponding [MEKAL]{} components, and the low temperature was fixed to $0.1 \keV$.
### Thermodynamic Quantities
Several thermodynamic quantities can be calculated directly from the best fit [MEKAL]{} model parameters. The electron density ($n_{e}$), pressure ($P$), and entropy ($S$) of the ICM are derived from the [MEKAL]{} temperature ($kT$) and normalization ($K$) as $$\label{NormDensity}
{\ensuremath{n_{\mathrm{e}}}}^{2}=\frac{4\pi \times 10^{14}\left(1+z\right)^{2}{\ensuremath{D_{\mathrm{A}}}}^{2}K}{1.2V}$$
$$P=kT{\ensuremath{n_{\mathrm{e}}}}$$
$$S=kT{\ensuremath{n_{\mathrm{e}}}}^{-2/3}$$
where the cosmological value of the angular diameter distance ${\ensuremath{D_{\mathrm{A}}}}$ is 1016 at the cluster redshift. The volume of the region, $V$, is given in units of $\cm^{3}$. If the region had not been deprojected (in the case of the thermodynamic maps), the volume of the region was estimated as
$$V={\ensuremath{D_{\mathrm{A}}}}^{3}\Omega \sqrt{{\ensuremath{\theta_{\mathrm{max}}}}^{2}-{\ensuremath{\theta_{\mathrm{min}}}}^{2}}$$
where ${\ensuremath{\theta_{\mathrm{max,min}}}}$ are the maximum and minimum angular distances of any point in the region to the center of the cluster, respectively, and $\Omega$ is the solid-angle extent of the region in the sky [@Henry2004; @Mahdavi2005]. If the spectrum had been fully deprojected (true for the annular profiles), the volume was calculated as $$V=\frac{4}{3}\pi{\ensuremath{D_{\mathrm{A}}}}^{3}\left({\ensuremath{\theta_{\mathrm{max}}}}^3-{\ensuremath{\theta_{\mathrm{min}}}}^{3}\right)$$ Since the uncertainties on the temperature are considerably larger than those on the density ($> 10\% $ for the temperature as compared to $\sim$2-5% for the densities), the fractional uncertainties on the pressure and entropy given are similar to the corresponding temperature measurement.
Spectral Results {#thermodyn}
----------------
### Thermodynamic Mapping {#thermomapsec}
The maps of temperature, density, pressure, and entropy in are shown in Fig. \[ThermoMaps\]. One of the most notable features is the spiral of low temperature gas wrapping to the east and north of the central AGN seen in the large scale temperature map. This is spatially coincident with the surface brightness excess seen in Fig. \[EllipBeta\]. Such spirals can be due to either ram pressure stripping of an infalling subcluster core, or merger induced oscillatory motion of the cluster core. Both scenarios are seen in many nearby systems and hydrodynamic simulations [e.g. @Churazov2003; @Ascasibar2006; @Dupke2007; @Lagana2009; @Owers2009; @ZuHone2009; @Million2010].
Zooming in on the temperature and entropy structure surrounding the AGN, Figs. \[ZoomTemp\] and \[ZoomEntropy\] show that there are also two small regions of higher temperature to the east and west of the central AGN spatially coincident with the depressions in X-ray emission noted in Fig. \[Butterworth\]. Although the temperature of these regions is up to 2 higher than their surroundings, the statistical significance is modest. The lowest temperature and entropy gas is located 25-30 to the north of the AGN and spatially coincident with the X-ray bright northern ridge. There is a similar ridge to the south which also has lower temperature and entropy gas, but at lower significance.
### Spectroscopic Cooling in the Bright Northern and Southern Ridges {#ridges}
Since cooling flows are only expected in the inner, densest regions of the ICM, we have searched for the presence of spectroscopic cooling in the bright ridges to the north and south of the central AGN. The regions defined as the northern and southern ridges are shown in Fig. \[RidgeRegions\]. These regions were constructed to surround the brightest emission in the image shown in Fig. \[MACSZoom\].
The two ridges are approximately equal distances from the central AGN, but have different spectral properties. The northern ridge has an emission weighted temperature of $kT=4.78 \pm 0.64$ and a metallicity $Z=0.53 \pm 0.11{{\ensuremath{Z_{\mathrm{\odot}}}}}$. This metallicity is higher than the average of this cluster of $Z=0.36 \pm 0.03 {{\ensuremath{Z_{\mathrm{\odot}}}}}$, discussed in more detail in Section \[metalprofsec\]. Adding a cooling flow component to the spectral model of the northern ridge, we measure a spectroscopic mass deposition rate of $\dot{M}=165 \pm 56$ . Comparing the luminosity of the composite [MEKAL+MKCFLOW]{} model with the [MKCFLOW]{} component alone, the spectroscopic cooling of gas down to $kT \sim$0.1 $\keV$ contributes $\sim$30% of the emission from the northern ridge region. The southern bright ridge has a similar overall temperature (emission weighted $kT=5.87^{+1.30}_{-0.46}$), but a lower metallicity ($Z=0.22 \pm 0.09$) and a 90% confidence upper limit on the spectral cooling rate of $\dot{M} <83$ . These results are summarized in Table \[speccoolingtable\].
We note that the [MEKAL+MKCFLOW]{} model offered a significant improvement to the spectral fit of the northern ridge ($\Delta C \sim$10 with one additional fit parameter, a significant improvement at the $\sim$99.9% confidence level), almost identical to the $\Delta C$ improvement obtained with a two temperature ([MEKAL+MEKAL]{} ) fit, which introduces two additional fit parameters to the default [MEKAL]{} model. We also attempted to further resolve the spectroscopic cooling flow by using 2 or more [MKCFLOW]{} components with different temperature ranges, but none of these more complicated models provided a further statistically significant improvement to the fit. For the southern ridge, the fit was not improved by the addition of either a cooling flow or a second [MEKAL]{} component.
[ c c c ]{}\
Measurement & North & South\
$kT$ & $4.78 \pm 0.64 $ & $5.87 ^{+1.30}_{-0.46}$\
\
$Z$ & $0.53 \pm 0.11$ & $ 0.22 \pm 0.09 $\
\
$\dot{M}$ & $165^{+45}_{-67} $ & $ <83$\
\
### Azimuthally Averaged Thermodynamic Profiles {#thermoprofsec}
The azimuthally averaged temperature profile for is shown in Figure \[TempWRidges\]. It is clear in this profile that the temperature does not decrease monotonically towards the center as is expected in a cool core cluster [e.g. @Allen2001; @Vikhlinin2005], but instead appears to increase in the central-most regions. This central jump in temperature is most clear in deprojection, where the innermost temperature is several higher than the measured temperature in the adjacent region. Such a sharp discontinuity in the temperature is often associated with shock heating, but the presence of the cooler ridges separated from the central AGN makes presence of shock heating ambiguous. To account for these cool ridge substructures, thermodynamic profiles were taken with the ridge regions excluded. These thermodynamic profiles are shown in Fig. \[ThermoProfs\]. These profiles more accurately describe the average thermodynamic structure of the ICM. After excluding the ridges, the deprojected temperature in the central-most region still does not decrease towards the center. The sharp temperature jump seen in the center of the deprojected temperature profile is less pronounced, but still present, and there are weak indications of discontinuities in the central bin of the density profile as well. All of the thermodynamic profiles of Figure \[ThermoProfs\] are consistent with shock heating occurring in the central 20 , but the small number of regions and relatively low signal-to-noise do not allow for us to claim an unambiguous detection. Such a jump in the central temperature could also be due to the presence of cavities devoid of ICM gas. In that case, the measured central temperature would be due to ICM gas in front of and behind the central AGN. This should lead to a flatter deprojected density profile in the central regions, a feature that is not observed. The innermost annular region of the thermodynamic profiles is also considerably larger than any apparent cavities in the X-ray emission, so it is unlikely that this temperature jump is due to projection effects involving cavities.
Both the temperature and density profiles show other discontinuities within the central 100 . The density profile decreases discontinuously at several locations while the temperature profile increases sharply at $r \sim$70 $\kpc$. These profiles lead to clear discontinuities in the entropy and an unusually flat pressure profile. Such discontinuities arise from the presence of either cold fronts or weak ($\mach < 2$) shocks. Since the lower temperature gas on the inner edges of these fronts has a higher density, this profile is more consistent with the presence of cold fronts than shock heating. Heating from weak shocks cannot be ruled out, however, since the state of the ICM at earlier times before any bulk motion or shock heating is unknown. Deeper data are required to distinguish between these possibilities.
### Metallicity Profiles {#metalprofsec}
Surprisingly, the metallicity profile of shown in Fig. \[Metals\] exhibits no deviations from a constant metallicity of $Z=0.36{{\ensuremath{Z_{\mathrm{\odot}}}}}$ out to distances as large as 400 . All attempts to fit the data to a multi-temperature [MEKAL]{} model did not lead to any significant increases in the central metallicity. A central metallicity peak is almost always observed in the profiles for cool core clusters [@Allen1998; @DeGrandi2001; @Leccardi2008; @Leccardi2010; @Ehlert2009]. The only region in the cluster that has a significant metallicity enhancement is the bright ridge of X-ray gas to the north of the central AGN, where the metallicity is closer to $Z=0.5{{\ensuremath{Z_{\mathrm{\odot}}}}}$. The unusual metallicity profile of argues that the cool core may have undergone substantial stripping, likely associated with its bulk motion. A dramatic example of stripping of a cool core due to bulk motion has been recently reported in the nearby Ophiuchus Cluster [@Million2010].
![The regions defined as the bright northern and southern ridges, drawn over the image of Fig. \[MACSZoom\] in blue. These encapsulate the regions of the brightest X-ray emission in the cluster apart from the central AGN. The spectral properties of these regions are discussed in detail in Section \[ridges\]. []{data-label="RidgeRegions"}](RidgesFull.ps){width="47.00000%" height="53.00000%"}
![\[TempWRidges\] The azimuthally averaged temperature profile of . The red triangles denote the projected temperature profile, while the black squares denote the deprojected profile. The central region has a much higher temperature than the adjacent regions, which suggests that shock heating might be present. The radius of the temperature minimum corresponds to the distance of the northern and southern ridges.](TempWithRidges.ps){width="47.00000%"}
Radio Observations with {#radiodata}
========================
1.4 GHz radio observations were made with the Very Large Array (VLA) of the National Radio Astronomy Observatory on 2006 April 14. The data were obtained in A configuration, and the time on source was 54 minutes. A central radio source with flux density $\sim$70 mJy is clearly associated with the central AGN and the core of the elliptical galaxy. Approximately 45" to the south is a Narrow Angle Tail (NAT) radio galaxy with flux density $\sim$135 mJy. As shown in Fig. \[RadioWide\], the tails of the NAT are swept back to the south, in the same north-south orientation as the major axis of the central elliptical. The radio morphology of the central AGN (seen in Fig. \[RadioZoom\]) is amorphous without the clearly defined jets or lobes that are found in many radio galaxies. Such amorphous radio structures have been seen associated with cD galaxies in cooling core clusters such as PKS0745-191 [@Baum1991; @Taylor1994], 3C317 in A2052 [@Zhao1993], and PKS 1246-410 in the Centaurus cluster [@Taylor2002]. Most likely the radio jets have been disrupted on small scales by dense gas.
The Central AGN: Power and Accretion {#centralagn}
====================================
The energy of the outburst occurring at the central AGN manifests itself as both radiative emission and as a jet that inflates cavities filled with radio plasma. The energy being input into both of these channels can be estimated from the observations.
Estimating the Radiative Power {#AGNRad}
------------------------------
The central AGN is sufficiently bright in X-rays to directly measure its spectrum and luminosity. Since its spectrum is presumed to be non-thermal in origin, we have modeled it with a variety of power law models.
Source counts are extracted from a 2 region centered on the AGN. The spectral background is extracted from regions that are chosen to be near the AGN and avoid the bright ridges to the north and south, which are likely poor representations of the true background surrounding the AGN. The source and background regions are shown in Fig. \[AGNFig\](a). The net spectrum for the AGN using these regions is shown in Fig \[AGNFig\](b). Three different assumptions about the absorption have been examined: 1) that the absorption is due only to Galactic contributions, and is fixed at the value of @Kalberla2005; 2) that the absorption is only due to Galactic contributions, but the column density [$n_{\mathrm{H}}$]{} is a free parameter in the fit; and 3) that the absorption includes both a fixed Galactic component and an intrinsic component at the redshift of the cluster with the intrinsic column density [$n_{\mathrm{H}}$]{} a free parameter. The results of these fits including the model fluxes and luminosities between 0.7 and 8.0 are given in Table \[AGNTable\]. It is clear from the fit statistics that a larger absorption column density is favored ($\Delta C >12$, a significant improvement at a confidence level well above 99.9%). We conclude that the central AGN has a luminosity in the energy range of 0.7-8.0 of $\sim$8 $\times 10^{43}$ . Systematic uncertainties arising from the choice of the background region are estimated at $\sim$30%.
Estimating the Jet Power {#AGNJet}
------------------------
The radio and X-ray emission both suggest that X-ray cavities filled with radio plasma reside to the east and west of the central AGN. The energy and power required to inflate and fill these cavities ($4PV$) can be estimated using the methods described by @Allen2006, [see also @Dunn2004; @Dunn2005; @Churazov2002].
![\[Metals\] Azimuthally averaged metallicity profiles of . The black curve is the mean profile derived for the Low Entropy Core (LEC) sample from @Leccardi2010 scaled to the estimated ${\ensuremath{r_{\mathrm{180}}}}$ for . The blue circles are the projected metallicity profile, while the red triangles are the deprojected metallicity profile. ](MetalProfsThick.ps){width="\columnwidth"}
Neither the X-ray nor radio image provide an well-defined, unambiguous choice for cavity regions. The X-ray image shows two regions to the east and west of the central AGN that look like cavities, although the outer boundaries are difficult to determine, particularly given the presence of the bright ridges to the north and south. In nearby systems, depressions in the X-ray emission associated with the full extent of the radio plasma-filled cavities are typically visible only in very deep X-ray observations [@Birzan2008]. The radio emission, on the other hand, has a much more naturally defined outer boundary, but our 1.4 GHz radio data for this $z=0.35$ cluster do not show clear lobe or jet structure to the radio source. The boundaries of this radio emission suggest that it is confined by the surrounding X-ray gas, but we caution that if the radio plasma were ‘leaking out beyond’ the boundary of the cavities then estimating cavity volumes based on the radio emission would lead to overestimating the energy in the cavities [@Finoguenov2008]. With these uncertainties in mind, identical calculations were performed with two different sets of cavities: one set based on the structure of the radio emission (hereafter the radio cavities) and another set based on the apparent cavities in the X-ray images (hereafter the minimal X-ray cavities). The true cavity volumes, $4PV$ enthalpy, and jet power are expected to lie somewhere between the values calculated from these two sets of cavities, both shown in Fig. \[Bubbles\].
Both the radio and minimal X-ray cavities were modeled as tri-axial ellipsoids with volume $V=(4/3)\pi{\ensuremath{r_{\mathrm{l}}}}{\ensuremath{r_{\mathrm{w}}}}{\ensuremath{r_{\mathrm{d}}}}$. The measured lengths ${\ensuremath{r_{\mathrm{l}}}}$ and ${\ensuremath{r_{\mathrm{w}}}}$ are the lengths of the axes in the image plane along and perpendicular to the jet axis, respectively. The final length, [$r_{\mathrm{d}}$]{}, is the axis of the cavity along the line of sight. The initial length of the axis along the line of sight was estimated as the smaller of the two planar axis lengths, but allowed to vary independently from them. Calculating the sound speed for the X-ray emitting gas with mean molecular weight $\mu=0.62$ and adiabatic index $\gamma=5/3$, we estimated the time scale for bubble formation as ${\ensuremath{t_{\mathrm{age}}}}=({\ensuremath{r_{\mathrm{l}}}}/{\ensuremath{c_{\mathrm{s}}}})$. From this, we calculated the power required to inflate the cavities as roughly ${\ensuremath{P_{\mathrm{jet}}}}=(4PV/{\ensuremath{t_{\mathrm{age}}}})$. A Monte Carlo analysis was performed that drew the temperature, density, and all three spatial axes from independent Gaussian distributions. The assumed uncertainties on [$r_{\mathrm{l}}$]{}, [$r_{\mathrm{w}}$]{}, and [$r_{\mathrm{d}}$]{} were 20%, 30%, and 30%, respectively, leading to a systematic uncertainty of $\sim$50% in the volume. Both the temperature and density were calculated from the major axis [$r_{\mathrm{l}}$]{} using a power-law parametrization of the profiles shown in Fig. \[ThermoProfs\] between 8 and 80 ($X(r)={\ensuremath{X_{\mathrm{0}}}}r^{\alpha}, X \in kT,{\ensuremath{n_{\mathrm{e}}}}$). From these variables the pressure, enthalpy, and jet power were then calculated. The prior assumptions and subsequent calculations of the cavity energetics are listed in Table \[AGNCalcs\]. Although each $4PV$ calculation includes systematic uncertainties of the particular cavity volume, the dominant uncertainty is identifying a particular set of cavities, which leads to an uncertainty in the $4PV$ enthalpy of approximately a factor of 8 and jet power uncertian within a factor of roughly 4.
We find that the total $4PV$ enthalpy between the two bubbles, after accounting for all uncertainties, is approximately $1-8 \times 10^{60} \erg$. This corresponds to a power input into the ICM from the jet of approximately ${\ensuremath{P_{\mathrm{jet}}}} \sim$ 4 – 14 $\times 10^{45}$. The mechanical energy going into inflating these cavities is nearly two orders of magnitude larger than the radiative emission of the AGN, and larger than the bolometric luminosity within the central 50 of the cluster. Based on the scaling relation of @Cavagnolo2010 and the radio luminosity, the inferred jet power is ${\ensuremath{P_{\mathrm{jet}}}} = 7.7 \times 10^{44}$ , consistent with the lower range of jet powers measured here after accounting for the scatter. The temperature profile shows evidence for heating that goes out to roughly the same distance as the radio emission, so it is possible that the bubbles are in fact as large as the radio emission even though the X-ray depressions are much smaller in scale. The jet power derived from the radio cavities is comparable to the power input measured in the 200 cavities of MS0735.6+7421 [@McNamara2005], which was previously the system with the most powerful jets measured.
{width="\hsize"}
{width="\hsize"}
{width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"} {width="0.33\hsize"}
Discussion
==========
As a larger, more luminous, higher redshift analog to nearby systems like the Perseus Cluster [@Fabian2003; @Fabian2006], provides an extreme example of a cluster with a rapidly cooling core and powerful AGN feedback. This powerful AGN outburst combined with merger-induced motion has led to a cool core undergoing destruction to an extent previously unobserved in galaxy clusters.
There is clear evidence that has undergone a merger event that induced large oscillatory motions of the core. On scales of $r \sim$200 $\kpc$, a spiral of cooler, denser gas seen in both the X-ray image and temperature map is observed to wrap around the core. Such spiral structures arise naturally from mergers and subsequent sloshing and are observable for several Gyr after the merger event [@Ascasibar2006]. Clear deviations from elliptical symmetry are seen in the isophotes, whose centroids shift with distance along the major axis to the north and south. On smaller scales near the core ($r \sim$50 $\kpc$), the X-ray data show that the core has a history of motion in the north-south direction. There are two bright ridges to the north and south of the central AGN. The northern ridge has a sharp northern edge (possibly a cold front), while the southern ridge has a diffuse tail of emission trailing its southern edge. Both of these features are consistent with the densest gas currently undergoing motion to the north.
[ c c c c c c c c ]{}\
& Galactic [$N_{\mathrm{H}}$]{} & Intrinsic [$N_{\mathrm{H}}$]{} & [Photon Index]{} & [Normalization]{} & Flux & Luminosity & $C/\nu$\
Fixed Galactic Absorption & 0.083 & 0.0 & $1.24 \pm 0.07$ & $(2.05 \pm 0.15) \times 10^{-5}$ & $1.76^{+0.14}_{-0.11}$ & $7.27^{0.58}_{-0.44} $ & 519.81/497\
\
Free Galactic Absorption & $0.42 \pm 0.11$ & 0.0 & $1.70 \pm 0.16$ & $(3.86^{+0.64}_{-0.96}) \times 10^{-5}$ & $1.97^{+0.18}_{-0.39}$ & $8.14^{+0.73}_{-1.63}$& 505.68/496\
\
Free Intrinsic Absorption & 0.083 & $0.71^{+0.27}_{-0.22}$ & $1.70 \pm 0.16 $ & $(3.83^{+0.88}_{-0.68}) \times 10^{-5}$ & $1.95^{+0.21}_{-0.23}$ & $8.05^{+0.89}_{-0.97}$& 506.42/496\
{width="0.95\hsize"}
{width="0.95\hsize"}
[ c c c c c ]{}\
& [East (Radio)]{} & [West (Radio)]{} & [East (X-Ray)]{} & [West (X-Ray)]{}\
[$r_{\mathrm{l}}$]{} () & $20.2 \pm 4.0$ & $30.0 \pm 6.0$ & $13.4 \pm 2.7$ & $12.9 \pm 2.6$\
[$r_{\mathrm{w}}$]{} () & $25.9 \pm 9.1$ & $24.5 \pm 7.3$ & $9.4 \pm 2.8$ & $10.3 \pm 3.1$\
[$r_{\mathrm{d}}$]{} () & $20.2 \pm 6.1$ & $24.5 \pm 7.3$ & $9.4 \pm 2.8$ & $10.3 \pm 3.1$\
[$kT_{\mathrm{0}}$]{} () & $5.30 \pm 0.25$ & $5.30 \pm 0.25$ & $5.30 \pm 0.25$ & $5.30 \pm 0.25$\
[$\alpha_{\mathrm{kT}}$]{} & 0 & 0 & 0 & 0\
[$n_{\mathrm{e,0}}$]{} ($\cm^{-3}$) & $3.47 \pm 0.28$ & $3.47 \pm 0.28$ & $3.47 \pm 0.28$ & $3.47 \pm 0.28$\
[$\alpha_{\mathrm{ne}}$]{} & $-1.17 \pm 0.02$ & $-1.17 \pm 0.02$ & $-1.17 \pm 0.02$ & $-1.17 \pm 0.02$\
\
$V$ ($10^{69} \cm^{-3}$) & $1.25_{-0.50}^{+0.83}$ & $1.64_{-0.65}^{+1.31}$ & $0.11 \pm 0.08$ & $0.14 \pm 0.08$\
\
$P$ ($\keV \cm^{-3}$) & $0.50^{+0.20}_{-0.10}$ & $0.33^{+0.11}_{-0.06}$ & $0.85 \pm 0.20$ & $0.81 \pm 0.25$\
\
$4PV$ ($10^{60} \erg$) & $4.62 \pm 2.16 $& $3.98^{+2.65}_{-1.51} $ & $0.64 \pm 0.32$ & $0.85 \pm 0.40$\
\
${\ensuremath{c_{\mathrm{s}}}}$ ($\km \s^{-1}$) & $1170 \pm 30 $ & $1170 \pm 30 $ & $1170 \pm 30 $ & $1170 \pm 30 $\
\
${\ensuremath{t_{\mathrm{age}}}}$ ($10^{6}$ yr) & $17.0 \pm 3.4 $ & $25.2 \pm 5.02 $ & $11.4 \pm 2.1$ & $10.7 \pm 2.1$\
\
${\ensuremath{P_{\mathrm{jet}}}}$ ($10^{45} \erg \s^{-1}$) & $8.50 \pm 4.61$ & $ 5.36 \pm 3.06$ & $1.90 \pm 1.22$ & $2.42 \pm 1.31$\
Our optical data independently suggest such a merger: the ICL around the cD galaxy is highly elongated in the north-south direction. [*B*]{}-band emission originating from a young stellar population is observed both to the north and south of the cD galaxy, while emission originating from ongoing star formation is observed predominantly at positions coincident with the northern ridge. The presence of a young stellar population to the south without any emission suggests that the primary region of star formation is moving northward.
In the midst of the motions of the cluster core, a powerful AGN outburst has taken place. The central AGN is bright in X-ray emission, with a luminosity in the energy band of 0.7-8.0 of $\sim$8 $\times 10^{43} $ . This AGN is surrounded by extended, amorphous 1.4 GHz radio emission. The major axis of this radio emission is spatially coincident with depressions in the X-ray emission. The physical extent of the cavities based on the observations is unclear, but estimating cavities based on the X-ray and radio emission gives a robust range for the $4PV$ enthalpy of the cavities and their corresponding jet power. The $4PV$ enthalpy of these cavities is sufficient to counteract the radiative losses from the central 50 for 30-250 Myr. The jet power ($\sim$4 – 14 $\times 10^{45}$ ) identifies among the most powerful cavity sources yet observed. The power input into inflating these cavities is approximately 100 times larger than the measured radiative power of the central AGN, and four to ten times larger than the bolometric luminosity of the cool core.
Unlike other more typical cool core clusters [@Allen1998; @DeGrandi2001; @Vikhlinin2005; @Pratt2007; @Werner2008], the azimuthally-averaged metallicity profile for shows no significant deviations from a constant value. Assuming that there once was a central metallicity peak in , this suggests that large masses of metal-rich gas has been stripped from the center of the cluster and displaced to the surrounding regions. The extent of transport required to account for the flat metallicity profile is strong evidence that the original cool core has undergone destructive stripping as it traversed from one side of the cD galaxy to the other. The only region with an exceptional metallicity is the northern central ridge. The central metallicity peak in cool core clusters is usually expected to be robust, even in clusters with powerful central AGN activity [@Bohringer2004; @Rasera2008].
![\[Bubbles\] Estimating the jet power using the X-ray cavities and radio emission. The image of Fig. \[Butterworth\] is shown with the radio contours overlaid in magenta, the radio cavities overlaid in blue, and the minimal X-ray cavities overlaid in black. The determination of the $4PV$ enthalpy and jet power are derived from a Monte Carlo analysis that generates plausible elliptical cavities based on these ellipses. The true cavity volumes are expected to reside between the volumes calculated from these two sets of cavities. ](BubbleContours2.ps){width="47.00000%"}
Merger and central AGN activity have resulted in the formation of two X-ray bright ridges roughly equal distances north and south of the central AGN. The X-ray bright ridge to the north of the central AGN has characteristics usually associated with a cool core. It is the location of the lowest temperature, highest density gas in the cluster. It is also the location of the most metal rich gas. The X-ray spectrum of this northern ridge has a spectroscopic cooling rate of $\dot{M}
\sim$165 down to $0.1 \keV$, in good agreement with the observed star formation rate estimated from the emission, $\sim$170 . This emission is expected to be almost entirely from star formation, as it is roughly 3 orders of magnitude larger than what expected from Case B Recombination [@Johnstone1986]. The spectroscopic cooling contributes a large fraction of the emission from this region, estimated at 30% based on the spectral modeling. The northern ridge might therefore be in the early stages of ’catastrophic cooling’ [e.g. @Fabian1977; @Peterson2006]. The cooler ICM gas to the south of the cD galaxy has a lower metallicity and a current spectroscopic cooling rate consistent with zero, but is also the location of a young stellar population. This appears consistent with sloshing-induced stripping of the cool core throughout its oscillations along the north-south axis. The asymmetric thermodynamic structure and different stellar populations of these ridges clearly indicate that there is no longer a single core of low entropy gas surrounding the cD galaxy. The extent to which AGN feedback contributed to the present-day thermodynamic structure of , apart from core sloshing, is beyond the scope of the current observations. The majority of the stripping and disruption of metal rich, low entropy gas from the original cool core could have been caused by the bulk motions, but it seems likely that the AGN outburst may have contributed to the separation of the preexisting core into two X-ray bright ridges roughly equal distances from the central AGN [@Guo2010]. This cluster’s star formation is exceptional, especially as it does not satisfy all of the empirical conditions for central star formation discussed in @Rafferty2008. Although the cooling time and entropy are below the thresholds put forward by @Rafferty2008, all of the systems with clear evidence for high star formation rates also have AGN jet powers smaller than the cooling luminosity, which is not the case in .
Although the extent of stripping and cool core disruption in is substantial, similar phenomena have also been observed in the nearby Ophiuchus Cluster [@Million2010]. Many of the morphological structures seen in are also seen in the Ophiuchus Cluster. Both clusters have clearly shifting isophotes, inner cold fronts, and comet-like diffuse emission trailing to one side of the inner cool core. The Ophiuchus Cluster also shows strong evidence for stripped core gas in the form of a metal rich ridge to the north of the cool core. These processes also appear to be occurring in , but with the added complications and energy of central AGN feedback.
The extent to which a cool core can be disrupted or even destroyed by AGN feedback and merger induced oscillations has important implications for cosmological studies with clusters [e.g. @Burns2008; @Mantz2009a; @Mantz2009b]. Further observations with X-ray, optical, and radio instruments could provide many new insights into the extent that this cool core has been disrupted. Radio observations at higher resolutions and lower frequencies could allow for a better understanding of the amorphous central radio source, in particular discerning the origin of the emission by measuring the radio spectrum. Higher resolution, deeper observations may also be able to resolve the jet and lobe structure of the central AGN source, which is critical for further constraining the magnitude and origin of the AGN outburst. Optical spectroscopy would enable more precise measurements of the star formation rate in the different regions surrounding the central AGN, and perhaps allow for measurements of the black hole mass. Optical spectroscopy would also allow for measurements of the emission line velocities within the cool core remnant, providing more details as to the extent of the disruption of the cool core. Deeper observations with would provide better measurements of the unusual metallicity structure and the extent that gas has been stripped from the cool core remnant. The distribution of regions that undergo cooling and the extent of that cooling could also be measured in more detail. Finally, a deeper X-ray exposure could shed new light on the thermodynamic structure of the cluster, in particular provide more compelling evidence for cold fronts and/or shock heating within the central 100 . Simulations designed to reconstruct the thermodynamic structure of may elucidate the nature of extreme sloshing and feedback, and perhaps also the future evolution of such a profoundly disrupted core.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was carried out with Observation Awards Number GO8-9119X & GO0-11139X. Support for this work was provided by the National Aeronautics and Space Administration through Chandra/Einstein Postdoctoral Fellowship Award Number PF8-90056 (N.W.) and PF9-00070 (A.S.) issued by the Chandra X-ray Observatory Center, which is operated for and on behalf of the National Aeronautics and Space Administration under contract NAS8-03060. G.G. is a postdoctoral researcher of the FWO-Vlaanderen (Belgium). RJHD acknowledges support from the Alexander von Humboldt Foundation. This research was supported by the DFG cluster of excellence ‘Origin and Structure of the Universe’ [www.universe-cluster.de](www.universe-cluster.de).
Further support for this work was provided by the Department of Energy Grant Number DE-AC02-76SF00515. The National Radio Astronomy Observatory is operated by Associated Universities, Inc., under cooperative agreement with the National Science Foundation. We also thank Silvano Molendi for providing the mean metallicity profile data from @Leccardi2010. We finally thank the anonymous referee for their suggestions and comments that improved this work.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study the asymptotic properties of the Bergman kernels associated to tensor powers of a positive line bundle on a compact Kähler manifold. We show that if the Kähler potential is in Gevrey class $G^a$ for some $a>1$, then the Bergman kernel accepts a complete asymptotic expansion in a neighborhood of the diagonal of shrinking size $k^{-\frac12+\frac{1}{4a+4{\varepsilon}}}$ for every ${\varepsilon}>0$. These improve the earlier results in the subject for smooth potentials, where an expansion exists in a $(\frac{\log k}{k})^{\frac12}$ neighborhood of the diagonal. We obtain our results by finding upper bounds of the form $C^m m!^{2a+2{\varepsilon}}$ for the Bergman coefficients $b_m(x, \bar y)$ in a fixed neighborhood by the method of [@BBS]. We also show that sharpening these upper bounds would improve the rate of shrinking neighborhoods of the diagonal $x=y$ in our results.'
address: 'Department of Mathematics, Johns Hopkins University, Baltimore, MA 21218, USA'
author:
- Hang Xu
title: Asymptotic properties of Bergman kernels for potentials with Gevrey regularity
---
Introduction
============
Let $(L,h) \to M$ be a positive Hermitian holomorphic line bundle over a compact complex manifold of dimension $n$. The metric $h$ induces the form ${\omega}= -\tfrac{\sqrt{-1}}{2} {\partial}{\bar{\partial}}\log(h)$ on $M$. For $k$ in $\mathbb N$, let $H^0(M,L^k)$ denote the space of holomorphic sections of $L^k$. The [Bergman projection]{} is the orthogonal projection $\Pi_k: {L}^{2}(M,L^k) \to H^0(M,L^k)$ with respect to the natural inner product induced by the metric $h^k$ and the volume form $\frac{ {\omega}^n }{n!}$. The *Bergman kernel* $K_k$, a section of $L^k{\otimes}\bar{L}^k$, is the distribution kernel of $\Pi_k$. Given $p \in M$, let $(V, e_L)$ be a local trivialization of $L$ near $p$. We write $| e_L |^2_{h}=e^{-\phi}$ and call $\phi$ a local potential. In the frame $ e_L^{k} {\otimes}{\bar{e}_L^{k}}$, the Bergman kernel $K_k(x,y)$ is understood as a function on $V \times V$. We note that on the diagonal $x=y$, the function $K_k(x,x)e^{-k\phi(x)}$ is independent of the choice of the local frame, hence it is a globally defined function on $M$ called the *Bergman function*, which is also equal to $| K_k(x, x) |_{h^k}$.
Zelditch [@Ze1] and Catlin [@Ca] proved that on the diagonal $x=y$, the Bergman kernel accepts a complete asymptotic expansion of the form $$\label{ZC}
K_k(x,x)e^{-k\phi(x)}\sim\frac{k^n}{\pi^n}\left(b_0(x,\bar{x})+\frac{b_1(x,\bar{x})}{k}+\frac{b_2(x,\bar{x})}{k^2}+\cdots\right).$$ *Near the diagonal*, i.e. in a $\sqrt{\frac{\log k}{k}}$-neighborhood of the diagonal, one has a scaling asymptotic expansion for the Bergman kernel (see [@ShZe; @MaMaBook; @MaMaOff; @LuSh; @HKSX]). For $d(x, y) \gg \sqrt{\frac{\log k}{{k}}}$, where $d$ is the Riemannian distance induced by $\omega$, no useful asymptotics are known for smooth metrics. However, there are off-diagonal upper bounds of Agmon type $$\label{Agmon} \left | K_k(x, y) \right |_{h^k} \leq C k^n e^{- c \sqrt{k} d(x, y)},$$ proved for smooth metrics in [@Ch1; @Del; @Lindholm; @Bern; @MaMaAgmon]. In fact as shown in [@Ch3; @HX], one has better decay estimates. More precisely, there exist positive constants $c, C$ and a function $f(k) \to \infty$ as $k \to \infty$ such that $$|K_k(x,y)|_{h^k} \leq
\begin{cases}
C k^{n} e^{-c \, k d(x,y)^2}, \quad & d(x,y)\leq f(k)\sqrt{\frac{\log k}{k}},\\
C k^{n} e^{-c \, f(k)\sqrt{k\log k} \,\, d(x, y)}, \quad & d(x,y) \geq f(k)\sqrt{\frac{\log k}{k}}.
\end{cases}$$ A quantitative version of the above estimate that relates the growth rate of $f(k)$ to the growth rate of the derivatives of the metric $h$ is obtained in [@HX]. In particular, when $h$ is in the Gevrey $a$ $(a\geq 1)$ class, we get $f(k)= \frac{k^{\frac{1}{4a-2}}}{\sqrt{\log k}}$.
This article generalizes the results in [@HLXanalytic] to the setting of Gevrey classes. To be precise, we prove an asymptotic expansion in a $k^{-\frac{1}{2}+\frac{1}{4a+4{\varepsilon}}}$ neighborhood of the diagonal for any ${\varepsilon}>0$ if the metric $h$ is in the Gevrey $a$ $(a>1)$ class. In particular, we show that in the Gevrey $a$ class, uniformly for all sequences $x_k$ and $y_k$ with $d(x_k, y_k) \leq k^{-\frac{1}{2}+\frac{1}{4a+4{\varepsilon}}}$, we have $$\left | K_k(x_k, y_k) \right |_{h^k} \sim \frac{k^n}{\pi^n} e^{- \frac{k D(x_k, y_k)}{2}}, \quad \text{as} \quad k \to \infty,$$ where $D(x, y)$ is Calabi’s diastasis function , which is controlled from above and below by $d^2(x, y)$. Before we state the results we must also mention that in [@BBS], there is an off-diagonal asymptotic expansion for the Bergman kernel of the form $$\label{ill}
K_k(x,y)=e^{k\psi(x,\bar y)}\frac{k^n}{\pi^n}\left (1+\sum _{j=1}^{ {N-1}}\frac{b_j(x, \bar y)}{k^j}\right )+e^{k\left(\frac{\phi(x)}{2}+\frac{\phi(y)}{2}\right)}k^{-N+ n}O_N(1),$$ which holds for all $d(x, y) \leq \delta$ for some $\delta >0$. Here, $\psi(x, \bar y)$ and $b_j(x, \bar y)$ are almost holomorphic extensions of $\phi(x)$ and $b_j(x, \bar x)$ from . However, note that this expansion is only useful when the term $e^{k\left(\frac{\phi(x)}{2}+\frac{\phi(y)}{2}\right)}k^{-N+ n}$ is a true remainder term, i.e. it is less than the principal term $k^n e^{k \psi(x,\bar y)}$ in size, which holds only in a neighborhodd $d(x, y) \leq {C} \sqrt{\frac{\log k}{{k}}}$ in general. In the case that $h$ is real analytic, this is valid in a larger neighborhood $d(x,y)\leq k^{-1/4}$ [@HLXanalytic]. In a recent preprint [@RSN], this is further improved to a fixed neighborhood independent of $k$.
We now state our main result and its corollaries.
\[Main\] Assume that the local potential $\phi$ is in the Gevrey class $G^a(V)$ for some $a>1$, meaning that for some $C_0$ and $C_1>0$, we have $$\|D_z^\alpha D_{\bar{z}}^\beta \phi(z)\|_{L^\infty (V)}\leq C_0C_1^{|\alpha|+|\beta|}(\alpha!\beta!)^a,
\hspace{12pt} \mbox{ for any multi-indices }\alpha,\beta\geq 0.$$ Then for every ${\varepsilon}>0$, there exist positive constants $\delta$ and $C$, and an open set $U \subset V$ containing $p$, such that for $N_0(k)=[(\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}]$ and uniformly for any $x,y\in U$, we have in the frame $e_L^k\otimes \bar{e}_L^k$ $$K_k(x,y)=e^{k\psi(x,\bar y)}\frac{k^n}{\pi^n}\left (1+\sum _{j=1}^{ N_0(k)-1}\frac{b_j(x, \bar y)}{k^j}\right )+e^{k\left(\frac{\phi(x)}{2}+\frac{\phi(y)}{2}\right)}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1),$$ where $\psi(x, z)$ is a certain almost holomorphic extension of $\phi(x)$ near the diagonal [^1] and $b_m(x , z)$ are certain almost holomorphic extensions (defined by ) of the Bergman kernel coefficients $b_m(x, \bar x)$.
As a first corollary of this theorem, we get a complete asymptotic expansion in a $k^{-\frac12+\frac{1}{4a+4{\varepsilon}}}$ neighborhood of the diagonal.
\[complete asymptotics\] Given the same assumptions and notations as in the above theorem, there exist positive constants $C$ and $\delta$, and an open set $U \subset V$ containing $p$, such that for all $k$ and $N\in \mathbb{N}$, we have for all $x,y\in U$ satisfying $d(x,y)\leq \delta k^{-\frac{1}{2}+\frac{1}{4a+4{\varepsilon}}}$, $$\label{complete formula}
K_k(x,y)=e^{-k\psi(x,\bar{y})}\frac{k^n}{\pi^n}\left(1+\sum_{j=1}^{{N-1}}\frac{b_j(x,\bar{y})}{k^j}+\mathcal{R}_N(x, \bar y,k)\right),$$ where $$\label{complete error}
\left|\mathcal{R}_N(x, \bar y ,k)\right|\leq {\frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}}.$$
And if we only assume $b_m(x,z)$ are arbitrary almost holomorphic extensions of Bergman kernel coefficients $b_m(x,\bar{x})$, then we still have . But the remainder term estimate will be weaker: $\left|\mathcal{R}_N(x, \bar y ,k)\right|\leq \frac{C_N}{k^N}$ for some constant $C_N$.
As another corollary, we obtain the following off-diagonal asymptotic in terms of Calabi’s diastasis [@Cal] function defined by $$\label{Diastatis}
D(x, y)= \phi(x) + \phi(y) - \psi(x, \bar y) - \psi(y, \bar x).$$ We point out that near a given point $p \in M$, we have $D(x, y) = |x -y|_p^2 + O(|x-p|_p^3+|y-p|_p^3)$, where $|z|^2_p:= \sum_{ i, j=1}^n \phi_{i \bar j}(p)z_i{\overline{z_j }}$. If we use *Bochner coordinates* at $p$ (introduced in [@Boc]), in which the potential admits the form $\phi(x)=|x|^2+O(|x|^4)$, we have $D(x, y) = |x -y|_p^2 + O(|x-p|_p^4+|y-p|_p^4)$.
\[Log\] Under the same assumptions and notations (and the same $\delta$ and same $U$) as in Theorem \[Main\], we have uniformly for all $x , y \in U$ satisfying $ D(x, y) \leq \frac12 \delta k^{-1+\frac{1}{2a+2{\varepsilon}}}$, $$\frac{1}{k}\log \left | K_k ( x, y ) \right | _{h^k}=-\frac{D(x, y)}{2} + \frac{n \log k}{k} -\frac{n\log\pi}{k}+ O \left (\frac{1}{k^2} \right ).$$
The following scaling asymptotic is then immediate:
\[Cor2\] In Bochner coordinates at $p$, we have uniformly for all $u,v \in \mathbb{C}^n$ with $| u|_p$ and $|v|_{p} < \frac{\sqrt{\delta}}{3}$, $$\frac{1}{k^{\frac{1}{2a+2{\varepsilon}}}}\log \left | K_k \left (\frac{u}{k^{\frac12-\frac{1}{4a+4{\varepsilon}}}},\frac{v}{k^{\frac12-\frac{1}{4a+4{\varepsilon}}} }\right ) \right |_{h^k}=-\frac{|u-v|_p^2}{2} + \frac{n \log k}{k^{\frac{1}{2a+2{\varepsilon}}}}-\frac{n\log\pi}{k^{\frac{1}{2a+2{\varepsilon}}}} + O \left ( \frac{1}{k^{1+\frac{1}{2a+2{\varepsilon}}}}\right ).$$
One of the key ingredients in our proofs is the following estimate on the Bergman kernel coefficients $b_m(x, z)$. We emphasize again that $b_m(x , z)$ are particular almost holomorphic extensions of the Bergman kernel coefficients $b_m(x, \bar x)$ appearing in the on-diagonal expansion of Zelditch [@Ze1] and Catlin [@Ca].
\[MainLemma\] Assume the potential $\phi$ is in Gevrey class $G^a(V)$ for some $a>1$. Let $b_m(x , z)$ be the almost holomorphic extensions (defined by ) of the Bergman kernel coefficients $b_m(x, \bar x)$. Then, there exists a neighborhood $U\subset V$ of $p$, such that for any $m \in \mathbb N$ we have $$\|b_m(x,z)\|_{L^\infty(U\times U)} \leq C^{m}m!^{2a+2{\varepsilon}},$$ where $C$ is a constant independent of $m$ but dependent on ${\varepsilon}$. Moreover, we have the following estimates on the derivatives of $b_m(x,z)$. Denote $v=(x,z)$. For any multi-indices $\alpha$ and $\beta$ and any $(x, z) \in U \times U$ $$\label{bmupperbound}
\left | D_v^\alpha D_{{\overline{v}}} ^\beta b_m(x,z) \right | \leq C^{m+|\alpha|+|\beta|}m!^{2a+2{\varepsilon}}\alpha!^{a+{\varepsilon}} \beta!^{a+{\varepsilon}} \exp{\left(-b (1- \delta_0(|\beta|)){|x-\bar{z}|^{-\frac{1}{a-1}}}\right)} ,$$ where $C$ is a constant independent of $m, \alpha, \beta$ but dependent on ${\varepsilon}$, and $\delta_0(|\beta|)=1$ only if $\beta=0$ and is zero otherwise. The constant $b$ is positive and is independent of $\alpha, \beta, m, {\varepsilon}$. In addition, when we are restricted to the diagonal $z= \bar x$, we can choose ${\varepsilon}=0$.
We conjecture that in the Gevrey $a$ case, there exist certain almost holomorphic extension $b_m(x,z)$ of the Bergman coefficients $b_m(x,\bar{x})$ such that $$\label{conj}
\left \| D_v^\alpha D_{{\overline{v}}} ^\beta b_m(x,z) \right \|_{L^\infty(U\times U)} \leq C^{m+|\alpha|+|\beta|}m!^{2a-1}\alpha!^{a} \beta!^{a} \exp{\left(-b (1- \delta_0(|\beta|)){|x-\bar{z}|^{-\frac{1}{a-1}}}\right)}$$ As we show in this paper, if this conjecture holds true, then all of the above results can be improved accordingly. In particular, the quantities $N_0(k)= [(k /C)^{\frac{1}{2a+2{\varepsilon}}} ]$ and $e^{- \delta k^\frac{1}{2a+2{\varepsilon}}}$ in the remainder estimate of Theorem \[Main\] would be replaced by $ [(k /C)^{\frac{1}{2a-1}}]$ and $e^{- \delta k^\frac{1}{2a-1}}$, moreover Corollary \[Log\] would hold for all $D(x, y) \leq \frac12 \delta k^{-1+\frac{1}{2a-1}}$. We expect is the best possible result one can seek, because by [@LuTian] the leading term in $b_m(x,\bar{x})$ is $\frac{m}{(m+1)!}\Delta^{m-1}\rho(x)$ where $\rho$ is the scalar curvature, so when the metric is in Gevrey class $G^a$, we have $\frac{m}{(m+1)!}\Delta^{m-1}\rho(x)\approx C^m m!^{2a-1}$. However we are unable to prove this conjecture for general Gevrey $a$ metrics using our method, which is based on a recursive formula of [@BBS]. In Section \[optimal\], we discuss the optimality and limitations of this method.
There is a huge literature on Bergman kernels on complex manifolds. Before closing the introduction we only list some related work that were not cited above: [@BoSj; @En; @Ch; @Loi; @LuTian; @Loi; @MaMa; @Liu; @LiuLu1; @Seto; @LuZe; @Ze3; @LS]. Applications of the Bergman kernel, and the closely related Szegö kernel, can be found in [@Do], [@BSZ], [@ShZe], [@YZ]. The book of Ma and Marinescu [@MaMaBook] contains an introduction to the asymptotic expansion of the Bergman kernel and its applications. See also the book review [@ZeBookReview] for more on the applications of Bergman kernels.
Organization of the paper {#organization-of-the-paper .unnumbered}
-------------------------
In Sections \[Sec Local\] and \[Sec MethodBBS\], we follow the construction of local Bergman kernel in [@BBS], but we obtain precise estimates for the error term by using the growth rate of Bergman coefficients $b_m(x,z)$ provided by Theorem \[MainLemma\]. In Section \[Sec Global\], we give the proofs of Theorem \[Main\] and Corollaries \[complete asymptotics\] and \[Log\]. The proof of Theorem \[MainLemma\] will be given in Section \[Sec ProofofMainLemma\]. Section \[optimal\] discusses the optimality of our bounds on Bergman coefficients. Section \[ProofOfGevreyStuff\] contains the proofs of the properties of almost homomorphic extensions of Gevrey functions.
Local Bergman kernels {#Sec Local}
=====================
In [@BBS], by using *good* complex contour integrals, Berman-Berndtsson-Sjöstrand constructed *local reproducing kernels* (mod $e^{-k\delta}$) for $U=B^n(0,1)\subset \mathbb{C}^n$, which reproduce holomorphic sections in $U$ up to $e^{-k\delta}$ error terms. These kernels are in general not holomorphic. By allowing more flexibility in choosing the amplitudes in the integral, the authors modified these local reproducing kernels to local Bergman kernels, which means that they are almost *holomorphic* local reproducing kernels mod $O(k^{-N})$. The global Bergman kernels are then approximated using the standard Hörmander’s $L^2$ estimates.
Throughout this paper, we assume that $\phi$ is in the Gevrey class $G^a(V)$ for some open neighborhood $V\subset M$ of a given point $p$. Let $B^n(0,r)$ be the ball of radius $r$ in $\mathbb C^n$. We identify $p$ with $0\in \mathbb{C}^n$ and $V$ with the ball $B^n(0,3)\subset\mathbb{C}^n$ and denote $U=B^n(0, 1)$. Let $e_L$ be a local holomorphic frame of $L$ over $V$ as introduced in the introduction. For each positive integer $k$, we denote $H_{k\phi}(U)$ to be the inner product space of $L^2$-holomorphic functions on $U$ with respect to $$\left ( u, v \right)_{k\phi}=\int_U u \bar v \, e^{-k\phi}d \text{Vol},$$ where $d\text{Vol}=\frac{\omega^n}{n!}$ is the natural volume form induced by the form $\omega= \frac{\sqrt{-1}}{2} \partial \bar \partial \phi$. So the norm of $u \in H_{k\phi}(U)$ is given by $$\|u\|^2_{k\phi}=\int_U |u|^2e^{-k\phi}d\text{Vol}.$$ Let $\chi\in C_0^\infty(B^n(0,1))$ be a smooth cut-off function such that $\chi=1$ in $B^n(0,\frac{1}{2})$ and vanishes outside $B^n(0,\frac{3}{4})$. The following result gives a refinement of the the result of [@BBS] by providing a more precise estimate for the error term when the Kähler potential is in Gevrey class $G^a$. The main ingredient of the proof is Theorem \[MainLemma\], whose proof is delayed to Section \[Sec ProofofMainLemma\].
\[LocalBergmanKernel\] For each $N \in \mathbb N$, there exist $K_{k,x}^{(N)}(y) \in H_{k\phi}(U)$ and a positive constant $C$ independent of $N$ and $k$, such that for all $u \in H_{k\phi}(U)$ we have $$\begin{aligned}
\label{eq 5.2}
\forall x \in B^n(0, 1/4): \quad u(x)=\left(\chi u, K^{(N)}_{k,x}\right)_{k\phi}+k^ne^{\frac{k\phi(x)}{2}}\mathcal R_{N+1}(\phi, k) \|u\|_{k\phi},\end{aligned}$$ where $$\label{eq 5.3}
\quad |\mathcal R_{N+1}(\phi, k)| \leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}.$$ The function $K_{k,x}^{(N)}$ is called a local Bergman kernel of order $N$.
In [@BBS], only the qualitative estimate $\mathcal R_{N+1}(\phi, k)= O_N(\frac{1}{k^{N+1}})$ is given.
To prove Proposition \[LocalBergmanKernel\], we first need to recall the techniques of [@BBS].
Review of the method of Berman-Berndtsson-Sjöstrand {#ReviewBBS}
---------------------------------------------------
The main idea is to construct the local almost holomorphic reproducing kernel (also called local Bergman kernel) by means of the calculus of contour pseudo-differential operators (contour $\Psi$DO for short) introduced by Sjöstrand [@Sj]. Before we introduce the notion of contour integrals we present some notations and definitions.
Suppose $\phi(x)$ is in Gevrey class $G^a(V)$ and $V=B^n(0, 3)$. By replacing $\phi(x)$ by $\phi(x) - \phi(0)$, we can assume that $\phi(0)=0$. We then denote $\psi(x,z)=F(\phi)(x,z)$ defined later in Definition \[AHE\] to be one holomorphic extension of $\phi(x)$. Moreover, since $\phi(x)$ is real-valued, we have $ {\overline{\psi(x, z)}} = \psi( \bar z, \bar x)$. We also define $$\label{theta}
\theta(x,y,z)=\int_0^1(D_x\psi)(tx+(1-t)y,z)dt,$$ where the differential operator $D_x$ is the gradient operator defined by $$\begin{aligned}
D_x&=(D_{x_1},D_{x_2},\cdots,D_{x_n}).\end{aligned}$$ Note that $\theta(x, x, z)= \psi_x(x, z)$. It is easy to prove that the Jacobian of the map $(x,y,z)\rightarrow (x,y,\theta)$ at $(x,y,z)=(0,0,0)$ is non-singular. Thus the map is actually an almost biholomorphic map between two neighborhoods of the origin of $\mathbb C^{3n}$. As a result, we can use $(x,y,z)$ or $(x,y,\theta)$ as local coordinates interchangeably. Without loss of generality we can assume that $(x, y, z) \in B^n(0, 3) \times B^n(0, 3) \times B^n(0, 3)$ and $\theta \in W$, where $$W =\theta \left ( B^n(0, 3) \times B^n(0, 3) \times B^n(0, 3)\right ).$$ Note that $W$ contains the origin because by our assumption $\phi(0)=0$.
A fundamental idea of [@BBS] is to use the estimate $$\label{BBS 2.1}
u(x)= c_n \left(\frac{k}{2\pi}\right)^n\int_\Lambda e^{k\,\theta\cdot (x-y)}u(y)\chi(y)d\theta\wedge dy
+O(e^{-k\delta})e^{\frac{k\phi(x)}{2}}\|u\|_{k\phi}$$ which holds uniformly for $x \in B^n(0, \frac14)$, for any holomorphic function $u$ defined on $ B^n(0, 1)$. Here, $c_n=i^{-n^2}$, $\delta$ is a positive constant, and $\Lambda=\{(y,\theta):\theta=\theta(x,y)\}$ is a *good contour*, which means that there exists $\delta>0$ such that for any $x,y$ in a neighborhood of the origin, $$\begin{aligned}
\label{GoodContour}
2\operatorname{Re}\theta\cdot (x-y)\leq -\delta|x-y|^2-\phi(y)+\phi(x).\end{aligned}$$ One can easily verify that $$\label{contour} \Lambda=\{(y,\theta):\theta=\theta(x,y,\bar{y})\},$$ with $\theta(x,y,z)$ defined by , is a good contour by observing that $$\theta\cdot (x-y)=\psi(x,\bar{y})-\psi(y,\bar{y}).$$ To put into a useful perspective, one should think of the integral in as a contour $\Psi$DO defined as follows. Let ${a=a(x,y,\theta,k)}$ be an almost holomorphic symbol in $B^n(0, 3) \times B^n(0, 3) \times W$, with an asymptotic expansion of the form $$a(x,y,\theta,k)\sim a_0(x,y,\theta)+\frac{a_1(x,y,\theta)}{k}+\frac{a_2(x,y,\theta)}{k^2}+\cdots \; .$$ For simplicity, we will suppress the dependency on $k$ and write $a=a(x,y,\theta)$.
A $\Psi$DO associated to a good contour $\Lambda$ and an amplitude $a(x, y, \theta)$, is an operator on $C^\infty_0(U)$ defined by $$\text{Op}_\Lambda(a)\, u = c_n \left(\frac{k}{2\pi}\right)^n \int_\Lambda e^{k\,\theta\cdot (x-y)}a(x, y, \theta)\, u(y)\, d\theta\wedge dy.$$ Thus in this language means that for $x \in B^n(0, 1/4)$ $$(\chi u)(x)= \text{Op}_\Lambda(1) (\chi u)
+O(e^{-k\delta})e^{\frac{k\phi(x)}{2}}\|u\|_{k\phi}.$$ Roughly speaking this says that $\text{Op}_\Lambda (1)$ is the identity operator mod $O(e^{-k\delta})$. We define the integral kernel $K_{k, x}(y)$ of $\text{Op}_\Lambda(a)$ with respect to the inner product $( \cdot, \cdot )_{k \phi}$, by $$\text{Op}_\Lambda(a) u =\left(u, K_{k,x}\right)_{k\phi}.$$ The first observation is that the kernel $K_{k, x}(y)$ of $\text{Op}_\Lambda(1)$, associated to the contour , is not almost holomorphic. The idea of [@BBS] is to replace $\text{Op}_\Lambda(1)$ by $\text{Op}_\Lambda(1+a)$ where $a(x, y, \theta)$ is a *negligible amplitude* and the kernel of $\text{Op}_\Lambda(1+a)$ is almost holomorphic. An amplitude $a(x, y, \theta)$ is negligible if $$\text{Op}_\Lambda(a) (\chi u)= O(k^{-\infty})e^{\frac{k\phi(x)}{2}}\|u\|_{k\phi}.$$ To find a suitable condition for negligible amplitudes one formally writes $$\text{Op}_\Lambda(a) = \text{Op}_\Lambda(S a |_{x=y}),$$ where $S$ is a standard operator that is used in microlocal analysis to turn a symbol $a(x, y, \theta)$ of a $\Psi$DO to a symbol of the form $\widetilde a (x, \theta)$. The operator $S$ is formally defined by $$S=e^{\frac{D_\theta \cdot D_y}{k}}=\sum_{m=0}^\infty\frac{(D_\theta\cdot D_y)^m}{m!k^m}.$$ Then an amplitude $a$ is negligible if $S a |_{x=y} \sim 0$ as a formal power series. This implies that there exists an almost holomorphic vector field $A(x, y, \theta)$ with formal power series $$A(x,y,\theta)\sim
A_0(x,y,\theta)+\frac{A_1(x,y,\theta)}{k}+\frac{A_2(x,y,\theta)}{k^2}+\cdots.$$ such that $$\label{SA}
Sa\sim k(x-y)\cdot SA \mod \mathcal{I}^\infty,$$ where $\mathcal{I}^\infty$ is the set of functions $f$ such that for any multi-index $\alpha$, $D^\alpha f=0$ when $x=y=\bar{z}$. Here $a_m(x,y,\theta)$ are almost holomorphic functions and $A_m(x,y,\theta)$ are almost holomorphic vector fields in $\mathbb C^n$, defined on $B^{n}(0,3) \times B^{n}(0,3) \times W$.
One particular $SA$ can be solved as follows. First note that by we must have $(SA)_0=0$ and $A_0=0$. Then we put $$\label{def SA}
SA(x,y,z)= - \frac{1}{k}\int_0^1 (D_{y}Sa)(x,tx+(1-t)y,z)dt.$$ By taking $S^{-1}$, $A$ can be solved uniquely as $$\label{A}
A(x,y,z)= - \frac{1}{k}S^{-1}\int_0^1 (D_{y}Sa)(x,tx+(1-t)y,z)dt.$$ Then by the fundamental theorem of calculus we have $$Sa(x,y,z)=k(x-y)\cdot SA(x,y,z)- ({\overline{x-y}})\cdot \int_0^1 (D_{\bar{y}}Sa)(x,tx+(1-t)y,z)dt.$$
By using the inverse operator $S^{-1}$,we have $$a(x,y,z)= D_\theta\cdot A + k(x-y)\cdot A - ({\overline{x-y}})\cdot S^{-1}\int_0^1 (D_{\bar{y}}Sa)(x,tx+(1-t)y,z)dt.$$ We use $a^{(N)}$ and $A^{(N)}$ to denote the partial sums of $a$ and $A$ up to order $\frac{1}{k^N}$ respectively. And we denote $$\nabla A:=D_\theta\cdot A+k(x-y)\cdot A.$$ Since $A_0=0$, we obtain $$\begin{aligned}
\label{a and nabla A}
\begin{split}
a^{(N)}-\nabla \left(A^{(N+1)}\right)
=&\frac{D_\theta\cdot A_{N+1}}{k^{N+1}}
-({\overline{x-y}})\cdot \left(S^{-1}\int_0^1 (D_{\bar{y}}Sa)(x,tx+(1-t)y,z)dt\right)^{(N)}.
\end{split}\end{aligned}$$
Next, we observe that the integral kernel of $\text{Op}_\Lambda(1+a)$ is almost holomorphic if $$\label{aB}
1+a(x,y,\theta)\sim B(x,z(x,y,\theta))\Delta_0(x,y,\theta),$$ where $$\Delta_0(x,y,\theta)=\frac{\det \psi_{yz}(y,z)}{\det\theta_z(x,y,z)},$$ and $B(x, z)$ is almost holomorphic and has an asymptotic expansion of the form $$\label{B} B(x, z) \sim b_0(x,z)+\frac{b_1(x,z)}{k}+\frac{b_2(x,z)}{k^2}+\cdots,$$ where $b_m(x, z)$ are almost holomorphic. In fact, as it turns out, $b_m(x, z)$ are an almost holomorphic extensions of $b_m(x, \bar x)$, the Bergman kernel coefficients of the on-diagonal asymptotic expansion of Zelditch-Catlin .
If the amplitude $a$ is negligible, then by applying $S( \cdot) |_{x=y}$ to both sides of , we get $$S \left( B(x,z(x,y,\theta))\Delta_0(x,y,\theta) \right )|_{x=y} \sim 1.$$ From this, one gets the following recursive equations for Bergman kernel coefficients $b_m(x, z)$, which will play a key role in the proof of Theorem \[MainLemma\]: $$\label{Recursive1}
b_m(x,z(x,x,\theta))=-\sum_{l=1}^m\frac{(D_y\cdot D_\theta)^l}{l!}\big(b_{m-l}\left(x,z(x,y,\theta)\right)\Delta_0(x,y,\theta)\big) \Big|_{y=x}.$$ Additionally, by comparing the coefficients on both sides of , we have the following relations between $a_m$ and $b_{m}$: $$\label{ambm}
a_m(x,y,\theta)=
\begin{cases}
\Delta_0(x,y,\theta)-1 & \mbox{ when }m=0,\\
b_m(x,z(x,y,\theta))\Delta_0(x,y,\theta) & \mbox{ when }m\geq 1.
\end{cases}$$ These equations will be useful in estimating $a_m$ in terms of the bounds on $b_m$ from Theorem \[MainLemma\].
Almost Holomorphic Extensions of Gevrey functions {#AHEoGF}
-------------------------------------------------
In this section, we will review the Gevrey class and consider almost holomorphic extensions of functions in such a class. Indeed, there are many different ways to construct almost holomorphic functions. We will adapt the way in [@Ju1] to construct a particular one which is suitable for our analysis. Afterwards, various properties of such an extension are introduced, which will be used for the proof of Proposition \[LocalBergmanKernel\] in Section \[Sec MethodBBS\]. Although all the properties are natural and elementary, the proofs are however very lengthy. For the convenience of the readers, we shall only state the results we need and postpone the proofs to Section \[ProofOfGevreyStuff\].
We recall the definition of Gevrey class $G^a(U)$. For more details, we refer the readers to [@Ge]. Take $\alpha,\beta\in (\mathbb{Z}^{\geq 0})^n$. Here are some standard notations of multi-indices we shall use in the following.
- $|\alpha|=\alpha_1+\alpha_2+\cdots+\alpha_n$.
- $ \alpha \leq \beta $ if $\alpha_1\leq \beta_1, \alpha_2\leq \beta_2,\cdots,\alpha_n\leq \beta_n$.
- $\alpha< \beta$ if $ \alpha \leq \beta$ and $\alpha \ne \beta$.
- $ \alpha !=\alpha_1!\alpha_2!\cdots \alpha_n!$.
\[Gevrey\] Let $a\in (1,\infty)$ and $U$ be an open subset of $\mathbb C^n$. We denote by $G^{a}(U)$ the set of functions $f(x)\in C^\infty(U,\mathbb{C})$ such that there exists some constant $C_0=C_0(f)>0$ and $C_1=C_1(f)>0$, satisfying $$\|D_x^\alpha D_{\bar{x}}^\beta f\|_{L^\infty(U)} \leq C_0 C_1^{|\alpha|+|\beta|}(\alpha!\beta!)^a,$$ for any multi-indices $\alpha,\beta\geq 0$. The space $G^a(U)$ is called the Gevrey class of index $a$. Note that each class $G^a(U)$ forms an algebra which is closed under differentiation and integration.
For any $f\in G^a(U)$, an almost holomorphic extension $F(f)(x,z)$ is a smooth function on $U\times U$ such that $F(f)(x,\bar{x})=f(x)$ and the anti-holomorphic derivatives have infinite vanishing order along $x=\bar{z}$. We will use the way in [@Ju1] to construct a particular almost holomorphic extension. In fact the construction of [@Ju1] is adapted from Borel’s method (see also Hörmander[@HoAH] ). Here, we follow [@Ju1] but we use a cut-off function $\chi$ in the Gevrey class $\in G^{1+{\varepsilon}}(\mathbb{R})$ where ${\varepsilon}$ is an arbitrary positive constant, and $$\label{Gevreycutoff}
\chi(x)=
\begin{cases}
1 & |x|\leq \frac{1}{2},\\
0 & |x|\geq 1.
\end{cases}$$ To show the existence of such a cut-off function, one can use the fact that for any ${\varepsilon}>0$, the function defined as $$f_{\varepsilon}(x)=
\begin{cases}
\exp(-x^{-\frac{1}{{\varepsilon}}}) & x>0\\
0 & x\leq 0,
\end{cases}$$ belongs to $G^{1+{\varepsilon}}(\mathbb{R})$ (See [@CC] for more details). Then by the standard construction, we define $$g(x)=\begin{cases}
0 & x\leq 0,\\
\frac{\int_0^x f_{\varepsilon}(t)f_{\varepsilon}(\frac{1}{2}-t)dt}{\int_0^1 f_{\varepsilon}(t)f_{\varepsilon}(\frac{1}{2}-t)dt} & x\in (0,1),\\
1 & x\geq 1.
\end{cases}$$ We can take our cut-off function to be $\chi(x)=g(x+1)g(-x+1)$.
We now define our almost holomorphic extension of Gevrey functions.
\[AHE\] Let $a\in (1,\infty)$, $U$ be the unit ball $B(0,1)$ in $\mathbb C^n$, and $f(x)\in G^{a}(U)$. Let $C_1=C_1(f)$ be the constant in Definition \[Gevrey\]. Then for $(y,z)\in U\times U$, we define an almost holomorphic extension $$F(f)(y,z)=\sum_{\alpha,\beta\geq 0} \frac{D_x^\alpha D_{\bar{x}}^\beta f}{\alpha!\beta!}\left(\frac{y+\bar{z}}{2}\right)\left(\frac{y-\bar{z}}{2}\right)^\alpha \left(\frac{z-\bar{y}}{2}\right)^\beta\chi\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right).$$
We will justify that $F(f)$ defined as above is genuinely an almost holomorphic extension of $f$ along $y=\bar{z}$. It is easy to see $F(f)(x,\bar{x})=f(x)$. And in the next lemma, we will verify that $D_{\bar{y}}F(f)(y,z)=O(|y-\bar{z}|^\infty)$, and $D_{\bar{z}}F(f)(y,z)=O(|y-\bar{z}|^\infty)$. To be more precise, we show that these quantities vanish at a certain exponential rate along $y=\bar{z}$.
\[lem almostholomorphic\] There exist positive constants $C$ and $b$ such that for any $y,z \in U$, the almost holomorphic extension $F(f)$ satisfies $$\begin{aligned}
\begin{split}
|D_{\bar{y}}F(f)(y,z)|\leq C \exp{\left(-b|y-\bar{z}|^{-\frac{1}{a-1}}\right)},\\
|D_{\bar{z}}F(f)(y,z)|\leq C \exp{\left(-b|y-\bar{z}|^{-\frac{1}{a-1}}\right)}.
\end{split}
\end{aligned}$$ In particular, $F(f)$ is almost holomorphic along $y=\bar{z}$.
Indeed, there are various ways to define an almost holomorphic extension besides Definition \[AHE\]. But they are all the same up to an $O(|y-\bar{z}|^\infty)$ error term.
\[AHE Uniqueness\] Let $U$ be the unit ball $B(0,1)$ in $\mathbb C^n$ and $f(x)\in C^\infty({\overline{U}})$. If $F(y,z), \widetilde{F}(y,z)\in C^\infty(U\times U)$ are both almost holomorphic extensions of $f$, then $$F(y,z)-\widetilde{F}(y,z)=O\left(|y-\bar{z}|^\infty\right).$$
Next, we show a more general version of Lemma \[lem almostholomorphic\], which gives estimates on all the derivatives of $F(f)$. It turns out that if $f\in G^a(U)$, then $F(f)\in G^{a+{\varepsilon}}(U\times U)$ and when the anti-holomorphic derivative appears, it always vanishes to infinite order along $y=\bar{z}$ at a certain exponential rate.
\[AHE2\] Take $f\in G^a(U)$. Let $C_0(f)$ and $C_1(f)$ be the constants satisfying for $f$. Then for any ${\varepsilon}>0$, there exist positive constants $C_1=C_1({\varepsilon}, a, C_1(f))$, $b=b(a, C_1(f))$ and $A=A(a,n)$ such that for any multi-indices $\gamma,\delta,\xi,\eta\geq 0$, we have $$|D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(f)(y,z)|
\leq AC_0(f)C_1^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}},$$ hence $F(f)(y,z)\in G^{a,{\varepsilon}}(U\times U)$. Moreover, if $\xi+\eta>0$, then $$| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(f)(y,z)|
\leq AC_0(f)C_1^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}}
\, \exp\left ( {- b {|y-\bar{z}|^{-\frac{1}{a-1}}}} \right ).$$ In addition, when we are restricted to the diagonal $z=\bar{y}$, we can let ${\varepsilon}=0$ in the above estimates.
This motivates us to give the following definition.
Let $U$ be an open neighborhood of the origin in $\mathbb{C}^{2n}$ and let ${\varepsilon}>0$ be a constant. A function $F(y, z) \in C^\infty(U)$ is called $G^{a,{\varepsilon}}$-*almost holomorphic* along the diagonal $z=\bar{y}$ if there exist positive constants $C_0=C_0(F), C_1=C_1(F)$ and $b=b(F)$ such that for any multi-indices $\gamma,\delta,\xi,\eta\geq 0$, we have $$| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(y,z)|
\leq C_0 C_1^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}}.$$ And when $\xi+\eta>0$, we have $$\label{AHG}
| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(y,z)|
\leq C_0 C_1^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}}
\, \exp\left ( {- b {|y-\bar{z}|^{-\frac{1}{a-1}}}} \right ).$$ In addition, when we are restricted to the diagonal $z=\bar{y}$, we can let ${\varepsilon}=0$ in the above estimates.
We use $\mathcal A_{diag}^{a,{\varepsilon}}(U)$ for the class of such functions. And we also use $\mathcal I_{diag}^{a,{\varepsilon}}(U)$ for functions $F(y, z) \in C^\infty(U \times U)$ satisfying with no restrictions on $\xi$ and $\eta$ (i.e. holds even if $\xi=\eta=0$). And we say a vector belongs to $\mathcal A_{diag}^{a,{\varepsilon}}(U)$ or $\mathcal I_{diag}^{a,{\varepsilon}}(U)$ if each component function belongs to that class. Obviously, we have ${{\overline{\partial}}}\mathcal A_{diag}^{a,{\varepsilon}}(U)\subset \mathcal I_{diag}^{a,{\varepsilon}}(U)$.
Since the recursive formula for Bergman coefficients requires studying functions of three variables in $\mathbb C^n$, we also present the following definition for functions in $\mathbb C^{3n}$.
\[A\_theta\] Let ${\varepsilon}>0$ be a constant. Let $\theta(x, y, z)$ be a function on $U$ such that $ \Phi: (x, y, z) \to (x, y, \theta(x, y, z))$ is a diffeomorphism between $U$ and its image denoted by an open set $U'\subset \mathbb{C}^{3n}$. Take $f(x,y,\theta)\in C^\infty(U')$. Denote $v'=(x,y, \theta)$. We say $f(x,y,\theta)$ is *$G^{a,{\varepsilon}}$-almost holomorphic* along $x=y=\bar{z}$ under $\theta$, if there exist some positive constants $C_0=C_0(f), C_1=C_1(f)$, and $b=b(f)$, such that for any multi-indices $\alpha,\beta\geq 0$, we have $$\label{eq 4.13}
\left | \left ( D^\alpha_{v'}D^{\beta}_{\bar{v}'}f \right )(x, y, \theta(x, y, z)) \right |\leq C_0C_1^{|\alpha+\beta|}(\alpha!\beta!)^{a+{\varepsilon}} \, \exp\left(- b \, (1-\delta_0(|\beta|)) {\max \{ |x- \bar z|, |y-\bar{z} | \} ^{-\frac{1}{a-1}}} \right) ,$$ where $\delta_0(\cdot)$ is the delta function whose value is $1$ at $0$ and it is zero elsewhere. In addition, when we are restricted to $x=y=\bar{z}$, we can let ${\varepsilon}=0$ in the above estimate.
We use $\mathcal{A}^{a,{\varepsilon}}_{\theta}(U')$ to denote the set of all $G^{a,{\varepsilon}}$-almost holomorphic functions along $x=y=\bar{z}$ under $\theta$ in the above sense. We will also use $\mathcal I_{\theta}^{a,{\varepsilon}}(U)$ for smooth functions $f(x,y,\theta)$ such that for any multi-indices $\alpha,\beta\geq 0$, $$\left | \left ( D^\alpha_{v'}D^{\beta}_{\bar{v}'}f \right )(x, y, \theta(x, y, z) \right |\leq C_0C_1^{|\alpha+\beta|}(\alpha!\beta!)^{a+{\varepsilon}} \, \exp\left(- b \, {\max \{ |x- \bar z|, |y-\bar{z} | \} ^{-\frac{1}{a-1}}} \right) ,$$ And we say a vector belongs to $\mathcal A_{\theta}^{a,{\varepsilon}}(U)$ or $\mathcal I_{\theta}^{a,{\varepsilon}}(U)$ if each component function belongs to that class.
In the following, for simplicity we will use the notation $$\begin{aligned}
\label{Lambda}
\lambda_{b, |\beta|}(x, y, z)= \exp\left(- b \, (1-\delta_0(|\beta|)) {\max \{ |x- \bar z |, |y-\bar{z} | \} ^{-\frac{1}{a-1}}} \right)\end{aligned}$$
Note that by the above notation, $\mathcal{A}^{a,{\varepsilon}}_{z}$ (or $\mathcal{I}^{a,{\varepsilon}}_{z}$) means $\mathcal{A}^{a,{\varepsilon}}_{\theta}$ (or $\mathcal{I}^{a,{\varepsilon}}_{\theta}$) when $\theta(x, y, z)=z$, which corresponds to the case $\Phi =I$.
The space $\mathcal{A}^{a,{\varepsilon}}_\theta(U)$ is closed under algebraic operations and differentiations.
\[AHE CLOSED\] For each $\theta$ as described in the previous definition, $\mathcal{A}^{a,{\varepsilon}}_\theta(U)$ is closed under summation, subtraction, multiplication and differentiation. It is also closed under division if the denominator is uniformly away from zero in $U$.
In particular, suppose $f,g\in \mathcal{A}_\theta^{a,{\varepsilon}}(U)$. Then we can choose the constants appearing in for the product $fg\in \mathcal{A}_\theta^{a,{\varepsilon}}(U)$ as $$\begin{aligned}
C_0(fg)=C_0(f)C_0(g),\quad C_1(fg)=2\max\{C_1(f),C_1(g)\}, \quad\mbox{and}\quad b(fg)=\min\{b(f),b(g)\}.
\end{aligned}$$ And for the differentiation, we can choose the constants as $$\begin{aligned}
C_0(D_{v'}^\alpha D_{\bar{v}'}^\beta f)=C_0(f)(2^aC_1(f))^{|\alpha+\beta|},\quad
C_1(D_{v'}^\alpha D_{\bar{v}'}^\beta f)=2^aC_1(f),\quad
\mbox{and}\quad b(D_{v'}^\alpha D_{\bar{v}'}^\beta f)=b(f),
\end{aligned}$$ where $v'=(x,y,\theta)$.
We shall use the following lemma that $\mathcal{A}^{a,{\varepsilon}}_{z}$ is closed under certain integrals.
\[AHE Integral\] If $f(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$, then $g(x,y,z)=\int_0^1 f(x,tx+(1-t)y,z)dt \in \mathcal{A}^{a,{\varepsilon}}_z$. And we can choose the constants appearing in Definition \[A\_theta\] as $$\begin{aligned}
C_0(g)=C_0(f), \quad C_1(g)=2^{a+{\varepsilon}+1}C_1(f), \quad b(g)=b(f).
\end{aligned}$$ Similarly, If $f(x,y,z)\in \mathcal{I}^{a,{\varepsilon}}_z$, then $g(x,y,z)=\int_0^1 f(x,tx+(1-t)y,z)dt \in \mathcal{I}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0(g)=C_0(f),\quad C_1(g)=2^{a+{\varepsilon}+1}C_1(f), \quad b(g)=b(f).
\end{aligned}$$
The space $\mathcal{A}^{a,{\varepsilon}}_{\theta}$ is also closed under composition in the following sense.
\[AHE Composition\] Let $f(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z(U)$ be a function defined on $U\subset \mathbb{C}^{3n}$. Let $\theta(x,y,z)$ be a map on $U$ such that $\Phi: (x,y,z)\rightarrow (x,y,\theta(x,y,z))\in $ is a diffeomorphism between $U$ and its image denoted by an open set $U'\subset \mathbb{C}^{3n}$. Let $\Phi^{-1}: (x,y, \theta)\rightarrow (x,y, z(x,y,\theta))$ be the inverse map of $\Phi$. If $z=z(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta (U')$, then the composition function $\widetilde{f}(x,y,\theta)=f(x,y,z(x,y,\theta))\in \mathcal{A}^{a,{\varepsilon}}_\theta(U')$.
In particular, if we use $C_0(f),C_1(f)$ and $b(f)$ to denote the constants in for an the function $f$, then we can choose the constants for $\widetilde{f}$ as $$\begin{aligned}
C_0(\widetilde{f})=C_0(f),\quad C_1(\widetilde{f})=2^{a+{\varepsilon}+3m} m^{a+{\varepsilon}-1}C_0(z(x,y,\theta))C_1(f)C(z(x,y,\theta)),
\end{aligned}$$ and $$\begin{aligned}
b(\widetilde{f})=\min\{b(f),b(z(x,y,\theta))\},
\end{aligned}$$ where $m=3n$, $C_0(z(x,y,\theta))=\max_{1\leq i\leq n}C_0(z_i(x,y,\theta))$, $C_1(z(x,y,\theta))=\max_{1\leq i\leq n}C_1(z_i(x,y,\theta))$, and $b(z(x,y,\theta))=\min_{1\leq i\leq n}b(z_i(x,y,\theta))$.
\[I Composition\] In Lemma \[AHE Composition\], if we further assume that $f(x,y,z)\in \mathcal{I}_z^{a,{\varepsilon}}$, then the composition $\widetilde{f}$ belongs to $\mathcal{I}_\theta^{a,{\varepsilon}}$ with the same choice of constants.
Now suppose $U=B(0,1)\subset \mathbb{C}^n$ and the potential $\phi$ belongs to $G^a(U)$ and let $\psi=F(\phi)$ be the almost holomorphic extension of $\phi$ defined by . Then it is easy to see that $\psi(y,z)\in \mathcal{A}^{a,{\varepsilon}}_{z}(U)$. Further by using Lemma \[AHE Integral\], if we take $\theta(x,y,z)=\int_0^1(D_y\psi)(tx+(1-t)y,z)dt$, then $\theta\in \mathcal{A}^{a,{\varepsilon}}_{z}$. The following lemma says that the implicit functions $z=z(x,y,\theta)$ belong to $\mathcal{A}^{a,{\varepsilon}}_\theta$.
\[AHE Inverse\] Consider the following system of equations: $$\begin{aligned}
\label{eq 4.32}
\theta=\int_0^1(D_y\psi)(tx+(1-t)y,z)dt\, .
\end{aligned}$$ Then the implicit functions $z=z(x,y,\theta)$ determined by the above equations belong to $\mathcal{A}^{a,{\varepsilon}}_\theta$.
As we said at the beginning of this section, the proofs of all the above lemmas will be given in Section \[ProofOfGevreyStuff\].
We are now prepared to prove Proposition \[LocalBergmanKernel\].
The remainder estimates and the proof of Proposition \[LocalBergmanKernel\] {#Sec MethodBBS}
===========================================================================
Let $a_m$, $A_m$, and $b_m$ be given by , , and . Remember that $a^{(N)}$, $A^{(N)}$, and $B^{(N)}$ are the partial sums of $a$, $A$, and $B$ up to order ${k^{-N}}$. When we apply the method of Berman-Berndtsson-Sjöstrand, the remainder term is closely related to the growth rate of $a_m$, $A_m$ and their derivatives as we will see soon. So we will first make a series of lemmas on estimating $a_m$ and $A_m$ preparing for the proof of Proposition \[LocalBergmanKernel\].
Let’s begin with estimating $a_m$.
\[amxyz\] For each integer $m\geq 0$, we have $a_m(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. And we can choose the constants appearing in Definition \[A\_theta\] as $$\begin{aligned}
C_0(a_m)= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(a_m)=C, &&b(a_m)=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Recall the relations between $a_m$ and $b_m$ from : $$a_m(x,y,z)=
\begin{cases}
\Delta_0(x,y,z)-1 & \mbox{ when }m=0,\\
b_m(x,z)\Delta_0(x,y,z) & \mbox{ when }m\geq 1.
\end{cases}$$ Since $\phi \in G^a$ , the almost holomorphic extension $\psi(y,z)$ introduced as in definition \[AHE\] belongs to $\mathcal A^{a,{\varepsilon}}_{z}$. Recall that by Lemma \[AHE Integral\], $\theta(x,y,z)=\int_0^1 (D_y\psi)(tx+(1-t)y,z)dt\in \mathcal A^{a,{\varepsilon}}_{z}$. By Lemma \[AHE CLOSED\], we know $\mathcal{A}^{a,{\varepsilon}}_{z}$ is closed under certain algebraic operations and differentiation, $\Delta_0(x,y,z)=\frac{\det \psi_{yz}(y,z)}{\det \theta_z(x,y,z)}$ is therefore also contained in $\mathcal{A}^{a,{\varepsilon}}_{z}$. Since $a_m(x,y,z)=b_m(x,z)\Delta_0(x,y,z)$ for $m\geq 1$, by our Lemma \[AHE CLOSED\] on the multiplication, we can choose $$C_0(a_m)=C_0(b_m)C_0(\Delta_0)=C^{m+1}m!^{2a+2{\varepsilon}},$$ $$C(a_m)=2\max\{C_1(b_m),C_1(\Delta_0)\},$$ and $$b(a_m)=\min\{b(b_m),b(\Delta_0)\}.$$ Thus, the result follows as $C_0(b_m)=C^mm!^{2a+2{\varepsilon}}$ for some positive constant $C$ and $C_1(b_m), b(b_m)$ are both independent of $m$ by Theorem \[MainLemma\]. In addition, it is easy to see that when we are restricted to $x=y=\bar{z}$, ${\varepsilon}$ can be replaced by $0$.
\[amxytheta\] Denote $\widetilde{a}_m=a_m(x,y,\theta)=a_m(x,y,z(x,y,\theta))$. Then $a_m(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$ and we can choose $$\begin{aligned}
C_0(\widetilde{a}_m)= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(\widetilde{a}_m)=C, &&b(\widetilde{a}_m)=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
By Lemma \[AHE Inverse\], we have $z=z(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$. Since $\widetilde{a}_m$ is obtained from the composition of $a_m(x,y,z)$ and the map $z=z(x,y,\theta)$, by Lemma \[AHE Composition\] $$C_0(\widetilde{a}_m)=C_0(a_m),$$ $$C_1(\widetilde{a}_m)=2^{a+{\varepsilon}+9n} (3n)^{a-1+{\varepsilon}}C_0(z(x,y,\theta))C_1(a_m)C_1(z(x,y,\theta)),$$ $$b(\widetilde{a}_m)=\min\{b(a_m),b(z(x,y,\theta))\}.$$ So the result follows directly from Lemma \[amxyz\].
After we obtain the estimates on $a_m(x,y,z)$ and $a_m(x,y,\theta)$, now we proceed to $(Sa)_m(x,y,z)$ and $(Sa)_m(x,y,\theta)$.
\[Sam\] For each integer $m\geq 0$, $(Sa)_m(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$ and $(Sa)_m(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0((Sa)_m(x,y,\theta))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((Sa)_m(x,y,\theta))=C, &&b((Sa)_m(x,y,\theta))=b,
\end{aligned}$$ $$\begin{aligned}
C_0((Sa)_m(x,y,z))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((Sa)_m(x,y,z))=C, &&b((Sa)_m(x,y,z))=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Since $$Sa=\sum_{i=0}^\infty\frac{(D_\theta\cdot D_y)^i}{i!k^i}\sum_{j=0}^\infty \frac{a_j}{k^j}
=\sum_{m=0}^\infty\sum_{i+j=m}\frac{(D_\theta\cdot D_y)^ia_j}{i!k^m},$$ we have $$(Sa)_m(x,y,\theta)=\sum_{i+j=m}\frac{(D_\theta\cdot D_y)^ia_j}{i!}(x,y,\theta)=\sum_{i+j=m}\sum_{|\delta|=l}\frac{1}{i!}\binom{l}{\delta}D^{\delta}_y D_\theta^\delta a_j(x,y,\theta).$$ Denote $v'=(x,y,\theta)$. Therefore, $$\begin{aligned}
D_{v'}^{\alpha}D_{\bar{v}'}^{\beta}(Sa)_m(x,y,\theta)
=\sum_{i+j=m}\sum_{|\delta|=i}\frac{1}{i!}\binom{i}{\delta}D_{v'}^{\alpha+(0,\delta,\delta)}D_{\bar{v}'}^{\beta} a_j(x,y,\theta)
\end{aligned}$$ By Lemma \[amxytheta\], we have $$\begin{aligned}
&\left|D_{v'}^{\alpha}D_{\bar{v}'}^{\beta}(Sa)_m(x,y,\theta)\right|
\\
&\qquad
\leq \sum_{i+j=m}\sum_{|\delta|=i}\frac{1}{i!}\binom{i}{\delta}C^{|\alpha+\beta|+2|\delta|+j+1} j!^{2a+2{\varepsilon}}(\alpha+(0,\delta,\delta))!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)
\\
&\qquad
\leq \sum_{i+j=m}\frac{n^i}{i!}2^{(a+{\varepsilon})(|\alpha|+2i)}C^{|\alpha+\beta|+2i+j+1} j!^{2a+2{\varepsilon}}i!^{2a+2{\varepsilon}}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)
\\
&\qquad\leq C^{m+1+|\alpha+\beta|} \alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}m!^{2a+2{\varepsilon}}\lambda_{b,|\beta|}(x,y,z),
\end{aligned}$$ where $C$ is a constant independent to $m$, which may vary from line to line, and $b=b(a_m)$ is also independent to $m$. So we obtain the result on $(Sa)_m(x,y,\theta)$. Note $\theta(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. The result on $(Sa)_m(x,y,z)$ follows by Lemma \[AHE Composition\] and keeping track of these constants.
Next, we will estimate the growth of $(SA)_m(x,y,z)$ and $(SA)_m(x,y,\theta)$.
\[SAm\] For each integer $m\geq 0$, $(SA)_m(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$ and $(SA)_m(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0((SA)_m(x,y,\theta))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((SA)_m(x,y,\theta))=C, &&b((SA)_m(x,y,\theta))=b,
\end{aligned}$$ $$\begin{aligned}
C_0((SA)_m(x,y,z))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((SA)_m(x,y,z))=C, &&b((SA)_m(x,y,z))=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Recall $(SA)_0(x,y,z)=0$. By , for $m\geq 1$, we have $$\begin{aligned}
(SA)_m(x,y,z)=-\int_0^1 D_y(Sa)_{m-1}(x,tx+(1-t)y,z)dt.
\end{aligned}$$ By Lemma \[Sam\], we have $D_y(Sa)_{m-1}(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. Then by Lemma \[AHE Integral\] and Lemma \[AHE Composition\], we have $(SA)_m(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$ and $(SA)_m(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$. The remaining part follows straightforwardly by keeping track of the constants.
We define $$d_m(x,y,z)=\int_0^1 D_{\bar{y}}(Sa)_{m}(x,tx+(1-t)y,z)dt,$$ and $$d_m(x,y,\theta)=d_m(x,y,z(x,y,\theta)).$$
Since $(Sa)_m\in \mathcal{A}_z^{a,{\varepsilon}}$, we have $D_{\bar{y}}(Sa)_m\in \mathcal{I}_z^{a,{\varepsilon}}$. By using Lemma \[AHE Integral\] and Remark \[I Composition\], we can also similarly prove the following estimates on $d_m(x,y,z)$ and $d_m(x,y,\theta)$.
\[dm\] For each integer $m\geq 0$, $d_m(x,y,\theta)\in \mathcal{I}^{a,{\varepsilon}}_\theta$ and $d_m(x,y,z)\in \mathcal{I}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0(d_m(x,y,\theta))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(d_m(x,y,\theta))=C, &&b(d_m(x,y,\theta))=b,
\end{aligned}$$ $$\begin{aligned}
C_0(d_m(x,y,z))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(d_m(x,y,z))=C, &&b(d_m(x,y,z))=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Now that we have the estimates on $(SA)_m$ in Lemma \[SAm\], by using the operator $S^{-1}$, we obtain the following estimates on $A_m$.
\[Am\] For each integer $m\geq 0$, $A_m(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$ and $A_m(x,y,z)\in \mathcal{A}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0(A_m(x,y,\theta))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(A_m(x,y,\theta))=C, &&b(A_m(x,y,\theta))=b,
\end{aligned}$$ $$\begin{aligned}
C_0(A_m(x,y,z))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1(A_m(x,y,z))=C, && b(A_m(x,y,z))=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Formally, we write $d=\sum_{m=0}^\infty \frac{d_m}{k^m}$. Similarly, by using Lemma \[dm\] and the operator $S^{-1}$, we obtain the estimates on $(S^{-1}d)_m$.
\[S-1dm\] For each integer $m\geq 0$, $(S^{-1}d)_m(x,y,\theta)\in \mathcal{I}^{a,{\varepsilon}}_\theta$ and $(S^{-1}d)_m(x,y,z)\in \mathcal{I}^{a,{\varepsilon}}_z$. And we can choose $$\begin{aligned}
C_0((S^{-1}d)_m(x,y,\theta))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((S^{-1}d)_m(x,y,\theta))=C, &&b((S^{-1}d)_m(x,y,\theta))=b,
\end{aligned}$$ $$\begin{aligned}
C_0((S^{-1}d)_m(x,y,z))= C^{m+1}m!^{2a+2{\varepsilon}}, && C_1((S^{-1}d)_m(x,y,z))=C, && b((S^{-1}d)_m(x,y,z))=b,
\end{aligned}$$ where $C$ and $b$ are some positive constants independent of $m$.
Since the proof of Lemma \[Am\] and \[S-1dm\] follow in the same way as that of Lemma \[Sam\], we omit them here.
We are now ready to estimate $A_N$, $A^{(N)}$ and $D_\theta \cdot A_N$ on the good contour $\Lambda=\{(y,\theta(x,y,\bar{y})): x,y\in B^n(0,1)\}$. For any smooth function $f(x,y,\theta)$, we denote $$\|f(x,y,\theta)\|_{L^\infty(\Lambda)}:=\|f(x,y,\theta(x,y,\bar{y}))\|_{L^\infty (B^{n}(0,1) \times B^{n}(0,1))}.$$
\[error term 1 for Gevrey\] We have $$\|A^{(N)}(x,y,\theta(x, y, \bar y))\|_{L^\infty(\Lambda)}\leq C k + \frac{C^{N} N!^{2a+2{\varepsilon}}}{k^N},$$ where $C$ is some constant independent of $N$ and $k$.
Note that by the definition of $A^{(N)}$ and estimates on each $A_m$ in Lemma \[Am\], we have $$\begin{aligned}
\|A^{(N)}\|_{L^\infty (\Lambda )} & \leq \frac{C1!^{2a+2{\varepsilon}}}{k}+\frac{C^2 2! ^{2a+2{\varepsilon}}}{k^2}+\cdots+\frac{C^N N! ^{2a+2{\varepsilon}}}{k^N}.\\
\end{aligned}$$ We need to study the function $ \frac{C^xx^{(2a+2{\varepsilon})x}}{e^{(2a+2{\varepsilon})x}k^x}$. To find the minimum of this function we consider $$f(x)=\log \frac{C^xx^{(2a+2{\varepsilon})x}}{e^{(2a+2{\varepsilon})x}k^x}=x\log C+(2a+2{\varepsilon})x\log x-(2a+2{\varepsilon})x-x\log k \hspace{12 pt}, \mbox{ for } x\in (0,\infty).$$ Since $$f'(x)=\log C+(2a+2{\varepsilon})\log x-\log k,$$ the only critical point of $f$ is $x_0=(\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}$, and the function $f$ is decreasing on the interval $(0, x_0]$ and increasing on the interval $[x_0, \infty)$. Hence if we take $N_0=[(\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}]$, then by using Stirling’s formula twice $$\begin{aligned}
\sum_{m=1}^N \frac{C^m m! ^{2a+2{\varepsilon}}}{k^m}
& \leq C' \sum_{m=1}^N m^{a+{\varepsilon}}\frac{C^m m^{m(2a+2{\varepsilon})}}{e^{m(2a+2{\varepsilon})}k^m}
\\
& \leq C' \left ( N_0^{a+{\varepsilon}+1} + N^{a+{\varepsilon}}(N-N_0)\frac{C^N N^{(2a+2{\varepsilon})N}}{e^{(2a+2{\varepsilon})N}k^N} \right )
\\
& \leq C' \left (\left(\frac{k}{C}\right)^{\frac{a+{\varepsilon}+1}{2a+2{\varepsilon}}} + N^{a+{\varepsilon}+1}\frac{C^N N^{(2a+2{\varepsilon})N}}{e^{(2a+2{\varepsilon})N}k^N} \right )
\\
& \leq C' \left (k + \frac{C''^N N!^{2a+2{\varepsilon}}}{k^N} \right ).
\end{aligned}$$ The result follows by replacing $C'$ and $C''$ by a larger constant $C$.
We also need the estimates on the anti-holomorphic derivatives of $A_N$ and $A^{(N)}$.
\[error term 4 for Gevrey\] There exists positive constants $C$ and $b$ independent of $N$ and $k$ such that for any $(x,y) \in B^{n}(0,1)\times B^{n}(0,1)$, we have $$\left | \left ( D_{\bar{\theta}}\cdot A^{(N)} \right ) (x,y,\theta(x, y, \bar y) \right | \leq \left(C k + \frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}\right)\exp\left(-b{|x-y|^{-\frac{1}{a-1}}}\right),$$ $$\left | \left ( D_{\bar{y}}\cdot A^{(N)} \right ) (x,y,\theta(x, y, \bar y) \right | \leq \left(C k + \frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}\right)\exp\left(-b{|x-y|^{-\frac{1}{a-1}}}\right).$$
We omit the proof of this lemma here since it follows in a similar way as the previous lemma by using Lemma \[Am\] and the only difference is the extra exponential factor that comes from the anti-holomorphic derivatives of $A_m$ since $A_m(x,y,\theta)\in \mathcal{A}_\theta^{a,{\varepsilon}}$.
Another key lemma is:
\[error term 3 for Gevrey\] There exists positive constants $C$ and $b$ independent of $N$ and $k$, such that for any $(x,y) \in B^{n}(0,1)\times B^{n}(0,1)$, we have $$\left|\left(S^{-1}\int_0^1 (D_{\bar{y}}Sa)(x,tx+(1-t)y,z)dt\right)_m (x,y,\bar y) \right|
\leq C^{m+1}m!^{2a+2{\varepsilon}}\exp\left(-b{|x-y|^{-\frac{1}{a-1}}}\right),$$ and $$\begin{aligned}
\left|\left(S^{-1}\int_0^1 (D_{\bar{y}}Sa)(x,tx+(1-t)y,z)dt\right)^{(N)} (x, y,\bar{y})\right|
\leq
\left(C k + \frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}\right)\exp\left(-b{|x-y|^{-\frac{1}{a-1}}}\right).
\end{aligned}$$
The first inequality directly follows from Lemma \[S-1dm\] and the second inequality follows by the same argument as in the proof of Lemma \[error term 1 for Gevrey\].
Recalling and using Lemma \[Am\] and \[error term 3 for Gevrey\] together, we obtain the following corollary.
\[prop 7.6\] There exists positive constants $C$ and $b$ independent of $N$, such that for any $(x,y) \in B^{n}(0,1)\times B^{n}(0,1)$, we have $$\begin{aligned}
\left|a^{(N)}-\nabla \left(A^{(N+1)}\right)\right|(x,y,\bar{y})
\leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}
+|x-y|\ \left(C k + \frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}\right)\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right).
\end{aligned}$$
Now we are ready to prove Proposition \[LocalBergmanKernel\].
We claim that $$\label{LocalBergmanKernel2}
u(x)=\text{Op}_\Lambda \left (1+a^{(N)} \right ) ( \chi u)
+k^n\mathcal R_{N+1}( \phi, k) e^{\frac{k\phi(x)}{2}}\|u\|_{k\phi},$$ where uniformly for $x \in B(0, \frac14 )$ we have $$\label{R}
|\mathcal R_{N+1}( \phi, k)| \leq\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}},$$ and the integral kernel of $\text{Op}_\Lambda \left (1+a^{(N)} \right )$ is almost holomorphic. The complex conjugate of this kernel is given by $$\label{KandB}
{\overline{K^{(N)}_{k,x}(y)}}=\left(\frac{k}{\pi}\right)^n e^{k\psi(x,\bar{y})}\left(1+ a^{(N)}(x, y, \theta(x, y, \bar y)\right) \Delta_0(x, y, \theta(x, y, \bar y))^{-1},$$ which by the relation is reduced to $${\overline{K^{(N)}_{k,x}(y)}}=\left(\frac{k}{\pi}\right)^n e^{k\psi(x,\bar{y})} B^{(N)}(x, \bar y).$$ Hence ${K^{(N)}_{k,x}(y)}$ is almost holomorphic in $y$ because $B(x, z)$ is almost holomorphic.
In the light of , to prove it suffices to show that $$\forall x \in B(0, \frac14): \quad \left | \text{Op}_\Lambda \left (a^{(N)} \right )( \chi u)(x) \right | \leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1-n}} e^{\frac{k\phi(x)}{2}}\|u\|_{k\phi}.$$ By definition, $$\text{Op}_\Lambda \left (a^{(N)} \right )( \chi u)(x)= c_n \left(\frac{k}{2\pi}\right)^n\int_\Lambda e^{k\theta\cdot (x-y)}u(y)\chi(y)\, a^{(N)} \, d\theta\wedge dy .$$ It is easy to see that using integration by parts (see for example the proof of Proposition 2.2 in [@BBS]), we get $$\begin{aligned}
\label{a^(N)}
\int_\Lambda &e^{k\theta\cdot (x-y)}u(y)\chi(y)a^{(N)}d\theta\wedge dy=
\\
&-\int_\Lambda d\chi\wedge u(y)e^{k\theta\cdot(x-y)}A^{(N+1)}\wedge dy
+\int_\Lambda e^{k\theta\cdot (x-y)}u(y)\chi(y)\left(a^{(N)}-\nabla \left(A^{(N+1)}\right) \right)d\theta\wedge dy
\\
&-\sum_{i,j}\int_{\Lambda} e^{k\theta\cdot (x-y)}u(y)\chi(y)\frac{\partial A_i}{\partial {\overline{\theta}}_j}d{\overline{\theta}}_j\wedge\widehat{d\theta_i}\wedge dy
-\sum_{i,j}\int_{\Lambda} e^{k\theta\cdot (x-y)}u(y)\chi(y)\frac{\partial A_i}{\partial {\overline{y}}_j}d{\overline{y}}_j\wedge\widehat{d\theta_i}\wedge dy\end{aligned}$$ In the first integral, we have identified the $n$-vector $A$ as an $(n-1, 0)$ form defined by $ A= \sum_{j=1}^n A_j \widehat{d\theta_j}$, where $\widehat{d\theta_j}$ is the wedge product of all $\{d \theta_k\}_{k \neq j}$ such that $d\theta_j \wedge \widehat{d\theta_j} = d\theta$.
We now estimate the integrals on the right hand side of the above equality. For the first integral, as $d\chi(y)=0$ for $y\in B^{n}(0,\frac{1}{2})$, we have $|x-y|\geq \frac{1}{4}$ for $x\in B^n(0,\frac{1}{4})$ or otherwise the integrand vanishes. If we take $\theta^*(x,y,z)=\int_0^1( D_{\bar{y}}\psi)(tx+(1-t)y,z)dt$, then by Taylor expansion we have $$2\operatorname{Re}\left(\theta\cdot (x-y)\right)+2\operatorname{Re}\left(\theta^*\cdot (x-y)\right)=2\operatorname{Re}\left(\psi(x,\bar{y})-\psi(y,\bar{y})\right)\leq \phi(x)-\phi(y)-\delta|x-y|^2,$$ where $ \theta^*=\theta^*(x,y,\bar{y})$ and $\delta$ is some positive constant. Note that $\theta^*(x,y,z)\in \mathcal{I}_z^{a,{\varepsilon}}$ by Lemma \[AHE Integral\]. We have $\theta^*(x,y,\bar{y})=O(|x-y|^{\infty})$. Thus by rescaling the unit ball, $\theta^*(x,y,\bar{y})$ can be absorbed by $\delta|x-y|^2$. Therefore, by changing $\delta$ to a smaller constant, the integrand of the first integral is bounded by some constant times $$|u(y)|e^{\frac{k\phi(x)}{2}-\frac{k\phi(y)}{2}-\delta k}\left|A^{(N+1)}\right|.$$ So by using Cauchy-Schwartz inequality, we obtain the first integral is bounded by some constant times $$\|A^{(N+1)}\|_{L^\infty(\Lambda)}e^{\frac{k\phi(x)}{2}} \|u\|_{k\phi}e^{-\delta k}.$$ By Lemma \[error term 1 for Gevrey\] and $ke^{-\delta k}\leq \frac{(N+2)!}{k^{N+1}\delta^{N+2}}$, the first integral is bounded by $$\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}e^{\frac{k\phi(x)}{2}} \|u\|_{k\phi}.$$
For the second term, the integrand is bounded by some constant times $$|u(y)|e^{\frac{k\phi(x)}{2}-\frac{k\phi(y)}{2}-k\delta|x-y|^2}\left|a^{(N)}-\nabla \left(A^{(N+1)}\right)\right|.$$ By Corollary \[prop 7.6\], we have $$\begin{aligned}
e^{-k\delta|x-y|^2}&\left|a^{(N)}-\nabla \left(A^{(N+1)}\right) \right|
\\
\leq& \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}+e^{-k\delta|x-y|^2} \left(C k + \frac{C^{N}N!^{2a+2{\varepsilon}}}{k^N}\right)\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right).\end{aligned}$$ Note for any positive integer $M$, $$\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right)\leq \left(\frac{a-1}{b}\right)^{M(a-1)}M!^{a-1}|x-y|^{M}.$$ Take $M=2N+4$. Since for any $x,y\in B^n(0,1)$, $$e^{-k\delta|x-y|^2} |x-y|^{2N+4}\leq \frac{(N+2)!}{(\delta k)^{N+2}},$$ which implies $$\begin{aligned}
ke^{-k\delta|x-y|^2}\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right)
\leq& \left(\frac{(a-1)^{2a-2}}{b^{2a-2}\delta}\right)^{N+2}(2N+4)!^{a-1}\frac{(N+2)!}{k^{N+1}}\\
\leq& \left(\frac{4^{a-1+{\varepsilon}}(a-1)^{2a-2}}{b^{2a-2}\delta}\right)^{N+2}\frac{(N+2)!^{2a-1}}{k^{N+1}}\\
\leq& \frac{C^{N+1}(N+1)!^{2a-1}}{k^{N+1}},\end{aligned}$$ for some constant $C$ independent to $N$. So it is easy to see that by choosing a larger constant $C$, we have $$e^{-k\delta|x-y|^2}\left|a^{(N)}-\nabla \left(A^{(N+1)}\right) \right|
\leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}.$$ Therefore, the second integral is also bounded by $\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}e^{\frac{k\phi(x)}{2}} \|u\|_{k\phi}$.
For the third and fourth terms, the integrands are respectively bounded by some constant times $$\begin{aligned}
|u(y)|e^{\frac{k\phi(x)}{2}-\frac{k\phi(y)}{2}-k\delta|x-y|^2}\left|D_{\bar{\theta}}\cdot A^{(N+1)}\right|,\quad
|u(y)|e^{\frac{k\phi(x)}{2}-\frac{k\phi(y)}{2}-k\delta|x-y|^2}\left|D_{\bar{y}}\cdot A^{(N+1)}\right|.\end{aligned}$$
By Lemma \[error term 4 for Gevrey\], we have $$\begin{aligned}
e^{-k\delta|x-y|^2}\left|D_{\bar{\theta}}\cdot A^{(N+1)}\right|
\leq e^{-k\delta|x-y|^2} \left(C k + \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}\right)\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right),\end{aligned}$$ and $$\begin{aligned}
e^{-k\delta|x-y|^2}\left|D_{\bar{y}}\cdot A^{(N+1)}\right|
\leq e^{-k\delta|x-y|^2} \left(C k + \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}\right)\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right),\end{aligned}$$ By the same argument as estimating the second term, we have $$\begin{aligned}
ke^{-k\delta|x-y|^2}\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right)
\leq \frac{C^{N+1}(N+1)!^{2a-1}}{k^{N+1}}.\end{aligned}$$ By choosing a larger constant $C$, $$e^{-k\delta|x-y|^2}\left|D_{\bar{\theta}}\cdot A^{(N+1)}\right|
\leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}.$$ Therefore, the third and fourth integral are also bounded by $\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}e^{\frac{k\phi(x)}{2}} \|u\|_{k\phi}$ and the result follows.
From local to global and the proof of Theorem \[Main\] {#Sec Global}
======================================================
Let $K_k(x,y)$ be the Bergman kernel of $(M,L^k)$. As we noted before, we also write $K_k(x, y)$ for the representation of the Bergman kernel in the local frame $e^k_L \otimes {\overline{e^k_L}}$ and we denote $K_{k,y}(x):= K_k(x,y)$. In the last section, we constructed the local Bergman kernel of order $N$, which we denoted by $K_{k}^{(N)}(x,y)=K_{k,y}^{(N)}(x)$. In this section, we show that $K_k(x,y)$ is equal to $K_{k}^{(N)}(x,y)$ up to order $k^{-N}$ when $x,y$ are sufficiently close to each other. Moreover, we will give a precise upper bound for the error term.
\[LocalGlobal\] There exists $ \delta >0$ such that whenever $d(x,y) < \delta$, we have $$\label{Cx}
K_k(x,y)=K_k^{(N)}(x,y)+ k^{\frac{3n}{2}} \widetilde{\mathcal R}_{N+1}(\phi, k) e^{\frac{k\phi(x)}{2}+\frac{k\phi(y)}{2}},$$ where $$\label{C}
| \widetilde{\mathcal R}_{N+1}(\phi, k)| \leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}},$$ and the constant $C$ is independent of $N$, $x$, $y$, and $k$.
We fix $x \in M$ and assume that $\phi$ is in Gevrey $a$ class in $B^n(x, 3)$. Let $\chi$ be a smooth cut-off function such that $$\chi(z)=
\begin{cases}
1 & z\in B^{n}(x,\frac{1}{2})\\
0 & z\notin B^{n}(x,\frac{3}{4})
\end{cases}.$$ We assume $y\in B^n(x,\frac{1}{4})$. We first observe that $$\label{S}
K_k(y,x)=\left(\chi K_{k,x},K^{(N)}_{k,y}\right)_{k\phi}+ \mathcal S_{N+1} (\phi, k) k^{\frac{3n}{2}}e^{k(\frac{\phi(x)}{2}+\frac{\phi(y)}{2})},$$ where $|\mathcal S_{N+1} (\phi, k)|\leq \frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}$. This is because, by Proposition \[LocalBergmanKernel\], we have $$\begin{aligned}
K_{k,x}(y)=\left(\chi K_{k,x}, K^{(N)}_{k,y}\right)_{k\phi}+k^n \mathcal S_{N+1} (\phi, k) e^{\frac{k\phi(y)}{2}}\|K_{k,x}\|_{k\phi},\end{aligned}$$ and by the reproducing property of Bergman kernel, we have $$\|K_{k,x}\|_{k\phi} e^{-k\frac{\phi(x)}{2}} \leq \left | \|K_{k,x}\|_{L^2(M,L^k)} \right |_{h^k}= \sqrt{| K_k(x,x)|_{h^k}}\leq Ck^{\frac{n}{2}}.$$ That why $| K_k(x,x)|_{h^k} \leq Ck^n$ follows from the extreme property of the Bergman function and also the sub-mean value inequality. For a simple proof see for example Lemma 4.1 of [@HKSX].
Next, we define$$\label{u}
u_{k,y}(z)= \chi(z) K_{k,y}^{(N)}(z)-\left(\chi K^{(N)}_{k,y}, K_{k,z}\right)_{k\phi}.$$ Our goal is to estimate $|u_{k, y}(x)|$. Since $\left(\chi K^{(N)}_{k,y}, K_{k,x}\right)_{k\phi}$ is the Bergman projection of $\chi K_{k,y}^{(N)}$, $u_{k,y}$ is the minimal $L^2$ solution to the equation $${{\overline{\partial}}}u={{\overline{\partial}}}(\chi K_{k,y}^{(N)}).$$ So by using Hörmander’s $L^2$ estimates [@Ho] (see [@B] for an exposition), we have $$\left\|u_{k,y}\right\|_{L^2}^2\leq \frac{C}{k}\left\|{{\overline{\partial}}}\left(\chi K_{k,y}^{(N)}\right)\right\|_{L^2}^2.$$ We have ${{\overline{\partial}}}(\chi K_{k,y}^{(N)})(z)={{\overline{\partial}}}\chi(z) K_{k,y}^{(N)}(z)+\chi(z) {{\overline{\partial}}}K_{k,y}^{(N)}(z)$. Recall that by $${K^{(N)}_{k,y}(z)}=\left(\frac{k}{\pi}\right)^n e^{k\psi(z,\bar{y})}B^{(N)}(z,\bar{y}).$$ For the first term, since ${{\overline{\partial}}}\chi(z)$ is supported in $d(z,x)\geq \frac{1}{2}$, if we choose $d(x,y)\leq \frac{1}{4}$, then using $$\operatorname{Re}\psi(z,\bar{y})\leq -\delta|z-y|^2+\frac{\phi(x)}{2}+\frac{\phi(y)}{2},$$ we have $$\left|{{\overline{\partial}}}\chi(z) K_{k,y}^{(N)}(z)\right|\leq Ck^n e^{-k\delta+k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\|B^{(N)}\|_{L^{\infty}(U\times U)}.$$ We can estimate $\|B^{(N)}\|_{L^\infty(U \times U)}$ using our Theorem \[MainLemma\] $$\begin{aligned}
\|B^{(N)}\|_{L^\infty ( U \times U)}
& \leq 1+ \frac{1}{k} \|b_1\|_{L^\infty ( U\times U)} + \dots \frac{1}{k^N} \|b_N\|_{L^\infty (U\times U)} \\
& \leq 1+\frac{C1!^{2a+2{\varepsilon}}}{k}+\frac{C^2 2! ^{2a+2{\varepsilon}}}{k^2}+\cdots+\frac{C^N N! ^{2a+2{\varepsilon}}}{k^N} \\
& \leq C\left(k+\frac{C^NN!^{2a+2{\varepsilon}}}{k^N}\right) .\end{aligned}$$ Hence, $$\begin{aligned}
\left|{{\overline{\partial}}}\chi(z) K_{k,y}^{(N)}(z)\right|
\leq& k^ne^{-k\delta+k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\left(Ck+\frac{C^NN!^{2a+2{\varepsilon}}}{k^N}\right)
\\
\leq& k^ne^{k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}.\end{aligned}$$ For the second term, we have $$\left|\chi(z) {{\overline{\partial}}}K_{k,y}^{(N)}(z)\right|\leq Ck^n e^{-k|z-y|^2+k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\left(k\left|{{\overline{\partial}}}\psi(z,\bar{y})\right|\left|B^{(N)}(z,\bar{y})\right|+\left|{{\overline{\partial}}}B^{(N)}(z,\bar{y})\right|\right).$$ By using the fact that $\psi(y,z)\in \mathcal{A}^{a,{\varepsilon}}_{diag}$ and Theorem \[MainLemma\], $$\begin{aligned}
k\left|{{\overline{\partial}}}\psi(z,\bar{y})\right|\left|B^{(N)}(z,\bar{y})\right|+\left|{{\overline{\partial}}}B^{(N)}(z,\bar{y})\right|
\leq Ck\left(k+\frac{C^NN!^{2a+2{\varepsilon}}}{k^N}\right)\exp(-b|z-y|^{-\frac{1}{a-1}}) .\end{aligned}$$ Then by the fact that $$\begin{aligned}
e^{-k\delta|x-y|^2}\exp\left(-b|x-y|^{-\frac{1}{a-1}}\right)
\leq \frac{C^{N+1}(N+1)!^{2a-1}}{k^{N+3}}, \end{aligned}$$ we have $$\begin{aligned}
\left|\chi(z) {{\overline{\partial}}}K_{k,y}^{(N)}(z)\right|\leq k^ne^{k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}.\end{aligned}$$ So $$\begin{aligned}
\label{dbarK}
\left|{{\overline{\partial}}}(\chi K_{k,y}^{(N)})(z)\right|
\leq k^ne^{k\frac{\phi(y)}{2}+k\frac{\phi(z)}{2}}\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}},\end{aligned}$$ and $$\begin{aligned}
\label{uL2norm}
\|u_{k,y}\|^2\leq \frac{C}{k} \|{{\overline{\partial}}}\left(\chi K_{k,y}^{(N)}\right)\|_{L^2}^2
\leq
C k^{2n}e^{k\phi(y)}\left(\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}\right)^2.\end{aligned}$$ By using Bochner-Martinelli formula in a small Euclidean ball $B^n(x,r)$, we have $$r^{2n-1}|u_{k,y}(x)|\leq C\int_{\partial B^n(x,r)} |u(z)|dS+C\int_{B^n(x,r)}|{{\overline{\partial}}}u(z)|\frac{r^{2n-1}}{|z-x|^{2n-1}}dV_0,$$ where $dS$ and $dV_0$ are respectively the standard volume forms of $\partial B^n(0,1)$ and $B^n(0,1)$ in Euclidean space. If we use the Bochner coordinates at $x$, then $\phi(z)-\phi(x)=O(|z|^2)$, and thus $$e^{k\left(\frac{\phi(z)}{2}-\frac{\phi(x)}{2}\right)}\leq C,\quad\mbox{for any } z\in B^n(x,\frac{1}{\sqrt{k}}).$$ By integrating the above inequality with respect to $r$ from $0$ to $\frac{1}{\sqrt{k}}$, we obtain $$\begin{aligned}
|u_{k,y}(x)|
\leq& Ck^n\int_{B^n(x,\frac{1}{\sqrt{k}})} |u(z)|dV_0+Ck^n\int_0^{\frac{1}{\sqrt{k}}}\int_{B^n(x,r)}|{{\overline{\partial}}}u(z)|\frac{r^{2n-1}}{|z-x|^{2n-1}}dV_0dr
\\
\leq& Ck^n\int_{B^n(x,\frac{1}{\sqrt{k}})} |u(z)|dV_0+C\|e^{-\frac{k\phi(z)}{2}}{{\overline{\partial}}}u(z)\|_{L^\infty}e^{\frac{k\phi(x)}{2}}\int_{B^n(x,\frac{1}{\sqrt{k}})}\frac{1}{|z-x|^{2n-1}}e^{k\left(\frac{\phi(z)}{2}-\frac{\phi(x)}{2}\right)}dV_0
\\
\leq &Ck^\frac{n}{2}e^{\frac{k\phi(x)}{2}}\left(\|u\|_{L^2}+\|e^{-\frac{k\phi(z)}{2}}{{\overline{\partial}}}u(z)\|_{L^\infty}\right).\end{aligned}$$
Therefore, by the estimates and , it follows that $$|u_{k,y}(x)|\leq C k^{\frac{3n}{2}}e^{k\frac{\phi(y)}{2}+k\frac{\phi(x)}{2}}\left(\frac{C^{N+1}(N+1)!^{2a+2{\varepsilon}}}{k^{N+1}}\right).$$ Combining this estimate with and recalling the definition of $u_{k, y}$ in , we get the result.
We point out that we have renewed the constant $C$ at each step, but the final constant is independent of $k$ and $N$. We also note that the constant $C$ may depend on the point $x$, however by a simple compactness argument one can see that each such $C$ can be bounded by a uniform constant independent of $x$.
Now we are ready to prove Theorem \[Main\] and its corollaries.
Proof of Theorem \[Main\]
-------------------------
By Proposition \[LocalGlobal\], we just need to show that with $N=N_0-1= [ (k/C)^{\frac{1}{2a+2{\varepsilon}}}]-1$, we have[^2] $$k^{\frac{3n}{2}} \mathcal R_{N_0}(\phi, k) e^{\frac{k\phi(x)}{2}+\frac{k\phi(y)}{2}} = e^{k\left(\frac{\phi(x)}{2}+\frac{\phi(y)}{2}\right)}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1).$$ However, by the same proposition we know that $$| R_{N_0}(\phi, k)| \leq \frac{C^{N_0}N_0!^{2a+2{\varepsilon}}}{k^{N_0}}.$$ Hence it is enough to show that $$\frac{C^{N_0}N_0!^{2a+2{\varepsilon}}}{k^{N_0}} = e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}} O(1).$$ By Stirling’s formula, $$\frac{C^{N_0}N_0!^{2a+2{\varepsilon}}}{k^{N_0}}
\leq C' N_0^{a+{\varepsilon}}\frac{C^{N_0}{N_0}^{(2a+2{\varepsilon})N_0}}{e^{(2a+2{\varepsilon})N_0}k^{N_0}}
\leq C' N_0^{a+{\varepsilon}}e^{-(2a+2{\varepsilon})N_0}
\leq C''k^{\frac{1}{2}}e^{- (2a+2{\varepsilon}) (\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}}.$$ Since $$k^{\frac{3n}{2}+\frac12} e^{- (2a+2{\varepsilon}) (\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}} \leq C''' e^{- (a+{\varepsilon}) (\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}},$$ $\delta =\frac{a+{\varepsilon}}{C^{\frac{1}{2a+2{\varepsilon}}}}$ would do the job.
Proof of Corollary \[complete asymptotics\]
-------------------------------------------
By Theorem \[Main\], uniformly for any $x,y\in U$, we have $$K_k(x,y)=e^{k\psi(x,\bar y)}\frac{k^n}{\pi^n}\left (1+\sum _{j=1}^{ N_0(k)-1}\frac{b_j(x, \bar y)}{k^j}\right )+e^{k\left(\frac{\phi(x)}{2}+\frac{\phi(y)}{2}\right)}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1).$$ For any given positive integer $N$, we rewrite the above formula as follows. $$K_k(x,y)=e^{-k\psi(x,\bar{y})}\frac{k^n}{\pi^n}\left(1+\sum_{j=1}^{N-1}\frac{b_j(x,\bar{y})}{k^j}+\sum_{j=N}^{N_0(k)-1}\frac{b_j(x,\bar{y})}{k^j}+e^{\frac{k}{2}\left(\phi(x)+\phi(y)-2\psi(x,\bar{y})\right)}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1)\right).$$ Our first observation is that, if $d(x,y)\leq \sqrt{\delta}k^{-\frac{1}{2}+\frac{1}{4a+4{\varepsilon}}}$, then $$\begin{aligned}
\left|e^{\frac{k}{2}\left(\phi(x)+\phi(y)-2\psi(x,\bar{y})\right)}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}}\right|=e^{\frac{k}{2}D(x,y)-\delta k^{\frac{1}{2a+2{\varepsilon}}}}
\leq e^{-\frac{1}{4}\delta k^{\frac{1}{2a+2{\varepsilon}}}}.
\end{aligned}$$ Now we estimate the term $\sum_{j=N}^{N_0-1}\frac{b_j(x,\bar{y})}{k^j}$. By Stirling’s formula, we have $$\begin{aligned}
\left|\frac{b_j(x,\bar{y})}{k^j}\right|\leq \frac{C^jj!^{2a+2{\varepsilon}}}{k^j}\leq C'j^{a+{\varepsilon}}\frac{C^jj^{(2a+2{\varepsilon})j}}{e^{(2a+2{\varepsilon})j}k^j}.
\end{aligned}$$ Since $\frac{C^jj^{(2a+2{\varepsilon})j}}{e^{(2a+2{\varepsilon})j}k^j}$ is monotonically decreasing for $1\leq j\leq N_0(k)-1$ (with the help of Stirling’s formula once more), we get $$\begin{aligned}
\left|\sum_{j=N}^{N_0-1}\frac{b_j(x,\bar{y})}{k^j}\right|
\leq& \frac{C^NN!^{2a+2{\varepsilon}}}{k^N}+\sum_{j=N+1}^{N_0-1}C'j^{a+{\varepsilon}}\frac{C^jj^{(2a+2{\varepsilon})j}}{e^{(2a+2{\varepsilon})j}k^j}
\\
\leq &\frac{C^NN!^{2a+2{\varepsilon}}}{k^N}+C'N_0^{a+{\varepsilon}+1}\frac{C^{N+1}(N+1)^{(2a+2{\varepsilon})(N+1)}}{e^{(2a+2{\varepsilon})(N+1)}k^{N+1}}
\\
\leq &\frac{C''^{N}N!^{2a+2{\varepsilon}}}{k^N}.
\end{aligned}$$ Therefore, $$\begin{aligned}
K_k(x,y)
=&e^{-k\psi(x,\bar{y})}\frac{k^n}{\pi^n}\left(1+\sum_{j=1}^{N-1}\frac{b_j(x,\bar{y})}{k^j}+\frac{C''^NN!^{2a+2{\varepsilon}}}{k^N}+e^{-\frac{1}{4}\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1)\right).
\end{aligned}$$ By the fact that $$\begin{aligned}
e^{-\frac{1}{4}\delta k^{\frac{1}{2a+2{\varepsilon}}}}
\leq \left(\frac{4}{\delta}\right)^{(2a+2{\varepsilon})N}\frac{\left((2a+2{\varepsilon})N\right)^{(2a+2{\varepsilon})N}}{e^{(2a+2{\varepsilon})N}k^N}
\leq \left(\frac{8a+8{\varepsilon}}{\delta}\right)^{(2a+2{\varepsilon})N}\frac{N!^{2a+2{\varepsilon}}}{k^N},
\end{aligned}$$ the first part of our result follows.
Now we prove the second part. Let $\widetilde{b}_m(x,z)$ be another almost holomorphic extension of $b_m(x,\bar{x})$. By Lemma \[AHE Uniqueness\], we have $$|b_m(x,\bar{y})-\widetilde{b}_m(x,\bar{y})|=O\left(|x-y|^\infty\right)=O\left(\frac{1}{k^\infty}\right).$$ The second equality follows from our assumption that $d(x,y)\leq \delta k^{-\frac{1}{2}+\frac{1}{4a+4{\varepsilon}}}$. So the result follows.
Let $\widetilde{b}_m$ be the almost holomorphic extension defined in Definition \[AHE\]. If we take $\tilde{b}_m$ in Corollary \[complete asymptotics\], then and hold. The reason is as follows.
For any $M\in \mathbb{N}$, there exists $C_M$ such that $$\left|b_m(x,\bar{y})-\widetilde{b}_m(x,\bar{y})\right|\leq C_M|x-y|^{M+1}.$$ And $C_M$ depends on the sup norm of the all the $(M+1)$th derivatives of $b_m(x,\bar{y})$ and $\widetilde{b}_m(x,\bar{y})$. By Theorem \[MainLemma\] and Lemma \[AHE2\], we have $$C_M\leq C^{m+M}m!^{2a+2{\varepsilon}}M!^{a+{\varepsilon}-1},$$ where $C$ is some positive constant independent of $m$. If we take $M=[(N-m)\frac{4a+4{\varepsilon}}{2a+2{\varepsilon}-1}]$, then $|x-y|^{M+1}\leq \frac{1}{k^{N-m}}$, whence by Stirling’s formula we obtain $$\begin{aligned}
\frac{\left|b_m(x,\bar{y})-\widetilde{b}_m(x,\bar{y})\right|}{k^m}
\leq& C^{m+M}m!^{2a+2{\varepsilon}}M^{M(a+{\varepsilon}-1)}k^{-N}\\
\leq& C'^{N}m!^{2a+2{\varepsilon}} (N-m)^{(N-m)(2a+2{\varepsilon})}k^{-N}\\
\leq& C''^{N}m!^{2a+2{\varepsilon}} (N-m)!^{2a+2{\varepsilon}}k^{-N},
\end{aligned}$$ where $C'$ and $C''$ are some positive constants independent of $m$. We rename $C''$ by $C$ and , follows by $$\begin{aligned}
\sum_{j=1}^{N-1}\frac{\left|b_j(x,\bar{y})-\widetilde{b}_j(x,\bar{y})\right|}{k^j}
\leq \sum_{j=1}^{N-1} \frac{C^NN!^{2a+2{\varepsilon}}}{k^N}\leq \frac{(2C)^NN!^{2a+2{\varepsilon}}}{k^N}.
\end{aligned}$$
Proof of Corollary \[Log\]
--------------------------
By Theorem \[Main\], we have $$K_k(x,y)=e^{k\psi(x, \bar y)}\frac{k^n}{\pi^n}\left(1+\sum _{j=1}^{N_0-1}\frac{b_j(x,\bar y)}{k^j}\right)+e^{\frac{k\phi(x)}{2}+\frac{k\phi(y)}{2}}e^{-\delta k^{\frac{1}{2a+2{\varepsilon}}}} O(1).$$ Recall that $D(x,y)=\phi(x)+\phi(y)-\psi(x,\bar y)-\psi(y, \bar x)$. Then $$\log|K_k(x,y)|_{h^k}=-\frac{kD(x,y)}{2}+n\log k -n\log \pi +\log\left|1+\sum _{j=1}^{N_0-1}\frac{b_j(x, \bar y)}{k^j}+e^{\frac{Q(x,y)}{2}k-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1)\right|,$$ where $Q(x,y)=\phi(x)+\phi(y)-2\psi(x,\bar{y})$. So it is sufficient to prove $$\log\left|1+\sum _{j=1}^{N_0-1}\frac{b_j(x,\bar{y})}{k^j}+e^{\frac{Q(x,y)}{2}k-\delta k^{\frac{1}{2a+2{\varepsilon}}}}O(1)\right|
=\log\left(1+O \left ( \frac{1}{k} \right )\right).$$ To do this we note that by our assumption $D(x, y) \leq \frac{1}{2} \delta k^{-1+\frac{1}{2a+2{\varepsilon}}}$, hence $$\left|e^{\frac{Q(x,y)}{2}k-\delta k^{\frac{1}{2a+2{\varepsilon}}}}\right|=
e^{\frac{D(x,y)}{2}k-\delta k^{\frac{1}{2a+2{\varepsilon}}}}
\leq e^{-\frac{3\delta}{4}k^{\frac{1}{2a+2{\varepsilon}}}}.$$ It remains to show that $$\left|\sum _{j=1}^{N_0-1}\frac{b_j(x,\bar{y})}{k^j}\right|=O \left (\frac{1}{k} \right ).$$ By the estimates on $b_j(x,\bar{y})$ from Theorem \[MainLemma\] and Stirling’s formula, we have $$\frac{|b_j(x,\bar{y})|}{k^j}\leq \frac{C^jj!^{2a+2{\varepsilon}}}{k^j}\leq C'j^{a+{\varepsilon}}\frac{C^j j^{(2a+2{\varepsilon})j}}{e^{2j}k^j}.$$ As shown in Lemma \[error term 1 for Gevrey\], the function $f(x)=\log \frac{C^xx^{(2a+2{\varepsilon})x}}{e^{2x}k^x}$ is decreasing on the interval $(0,(\frac{k}{C})^{\frac{1}{2a+2{\varepsilon}}}]$, thus for $j\in [2,N_0-1]$, $$\frac{|b_j(x,\bar{y})|}{k^j}\leq C'(N_0-1)^{a+{\varepsilon}}\frac{C^2 2^{4a+4{\varepsilon}}}{e^{4}k^2}\leq C''C^2\frac{N_0^{a+{\varepsilon}}}{k^2}.$$ Therefore, $$\left|\sum _{j=1}^{N_0-1}\frac{b_j(x,\bar{y})}{k^j}\right|\leq \frac{C}{k}+C''C^2\frac{N_0^{a+{\varepsilon}+1}}{k^2}\leq \frac{C}{k}+C''C\frac{1}{k}=O \left (\frac{1}{k} \right ).$$
Estimates on Bergman Kernel Coefficients {#Sec ProofofMainLemma}
========================================
As before, we assume the metric is in Gevrey class $G^a(U)$ for some neighborhood $U$ of $p$. We will estimate the growth rate of the Bergman kernel coefficients $b_m(x,z)$ as $m\rightarrow \infty$ for $x,z$ in $U$. Our goal is to prove Theorem \[MainLemma\].
The key ingredient for the proof is the following recursive formula[^3] on $b_m(x,z)$ established in [@BBS]. $$\label{Recursive}
b_m(x,z(x,x,\theta))=-\sum_{l=1}^m\frac{(D_y\cdot D_\theta)^l}{l!}\big(b_{m-l}\left(x,z(x,y,\theta)\right)\Delta_0(x,y,\theta)\big)\Big|_{y=x}.$$ We will break the proof of Theorem \[MainLemma\] into two steps. The first step is to derive from the recursive formula , a recursive inequality on $\|D^\mu_zD^\nu_{\bar{z}}b_m(x,z)\|_{L^\infty(U\times U)}$ for some neighborhood $U$. The second step is to estimate $\|D^\mu_zD^\nu_{\bar{z}}b_m(x,z)\|_{L^\infty(U\times U)}$ by induction.
In the following we shall use the following standard notations for multi-indicies.
- $ \mathbbm 1 = (1, 1, \cdots, 1)$.
- $\binom{\alpha}{\beta}=\binom{\alpha_1}{\beta_1}\binom{\alpha_2}{\beta_2}\cdots \binom{\alpha_n}{\beta_n}$.
- $\binom{l}{\delta_1,\delta_2,\cdots,\delta_n}=\frac{l!}{\delta_1!\delta_2!\cdots\delta_n!}$ for any non-negative integer $l$ and multi-index $\delta\geq 0$ such that $|\delta|=l$.
\[bmLemmaGevrey\] Suppose the potential $\phi\in G^a(U)$. Let $W=\{(x,z)\in U\times U: x\neq z\}$ and $$b=\min\{b(\Delta_0(x,y,\theta)),b(\psi_{x_i}(x,z)),b(z_i(x,y,\theta)): 1\leq i\leq n\}.$$ If we denote $v=(x,z)$ and $$b_{m,\mu\bar{\nu}}=\left\|\frac{1}{\lambda_{b,|\nu|}(x,x,z)}D_v^\mu D_{{\overline{v}}}^{\nu} b_m(x,z)\right\|_{L^\infty(W)},$$ then the exists some positive constant $C$ independent of $m,\mu,\nu$, such that $$\begin{aligned}
\label{bmGevrey}
\begin{split}
b_{m,\mu\bar{\nu}}
& \leq\sum_{l=1}^m
\sum_{|\delta|=l}\delta!^{2a+2{\varepsilon}-1}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi+\eta|\leq |\alpha+\beta|}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu} \frac{ b_{m-l,\widetilde{\xi}+\mu_0{\overline{\widetilde{\eta}+\nu_0}}}}{(\xi!\eta!)^{a+{\varepsilon}}}
C^{|\mu-\mu_0+\nu-\nu_0|+|\delta+\xi+\eta|}
\\&
\quad\quad \cdot
\binom{\mu}{\mu_0}\binom{\nu}{\nu_0}
(\mu-\mu_0)!^{a+{\varepsilon}}(\nu-\nu_0)!^{a+{\varepsilon}},
\end{split}
\end{aligned}$$ where $\xi,\eta\in (\mathbb{Z}^{\geq 0})^n$ and $\widetilde{\xi}=(0,\cdots 0,\xi), \widetilde{\eta}=(0, \cdots, 0,\eta)\in (\mathbb{Z}^{\geq 0})^{2n}$.
We first work on $(D_y\cdot D_\theta)^l\Big(b_{m-l}\left(x,z\left(x,y,\theta\right)\right)\Delta_0\left(x,y,\theta\right)\Big)$. We expand $(D_y\cdot D_\theta)^l$ and obtain $$\begin{aligned}
\label{DDl}
\begin{split}
(D_y&\cdot D_\theta)^l\big(b_{m-l}\left(x,z\left(x,y,\theta\right)\right)\Delta_0\left(x,y,\theta\right)\big)\\
=&\sum_{|\delta|=l}\binom{l}{\delta_1,\delta_2,\cdots,\delta_n}
D_y^\delta D_\theta^\delta
\big(b_{m-l}(x,z(x,y,\theta))\Delta_0(x,y,\theta)\big)\\
=&\sum_{|\delta|=l}\binom{l}{\delta_1,\delta_2,\cdots,\delta_n}
\sum_{\alpha,\beta\leq \delta}\binom{\delta}{\alpha}\binom{\delta}{\beta}D_y^\alpha D_\theta^\beta
\big(b_{m-l}(x,z(x,y,\theta))\big)D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0
\\
=& b_{m-l}(x,z(x,y,\theta))(D_y\cdot D_\theta)^l\Delta_0(x,y,\theta)
\\&+
\sum_{|\delta|=l}l!\delta!
\sum_{\substack{\alpha,\beta\leq \delta\\ \alpha+\beta>0}}
\sum_{1\leq|\xi+\eta|\leq |\alpha+\beta|} \frac{D_z^\xi D_{\bar{z}}^\eta b_{m-l}(x,z)}{\xi!\eta!}
\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!}
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!},
\end{split}
\end{aligned}$$ where the index set $A_{\alpha\beta\xi\eta}$ is defined by $$\begin{aligned}
\label{Index Set}
\begin{split}
A_{\alpha \beta\xi\eta}&
\\
=&\left \{ \{ \alpha_{ij},\beta_{ij} \}_{1\leq i \leq n, 1 \leq j \leq\xi_i}, \{ \alpha'_{ik}, \beta'_{ik}\}_{1\leq i \leq n, 1 \leq k \leq\eta_i}: \quad \begin{array}{ll} \sum_{1\leq i \leq n, 1 \leq j \leq \xi_i} \alpha_{ij}+\sum_{1\leq i \leq n, 1 \leq k \leq \eta_i} \alpha'_{ik}=\alpha, \\ \sum_{1\leq i \leq n, 1 \leq j \leq \xi_i} \beta_{ij}+\sum_{1\leq i \leq n, 1 \leq k \leq \eta_i} \beta'_{ik}=\beta, \\ \alpha_{ij}+\beta_{ij}>0,\quad \alpha'_{ik}+\beta'_{ik}>0 \end{array} \right \}.
\end{split}
\end{aligned}$$
We now substitute into equation and obtain $$\begin{aligned}
b_m&(x,z(x,x,\theta))\\
&=-\sum_{l=1}^m
\Bigg(\frac{1}{l!}b_{m-l}(x,z(x,x,\theta))(D_y\cdot D_\theta)^l\Delta_0(x,x,\theta)+\sum_{|\delta|=l}\delta!
\sum_{\substack{\alpha,\beta\leq \delta\\ \alpha+\beta>0}}\sum_{1\leq|\xi+\eta|\leq |\alpha+\beta|} \frac{D_z^\xi D_{\bar{z}}^\eta b_{m-l}(x,z(x,x,\theta))}{\xi!\eta!}
\\&
\qquad \cdot\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\theta)
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}(x,x,\theta)
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\theta)\Bigg).
\end{aligned}$$
The correspondence $(x,x,z)\leftrightarrow (x,x,\theta=\psi_x(x,z))$, turns this into $$\begin{aligned}
b_m&(x,z)\\
&=-\sum_{l=1}^m
\Bigg(\frac{1}{l!}b_{m-l}(x,z)(D_y\cdot D_\theta)^l\Delta_0(x,x,\psi_x(x,z))+\sum_{|\delta|=l}\delta!
\sum_{\substack{\alpha,\beta\leq \delta\\ \alpha+\beta>0}}\sum_{1\leq|\xi+\eta|\leq |\alpha+\beta|} \frac{D_z^\xi D_{\bar{z}}^\eta b_{m-l}(x,z)}{\xi!\eta!}
\\&
\qquad \cdot\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\psi_x(x,z))
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}(x,x,\psi_x(x,z))
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\psi_x(x,z))\Bigg).
\end{aligned}$$
Denote $v=(x,z)$. Note that in this recursive formula, the coefficients $b_m$ depend on not only the previous coefficients $b_{m-l}$ but also derivatives of $b_{m-l}$. Hence, we need to include derivatives of $b_m$ in our inductive argument. To do this we apply $D_v^{\mu}D_{{\overline{v}}}^{\nu}$ on both sides and obtain a recursive formula for the derivatives of $b_m$. $$\begin{aligned}
\label{bmrecursive}
\begin{split}
D_v^{\mu}&D_{{\overline{v}}}^{\nu} b_m(x,z)
\\
=&-\sum_{l=1}^m\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}\binom{\mu}{\mu_0}\binom{\nu}{\nu_0}
\Bigg\{\frac{1}{l!}D_v^{\mu_0}
D_{{\overline{v}}}^{\nu_0}b_{m-l}(x,z)D_v^{\mu-\mu_0} D_{{\overline{v}}}^{\nu-\nu_0}\left((D_y\cdot D_\theta)^l\Delta_0(x,x,\psi_x(x,z))\right)
\\
&+\sum_{|\delta|=l}\delta!
\sum_{\substack{\alpha,\beta\leq \delta\\ \alpha+\beta>0}}\sum_{1\leq|\xi+\eta|\leq |\alpha+\beta|} \frac{D_v^{\mu_0+\widetilde{\xi}} D_{{\overline{v}}}^{\nu_0+\widetilde{\eta}} b_{m-l}(x,z)}{\xi!\eta!}
\\& \quad \cdot D_v^{\mu-\mu_0}D_{{\overline{v}}}^{\nu-\nu_0} \left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\psi_x(x,z))
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\psi_x(x,z))\right)\Bigg\}.
\end{split}
\end{aligned}$$
Now we will estimate the factor $$D_v^{\mu-\mu_0} D_{{\overline{v}}}^{\nu-\nu_0}\left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\psi_x(x,z))
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\psi_x(x,z))\right).$$
Denote $\Phi(x,z)=(\varphi_1,\varphi_2,\cdots,\varphi_{3n})=(x,x,\psi_x(x,z))$, and $w=(x,y,\theta)$. In general, for any smooth function $f(x,y,\theta)$ and any multi-indices $\mu,\nu\in (\mathbb{Z}^{\geq 0})^{2n}$, we have $$\begin{aligned}
\frac{D_v^{\mu}D_{{\overline{v}}}^{\nu}}{\mu!\nu!}\left(f\left(\Phi(x,z)\right)\right)
=\sum_{0\leq |\rho+\tau|\leq |\mu+\nu|}\frac{D_w^\rho D_{{\overline{w}}}^\tau f}{\rho!\tau!}\sum_{A_{\mu\nu\rho\tau}}
\prod_{ij}\frac{D_v^{\mu_{ij}}D_{{\overline{v}}}^{\nu_{ij}}\varphi_i}{\mu_{ij}!\nu_{ij}!}
\prod_{ik}\frac{D_v^{\mu'_{ik}}D_{{\overline{v}}}^{\nu'_{ik}}{\overline{\varphi_i}}}{\mu'_{ik}!\nu'_{ik}!},
\end{aligned}$$ where the index set $A_{\mu\nu\rho\tau}$ is defined similar as in with a minor change that $1\leq i\leq 3n$. Applying this to our case, we obtain $$\begin{aligned}
\frac{D_v^{\mu}D_{{\overline{v}}}^{\nu}}{\mu!\nu!}&\left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\psi_x(x,z))
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\psi_x(x,z))\right)
\\
&
=\sum_{0\leq |\rho+\tau|\leq |\mu+\nu|}\frac{D_w^\rho D_{{\overline{w}}}^\tau}{\rho!\tau!}
\left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!}
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}\right)
\\
&\qquad\cdot \sum_{A_{\mu\nu\rho\tau}}
\prod_{ij}\frac{D_v^{\mu_{ij}}D_{{\overline{v}}}^{\nu_{ij}}\varphi_i}{\mu_{ij}!\nu_{ij}!}
\prod_{ik}\frac{D_v^{\mu'_{ik}}D_{{\overline{v}}}^{\nu'_{ik}}{\overline{\varphi_i}}}{\mu'_{ik}!\nu'_{ik}!}
\end{aligned}$$
We will use $C$ to denote a constant depending on constants ${\varepsilon},a,n$ and potential $\phi$ but independent with all the indices $m,\mu,\nu$, which may vary from line to line. Since $\Phi(x,z)\in \mathcal{A}^{a,{\varepsilon}}_{diag}$ for each $1\leq i\leq n$, we have $$\begin{aligned}
\label{diagonal 1.1}
\left|\sum_{A_{\mu\nu\rho\tau}}
\prod_{ij}\frac{D_v^{\mu_{ij}}D_{{\overline{v}}}^{\nu_{ij}}\varphi_i}{\mu_{ij}!\nu_{ij}!}
\prod_{ik}\frac{D_v^{\mu'_{ik}}D_{\bar{v}}^{\nu'_{ik}}{\overline{\varphi_i}}}{\mu'_{ik}!\nu'_{ik}!}\right|
\leq \sum_{A_{\mu\nu\rho\tau}} C^{|\mu+\nu|+|\rho+\tau|}\prod_{ij}\left(\mu_{ij}+\nu_{ij}\right)!^{a+{\varepsilon}-1}
\prod_{ik}\left(\mu'_{ik}+\nu'_{ik}\right)!^{a+{\varepsilon}-1}.
\end{aligned}$$ Apply the combinatorial lemma \[Combinatoric 2\] that we will prove later to the two products appearing above, we have $$\begin{aligned}
\prod_{ij}\left(\mu_{ij}+\nu_{ij}\right)!\leq \frac{(2n)^{|\rho|}}{\rho!}\left(\sum_{ij}(\mu_{ij}+\nu_{ij})\right)!,
\quad
\prod_{ik}\left(\mu'_{ik}+\nu'_{ik}\right)!\leq \frac{(2n)^{|\tau|}}{\tau!}\left(\sum_{ik}(\mu'_{ik}+\nu'_{ik})\right)!.
\end{aligned}$$ Therefore, $$\begin{aligned}
\left|\sum_{A_{\mu\nu\rho\tau}}
\prod_{ij}\frac{D_v^{\mu_{ij}}D_{{\overline{v}}}^{\nu_{ij}}\varphi_i}{\mu_{ij}!\nu_{ij}!}
\prod_{ik}\frac{D_v^{\mu'_{ik}}D_{{\overline{v}}}^{\nu'_{ik}}{\overline{\varphi_i}}}{\mu'_{ik}!\nu'_{ik}!}\right|
&\leq \sum_{A_{\mu\nu\rho\tau}} C^{|\mu+\nu|+|\rho+\tau|}\left(\frac{(\mu+\nu)!}{\rho!\tau!}\right)^{a+{\varepsilon}-1}
\\
&\leq C^{|\mu+\nu|+|\rho+\tau|}\left(\frac{\mu!\nu!}{\rho!\tau!}\right)^{a+{\varepsilon}-1}.
\end{aligned}$$ The last inequality follows from $$\begin{aligned}
\#A_{\mu\nu\rho\tau}\leq \binom{\mu+|\rho+\tau|\mathbbm{1}}{|\rho+\tau|\mathbbm{1}} \binom{\nu+|\rho+\tau|\mathbbm{1}}{|\rho+\tau|\mathbbm{1}}
\leq 2^{|\mu+\nu|+4n|\rho+\tau|}.
\end{aligned}$$
As $z(x,y,\theta), \Delta_0(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_{\theta}$ by Lemma \[AHE Composition\] and Lemma \[AHE Inverse\], after a straightforward calculation, we have $$\begin{aligned}
\label{diagonal 1.2}
\begin{split}
&\left|\frac{D_w^\rho D_{{\overline{w}}}^\tau}{\rho!\tau!}
\left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!}
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}\right)\right|
\\
&\qquad\leq C^{|\rho+\tau|+|\delta+\xi+\eta|+1}\left(\frac{\delta!^2\rho!\tau!}{\xi!\eta!}\right)^{a+{\varepsilon}-1}
\binom{\alpha+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}
\binom{\beta+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}.
\end{split}
\end{aligned}$$ Therefore, for any $\mu,\nu\geq 0$, we obtain $$\begin{aligned}
\begin{split}
&\left|\frac{D_v^{\mu}D_{{\overline{v}}}^{\nu}}{\mu!\nu!}\left(\frac{D_y^{\delta-\alpha}D_\theta^{\delta-\beta}\Delta_0}{(\delta-\alpha)!(\delta-\beta)!} (x,x,\psi_x(x,z))
\sum_{A_{\alpha\beta\xi\eta}}
\prod_{ij}\frac{D_y^{\alpha_{ij}}D_{\theta}^{\beta_{ij}}z_i}{\alpha_{ij}!\beta_{ij}!}
\prod_{ik}\frac{D_y^{\alpha'_{ik}}D_{\theta}^{\beta'_{ik}}{\overline{z_i}}}{\alpha'_{ik}!\beta'_{ik}!}(x,x,\psi_x(x,z))\right)\right|
\\
&
\qquad\leq \sum_{0\leq |\rho+\tau|\leq |\mu+\nu|}C^{|\mu+\nu|+|\delta+\xi+\eta|+1}\left(\frac{\delta!^2\mu!\nu!}{\xi!\eta!}\right)^{a+{\varepsilon}-1}
\binom{\alpha+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}
\binom{\beta+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}
\\
&
\qquad
\leq C^{|\mu+\nu|+|\delta+\xi+\eta|+1}\left(\frac{\delta!^2\mu!\nu!}{\xi!\eta!}\right)^{a+{\varepsilon}-1}.
\end{split}
\end{aligned}$$ The last inequality follows from the fact that $\alpha,\beta\leq \delta$, $$\begin{aligned}
\binom{\alpha+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}},\quad
\binom{\beta+|\xi+\eta|\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}
\leq 2^{|\delta|+|\xi+\eta|n}.
\end{aligned}$$ Similarly, for any $\mu,\nu\geq 0$, we also have $$\begin{aligned}
\left| \frac{1}{l!}\frac{D_v^{\mu} D_{{\overline{v}}}^{\nu}}{\mu!\nu!}\left((D_y\cdot D_\theta)^l\Delta_0(x,x,\psi_x(x,z))\right)\right|
\leq \sum_{|\delta|=l}\delta! C^{|\mu+\nu|+|\delta|+1}\left(\delta!^2\mu!\nu!\right)^{a+{\varepsilon}-1}
\end{aligned}$$ Then implies the following inequality $$\begin{aligned}
\label{diagonal 2}
\begin{split}
\left|D_v^\mu D_{{\overline{v}}}^\nu b_m(x,z)\right|
&\leq\sum_{l=1}^m
\sum_{|\delta|=l}\delta!^{2a+2{\varepsilon}-1}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi+\eta|\leq |\alpha+\beta|}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu} \frac{\left|D_v^{\widetilde{\xi}+\mu_0} D_{{\overline{v}}}^{\widetilde{\eta}+\nu_0} b_{m-l}(x,z)\right|}{(\xi!\eta!)^{a+{\varepsilon}}}
\\&
\quad\quad \cdot
C^{|\mu-\mu_0+\nu-\nu_0|+|\delta+\xi+\eta|}\binom{\mu}{\mu_0}\binom{\nu}{\nu_0}
(\mu-\mu_0)!^{a+{\varepsilon}}(\nu-\nu_0)!^{a+{\varepsilon}}.
\end{split}
\end{aligned}$$
Now we will change all the derivatives $\left|D_v^\mu D_{{\overline{v}}}^\nu b_m(x,z)\right|$ to the notation $b_{m,\mu\bar{\nu}}$ in the above inequality. Note that on the right hand side of , when $\nu\neq 0$, the anti-holomorphic derivative will hit on either $b_{m-l}(x,z)$ or at least one of the these functions $z_i(x,y,\theta), \Delta_0(x,y,\theta) \in \mathcal{A}^{a,{\varepsilon}}_{\theta}$ and $\varphi_i(x,z)\in \mathcal{A}^{a,{\varepsilon}}_{diag}$ for $1\leq i\leq n$. We will consider each case in the following. If the anti-holomorphic derivative hits on $z_i(x,y,\theta)$ or $\Delta_0(x,y,\theta)$, since the derivatives are evaluated at $(x,x,\psi_x(x,z))$, we will have the extra factor $\exp(-b|x-\bar{z}|^{-\frac{1}{a-1}})$ on the right hand side. If the anti-holomorphic derivative hits on $\varphi_i(x,z)$, we also have the extra factor $\exp(-b|x-\bar{z}|^{-\frac{1}{a-1}})$ since $\varphi_i(x,z)\in \mathcal{A}^{a,{\varepsilon}}_{diag}$. The last case is that the anti-holomorphic derivative hits on $b_{m-l}(x,z)$, which means $|\eta+\nu_0|\neq 0$. We again have the extra factor $\exp(-b|x-\bar{z}|^{-\frac{1}{a-1}})$ when we change $D_z^{\xi+\mu_0}D_{\bar{z}}^{\eta+\nu_0}b_{m-l}(x,z)$ into $b_{m-l,\widetilde{\xi}+\mu_0{\overline{\widetilde{\eta}+\nu_0}}}$. So no matter in which case, at least one $\exp(-b|x-\bar{z}|^{-\frac{1}{a-1}})$ will appear on the right side when $\nu\neq 0$. And thus the desired result follows.
Next we use this lemma to prove Theorem \[MainLemma\].
Proof of Theorem \[MainLemma\]
------------------------------
For convenience, we define $$a_{m,\mu\bar{\nu}}=\frac{b_{m,\mu\bar{\nu}}}{(2m+1)!^{a+{\varepsilon}}\mu!^{a+{\varepsilon}}\nu!^{a+{\varepsilon}}}.$$ Then by Lemma \[bmLemmaGevrey\] $$\begin{aligned}
\label{amrecursiveGevrey}
\begin{split}
&a_{m,\mu\bar{\nu}}
\\
&\leq\sum_{l=1}^m
\sum_{|\delta|=l}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi+\eta|\leq |\alpha+\beta|}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu} \frac{ a_{m-l,\widetilde{\xi}+\mu_0{\overline{\widetilde{\eta}+\nu_0}}}}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}
C^{|\mu-\mu_0+\nu-\nu_0+|\delta+\xi+\eta|},
\end{split}
\end{aligned}$$ where $\xi,\eta\in (\mathbb{Z}^{\geq 0})^n$ and $\widetilde{\xi}=(0,\cdots 0,\xi), \widetilde{\eta}=(0, \cdots, 0,\eta)\in (\mathbb{Z}^{\geq 0})^{2n}$. Since $b_0(x,z)=1$, we have $$\label{a0Gevrey}
a_{0,\mu\bar{\nu}}=
\begin{cases}
1 & \mu=\nu=(0,0,\cdots,0),\\
0 & \mbox{otherwise}.
\end{cases}$$ We will argue by induction on $m$ to prove that for any integer $m\geq 0$ and multi-index $\mu,\nu\geq 0$, $$\label{amupperboundGevrey}
a_{m,\mu\bar{\nu}}\leq\binom{2m+|\mu+\nu|}{|\mu+\nu|}^{a+{\varepsilon}} A^{m}(2C)^{|\mu+\nu|},$$ where $C$ is the same constant which appears on the right hand side of and $A$ is a bigger constant to be selected later. Without losing of generality, we assume $C\geq 1$. Obviously implies that holds for $m=0$ and any $\mu,\nu\geq 0$. Assume that holds up to $m-1$ and we proceed to $m$. By , we have $$\begin{aligned}
a_{m,\mu\bar{\nu}}
&\leq\sum_{l=1}^m
\sum_{|\delta|=l}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi+\eta|\leq |\alpha+\beta|}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
\frac{A^{m-l}}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}
\\&
\qquad \cdot
2^{|\xi+\eta|+|\mu_0+\nu_0|}C^{|\mu+\nu|+|\delta+2\xi+2\eta|} \binom{2(m-l)+|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}{|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}^{a+{\varepsilon}}
\\
& \leq\sum_{l=1}^m
\sum_{|\delta|=l}
\sum_{|\xi+\eta|\leq 2l}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
\frac{A^{m-l}}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}
\\&
\qquad\cdot
2^{|\xi+\eta|+|\mu_0+\nu_0|}C^{|\mu+\nu|+|\delta+2\xi+2\eta|}
\binom{2(m-l)+|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}{|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}^{a+{\varepsilon}}
\cdot\#\{\alpha\leq \delta\} \cdot \#\{\beta\leq \delta\}.
\end{aligned}$$
Due to the fact $$\#\{\alpha\leq \delta\}
=\#\{\beta\leq \delta\}
\leq 2^{|\delta|},$$ it follows that $$\begin{aligned}
a_{m,\mu\bar{\nu}}
& \leq\sum_{l=1}^m
\sum_{|\delta|=l}
\sum_{|\xi+\eta|\leq 2l}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
\frac{A^{m-l}}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}
\\&
\qquad\cdot
2^{|\xi+\eta|+|\mu_0+\nu_0|+2|\delta|}C^{|\mu+\nu|+|\delta+2\xi+2\eta|}
\binom{2(m-l)+|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}{|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}^{a+{\varepsilon}}
\\
&\leq
A^m(2C)^{|\mu+\nu|}\sum_{l=1}^m
\sum_{|\delta|=l}
\sum_{|\xi+\eta|\leq 2l}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
\frac{1}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}A^{-l}
\\&
\qquad\cdot
2^{4l+|\mu_0+\nu_0|-|\mu+\nu|}C^{5l}
\binom{2(m-l)+|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}{|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}^{a+{\varepsilon}}
\end{aligned}$$
Moreover, since $$\begin{aligned}
\#\{|\delta|=l\}=\binom{l+n-1}{n-1}\leq 2^{l+n-1}\leq 2^{nl},
\end{aligned}$$ we have $$\begin{aligned}
\label{am2Gevrey}
\begin{split}
a_{m,\mu\bar{\nu}}
&\leq
A^m(2C)^{|\mu+\nu|}\sum_{l=1}^m
\sum_{|\xi+\eta|\leq 2l}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
\frac{1}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{\widetilde{\xi}+\mu_0}{\widetilde{\xi}}^{a+{\varepsilon}}
\binom{\widetilde{\eta}+\nu_0}{\widetilde{\eta}}^{a+{\varepsilon}}
\\&
\qquad\cdot
\binom{2(m-l)+|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}{|\widetilde{\xi}+\mu_0+\widetilde{\eta}+\nu_0|}^{a+{\varepsilon}} 2^{|\mu_0+\nu_0|-|\mu+\nu|}\left(\frac{2^{n+4}C^5}{A}\right)^l
\end{split}
\end{aligned}$$
In the next step we apply the combinatorial inequality $$\binom{\widetilde{\xi}+\mu_0}{\mu_0}\binom{\widetilde{\eta}+\nu_0}{\nu_0}
\leq \binom{\widetilde{\xi}+\widetilde{\eta}+\mu_0+\nu_0}{\mu_0+\nu_0}
\leq \binom{|\xi+\eta|+|\mu_0+\nu_0|}{|\mu_0+\nu_0|}$$ and $$\begin{aligned}
\binom{2(m-l)+|\xi+\eta|+|\mu_0+\nu_0|}{|\xi+\eta|+|\mu_0+\nu_0|}\binom{|\xi+\eta|+|\mu_0+\nu_0|}{|\mu_0+\nu_0|}
=\binom{2(m-l)+|\xi+\eta|+|\mu_0+\nu_0|}{|\mu_0+\nu_0|}\binom{2m-2l+|\xi+\eta|}{2m-2l}.
\end{aligned}$$ Observe that, since $|\xi+\eta|\leq 2l, \mu_0\leq \mu,\nu_0\leq\nu$, we have $$\begin{aligned}
\binom{2(m-l)+|\xi+\eta|+|\mu_0+\nu_0|}{|\mu_0+\nu_0|}
\leq \binom{2m+|\mu+\nu|}{|\mu+\nu|}.
\end{aligned}$$ Plugging these into , we obtain $$\begin{aligned}
\begin{split}
a_{m,\mu\bar{\nu}}
&\leq
A^m(2C)^{|\mu+\nu|}\sum_{l=1}^m
\frac{1}{\binom{2m+1}{2l}^{a+{\varepsilon}}}
\binom{2m+|\mu+\nu|}{|\mu+\nu|}^{a+{\varepsilon}}
\sum_{|\xi+\eta|\leq 2l}
\binom{2m-2l+|\xi+\eta|}{2m-2l}^{a+{\varepsilon}}
\\&
\qquad\qquad\qquad\quad\cdot
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu}
2^{|\mu_0+\nu_0|-|\mu+\nu|}\left(\frac{2^{n+4}C^5}{A}\right)^l
\end{split}
\end{aligned}$$
Again since $$\#\{|\xi+\eta|=k\}=\binom{k+2n-1}{2n-1}\leq 2^{k+2n-1},$$ the sum over index $\xi,\eta$ on the right hand side can be estimated as $$\begin{aligned}
\sum_{|\xi+\eta|\leq 2l} \binom{2m-2l+|\xi+\eta|}{2m-2l}^{a+{\varepsilon}}
=\sum_{k=0}^{2l}\sum_{|\xi+\eta|=k} \binom{2m-2l+k}{2m-2l}^{a+{\varepsilon}}
\leq 2^{2l+2n-1}\binom{2m+1}{2m-2l+1}^{a+{\varepsilon}}.
\end{aligned}$$ Therefore, $$\begin{aligned}
a_{m,\mu\bar{\nu}}
\leq&
A^m(2C)^{|\mu+\nu|} \binom{2m+|\mu+\nu|}{|\mu+\nu|}^{a+{\varepsilon}}
\sum_{l=1}^m
\left(\frac{2^{3n+6}C^5}{A}\right)^l
\sum_{\mu_0\leq \mu}2^{|\mu_0|-|\mu|}\sum_{\nu_0\leq \nu}
2^{|\nu_0|-|\nu|}
\\
\leq&
A^m(2C)^{|\mu+\nu|} \binom{2m+|\mu+\nu|}{|\mu+\nu|}^{a+{\varepsilon}}
\sum_{l=1}^m
\left(\frac{2^{3n+6}C^5}{A}\right)^l2^{4n}
\end{aligned}$$ By taking $A= 2^{7n+7}C^5$, we surely have $\sum_{l=1}^m\left(\frac{2^{3n+6}C^5}{A}\right)^l2^{4n}<1$, which implies that $$a_{m,\mu\bar{\nu}}\leq A^m(2C)^{|\mu+\nu|}\binom{2m+|\mu+\nu|}{|\mu+\nu|}^{a+{\varepsilon}}.$$ So if we write $a_{m,\mu\bar{\nu}}$ in terms of $D_v^\mu D_{{\overline{v}}}^{\nu}b_m(x,z)$, then by the continuity of each $D_v^\mu D_{{\overline{v}}}^{\nu}b_m(x,z)$ for all $x,z\in U$, we have $$\begin{aligned}
\begin{split}
|D_v^\mu D_{{\overline{v}}}^{\nu}b_m(x,z)|&\leq (2m+1)!^{a+{\varepsilon}}\mu!^{a+{\varepsilon}}\nu!^{a+{\varepsilon}} a_{m,\mu\bar{\nu}}
\lambda_{b,|\nu|}(x,x,z)
\\
&\leq (64^{a+{\varepsilon}}A)^{m+|\mu+\nu|}m!^{2a+2{\varepsilon}}\mu!^{a+{\varepsilon}}\nu!^{a+{\varepsilon}}
\lambda_{b,|\nu|}(x,x,z).
\end{split}
\end{aligned}$$ Thus follows by renaming $64^{a+{\varepsilon}}A$ to $C$.
In particular, when we are restricted to diagonal $z=\bar{x}$, $$D_v^\mu D_{{\overline{v}}}^{\nu}b_m(x,\bar{x})=0, \qquad \mbox{ for any multi-indices $\mu\geq 0$ and $|\nu|\neq 0$}.$$ And thus when $z=\bar{x}$, the recursive inequality reduces to $$\begin{aligned}
\left|D_v^\mu b_m(x,\bar{x})\right|
\leq\sum_{l=1}^m
\sum_{|\delta|=l}\delta!^{2a+2{\varepsilon}-1}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi|\leq |\alpha+\beta|}
\frac{\left|D_v^{\widetilde{\xi}+\mu_0} b_{m-l}(x,\bar{x})\right|}{(\xi!\eta!)^{a+{\varepsilon}}}
C^{|\mu-\mu_0|+|\delta+\xi|}\binom{\mu}{\mu_0}
(\mu-\mu_0)!^{a+{\varepsilon}}.
\end{aligned}$$ Note that the constant ${\varepsilon}$ only comes from and because of the derivatives of $\varphi_i(x,z)\in \mathcal{A}^{a,{\varepsilon}}_{diag}$ for $1\leq i\leq n$ and $z(x,y,\theta), \Delta_0(x,y,\theta)\in \mathcal{A}^{a,{\varepsilon}}_\theta$. By the definitions of $\mathcal{A}^{a,{\varepsilon}}_{diag}$ and $\mathcal{A}^{a,{\varepsilon}}_\theta$, ${\varepsilon}$ can be replaced by $0$ when we are restricted to $x=y=\bar{z}$. Therefore, further reduces to $$\begin{aligned}
\left|D_v^\mu b_m(x,\bar{x})\right|
\leq\sum_{l=1}^m
\sum_{|\delta|=l}\delta!^{2a-1}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi|\leq |\alpha+\beta|}
\frac{\left|D_v^{\widetilde{\xi}+\mu_0} b_{m-l}(x,\bar{x})\right|}{(\xi!\eta!)^{a+{\varepsilon}}}
C^{|\mu-\mu_0|+|\delta+\xi|}\binom{\mu}{\mu_0}
(\mu-\mu_0)!^{a+{\varepsilon}}.
\end{aligned}$$ By using a similar inductive argument as that of estimating $\left|D_v^\mu D_{{\overline{v}}}^\nu b_m(x,z)\right|$, we obtain for any $x\in U$, $$\begin{aligned}
\left|D_v^\mu b_m(x,\bar{x})\right|\leq C^{m+|\mu|}m!^{2a}\mu!^{a}.
\end{aligned}$$
Optimality of the upper bounds on Bergman coefficients $b_m$ {#optimal}
============================================================
In this section, we will show that although it would be desirable to improve the estimate to $$\label{bmconjecture}
\left | D_v^\alpha D_{{\overline{v}}} ^\beta b_m(x,z) \right | \leq C^{m+|\alpha|+|\beta|}m!^{2a+2{\varepsilon}-1}\alpha!^{a+{\varepsilon}} \beta!^{a+{\varepsilon}} \exp{\left(-b (1- \delta_0(|\beta|)){|x-\bar{z}|^{-\frac{1}{a-1}}}\right)}\lambda_{b,|\eta|}(x,x,z),$$ it is not possible to prove it by the recursive inequality . Here we provide an example which satisfies while fails . For simplicity, we assume $C=1$ in . Let’s consider the worse case when the equality holds, i.e. $$\begin{aligned}
\label{bmoptimalGevrey}
\begin{split}
&b_{m,\mu\bar{\nu}}
\\
&=\sum_{l=1}^m
\sum_{|\delta|=l}\delta!^{2a+2{\varepsilon}-1}
\sum_{\alpha,\beta\leq \delta}
\sum_{|\xi+\eta|\leq |\alpha+\beta|}
\sum_{\mu_0\leq \mu}\sum_{\nu_0\leq \nu} \frac{ b_{m-l,\widetilde{\xi}+\mu_0{\overline{\widetilde{\eta}+\nu_0}}}}{(\xi!\eta!)^{a+{\varepsilon}}}
\binom{\mu}{\mu_0}\binom{\nu}{\nu_0}
(\mu-\mu_0)!^{a+{\varepsilon}}(\nu-\nu_0)!^{a+{\varepsilon}},
\end{split}\end{aligned}$$ One can easily check that this recursive equation uniquely defines $\{b_{m,\mu\bar{\nu}} \}$ given an initial data $\{ b_{0, \mu\bar{\nu}} \}$. We shall only focus on the terms $b_{m,k\widetilde{e_1}}$ where $e_1=(1,0,\cdots,0)\in \mathbb{R}^n$, $\widetilde{e}_1=(0,\cdots, 0, e_1)\in \mathbb{R}^{2n}$ and show by induction that $$\label{bmLB}
b_{m,k\widetilde{e_1}}\geq 2^{-(a+{\varepsilon})m}(2m-2+k)!^{a+{\varepsilon}} \hspace{12pt} \mbox{ for any }m\geq 1,k\geq 0.$$ First let’s check $b_{1,k\widetilde{e_1}}$. Since we know $$b_{0,\mu\bar{\nu}}=
\begin{cases}
1 & \mu=\nu=(0,0,\cdots,0),\\
0 & \mbox{otherwise},
\end{cases}$$ by we have $$\begin{aligned}
\label{eq 3.15}
b_{1,\mu\bar{\nu}}
=
\sum_{|\delta|=1}
\sum_{\alpha,\beta\leq \delta}
\mu!^{a+{\varepsilon}}\nu!^{a+{\varepsilon}}
\geq \mu!^{a+{\varepsilon}}\nu!^{a+{\varepsilon}}.\end{aligned}$$ Therefore holds for $b_{1,k\widetilde{e_1}}$. Assume that holds for $b_{1,k\widetilde{e_1}},b_{2,k\widetilde{e_1}},\cdots b_{m-1,k\widetilde{e_1}}$. Then by only considering the terms with index $l=|\alpha|=|\beta|=1$, $\mu_0=\mu$ and $\xi=2e_1$, $\eta=0$ in , we obtain for $m\geq 2$ $$\begin{aligned}
b_{m,k\widetilde{e}_1}
\geq
\sum_{|\delta|=1}
\sum_{|\alpha|=|\beta|=1}
\frac{ b_{m-1,(k+2)\widetilde{e_1}}}{2^{a+{\varepsilon}}}
\geq
\frac{ b_{m-1,(k+2)\widetilde{e_1}}}{2^{a+{\varepsilon}}}
\geq
2^{-(a+{\varepsilon})m}(2m-2+k)!^{a+{\varepsilon}}. \end{aligned}$$ Note that if in particular we put $k=0$ into , then we get $$\begin{aligned}
b_{m,0}\geq \left(\frac{1}{8}\right)^{(a+{\varepsilon})m}m!^{2a+2{\varepsilon}},\end{aligned}$$ which show that up to an exponential factor $C^m$, $m!^{2a+2{\varepsilon}}$ is the best upper bound one can hope from the recursive inequality .
Proofs of main lemmas on almost holomorphic extensions of Gevrey functions {#ProofOfGevreyStuff}
==========================================================================
In this section, we will complete all the proofs skipped in Section \[AHEoGF\].
We will prove the estimate on $ D_{\bar{z}}F(f)$. The other one follows in the same way. For simplicity, we denote $$\widetilde{\chi}(|\alpha+\beta|)=\chi\big(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\big).$$ For any $1\leq i\leq n$, we have $$\begin{aligned}
D_{\bar{z}_i}&F(f)\\
&=\frac{1}{2}\sum_{\alpha,\beta\geq 0} \frac{ D_x^{\alpha+e_i} D_{\bar{x}}^{\beta} f}{\alpha!\beta!}\left(\frac{y-\bar{z}}{2}\right)^\alpha
\left(\frac{z-\bar{y}}{2}\right)^\beta \Big(\widetilde{\chi}(|\alpha+\beta|)-\widetilde{\chi}\left(|\alpha+\beta+e_i|\right)\Big)
\\&
+2\sum_{\alpha,\beta\geq 0} \frac{ D_x^\alpha D_{\bar{x}}^\beta f}{\alpha!\beta!}\left(\frac{y-\bar{z}}{2}\right)^\alpha \left(\frac{z-\bar{y}}{2}\right)^{\beta+e_i}|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\chi'\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right).
\\
&:=I+II.
\end{aligned}$$ We will use $C$ to denote a constant depend on $a, {\varepsilon}, C_1(f)$ and the cut-off function $\chi$, which may change from line to line. Since $f\in G^a(U)$ and by Stirling’s formula, each term in $I$ is bounded by $$\begin{aligned}
C_0C_1^{|\alpha+\beta|+1}(\alpha_i+1)^a&(\alpha!\beta!)^{a-1}\left|\frac{y-\bar{z}}{2}\right|^{|\alpha+\beta|}
\\
&\leq
CC_0C_1^{|\alpha+\beta|}\left(\left|\alpha+\beta\right|+1\right)^{a+\frac{a-1}{2}}\left(\frac{|\alpha+\beta|}{e}\right)^{(a-1)|\alpha+\beta|}\left|\frac{y-\bar{z}}{2}\right|^{|\alpha+\beta|}.
\end{aligned}$$
Note that the difference of cut-off functions in $I$ is zero unless $$\label{IndexRange}
\left|\alpha+\beta\right|\in \left[\frac{1}{2}\left(\sqrt{2}C_1|y-\bar{z}|\right)^{-\frac{1}{a-1}}-1,\frac{1}{2}\left(C_1|y-\bar{z}|\right)^{-\frac{1}{a-1}}\right].$$ It implies that each term in $I$ is bounded by $$CC_0(|\alpha+\beta|+1)^{a+\frac{a-1}{2}}2^{-|\alpha+\beta|}e^{-(a-1)|\alpha+\beta|}
\leq
CC_0e^{-\frac{a-1}{2}(\sqrt{2}C_1|y-\bar{z}|)^{-\frac{1}{a-1}}}.$$ Since there are less than $\left(\frac{1}{2}(C_1|y-\bar{z}|)^{-\frac{1}{a-1}}+1\right)^{2n}$ many terms in $I$, we have $$|I|\leq CC_0 e^{-b|y-\bar{z}|^{-\frac{1}{a-1}}},$$ where $b$ is a positive constant depending on $a,C_1=C_1(f)$.
For the second term $II$, similarly we have or $\chi'$ vanishes otherwise. And thus each term is bounded by $$\begin{aligned}
CC_0 C_1&^{|\alpha+\beta|+2}\left(\frac{|\alpha+\beta|}{e}\right)^{(a-1)|\alpha+\beta|}\left|\frac{y-\bar{z}}{2}\right|^{|\alpha+\beta|+1}|\alpha+\beta|^{\frac{5}{2}(a-1)}
\\
\leq &
CC_0 e^{-(a-1)|\alpha+\beta|}|\alpha+\beta|^{\frac{3}{2}(a-1)}
\\
\leq &
CC_0 e^{-\frac{a-1}{4}(\sqrt{2}C_1|y-\bar{z}|)^{-\frac{1}{a-1}}}
\\
\leq &
CC_0 e^{-b|y-\bar{z}|^{-\frac{1}{a-1}}}.
\end{aligned}$$ So we have $$| D_{\bar{z}_i}F(f)(y,z)|\leq CC_0 \exp\left({-b|y-\bar{z}|^{-\frac{1}{a-1}}}\right) \qquad \mbox{for } 1\leq i\leq n.$$ Thus the result follows.
Since $F$ is an almost holomorphic extension of $f$, by taking the Taylor expansion, for any $N\in \mathbb{N}$ we have $$\begin{aligned}
F(x+y,\bar{x}+z)&=\sum_{|\alpha+\beta|\leq N-1}\frac{D_y^{\alpha} D_z^{\beta}F}{\alpha!\beta!}\left(x,\bar{x}\right)y^\alpha z^\beta+O\left(|(y,z)|^{N}\right)\\
&=\sum_{|\alpha+\beta|\leq N-1}\frac{D_x^\alpha D_{\bar{x}}^{\beta}f}{\alpha!\beta!}\left(x\right)y^\alpha z^\beta+O\left(|(y,z)|^{N}\right).
\end{aligned}$$ If we take $x=\frac{y+\bar{z}}{2}$ and replace $(y,z)$ by $(\frac{y-\bar{z}}{2},\frac{z-\bar{y}}{2})$, then $$F(y,z)=\sum_{|\alpha+\beta|\leq N-1}\frac{D_x^\alpha D_{\bar{x}}^{\beta}f}{\alpha!\beta!}\left(\frac{y+\bar{z}}{2}\right)\left(\frac{y-\bar{z}}{2}\right)^\alpha \left(\frac{z-\bar{y}}{2}\right)^\beta+O\left(\left|y-\bar{z}\right|^{N}\right).$$ Similarly, the same identity holds for $\widetilde{F}(y,z)$. Therefore, for any $N\in \mathbb{N}$, we have $$F(y,z)-\widetilde{F}(y,z)=O\left(|y-\bar{z}|^N\right).$$
To prove this lemma we first need to obtain some estimates on the derivatives of our cut-off function $\chi$.
Let ${\varepsilon}>0$ be a constant and $\chi\in G^{1+{\varepsilon}}(\mathbb{R})$ be the cut-off function constructed in . Then there exists some positive constant $C=C(\chi)$ such that for any multi-indices $\gamma,\delta,\xi,\eta\geq 0$, we have $$\begin{aligned}
\| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^\eta &\left(\chi\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right)\right)\|_{L^\infty(\mathbb{C}^{2n})}
\\
& \leq \left(2^{a-1}CC_1|\alpha+\beta|^{a-1}\right)^{|\gamma+\eta+\xi+\delta|}
(\gamma+\eta+\xi+\delta)!^{1+{\varepsilon}}.
\end{aligned}$$
By a straightforward calculation, we have $$\begin{aligned}
D_y^\gamma \left(\chi\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right)\right)
=\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2(\bar{y}-z)\right)^\gamma\chi^{(|\gamma|)},
\end{aligned}$$ $$\begin{aligned}
D_y^\gamma D_{\bar{z}}^\eta \left(\chi\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right)\right)
=(-1)^{|\eta|}\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2(\bar{y}-z)\right)^{\gamma+\eta}
\chi^{(|\gamma+\eta|)}.
\end{aligned}$$ Therefore, $$\begin{aligned}
& D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^\eta \left(\chi((\alpha+\beta)^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2)\right)
\\
=&\sum_{\xi_0\leq \xi,\delta_0\leq \delta}\binom{\xi}{\xi_0}\binom{\delta}{\delta_0}(-1)^{|\eta+\xi_0+\delta-\delta_0|}\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\right)^{|\gamma+\eta+\xi_0+\delta_0|}
\\&
\frac{(\gamma+\eta)!}{(\gamma+\eta+\xi_0+\delta_0-\xi-\delta)!}
(\bar{y}-z)^{\gamma+\eta-(\xi-\xi_0)-(\delta-\delta_0)}
(\bar{z}-y)^{\xi_0+\delta_0}
\chi^{(|\gamma+\eta+\xi_0+\delta_0|)}.
\end{aligned}$$ Since the cut-off function $\chi$ is in $G^{1+{\varepsilon}}(R)$, there exists a positive constant $C=C(\chi)$, which may vary from line to line, such that $$\begin{aligned}
&\left| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^\eta \left(\chi(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2)\right)\right|
\\
\leq&\sum_{\xi_0\leq \xi,\delta_0\leq \delta}\binom{\xi}{\xi_0}\binom{\delta}{\delta_0}\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\right)^{|\gamma+\eta+\xi_0+\delta_0|}C^{|\gamma+\eta+\xi_0+\delta_0|}
\\&
\frac{(\gamma+\eta)!}{(\gamma+\eta+\xi_0+\delta_0-\xi-\delta)!}
\left|\bar{y}-z\right|^{|\gamma+\eta-\xi+2\xi_0-\delta+2\delta_0|}
|\gamma+\eta+\xi_0+\delta_0|!^{1+{\varepsilon}}
\\
\leq&C^{|\gamma+\delta+\xi+\eta|}(\gamma+\eta+\xi+\delta)!^{1+{\varepsilon}}\\
&\sum_{\xi_0\leq \xi,\delta_0\leq \delta}\binom{\xi}{\xi_0}\binom{\delta}{\delta_0}\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\right)^{|\gamma+\eta+\xi_0+\delta_0|}
\left|\bar{y}-z\right|^{|\gamma+\eta-\xi+2\xi_0-\delta+2\delta_0|}.
\end{aligned}$$ Our result follows by using that for any $y,z\in\mathbb{C}^n$, $$|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\leq 1.$$
We now estimate the derivatives on $F(f)$. By a straightforward calculation, we have $$\begin{aligned}
& D_y^\gamma D_z^\delta D_{\bar{y}}^{\xi} D_{\bar{z}}^{\eta} F(f)(y,z)
\\
=&\sum_{\substack{\gamma_0+\gamma_1+\gamma_2=\gamma\\\delta_0+\delta_1+\delta_2=\delta}}
\sum_{\substack{\xi_0+\xi_1+\xi_2=\xi\\\eta_0+\eta_1+\eta_2=\eta}}
\sum_{\substack{\alpha\geq \gamma_1+\eta_1,\\ \beta\geq \xi_1+\delta_1}}
\binom{\gamma}{\gamma_0,\gamma_1,\gamma_2}
\binom{\delta}{\delta_0,\delta_1,\delta_2}
\binom{\xi}{\xi_0,\xi_1,\xi_2}
\binom{\eta}{\eta_0,\eta_1,\eta_2}
\\&
\cdot \left(\frac{1}{2}\right)^{|\gamma_0+\gamma_1+\delta_0+\delta_1+\xi_0+\xi_1+\eta_0+\eta_1|}
(-1)^{|\xi_1+\eta_1|}
\frac{ D_x^{\alpha+\gamma_0+\eta_0} D_{\bar{x}}^{\beta+\delta_0+\xi_0} f}{(\alpha-\gamma_1-\eta_1)!(\beta-\delta_1-\xi_1)!}
\\
&
\cdot
\left(\frac{y-\bar{z}}{2}\right)^{\alpha-\gamma_1-\eta_1} \left(\frac{z-\bar{y}}{2}\right)^{\beta-\delta_1-\xi_1}
D_y^{\gamma_2} D_z^{\delta_2} D_{\bar{y}}^{\xi_2} D_{\bar{z}}^{\eta_2} \left(\chi\left(|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\right)\right).
\end{aligned}$$ Let $C_0=C_0(f)$ introduced in Definition \[Gevrey\]. We use $C$ to denote a constant which depends on $a$, $C_1(f)$ introduced in Definition \[Gevrey\] and the cut-off function $\chi$, which may vary from line to line. By the fact that $f\in G^a(U)$ and the previous lemma, it follows that $$\begin{aligned}
&\left| D_y^\gamma D_z^\delta D_{\bar{y}}^{\xi} D_{\bar{z}}^{\eta} F(f)(y,z)\right|
\\
&\leq
C_0C^{|\gamma+\delta+\xi+\eta|}
\sum_{\substack{\gamma_0+\gamma_1+\gamma_2=\gamma\\\delta_0+\delta_1+\delta_2=\delta}}
\sum_{\substack{\xi_0+\xi_1+\xi_2=\xi\\\eta_0+\eta_1+\eta_2=\eta}}
\sum_{\substack{\alpha\geq \gamma_1+\eta_1\\\beta\geq \xi_1+\delta_1}}
\binom{\gamma}{\gamma_0,\gamma_1,\gamma_2}
\binom{\delta}{\delta_0,\delta_1,\delta_2}
\binom{\xi}{\xi_0,\xi_1,\xi_2}
\binom{\eta}{\eta_0,\eta_1,\eta_2}
\\
& \qquad\cdot C_1^{|\alpha+\beta|}\frac{(\alpha+\gamma_0+\eta_0)!^a(\beta+\delta_0+\xi_0)!^a}{(\alpha-\gamma_1-\eta_1)!(\beta-\delta_1-\xi_1)!}\left|\frac{y-\bar{z}}{2}\right|^{|\alpha-\gamma_1-\eta_1+\beta-\delta_1-\xi_1|}
\\
&\qquad \cdot
(\gamma_2+\eta_2+\xi_2+\delta_2)!^{1+{\varepsilon}}
\left(2^{a-1}|\alpha+\beta|^{a-1}C_1\right)^{|\gamma_2+\eta_2+\xi_2+\delta_2|}.
\end{aligned}$$ Using the fact $|\alpha+\beta|^{2(a-1)}4^{a-1}C_1^2\left|y-\bar{z}\right|^2\leq 1$ and Stirling’s formula, it is bounded by $$\begin{aligned}
&
C_0C^{|\gamma+\delta+\xi+\eta|}
\sum_{\substack{\gamma_0+\gamma_1+\gamma_2=\gamma\\\delta_0+\delta_1+\delta_2=\delta}}
\sum_{\substack{\xi_0+\xi_1+\xi_2=\xi\\\eta_0+\eta_1+\eta_2=\eta}}
\sum_{\substack{\alpha\geq \gamma_1+\eta_1\\\beta\geq \xi_1+\delta_1}}
\binom{\gamma}{\gamma_0,\gamma_1,\gamma_2}
\binom{\delta}{\delta_0,\delta_1,\delta_2}
\binom{\xi}{\xi_0,\xi_1,\xi_2}
\binom{\eta}{\eta_0,\eta_1,\eta_2}
\\
&\cdot \frac{(\alpha+\gamma_0+\eta_0)!^a(\beta+\delta_0+\xi_0)!^a}{(\alpha-\gamma_1-\eta_1)!(\beta-\delta_1-\xi_1)!}\left(\frac{C_1}{2}\right)^{|\alpha+\beta|} (\gamma_2+\eta_2+\xi_2+\delta_2)!^{1+{\varepsilon}}
\\
&\cdot
\left(2^{a-1}|\alpha+\beta|^{a-1}C_1\right)^{|\gamma_2+\eta_2+\xi_2+\delta_2+\gamma_1+\delta_1+\xi_1+\eta_1|-|\alpha+\beta|}
\\
\leq&
C_0C^{|\gamma+\delta+\xi+\eta|}
\sum_{\substack{\gamma_0+\gamma_1+\gamma_2=\gamma\\\delta_0+\delta_1+\delta_2=\delta}}
\sum_{\substack{\xi_0+\xi_1+\xi_2=\xi\\\eta_0+\eta_1+\eta_2=\eta}}
\sum_{\substack{\alpha\geq \gamma_1+\eta_1\\\beta\geq \xi_1+\delta_1}}
\binom{\gamma}{\gamma_0,\gamma_1,\gamma_2}
\binom{\delta}{\delta_0,\delta_1,\delta_2}
\binom{\xi}{\xi_0,\xi_1,\xi_2}
\binom{\eta}{\eta_0,\eta_1,\eta_2}
\\
&\cdot
\frac{(\alpha+\gamma_0+\eta_0)!^a(\beta+\delta_0+\xi_0)!^a}{(\alpha-\gamma_1-\eta_1)!(\beta-\delta_1-\xi_1)!}2^{-a(|\alpha+\beta|)} (\gamma_2+\eta_2+\xi_2+\delta_2)!^{1+{\varepsilon}}
\\
&
\cdot
|\alpha+\beta|^{(a-1)|\gamma_2+\eta_2+\xi_2+\delta_2+\gamma_1+\delta_1+\xi_1+\eta_1|-(a-1)|\alpha+\beta|}
\\
\leq&
C_0C^{|\gamma+\delta+\xi+\eta|}
\sum_{\alpha,\beta\geq 0}\sum_{\substack{\gamma_0+\gamma_1+\gamma_2=\gamma\\\delta_0+\delta_1+\delta_2=\delta}}
\sum_{\substack{\xi_0+\xi_1+\xi_2=\xi\\\eta_0+\eta_1+\eta_2=\eta}}
\binom{\gamma}{\gamma_0,\gamma_1,\gamma_2}
\binom{\delta}{\delta_0,\delta_1,\delta_2}
\binom{\xi}{\xi_0,\xi_1,\xi_2}
\binom{\eta}{\eta_0,\eta_1,\eta_2}
\\&\cdot
(\gamma_0+\eta_0)!^a(\delta_0+\xi_0)!^a(\gamma_1+\eta_1)!(\delta_1+\xi_1)! (\gamma_2+\eta_2+\xi_2+\delta_2)!^{1+{\varepsilon}}
\\&
\cdot
e^{-(a-1)|\alpha+\beta|}
|\alpha+\beta|^{(a-1)(|\gamma_2+\eta_2+\xi_2+\delta_2+\gamma_1+\delta_1+\xi_1+\eta_1|+1)}
\end{aligned}$$
For any $\alpha,\beta\geq 0$, we have $$\begin{aligned}
|\alpha+\beta|&^{(a-1)(|\gamma_1+\delta_1+\xi_1+\eta_1+\gamma_2+\delta_2+\xi_2+\eta_2|+1)}e^{-\frac{1}{2}(a-1)|\alpha+\beta|}
\\
&\leq 2^{(a-1)(|\gamma_1+\delta_1+\xi_1+\eta_1+\gamma_2+\delta_2+\xi_2+\eta_2|+1)}(|\gamma_1+\delta_1+\xi_1+\eta_1+\gamma_2+\delta_2+\xi_2+\eta_2|+1)!^{a-1}.
\end{aligned}$$ Therefore, $$\begin{aligned}
| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(f)(y,z)|
\leq 2^{a-1}C_0C^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}}
\sum_{\alpha,\beta\geq 0}
e^{-\frac{1}{2}(a-1)|\alpha+\beta|}.
\end{aligned}$$ Note that when $\xi+\eta>0$, we have . So $$\begin{aligned}
| D_y^\gamma D_z^\delta D_{\bar{y}}^\xi D_{\bar{z}}^{\eta} F(f)(y,z)|
\leq 2^{a-1}C_0C^{|\gamma+\delta+\xi+\eta|}(\gamma!\delta!\xi!\eta!)^{a+{\varepsilon}}
e^{-\frac{a-1}{8}(\sqrt{2}C_1|y-\bar{z}|)^{-\frac{1}{a-1}}}
\sum_{\alpha,\beta\geq 0}
e^{-\frac{1}{4}(a-1)|\alpha+\beta|}.
\end{aligned}$$ The result follows as $\sum_{\alpha,\beta\geq 0}
e^{-\frac{1}{4}(a-1)|\alpha+\beta|}\leq \left(\frac{4}{a-1}\right)^{2n}e^{\frac{(a-1)n}{2}}$.
In addition, when $z=\bar{y}$, note all the derivatives of $\chi$ vanish and $|\chi|\leq 1$, whence we can replace ${\varepsilon}$ by zero.
It is easy to see that $\mathcal{A}^{a,{\varepsilon}}_\theta$ is closed under summation, subtraction and differentiation. Now we consider multiplication. Take $f,g \in \mathcal{A}^{a,{\varepsilon}}_{\theta}$. We will use $C_0(f),C_1(f),b(f)$ and $C_0(g),C_1(g),b(g)$ to denote the constants in corresponding to $f,g$ respectively. Take $C_1=\max\{C_1(f), C_1(g)\}$ and $b=\min\{b(f),b(g)\}$. Let $v'=(x,y,\theta)$. Then $$\begin{aligned}
&\left|D^{\alpha}_{v'}D^{\beta}_{\bar{v}'}(fg)(x,y,\theta(x,y,z))\right|
\\
\leq& \sum_{\alpha_0\leq \alpha,\beta_0\leq \beta} \binom{\alpha}{\alpha_0}\binom{\beta}{\beta_0} \left|D^{\alpha_0}_{v'}D^{\beta_0}_{\bar{v}'}f(x,y,\theta(x,y,z))\right| \left|D^{\alpha-\alpha_0}_{v}D^{\beta-\beta_0}_{\bar{v}'}g(x,y,\theta(x,y,z))\right|
\\
\leq &
\sum_{\alpha_0\leq \alpha,\beta_0\leq \beta}C_0(f)C_0(g) \binom{\alpha}{\alpha_0}\binom{\beta}{\beta_0} C_1^{|\alpha+\beta|}
(\alpha!\beta!)^{a+{\varepsilon}}
\lambda_{b,|\beta|}(x,y,z)
\\
= &
C_0(f)C_0(g)(2C_1)^{|\alpha+\beta|}
(\alpha!\beta!)^{a+{\varepsilon}}
\lambda_{b,|\beta|}(x,y,z).
\end{aligned}$$ In addition, when we are restricted to $x=y=\bar{z}$, it is easy to see that we can replace ${\varepsilon}$ by $0$. Therefore, $fg\in \mathcal{A}^{a,{\varepsilon}}_{\theta}$. And we can choose $C_0(fg)=C_0(f)C_0(g)$, $C_1(fg)=2\max\{C_1(f),C_1(g)\}$ and $b(fg)=\min\{b(f),b(g)\}$.
To prove $\mathcal{A}^{a,{\varepsilon}}_{\theta}$ is closed under division, we will verify that if $f\in \mathcal{A}^{a,{\varepsilon}}_{\theta}(U')$ and $\inf_{U'}|f|\geq C_2>0$, then the reciprocal $\frac{1}{f}\in \mathcal{A}^{a,{\varepsilon}}_{\theta}(U')$. Define $h(w)=\frac{1}{w}$ for $w\in \mathbb{C}\setminus\{0\}$. Then for any $v_0\in U'$, we have $$\begin{aligned}
D_{v'}^\alpha D_{\bar{v}'}^{\beta}(h\circ f)(v_0)=\alpha!\beta!\sum_{k=0}^{|\alpha+\beta|} \frac{(-1)^k}{f(v_0)^{k+1}}\sum_{\substack{\alpha_1+\alpha_2+\cdots+\alpha_k=\alpha\\\beta_1+\beta_2+\cdots+\beta_k=\beta\\\alpha_1+\beta_1>0,\cdots,\alpha_k+\beta_k>0}}\frac{D_{v'}^{\alpha_1}D_{\bar{v}'}^{\beta_1}f(v_0)}{\alpha_1!\beta_1!}\cdots \frac{D_{v'}^{\alpha_k}D_{\bar{v}'}^{\beta_k}f(v_0)}{\alpha_k!\beta_k!}.
\end{aligned}$$ We use $C_0=C_0(f), C_1=C_1(f)$ and $b=b(f)$ to denote the constants in for $f$. Without losing of generality, we can assume $C_0>1$ and $C_2<1$. Then $$\begin{aligned}
\Big|D_{v'}^\alpha D_{\bar{v}'}^{\beta}(h\circ f)(x,y,&\theta(x,y,z))\Big|
\\
\leq &\sum_{k=0}^{|\alpha+\beta|} \frac{1}{C_2^{k+1}}\sum_{\substack{\alpha_1+\alpha_2+\cdots+\alpha_k=\alpha\\\beta_1+\beta_2+\cdots+\beta_k=\beta\\\alpha_1+\beta_1>0,\cdots,\alpha_k+\beta_k>0}}
(\alpha!\beta!)^{a+{\varepsilon}}C_0^{k}C_1^{|\alpha+\beta|}\lambda_{b,|\beta|}(x,y,z)
\\
\leq &\sum_{k=0}^{|\alpha+\beta|}\binom{\alpha+k\mathbbm{1}}{k\mathbbm{1}} \binom{\beta+k\mathbbm{1}}{k\mathbbm{1}}
\frac{1}{C_2}\left(\frac{C_0C_1}{C_2}\right)^{|\alpha+\beta|}(\alpha!\beta!)^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)
\\
\leq &\frac{1}{C_2} \left(\frac{2^{6n+2}C_0C_1}{C_2}\right)^{|\alpha+\beta|}(\alpha!\beta!)^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z).
\end{aligned}$$ In addition, when we are restricted to $x=y=\bar{z}$, it is easy to see that we can replace ${\varepsilon}$ by $0$. Therefore, $\frac{1}{f}=h\circ f\in \mathcal{A}^{a,{\varepsilon}}_{\theta}(U')$.
Denote $v=(x,y,z)$ and $C_0=C_0(f), C_1=C_1(f), b=b(f)$. As $t\in [0,1]$, we have $$\begin{aligned}
\left|D_v^\alpha D_{\bar{v}}^{\beta} g(x,y,z)\right|
\leq& \max_{t\in [0,1]} \left|D_v^\alpha D_{\bar{v}}^{\beta} \big(f(x,tx+(1-t)y,z)\big)\right|
\end{aligned}$$ We write $D_v^\alpha=D_x^{\alpha_1}D_y^{\alpha_2}D_z^{\alpha_3}$ and $D_{\bar{v}}^\beta=D_{\bar{x}}^{\beta_1}D_{\bar{y}}^{\beta_2}D_{\bar{z}}^{\beta_3}$. Then $$\begin{aligned}
\Big|D_v^\alpha D_{\bar{v}}^{\beta} &g(x,y,z)\Big|
\\
\leq& \max_{t\in [0,1]} \left|\sum_{\alpha'_1\leq\alpha_1}\sum_{\alpha'_1\leq\alpha_1} \binom{\alpha_1}{\alpha_1'}\binom{\beta_1}{\beta_1'}
D_x^{\alpha_1'}D_y^{\alpha_2+\alpha_1-\alpha_1'}D_z^{\alpha_3}
D_{\bar{x}}^{\beta_1'}D_{\bar{y}}^{\beta_2+\beta_1-\beta_1'}D_{\bar{z}}^{\beta_3}
f(x,tx+(1-t)y,z)\right|
\\
\leq&
\sum_{\alpha'_1\leq\alpha_1}\sum_{\alpha'_1\leq\alpha_1} \binom{\alpha_1}{\alpha_1'}\binom{\beta_1}{\beta_1'}
C_0\left(2^{a+{\varepsilon}}C_1\right)^{|\alpha+\beta|}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}
\max_{t\in [0,1]}\lambda_{b,|\beta|}(x,tx+(1-t)y,z)
\\
\leq& C_0\left(2^{a+{\varepsilon}+1}C_1\right)^{|\alpha+\beta|}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z).
\end{aligned}$$ The last inequality follows from $|tx+(1-t)y-\bar{z}|\leq \max\{|x-\bar{z}|,|y-\bar{z}|\}$ for any $t\in [0,1]$. In addition, as $f\in \mathcal{A}^{a,{\varepsilon}}_\theta$, when restricted to $x=y=\bar{z}$, we can replace ${\varepsilon}$ by $0$. So we get the first part of the lemma. The second part follows by the same argument.
Let $m=3n$. We denote $v=(x,y,z)$, $v'=(x,y,\theta)$. By a straightforward calculation, we have $$\begin{aligned}
\frac{D_{v'}^\alpha D_{\bar{v}'}^\beta}{\alpha!\beta!}&\widetilde{f}
\\
=&
\sum_{0\leq |\xi+\eta|\leq |\alpha+\beta|}
\frac{D_{v}^\xi D_{\bar{v}}^\eta f}{\xi!\eta!}
\sum_{A_{\alpha\beta\xi\eta}} \frac{D_{v'}^{\alpha_{11}}D_{\bar{v}'}^{\beta_{11}}v_1}{\alpha_{11}!\beta_{11}!}
\frac{D_{v'}^{\alpha'_{11}}D_{\bar{v}'}^{\beta'_{11}}{\overline{v_1}}}{\alpha'_{11}!\beta'_{11}!}
\cdots
\frac{D_{v'}^{\alpha_{m\xi_m}}D_{\bar{v}'}^{\beta_{m\xi_m}}v_m}{\alpha_{m\xi_m}!\beta_{m\xi_m}!}
\frac{D_{v'}^{\alpha'_{m\eta_m}}D_{\bar{v}'}^{\beta'_{m\eta_m}}{\overline{v_m}}}{\alpha'_{m\eta_m}!\beta'_{m\eta_m}!},
\end{aligned}$$ where $A_{\alpha\beta\xi\eta}$ is defined in . Then since $f\in \mathcal{A}^{a,{\varepsilon}}_z$ and $v=(x,y,z(x,y,\theta))\in \mathcal{A}^{a,{\varepsilon}}_\theta$, by taking $b=\min\{b(f),b(v)\}$, $C_0(v)=\max_{i}C_0(v_i)$ and $C_1(v)=\max_{i}C_1(v_i)$, we have $$\begin{aligned}
&\left|\frac{D_{v'}^\alpha D_{\bar{v}'}^\beta}{\alpha!\beta!}\widetilde{f}(x,y,\theta(x,y,z))\right|
\\
\leq&
\sum_{0\leq |\xi+\eta|\leq |\alpha+\beta|}
C_0(f)C_1(f)^{|\xi+\eta|}
\sum_{A_{\alpha\beta\xi\eta}}C_0(v)^{|\xi+\eta|}C_1(v)^{|\alpha+\beta|}\left(\xi!\eta!\prod_{i,j}\alpha_{ij}!\beta_{ij}!\prod_{ik}\alpha'_{ik}!\beta'_{ik}!\right)^{a-1+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z).
\end{aligned}$$
Now we prove two combinatorial lemmas to estimate of $\xi!\eta!\prod_{i,j}\alpha_{ij}!\beta_{ij}!\prod_{i,k}\alpha'_{ik}!\beta'_{ik}!$ appearing in the above inequality.
\[lem 4.6\] For any integers $k,i_1,i_2\cdots i_k\in \mathbb{Z}^+$, we have $$k!i_1!i_2!\cdots i_k!\leq (i_1+i_2+\cdots+i_k)!.$$
We will do induction on $k$. When $k=1$, the result follows trivially. Assume it is true for $k-1$ and we proceed to the case $k$. For simplicity, we denote $i=i_1+i_2+\cdots+i_k$. Then by using the result by induction, we have $$k!i_1!i_2!\cdots i_k!\leq k(i_1+i_2+\cdots+i_{k-1})!i_k!=\frac{k}{\binom{i}{i_k}}i!.$$ Note that $\binom{i}{i_k}\geq i\geq k$ and thus the result follows.
\[Combinatoric 2\] For any multi-indices, $\alpha_1,\alpha_2,\cdots, \alpha_k\in (\mathbb{Z}^{\geq 0})^n$, if $|\alpha_i|>0$ for each $1\leq i\leq k$, then we have $$\alpha_1!\alpha_2!\cdots\alpha_k!\leq \frac{(\alpha_1+\alpha_2+\cdots+\alpha_k)!}{k!}n^k.$$
We denote $\alpha_i=(\alpha_{i1},\alpha_{i2},\cdots,\alpha_{in})$ and define $l_j=\#\{1\leq i\leq k, \alpha_{ij}\neq 0\}$ for $1\leq j\leq n$. Then by applying Lemma \[lem 4.6\] to the $j$-th component of each $\alpha_i$ for $1\leq i\leq k$, we have $$\begin{aligned}
\alpha_{1j}!\alpha_{2j}!\cdots\alpha_{kj}!\leq \frac{(\alpha_{1j}+\alpha_{2j}+\cdots+\alpha_{kj})!}{l_j!}.
\end{aligned}$$ Therefore, $$\begin{aligned}
\label{eq 4.16}
\alpha_1!\alpha_2!\cdots\alpha_k!\leq \frac{(\alpha_1+\alpha_2+\cdots+\alpha_k)!}{l_1!l_2!\cdots l_n!}.
\end{aligned}$$ Since $|\alpha_i|\neq 0$ for each $i$, we have $l:=l_1+l_2+\cdots l_n\geq k$. Then we can find nonnegative integers $k_1,k_2,\cdots,k_n$ such that $k_1+k_2+\cdots+k_n=k$ and $k_j\leq l_j$ for $1\leq j\leq n$. Therefore $$l_1!l_2!\cdots l_n!\geq k_1!k_2!\cdots k_n!=\frac{k!}{\binom{k}{k_1,k_2,\cdots,k_n}}\geq \frac{k!}{n^k}.$$ Plug this back into , we have the result.
By using Lemma \[Combinatoric 2\], we have the upper bound for the factorials on the right hand side as $$\begin{aligned}
\xi!\eta!\prod_{i,j}\alpha_{ij}!\beta_{ij}!\prod_{ik}\alpha'_{ik}!\beta'_{ik}!
\leq m^{|\xi+\eta|}\left(\sum_{ij}\alpha_{ij}+\beta_{ij}\right)! \left(\sum_{ik}\alpha'_{ik}+\beta'_{ik}\right)!
\leq m^{|\xi+\eta|}(\alpha+\beta)!.
\end{aligned}$$
Therefore, $$\begin{aligned}
&\left|\frac{D_{v'}^\alpha D_{\bar{v}'}^\beta}{\alpha!\beta!}\widetilde{f}(x,y,\theta(x,y,z))\right|
\\
&\quad\leq
\sum_{0\leq |\xi+\eta|\leq |\alpha+\beta|}
C_0(f)C_1(f)^{|\xi+\eta|}
\sum_{A_{\alpha\beta\xi\eta}}C_0(v)^{|\xi+\eta|}C_1(v)^{|\alpha+\beta|}\left(m^{|\xi+\eta|}(\alpha+\beta)!\right)^{a-1+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)
\\
&\quad\leq
\sum_{0\leq |\xi+\eta|\leq |\alpha+\beta|}
C_0(f)\left(m^{a-1+{\varepsilon}}C_1(f)C_0(v)C_1(v)\right)^{|\alpha+\beta|}
(\alpha+\beta)!^{a-1+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)\cdot\#A_{\alpha\beta\xi\eta}.
\end{aligned}$$ Note the cardinality of $A_{\alpha\beta\xi\eta}$ has the following upper bounded $$\#A_{\alpha\beta\xi\eta}\leq \binom{\alpha+(|\xi+\eta|)\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}\binom{\beta+(|\xi+\eta|)\mathbbm{1}}{|\xi+\eta|\mathbbm{1}}.$$ So we get $$\begin{aligned}
&\left|\frac{D_{v'}^\alpha D_{\bar{v}'}^\beta}{\alpha!\beta!}\widetilde{f}(x,y,\theta(x,y,z))\right|
\\
&\quad\leq
\sum_{0\leq |\xi+\eta|\leq |\alpha+\beta|}
C_0(f)\Big(2^{a+{\varepsilon}+2m}m^{a-1+{\varepsilon}}C_1(f)C_0(v)C_1(v)\Big)^{|\alpha+\beta|}
(\alpha!\beta!)^{a-1+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z)
\\
&\quad\leq
C_0(f)\Big(2^{a+3m+{\varepsilon}}m^{a-1+{\varepsilon}}C_1(f)C_0(v)C_1(v)\Big)^{|\alpha+\beta|}
(\alpha!\beta!)^{a-1+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z).
\end{aligned}$$ And if we keep track of the constant ${\varepsilon}$, it is easy to see that ${\varepsilon}$ comes from derivatives of since $f$ and $v=(x,y,z(x,y,\theta))$. Since $f\in \mathcal{A}^{a,{\varepsilon}}_z$ and $v=(x,y,z(x,y,\theta))\in \mathcal{A}^{a,{\varepsilon}}_\theta$, we can replace ${\varepsilon}$ by $0$ when restricted to $x=y=\bar{z}$. Therefore, we can take $C_0(\widetilde{f})=C_0(f)$, $C_1(\widetilde{f})=2^{a+3m+{\varepsilon}} m^{a-1+{\varepsilon}}C_0(v)C_1(f)C_1(v)$.
We are going to prove the following more general lemma. Note that we can assume $\psi(y,z)=yz+O(|(y,z)|^4)$ by using the Bochner coordinates at $0$. Then the Lemma \[AHE Inverse\] follows directly by taking $F(x,y,z,\theta)=\int_0^1(D_y\psi)(tx+(1-t)y,z)dt-\theta$ and $\theta(x,y,z)=\int_0^1(D_y\psi)(tx+(1-t)y,z)dt$.
Consider smooth maps $\theta(x,y,z)=(\theta_1(x,y,z),\theta_2(x,y,z),\cdots,\theta_n(x,y,z))$ and $F(x,y,z,\theta)=(F_1(x,y,z,\theta),F_2(x,y,z,\theta),\cdots,F_n(x,y,z,\theta))$ satisfying the system of equations $F(x,y,z,\theta(x,y,z))=0$. Assume that for any $x,y,z\in U=B^n(0,1)$ and multi-indices $\alpha,\beta\geq 0$, we have $$\begin{aligned}
\label{F}
\left|\left(D_{(x,y,z,\theta)}^\alpha D_{(\bar{x},\bar{y},\bar{z},\bar{\theta})}^{\beta}F\right)(x,y,z,\theta(x,y,z))\right|\leq C_0C_1^{|\alpha+\beta|}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z),
\end{aligned}$$ where $C_0=C_0(F), C_1=C_1(F)$ and $b=b(F)$ are some positive constants. And ${\varepsilon}$ can be replaced by $0$, when we are restricted to $x=y=\bar{z}$.
If the $2n\times2n$ matrix $\begin{pmatrix}
\frac{\partial F}{\partial z} & \frac{\partial F}{\partial {\overline{z}}}\\
\frac{\partial {\overline{F}}}{\partial z} & \frac{\partial {\overline{F}}}{\partial {\overline{z}}}
\end{pmatrix}$ is the identity matrix at $(x_0,y_0,z_0,\theta_0)$, then the implicit functions $z=z(x,y,\theta)$ near $(x_0,y_0,\theta_0)$ determined by the equation $F(x,y,z,\theta)=0$ belong to $\mathcal{A}^{a,{\varepsilon}}_\theta$.
We first consider a special case when $F$ is a function in the following lemma.
Consider smooth maps $\theta(x,y,z)=(\theta_1(x,y,z),\theta_2(x,y,z),\cdots,\theta_n(x,y,z))$ and function $f(x,y,z,\theta)$ such that $f(x,y,z,\theta(x,y,z))=0$. And for any $x,y,z\in U$ and multi-indices $\alpha,\beta\geq 0$, we have $$\begin{aligned}
\label{f}
\left|\left(D_{(x,y,z,\theta)}^\alpha D_{(\bar{x},\bar{y},\bar{z},\bar{\theta})}^{\beta}f\right)(x,y,z,\theta(x,y,z))\right|\leq C_0C_1^{|\alpha+\beta|}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b,|\beta|}(x,y,z),
\end{aligned}$$ where $C_0, C_1$ and $b$ are some positive constants. And ${\varepsilon}$ can be replace by $0$, when we are restricted to $x=y=\bar z$.
Assume at $(x_0,y_0,z_0, \theta_0=\theta(x_0,y_0,z_0))$, the matrix $\begin{pmatrix}
\frac{\partial f}{\partial z_n} & \frac{\partial f}{\partial {\overline{z_n}}}\\
\frac{\partial {\overline{f}}}{\partial z_n} & \frac{\partial {\overline{f}}}{\partial {\overline{z_n}}}
\end{pmatrix}$ is non-singular. Then the implicit function $z_n=z_n(x,y,z_1,z_2\cdots,z_{n-1},\theta)$ determined by the equation $f(x,y,z,\theta)=0$ satisfies that for any multi-indices $\alpha,\beta\geq 0$, $$\label{zn}
\left|\left(D_{(x,y,z',\theta)}^\alpha D_{(\bar{x},\bar{y},\bar{z}',\bar{\theta})}^{\beta}z_n\right)(x,y,z',\theta(x,y,z))\right|\leq C'_0{C'_1}^{|\alpha+\beta|}\alpha!^{a+{\varepsilon}}\beta!^{a+{\varepsilon}}\lambda_{b',\beta}(x,y,z),$$ where $C'_0,C'_1$ and $b'$ are some positive constants and $z'=(z_1,z_2,\cdots,z_{n-1})$. In addition, when we are restricted to $x=y=\bar z$, ${\varepsilon}$ can be replace by $0$.
For simplicity, we denote $v=(x,y,z_1,z_2\cdots,z_{n-1},\theta)$. Near some point $(v, z_n)$, we have the Taylor series of $f$ as $$f(v',z'_n)=\sum_{\alpha,\beta\geq 0,i,j\geq 0} a_{\alpha\bar{\beta}i\bar{j}}(v'-v)^\alpha({\overline{v'-v}})^{\beta}(z'_n-z_{n})^i({\overline{z'_n-z_{n}}})^j,$$ where $a_{\alpha\bar{\beta}i\bar{j}}=\frac{ D_v^{\alpha} D_{\bar{v}}^\beta D_{z_n}^i D_{{\bar{z}_n}}^jf}{\alpha!\beta!i!j!}(v,z_{n})$. The equation $f=0$ implies $$\begin{aligned}
\label{eq 4.15}
\begin{split}
&a_{0\bar{0}0\bar{0}}+a_{0\bar{0}1\bar{0}}(z'_n-z_{n})+a_{0\bar{0}0\bar{1}}({\overline{z'_n-z_{n}}})
\\=&-\sum_{|\alpha+\beta|>0}\left(a_{\alpha\bar{\beta}0\bar{0}}+a_{\alpha\bar{\beta}1\bar{0}}(z'_n-z_{n})+a_{\alpha\bar{\beta}0\bar{1}}({\overline{z'_n-z_{n}}})\right)(v'-v)^\alpha({\overline{v'-v}})^{\beta}
\\&
-\sum_{\alpha,\beta,i+j\geq 2}a_{\alpha\bar{\beta}i\bar{j}}(v'-v)^\alpha({\overline{v'-v}})^{\beta}(z'_n-z_{n})^i({\overline{z'_n-z_{n}}})^j.
\end{split}
\end{aligned}$$ Assume near $v$, the Taylor series of $z_n=z_n(v)$ is as follows. $$z'_n-z_{n}=\sum_{|\gamma+\delta|>0}b_{\gamma\bar{\delta}}(v'-v)^\gamma({\overline{v'-v}})^\delta,$$ where $b_{\gamma\bar{\delta}}=\frac{ D_v^\gamma D_{\bar{v}}^\delta z_n}{\gamma!\delta!}(v)$.
We define the following index sets for simplicity. $$\begin{aligned}
A_{\alpha\beta\gamma\delta}
=\left \{ i,j, \{ \xi_k,\eta_k \}_{1\leq k \leq i}, \{ \xi'_{l}, \eta'_{l}\}_{1\leq l \leq j}: \quad \begin{array}{ll} \alpha+\sum_{1\leq k \leq i} \xi_{k}+\sum_{1\leq l \leq j} \xi'_{l}=\gamma, \\ \beta+\sum_{1\leq k \leq i} \eta_{k}+\sum_{1\leq l \leq j} \eta'_{l}=\delta, \\ \xi_{k}+\xi'_{k}>0,\eta_{l}+\eta'_{l}>0, i+j\geq 2 \end{array} \right \}.
\end{aligned}$$ $$\begin{aligned}
B_{\gamma\delta}=\left\{\alpha,\beta,\xi,\eta:
\alpha+\xi=\gamma, \beta+\eta=\delta, |\alpha+\beta|>0\right\}.
\end{aligned}$$ When restrict to points $(x,y,z,\theta(x,y,z))$, $a_{0\bar{0}0\bar{0}}=0$. By comparing the coefficients of , for any multi-indices $|\gamma+\delta|>0$, we have $$\begin{aligned}
\begin{split}
&a_{0\bar{0}1\bar{0}}b_{\gamma\bar{\delta}}+a_{0\bar{0}0\bar{1}}{\overline{b_{\delta\bar{\gamma}}}}
\\&\quad=-a_{\gamma\bar{\delta}0\bar{0}}-\sum_{B_{\gamma\delta}}a_{\alpha\bar{\beta}1\bar{0}}b_{\xi\bar{\eta}}
-\sum_{B_{\gamma\delta}}a_{\alpha\bar{\beta}0\bar{1}}{\overline{b_{\eta\bar{\xi}}}}
-\sum_{A_{\alpha\beta\gamma\delta}}a_{\alpha\bar{\beta}i\bar{j}}b_{\xi_1\bar{\eta}_1}\cdots b_{\xi_i\bar{\eta}_i}{\overline{b_{\eta_1'\bar{\xi}_1'}}}\cdots{\overline{b_{\eta_j'\bar{\xi}_j'}}}.
\end{split}
\end{aligned}$$ Taking the conjugate and switch the multi-indices $\gamma$ and $\delta$, we have $$\begin{aligned}
\begin{split}
&{\overline{a_{0\bar{0}0\bar{1}}}}b_{\gamma\bar{\delta}}+{\overline{a_{0\bar{0}1\bar{0}}}}{\overline{b_{\delta\bar{\gamma}}}}
\\&\quad =-{\overline{a_{\delta\bar{\gamma}0\bar{0}}}}-\sum_{B_{\delta\gamma}}{\overline{a_{\alpha\bar{\beta}1\bar{0}}}}{\overline{b_{\xi\bar{\eta}}}}
-\sum_{B_{\delta\gamma}}{\overline{a_{\alpha\bar{\beta}0\bar{1}}}}b_{\eta\bar{\xi}}
-\sum_{A_{\alpha\beta\delta\gamma}}{\overline{a_{\alpha\bar{\beta}i\bar{j}}}}{\overline{b_{\xi_1\bar{\eta}_1}}}\cdots {\overline{b_{\xi_i\bar{\eta}_i}}}b_{\eta_1'\bar{\xi}_1'}\cdots b_{\eta_j'\bar{\xi}_j'}.
\end{split}
\end{aligned}$$ Then for any $|\gamma+\delta|>0$, by solving $b_{\gamma\bar{\delta}}$, we obtain the following recursive formula on the coefficients $b_{\gamma\bar{\delta}}$. $$\begin{aligned}
\label{eq 4.19}
\begin{split}
&b_{\gamma\bar{\delta}}
\\=&-\frac{{\overline{a_{0\bar{0}1\bar{0}}}}}{|a_{0\bar{0}1\bar{0}}|^2-|a_{0\bar{0}0\bar{1}}|^2}
\left(a_{\gamma\bar{\delta}0\bar{0}}+\sum_{B_{\gamma\delta}}a_{\alpha\bar{\beta}1\bar{0}}b_{\xi\bar{\eta}}
+\sum_{B_{\gamma\delta}}a_{\alpha\bar{\beta}0\bar{1}}{\overline{b_{\eta\bar{\xi}}}}
+\sum_{A_{\alpha\beta\gamma\delta}}a_{\alpha\bar{\beta}i\bar{j}}b_{\xi_1\bar{\eta}_1}\cdots b_{\xi_i\bar{\eta}_i}{\overline{b_{\eta_1'\bar{\xi}_1'}}}\cdots {\overline{b_{\eta_j'\bar{\xi}_j'}}}\right)
\\
&+\frac{a_{0\bar{0}0\bar{1}}}{|a_{0\bar{0}1\bar{0}}|^2-|a_{0\bar{0}0\bar{1}}|^2}
\left({\overline{a_{\delta\bar{\gamma}0\bar{0}}}}+\sum_{B_{\delta\gamma}}{\overline{a_{\alpha\bar{\beta}1\bar{0}}}}{\overline{b_{\xi\bar{\eta}}}}
+\sum_{B_{\delta\gamma}}{\overline{a_{\alpha\bar{\beta}0\bar{1}}}}b_{\eta\bar{\xi}}
+\sum_{A_{\alpha\beta\delta\gamma}}{\overline{a_{\alpha\bar{\beta}i\bar{j}}}}{\overline{b_{\xi_1\bar{\eta}_1}}}\cdots {\overline{b_{\xi_i\bar{\eta}_i}}}b_{\eta_1'\bar{\xi}_1'}\cdots b_{\eta_j'\bar{\xi}_j'}\right).
\end{split}
\end{aligned}$$ By , when $\theta=\theta(x,y,z)$, the Taylor coefficients $a_{\alpha\bar{\beta}k\bar{l}}$ satisfies that $$|a_{\alpha\bar{\beta}k\bar{l}}|\leq C_0C_1^{|\alpha+\beta|+k+l}(\alpha!\beta!k!l!)^{a-1+{\varepsilon}}\lambda_{b,|\beta|+l}(x,y,z),$$ where $\lambda$ is as defined in . We normalized $a_{\alpha\bar{\beta}k\bar{l}}$ to $\widetilde{a}_{\alpha\beta kl}$ as $$\label{atilde}
\widetilde{a}_{\alpha\beta kl}
=\frac{|a_{\alpha\bar{\beta}k\bar{l}}|}{(\alpha!\beta!k!l!)^{a-1+{\varepsilon}}\lambda_{b,|\beta|+l}(x,y,z)},$$ which is dominated by $C_0C_1^{|\alpha+\beta|+k+l}$. Similarly, we define $$\label{btilde}
\widetilde{b}_{\gamma\delta}=
\frac{|b_{\gamma\bar{\delta}}|}{(\gamma!\delta!)^{a-1+{\varepsilon}}\lambda_{b,|\beta|+l}(x,y,z)}.$$ Since the matrix $\begin{pmatrix}
\frac{\partial f}{\partial z_n} & \frac{\partial f}{\partial {\overline{z_n}}}\\
\frac{\partial {\overline{f}}}{\partial z_n} & \frac{\partial {\overline{f}}}{\partial {\overline{z_n}}}
\end{pmatrix}$ is non-singular at $(x_0,y_0,z_0,\theta_0)$, by choosing a sufficiently small neighborhood $U$ of $(x_0,y_0,z_0,\theta_{n0})$, we have $\inf_{U}\left||a_{0\bar{0}1\bar{0}}|^2-|a_{0\bar{0}0\bar{1}}|^2\right|\geq A>0$. By , and using triangle inequalities and Lemma \[Combinatoric 2\], we write the following recursive inequality on $b_{\gamma\delta}$. $$\begin{aligned}
\label{eq 4.28}
\begin{split}
\widetilde{b}_{\gamma\delta}
&\leq \frac{C_0C}{A}
\left(\widetilde{a}_{\gamma\delta00}+\sum_{B_{\gamma\delta}}\widetilde{a}_{\alpha\beta10}\widetilde{b}_{\xi\eta}
+\sum_{B_{\gamma\delta}}\widetilde{a}_{\alpha\beta01}\widetilde{b}_{\eta\xi}
+\sum_{A_{\alpha\beta\gamma\delta}}\widetilde{a}_{\alpha\beta ij}\widetilde{b}_{\xi_1\eta_1}\cdots \widetilde{b}_{\eta_j'\xi_j'}(6n)^{i+j}\right)
\\
&\quad +\frac{C_0C}{A}
\left(\widetilde{a}_{\delta\gamma00}+\sum_{B_{\delta\gamma}}\widetilde{a}_{\alpha\beta10}\widetilde{b}_{\xi\eta}
+\sum_{B_{\delta\gamma}}\widetilde{a}_{\alpha\beta01}\widetilde{b}_{\eta\xi}
+\sum_{A_{\alpha\beta\delta\gamma}}\widetilde{a}_{\alpha\beta ij}\widetilde{b}_{\xi_1\eta_1}\cdots \widetilde{b}_{\eta_j'\xi_j'}(6n)^{i+j}\right).
\end{split}
\end{aligned}$$
Recall the definition of *majorant* for power series as follows.
Consider two power series in variables $x\in \mathbb{R}^n$. $$\begin{aligned}
f(x)\sim \sum_{\alpha\geq 0} a_{\alpha}x^{\alpha}, \quad g(x)\sim \sum_{\alpha\geq 0} b_{\alpha}x^{\alpha}.
\end{aligned}$$ We say that $g$ is a majorant of $f$, or $b_\alpha$ is a majorant of $a_\alpha$, if $|a_\alpha|\leq b_\alpha$ for any $\alpha\geq 0$. And we denote this by $f<<g$
For multi-indices $|{\alpha}+\delta|>0$, we define $d_{{\alpha}\delta}$ recursively as $$\begin{aligned}
\label{eq 4.29}
\begin{split}
d_{\gamma\delta}
&=\frac{C_0^2C_1}{A}
\left(C_2^{|\gamma+\delta|}+\sum_{B_{\gamma\delta}}C_2^{|\alpha+\beta|+1}d_{\xi\eta}
+\sum_{B_{\gamma\delta}}C_2^{|\alpha+\beta|+1}d_{\eta\xi}
+\sum_{A_{\alpha\beta\gamma\delta}}C_2^{|\alpha+\beta|+i+j}d_{\xi_1\eta_1}\cdots d_{\eta_j'\xi_j'}\right)
\\
&\quad+\frac{C_0^2C_1}{A}
\left(C_2^{|\delta+\gamma|}+\sum_{B_{\delta\gamma}}C_2^{|\alpha+\beta|+1}d_{\xi\eta}
+\sum_{B_{\delta\gamma}}C_2^{|\alpha+\beta|+1}d_{\eta\xi}
+\sum_{A_{\alpha\beta\delta\gamma}}C_2^{|\alpha+\beta|+i+j}d_{\xi_1\eta_1}\cdots d_{\eta_j'\xi_j'}\right),
\end{split}
\end{aligned}$$ where $C_2=6nC_1$. Since $C_0C_2^{|\alpha+\beta|+i+j}$ is a majorant of $\widetilde{a}_{\alpha\beta ij}$, $d_{\gamma\delta}$ defined as above is a majorant of $\widetilde{b}_{\gamma\delta}$ in for any $|\gamma+\delta|>0$. Now we will solve $d_{\gamma\delta}$ by the recursive equation . Formally, we define $d(u,v)=\sum_{|\gamma+\delta|>0}d_{\gamma\delta} u^{\gamma}v^{\delta}$. Then is equivalent to $$\begin{aligned}
d&(u,v)\\
&=\frac{2C_0^2C_1}{A}\left(\frac{1}{1-C_2d(u,v)}\frac{1}{1-C_2d(v,u)}\prod_{i=1}^m\frac{1}{(1-C_2u_i)(1-C_2v_i)}-1-C_2d(u,v)-C_2d(v,u)\right),
\end{aligned}$$ where $m=3n$. It is easy to see that $d(u,v)=d(v,u)$ and thus $$d(u,v)=\frac{2C_0^2C_1}{A}\left(\frac{1}{(1-C_2d(u,v))^2}\prod_{i=1}^m\frac{1}{(1-C_2u_i)(1-C_2v_i)}-1-2C_2d(u,v)\right).$$ Observe that $(u,v,d)=0$ satisfies the equation and there is no linear term of $d$ on the right hand side. By the Implicit Theorem for real analytic functions (See [@KrPa] for more details), it follows that $d(u,v)$ is real analytic near the origin. Therefore, there exists some constant $C_3$ such that $\widetilde{b}_{\gamma\delta}\leq d_{\gamma\delta}\leq C_3^{\gamma+\delta}$ for any $|\gamma+\delta|>0$. By using , we obtain the desired bounds for $b_{\gamma\bar{\delta}}$. In addition, note that the constant ${\varepsilon}$ only comes from the estimate of $f$ in . Therefore, when we are restricted to $x=y=\bar{z}$, the constant ${\varepsilon}$ can be replaced by $0$.
Now we will do induction on the dimension $n$. When $n=1$, the result directly follows from the previous lemma. We assume the result holds for $n-1$ and proceed to $n$. First, we consider the equation $F_n(x,y,z,\theta)=0$. Since the matrix $\begin{pmatrix}
\frac{\partial F_n}{\partial z_n} & \frac{\partial F_n}{\partial {\overline{z_n}}}\\
\frac{\partial {\overline{F}}_n}{\partial z_n} & \frac{\partial {\overline{F}}_n}{\partial {\overline{z_n}}}
\end{pmatrix}$ is identity at $(x_0,y_0,z_0,\theta_0)$, by using the previous lemma again, we have the implicit function $z_n=h_n(x,y,z',\theta)$, which satisfies $$F_n(x,y,z',h_n,\theta)=0.$$ Take the derivative with respect to $z_j$ and $\bar{z}_j$ for $1\leq j\leq n-1$. Then at $(x_0,y_0,z_0,\theta_0)$, $$\begin{aligned}
\frac{\partial h_n}{\partial z_j}=
\frac{\partial h_n}{\partial \bar{z}_j}=0.
\end{aligned}$$ Define $G_{i}(x,y,z',\theta)=F_{i}(x,y,z',h_n(x,y,z',\theta))$ for $1\leq i\leq n-1$. Since functions $F$ and $h_n$ satisfy and respectively, the composition function $G_i(x,y,z',\theta)$ for $1\leq i\leq n-1$ also satisfy the estimates on the derivatives by a similar argument as in the proof of Lemma \[AHE Composition\]. On the other hand, we have for any $1\leq i,j\leq n-1$, at $(x_0,y_0,z_0,\theta_0)$ $$\begin{aligned}
\frac{\partial G_i}{\partial z_j}=\frac{\partial F_i}{\partial z_j}=\delta_{ij}, \quad \frac{\partial G_i}{\partial \bar{z}_j}=\frac{\partial F_i}{\partial \bar{z}_j}=0.
\end{aligned}$$ Therefore, the matrix$\begin{pmatrix}
\frac{\partial G}{\partial z'} & \frac{\partial G}{\partial {\overline{z}}'}\\
\frac{\partial {\overline{G}}}{\partial z'} & \frac{\partial {\overline{G}}}{\partial {\overline{z}}'}
\end{pmatrix}$ is identity at $(x_0,y_0,z_0,\theta_0)$. Using the conclusion from the induction, we have the implicit functions $z_i=h_i(x,y,\theta)$ of the equations $G_i(x,y,z',\theta)=0$ for $1\leq i\leq n-1$. It is easy to verify that $z_i=h_i(x,y,\theta)$ for $1\leq i\leq n-1$ and $z_n=h_n(x,y,h_1,h_2,\cdots, h_{n-1},\theta_n)$ satisfy all the requirements and our result follows.
Acknowledgements {#acknowledgements .unnumbered}
================
The author is very grateful to Prof. Hamid Hezari for many stimulating conversations and valuable suggestions. The author also thanks Prof. Zhiqin Lu and Prof. Bernard Shiffman for their constant support and mentoring.
[HHHH]{}
Berman, R., Berndtsson, B., Sjöstrand, J., *A direct approach to Bergman kernel asymptotics for positive line bundles*. Ark. Mat. **46(2)**, 197–217 (2008).
Berndtsson, B. *Bergman kernels related to Hermitian line bundles over compact complex manifolds*, Explorations in complex and Riemannian geometry, Contemp. Math., **332**, 1–17, Amer. Math. Soc., Providence, RI, 2003.
Berman, R., *Sharp asymptotics for [T]{}oeplitz determinants and convergence towards the [G]{}aussian free field on [R]{}iemann surfaces*. Int. Math. Res. Not. **22**, 5031–5062 (2012).
Berndtsson, B. *An introduction to things ${{\overline{\partial}}}$. Analytic and Algebraic Geometry*, McNeal, 7–76 (2010).
Bleher, P., Shiffman, B., Zelditch, S., *Universality and scaling of correlations between zeros on complex manifolds*. Invent. Math. **142** (2), 351–395 (2000).
Bochner, S. *Curvature in [H]{}ermitian metric*. Bull. Amer. Math. Soc. 179–195 (1947).
Boutet de Monvel, L., Sjöstrand, J. *Sur la singularité des noyaux de Bergman et de Szegö*. Équations aux Dérivées Partielles de Rennes, Asterisque **34-35**, 123–164 (1976), Soc. Math. France, Paris.
Calabi, E., *Isometric imbedding of complex manifolds.* Ann. of Math. **58** (2), 1–23 (1953).
Catlin, D., *The Bergman kernel and a theorem of Tian. Analysis and Geometry in Several Complex Variables*, Katata, Trends Math., 1–23. Birkhäuser, Boston (1999).
Christ, M. *On the $\bar \partial$ equation in weighted $L^2$ norms in $\mathbb C^1$.* J. Geom. Anal., **1**(3), 193–230 (1991).
Christ, M., *Slow off-diagonal decay for Szegö kernels associated to smooth Hermitian line bundles.* Harmonic analysis at Mount Holyoke (South Hadley, MA, 2001), 77–89, Contemp. Math., **320**, Amer. Math. Soc., Providence, RI, 2003.
Christ, M., *Upper bounds for Bergman kernels associated to positive line bundles with smooth Hermitian metrics*, unpublished (2013), arXiv:1308.0062.
Christ, M., *Off-diagonal decay of Bergman kernels: On a conjecture of Zelditch*, unpublished (2013), arXiv:1308.5644.
Chung, S.-Y. and Chung, J., *There exist no gaps between [G]{}evrey differentiable and nowhere [G]{}evrey differentiable*. Proc. Amer. Math. Soc. **133**, 859–863 (2005).
Dai, X., Liu, K., Ma, X., *On the asymptotic expansion of [B]{}ergman kernel.* J. Differential Geom. **72**, 1–41(2006).
Donaldson S. K. *Scalar curvature and projective embeddings. I*, J. Differential Geom. **59**(3), 479–522 (2001).
Delin, H. *Pointwise estimates for the weighted Bergman projection kernel in $\mathbb C^n$, using a weighted $L^2$ estimate for the $\bar \partial$ equation.* Ann. Inst. Fourier (Grenoble), **48**(4), 967–997 (1998).
Engliš, M., *The asymptotics of a Laplace integral on a Kähler manifold.* Trans. Amer. Math. Soc. **528**, 1–39 (2000).
Fefferman, C., *The Bergman kernel and biholomorphic mappings of psuedoconvex domains.* Invent. Math. **26**, 1–66 (1974).
Gevrey, M., *Sur la nature analytique des solutions des équations aux dérivées partielles.* [P]{}remier mémoire. (French) Ann. Sci. École Norm. Sup. **35**(3), 129–190 (1918).
Hezari, H., Kelleher, C., Seto, S., Xu, H., *Asymptotic expansion of the Bergman kernel via perturbation of the Bargmann-Fock model,* Journal of Geometric Analysis, **26**(4), 2602–2638 (2016).
Hezari, H., Lu, Z., and Xu, H., *Off-diagonal asymptotic properties of bergman kernels associated to analytic Kähler potentials*, (2018).
Hezari, H., Xu, H., *Quantitative upper bounds for Bergman kernels associated to smooth potentials,* arXiv:1807.00204 (2018).
Hörmander, L., *An introduction to complex analysis in several variables.* D. Van Nostrand Co., Inc., Princeton, N.J.-Toronto, Ont.-London 1966.
Hörmander, L., *Lecture notes at the Nordic Sumer School of Mathematics*, 1968.
Jung, K., *The Adiabatic Theorem for Switching Processes with Gevrey Class Regularity.* A note in English from author’s thesis titled: Adiabatik und Semiklassik bei Regularität vom Gevrey-Typ, Berlin, Techn. Universitat, Diss., 1997.
Jung, K., *Phase space tunneling for operators with symbols in a Gevrey class.* J. Math. Phys. **41**(7), 4478–4496 (2000).
Karabegov, A., Schlichenmaier, M. *Identification of Berezin-Toeplitz quantization*, J. Reine Angew. Math. **540**, 49–76 (2001).
Krantz, Steven G.; Parks, Harold R., *A primer of real analytic functions, Second edition.* Birkhäuser Advanced Texts: Basler Lehrbücher. \[Birkhäuser Advanced Texts: Basel Textbooks\] Birkhäuser Boston, Inc., Boston, MA, 2002.
Lindholm, N. *Sampling in weighted $L^p$ spaces of entire functions in $\mathbb C^n$ and estimates of the Bergman kernel.* J. Funct. Anal. **182**(2), 390–426 (2001).
Liu, C.-J., *The asymptotic [T]{}ian-[Y]{}au-[Z]{}elditch expansion on [R]{}iemann surfaces with constant curvature*, Taiwanese J. Math., 1665–1675 (2010).
Liu, C.-J., Lu, Z., *Uniform asymptotic expansion on [R]{}iemann surfaces*. Analysis, complex geometry, and mathematical physics: in honor of [D]{}uong [H]{}. [P]{}hong, Contemp. Math. **644**, 159–173 (2015).
Liu, C.-J., Lu, Z., *Abstract [B]{}ergman kernel expansion and its applications*, Trans. Amer. Math. Soc. **368**, 1467–1495 (2016).
Loi, A., *The Tian-Yau-Zelditch asymptotic expansion for real analytic Kähler metrics.* Int. J. Geom. Methods Mod. Phys. **1**(3), 253–263 (2004).
Lu, Z., *On the Lower Order Terms of the Asymptotic Expansion of Tian-Yau-Zelditch*, American Journal of Mathematics, **122**(2), 235–273 (2000).
Lu, Z., Seto S. *Agmon type estimates of the Bergman Kernel for non-compact manifolds*, preprint, 2017.
Lu, Z., Shiffman, B., *Asymptotic Expansion of the Off-Diagonal Bergman Kernel on Compact Manifolds*, Journal of Geometric Analysis, **25**(2), 761–782 (2015).
Lu, Z., Tian, G., *The log term of the [S]{}zegö kernel*, Duke Math. J., **125**(2), 351–387 (2004).
Lu, Z., Zelditch, S., *Szegö kernels and [P]{}oincaré series*, J. Anal. Math., **130**, 167–184 (2016).
Ma, X. and Marinescu, G., *Holomorphic Morse inequalities and Bergman kernels*, Progress in Math., **254**, Birkhäuser, Basel, 2007.
Ma, X., Marinescu, G., *Generalized [B]{}ergman kernels on symplectic manifolds*, Adv. Math., **217**(4), 1756-1815 (2008).
Ma, X., Marinescu, G., *Remark on the Off-Diagonal Expansion of the Bergman Kernel on Compact Kähler Manifolds*, Communications in Mathematics and Statistics, **1**(1), 37–41 (2013).
Ma, X., Marinescu, G., *Exponential estimate for the asymptotics of Bergman kernels.* Math. Ann. **362**(3-4), 1327–1347 (2015).
Rouby, O., Sjöstrand, J., Ngoc, S.,*Analytic Bergman operators in the semiclassical limit,* arXiv:1808.00199v1 (2018).
Seto, S. *On the asymptotic expansion of the Bergman kernel*, Thesis (Ph.D.)-University of California, Irvine. (2015).
Shiffman, B., Zelditch, S., *Asymptotics of almost holomorphic sections of ample line bundles on symplectic manifolds*, J. Reine Angew. Math. **544**, 181-222 (2002).
Sjöstrand, Singularités analytiques microlocales, Astérisque, **95**(1982), 1–166, Soc. Math. France, Paris.
Tian, G., *On a set of polarized metrics on algebraic manifolds*, J. Differ. Geom. **32**(1), 99–130 (1990).
Xu, H., *A closed formula for the asymptotic expansion of the Bergman kernel.* Comm. Math. Phys. **314**(3), 555–585 (2012).
Yuan, Y., Zhu, J. *Holomorphic line bundles over a tower of coverings.* J. Geom. Anal. **26**(3), 2013–2039 (2016).
Zelditch, S., *Szegö kernels and a theorem of Tian*, Internat. Math. Res. Notices **6**, 317–331 (1998).
Zelditch, S. *Book review of “Holomorphic Morse inequalities and Bergman kernels” (by Xiaonan Ma and George Marinescu)*. Bulletin of the American Mathematical Society **46**, 349–361 (2009).
Zelditch, S., *Pluri-potential theory on [G]{}rauert tubes of real analytic [R]{}iemannian manifolds, [I]{}*. Spectral geometry, Proc. Sympos. Pure Math. **84**, 299–339 (2012).
Zelditch, S., *Off-diagonal decay of toric Bergman kernels*, Lett. Math. Phys. **106**(12), 1849–1864 (2016). Volume in Memory of Louis Boutet de Monvel.
[^1]: In the sense of Borel and Hörmander [@HoAH]; see our definition .
[^2]: For convenience, we use $N_0$ for $N_0(k)= {[(k/C)^{\frac{1}{2a+2{\varepsilon}}}]}$.
[^3]: We discussed its proof in .
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We prove that the complete $L$-functions of classical holomorphic newforms have infinitely many simple zeros.'
address: 'School of Mathematics, University of Bristol, University Walk, Bristol, BS8 1TW, United Kingdom'
author:
- 'Andrew R. Booker'
title: 'Simple zeros of degree $2$ $L$-functions'
---
[^1]
Introduction
============
Let $\pi$ be a cuspidal automorphic representation of $\operatorname{GL}_n({\mathbb{A}}_{\mathbb{Q}})$ with corresponding $L$-function $\Lambda(s,\pi)$. The Grand Riemann Hypothesis (GRH) and Grand Simplicity Hypothesis (GSH) predict that the zeros of $\Lambda(s,\pi)$ lie on the line $\Re(s)=\frac12$ and are simple, apart from at most one multiple zero if $\pi$ is associated to a geometric motive (cf. the BSD conjecture). These conjectures have not yet been shown to hold for a single example, and most partial evidence in their favor has been for $n=1$, i.e. the Dirichlet $L$-functions. In particular, for $n>1$, the only cuspidal representation that we are aware of for which it is known that $\Lambda(s,\pi)$ has infinitely many simple zeros is the one associated to the Ramanujan $\Delta$ modular form, which is a theorem of Conrey and Ghosh [@conrey-ghosh] from 1988.
As Conrey and Ghosh remark in their paper, most of their arguments would apply to any degree $2$ $L$-function, but they were unable to conclude the proof without assuming *a priori* the existence of at least one simple zero (which they verified directly for the $L$-function associated to $\Delta$). In this paper, we analyze their method from a structural point of view, along the lines of [@bk-weil] and [@kp], to prove the following:
\[mainthm\] Let $f\in S_k(\Gamma_1(N))^{\rm new}$ be a normalized Hecke eigenform of arbitrary weight and level. Then the complete $L$-function $\Lambda_f(s)=\int_0^\infty f(iy)y^{s-1}\,dy$ has infinitely many simple zeros.
As our proof will show, a lack of simple zeros leads to inconsistencies unless the local $L$-factor of $\Lambda_f(s)$ is a square at every unramified prime (which cannot happen for holomorphic modular forms). In effect, we establish a connection (albeit a very loose one) between the zeros of the global $L$-function and those of its local factor polynomials.
Recently, Cho [@cho] has generalized [@conrey-ghosh] to prove that the $L$-functions of the first few Maass cusp forms of level $1$ have infinitely many simple zeros. Our proof could be modified in an analogous fashion to extend Theorem \[mainthm\] to all cuspidal Maass newforms. Moreover, the assumption that $f$ is a cusp form is also unnecessary, so in fact the method could be generalized to show that if $\chi_1$ and $\chi_2$ are primitive Dirichlet characters and $t\in{\mathbb{R}}$ then $\Lambda(s,\chi_1)\Lambda(s+it,\chi_2)$ has infinitely many simple zeros unless $\chi_1=\chi_2$ and $t=0$. However, stronger results of this type may be obtained by other methods, e.g. [@cgg].
Finally we note that Conrey and Ghosh’s result for $f=\Delta$ is a bit stronger than the conclusion of Theorem \[mainthm\] for that case. Precisely, if $N_f(T)$ denotes the number of simple zeros of $\Lambda_f(s)$ with imaginary part in $[0,T]$, they showed that for every $\varepsilon>0$, the inequality $N_\Delta(T)\ge
T^{\frac16-\varepsilon}$ holds for some arbitrarily large values of $T$. With Theorem \[mainthm\] in hand, it seems likely that their proof of this estimate would generalize at least to all eigenforms of level $1$. However, in this paper we content ourselves with the qualitative statement of Theorem \[mainthm\].
Acknowledgements {#acknowledgements .unnumbered}
----------------
I thank Brian Conrey for helpful comments.
Notation {#notation .unnumbered}
--------
Let $f$ be as in the statement of Theorem \[mainthm\], and let $\xi$ denote its nebentypus character. Let $$L_f(s)=\sum_{n=1}^\infty a_f(n)n^{-s}
=\prod_p\frac1{1-a_f(p)p^{-s}+\xi(p)p^{k-1-2s}}$$ be the finite $L$-function of $f$, and $\Lambda_f(s)=(2\pi)^{-s}\Gamma(s)L_f(s)$ the completed version. Then we have the functional equation $$\label{funceq}
\Lambda_f(s)=\epsilon N^{\frac{k}2-s}\Lambda_{\bar{f}}(k-s),$$ where $\bar{f}\in S_k(\Gamma_0(N),\overline{\xi})$ is the dual of $f$, and $\epsilon\in{\mathbb{C}}$ is the root number. We define $$D_f(s)=L_f(s)\frac{d^2}{ds^2}\log L_f(s)=\sum_{n=1}^{\infty}c_f(n)n^{-s}.$$ Note that $D_f(s)$ continues meromorphically to ${\mathbb{C}}$, with poles precisely at the simple zeros of $L_f(s)$ (including the trivial zeros $s=0, -1, -2, \ldots$).
Next, for any $\alpha\in{\mathbb{Q}}^\times$, we define the additive twists $$L_f(s,\alpha)=\sum_{n=1}^\infty a_f(n)e(\alpha n)n^{-s}
\quad\mbox{and}\quad
D_f(s,\alpha)=\sum_{n=1}^\infty c_f(n)e(\alpha n)n^{-s}.$$ By Deligne’s bound $|a_f(p)|\le 2p^{\frac{k-1}2}$, we see that each of these is holomorphic for $\Re(s)>\frac{k+1}2$. Moreover, it follows from [@bk-converse Prop. 3.1] that $L_f(s,\alpha)$ continues to an entire function. One could similarly prove that $D_f(s,\alpha)$ has meromorphic continuation to ${\mathbb{C}}$ for every $\alpha$, but it turns out to be enough for our purposes to consider $\alpha=1/q$, where $q$ is a prime number not dividing $N$. In this case, we have the following expansion of the exponential function in terms of Dirichlet characters: $$e\!\left(\frac{n}{q}\right)
=1-\frac{q}{q-1}\chi_0(n)
+\frac1{q-1}\sum_{\substack{\chi\;(\text{mod }q)\\\chi\ne\chi_0}}
\tau(\overline{\chi})\chi(n),$$ where $\chi_0\;(\text{mod }q)$ is the trivial character, the sum ranges over all non-trivial $\chi\;(\text{mod }q)$, and $\tau(\overline{\chi})$ denotes the Gauss sum of $\overline{\chi}$. Multiplying both sides by $c_f(n)n^{-s}$ and summing over $n$, we thus see that $$D_f\!\left(s,\frac1q\right)=D_f(s)
-\frac{q}{q-1}D_f(s,\chi_0)
+\frac1{q-1}\sum_{\substack{\chi\;(\text{mod }q)\\\chi\ne\chi_0}}
\tau(\overline{\chi})D_f(s,\chi),$$ where, for each $\chi$, $D_f(s,\chi)$ denotes the multiplicative twist $$D_f(s,\chi)=\sum_{n=1}^\infty c_f(n)\chi(n)n^{-s}.$$
By the known non-vanishing results for automorphic $L$-functions [@jacquet-shalika], all poles of $D_f(s)/\Gamma(s)$ and $D_f(s,\chi)/\Gamma(s)$ for $\chi\ne\chi_0$ are confined to the critical strip $\bigl\{s\in{\mathbb{C}}:\Re(s)\in\bigl(\frac{k-1}2,\frac{k+1}2\bigr)\bigr\}$. On the other hand, from the formula $$\sum_{n=1}^{\infty}a_f(n)\chi_0(n)n^{-s}
=(1-a_f(q)q^{-s}+\xi(q)q^{k-1-2s})L_f(s),$$ it follows that $D_f(s,\chi_0)$ has a pole at every simple zero of the local Euler factor polynomial $1-a_f(q)q^{-s}+\xi(q)q^{k-1-2s}$, except possibly at $s=0$ when $k=1$. By Deligne, the zeros of this polynomial occur on the line $\Re(s)=\frac{k-1}2$, and they are simple if and only if the polynomial is not a square. By the above, we see that $D_f(s,1/q)$ inherits these poles when they occur.
Proof of Theorem \[mainthm\]
============================
The main tool used in the proof is the following proposition, whose proof we defer until the final section.
\[mainprop\] Suppose that $\Lambda_f(s)$ has at most finitely many simple zeros. Then, for any $\alpha\in{\mathbb{Q}}^\times$ and $M\in{\mathbb{Z}}_{\ge0}$, $$\label{maineq}
\begin{aligned}
D_f(s,\alpha)
-\epsilon(i\operatorname{sgn}(\alpha))^k
(N\alpha^2)^{s-\frac{k}2}
\sum_{m=0}^{M-1}
&m!\left(\frac{iN\alpha}{2\pi}\right)^m
{{s+m-1}\choose{m}}
{{s+m-k}\choose{m}}\\
&\cdot D_{\bar{f}}\!\left(s+m,-\frac1{N\alpha}\right)
\end{aligned}$$ continues to a holomorphic function for $\Re(s)>\frac{k+1}2-M$.
From now on we will assume that $\Lambda_f(s)$ has at most finitely many simple zeros and attempt to reach a contradiction. To that end, let $M$ be a positive integer and $q$ a prime not dividing $N$. By Dirichlet’s theorem, there are distinct primes $q_1,\ldots,q_M\in q+N{\mathbb{Z}}$, and it follows that $D_{\bar{f}}(s,-q_j/N)=D_{\bar{f}}(s,-q/N)$ for every $j$. Thus, applying Prop. \[mainprop\] with $\alpha=1/q_j$, we obtain that $$\label{qjhol}
\left(\frac{N}{q_j^2}\right)^{\frac{k}2-s}
D_f\!\left(s,\frac1{q_j}\right)
-\epsilon i^k\sum_{m=0}^{M-1}
m!\left(\frac{iN}{2\pi q_j}\right)^m
{{s+m-1}\choose{m}}
{{s+m-k}\choose{m}}
D_{\bar{f}}\!\left(s+m,-\frac{q}{N}\right)$$ is holomorphic for $\Re(s)>\frac{k+1}2-M$.
Next let $m_0\in{\mathbb{Z}}$ with $0\le m_0<M$. By the Vandermonde determinant, there are numbers $c_1,\ldots,c_M\in{\mathbb{Q}}$ such that $$\sum_{j=1}^M c_jq_j^{-m}=
\begin{cases}
1&\text{if }m=m_0,\\
0&\text{if }m\ne m_0
\end{cases}
\quad\text{for every }
m\in{\mathbb{Z}}\cap[0,M).$$ Multiplying by $-c_j$, summing over $j$ and replacing $s$ by $s-m_0$, we find that $$\epsilon i^k
m_0!\left(\frac{iN}{2\pi}\right)^{m_0}
{{s-1}\choose{m_0}}
{{s-k}\choose{m_0}}
D_{\bar{f}}\!\left(s,-\frac{q}{N}\right)
-\sum_{j=1}^Mc_j
\left(\frac{N}{q_j^2}\right)^{\frac{k}2+m_0-s}
D_f\!\left(s-m_0,\frac1{q_j}\right)$$ is holomorphic for $\Re(s)>m_0+\frac{k+1}2-M$. This establishes the meromorphic continuation of $D_{\bar{f}}(s,-q/N)$ to that region. Moreover, since $D_f(s,1/q_j)$ is holomorphic on $\{s\in{\mathbb{C}}:\Re(s)<\frac{k-1}2\}\setminus{\mathbb{Z}}$ for each $j$, we see that $D_{\bar{f}}(s,-q/N)$ is holomorphic on $\{s\in{\mathbb{C}}:\Re(s)\in(m_0+\frac{k+1}2-M,m_0+\frac{k-1}2)\}\setminus{\mathbb{Z}}$. Thus, choosing $m_0=2$ and $M$ arbitrarily large, we find that $D_{\bar{f}}(s,-q/N)$ has meromorphic continuation to ${\mathbb{C}}$, with poles possible only at integer points.
Hence, applying Prop. \[mainprop\] again with $\alpha=1/q$ and $M=2$, we learn that $D_f(s,1/q)$ can only have poles at integer points. However, we have already seen that $D_f(s,1/q)$ has a pole at every simple zero (except possibly $s=0$) of the local Euler factor polynomial $1-a_f(q)q^{-s}+\xi(q)q^{k-1-2s}$. This polynomial, in turn, has infinitely many simple zeros along the line $\Re(s)=\frac{k-1}2$ if and only if $|a_f(q)|<2q^{\frac{k-1}2}$. By the Rankin–Selberg method, the average value of $|a_f(q)|^2/q^{k-1}$ is $1$, so such primes $q$ exist in abundance. This concludes the proof of Theorem \[mainthm\].
Proof of Proposition \[mainprop\]
=================================
Let $\Delta_f(s)=(2\pi)^{-s}\Gamma(s)D_f(s)$. Taking the logarithm of and differentiating twice, we find $$\psi'(s)+\frac{d^2}{ds^2}\log L_f(s)
=\psi'(k-s)+\frac{d^2}{ds^2}\log L_{\bar f}(k-s),$$ where $\psi(s)=\frac{\Gamma'}{\Gamma}(s)$ is the digamma function. Thus, it follows that $$\label{dfunceq}
\Delta_f(s)+\Lambda_f(s)(\psi'(s)-\psi'(k-s))
=\epsilon N^{\frac{k}2-s}\Delta_{\bar{f}}(k-s).$$ Next, since $\Lambda_f(s)$ has at most finitely many simple zeros, there is a rectangle $\mathcal{C}$ contained within the critical strip $\{s\in{\mathbb{C}}:\Re(s)\in(\frac{k-1}2,\frac{k+1}2)\}$ which encloses all simple zeros. For $z\in{\mathbb{H}}=\{z\in{\mathbb{C}}:\Im(z)>0\}$, we define $$F(z)=\sum_{n=1}^\infty c_f(n)e(nz),\quad
\overline{F}(z)=\sum_{n=1}^\infty c_{\bar{f}}(n)e(nz),$$ $$A(z)=\frac1{2\pi i}\int_{\Re(s)=k-\frac12}
\bigl(\psi'(s)+\psi'(s+1-k)\bigr)\Lambda_f(s)(-iz)^{-s}\,ds,$$ and $$B(z)=\frac1{2\pi i}\int_{\mathcal{C}}\Delta_f(s)(-iz)^{-s}\,ds
+\frac1{2\pi i}\int_{\Re(s)=k-\frac12}\frac{\pi^2}{\sin^2(\pi s)}
\Lambda_f(s)(-iz)^{-s}\,ds.$$ Here $\mathcal{C}$ is given counter-clockwise orientation, and $(-iz)^{-s}$ is defined as $e^{-s\log(-iz)}$ using the principal branch of the logarithm.
These functions are related as follows:
\[mainidlemma\] We have $$\label{mainidentity}
F(z)+A(z)
=\epsilon(-i\sqrt{N}z)^{-k}\overline{F}\!\left(-\frac1{Nz}\right)+B(z)$$ for all $z\in{\mathbb{H}}$.
By Mellin inversion, we have $$F(z)=\frac1{2\pi i}\int_{\Re(s)=\frac{k}2+1}\Delta_f(s)(-iz)^{-s}\,ds$$ and $$\epsilon(-i\sqrt{N}z)^{-k}\overline{F}\!\left(-\frac1{Nz}\right)
=\frac{\epsilon N^{k/2}}{2\pi i}
\int_{\Re(s)=\frac{k}2+1}\Delta_{\bar{f}}(s)(-iNz)^{s-k}\,ds.$$ Since $\Lambda_f(s)$ has at most finitely many simple zeros, there is a $\delta>0$ such that $\Delta_{\bar{f}}(s)$ is holomorphic for $\Re(s)>\frac{k+1}2-\delta$. Moreover, it follows from the Phragmén–Lindelöf convexity principle that for any fixed $z$, $\Delta_{\bar{f}}(s)(-iNz)^{s-k}$ decays rapidly as $|\Im(s)|\to\infty$ in any fixed vertical strip. Hence, we may shift the contour of the last line to $\Re(s)=\frac{k+1-\delta}2$ and apply to obtain $$\label{fbarint}
\begin{aligned}
&\frac{\epsilon N^{k/2}}{2\pi i}
\int_{\Re(s)=\frac{k+1-\delta}2}\Delta_{\bar{f}}(s)(-iNz)^{s-k}\,ds
=\frac{\epsilon N^{k/2}}{2\pi i}
\int_{\Re(s)=\frac{k-1+\delta}2}
\Delta_{\bar{f}}(k-s)(-iNz)^{-s}\,ds\\
&=\frac1{2\pi i}\int_{\Re(s)=\frac{k-1+\delta}2}\Delta_f(s)(-iz)^{-s}\,ds
+\int_{\Re(s)=\frac{k-1+\delta}2}\Lambda_f(s)[\psi'(s)-\psi'(k-s)]
(-iz)^{-s}\,ds.
\end{aligned}$$
Note that $$\frac1{2\pi i}\int_{\Re(s)=\frac{k}2+1}\Delta_f(s)(-iz)^{-s}\,ds
-\frac1{2\pi i}\int_{\Re(s)=\frac{k-1+\delta}2}\Delta_f(s)(-iz)^{-s}\,ds
=\frac1{2\pi i}\int_{\mathcal{C}}\Delta_f(s)(-iz)^{-s}\,ds,$$ which is the first term of $B(z)$. Next, since $\psi'(s)-\psi'(k-s)$ is holomorphic for $\Re(s)\in(0,k)$, we may shift the contour of the last integral in to $\Re(s)=k-\frac12$. Using the reflection formula $\psi'(1-s)+\psi'(s)=\pi^2/\sin^2(\pi s)$, we have $$\psi'(s)-\psi'(k-s)=\psi'(s)+\psi'(s+1-k)-\frac{\pi^2}{\sin^2(\pi s)}.$$ This yields $A(z)$ and the remaining term of $B(z)$.
Now, the main idea of the proof of Prop. \[mainprop\] is to compute $(2\pi)^s/\Gamma(s)$ times the Mellin transform of both sides of along the line $\Re(z)=\alpha\in{\mathbb{Q}}^\times$. For $F(z)$, we have $$\label{Fmellin}
\begin{aligned}
\frac{(2\pi)^s}{\Gamma(s)}\int_0^\infty F(\alpha+iy)y^s\frac{dy}y
&=\frac{(2\pi)^s}{\Gamma(s)}\int_0^\infty
\sum_{n=1}^\infty c_f(n)e(\alpha n)e^{-2\pi ny}y^s\frac{dy}y\\
&=\sum_{n=1}^\infty c_f(n)e(\alpha n)n^{-s}
=D_f(s,\alpha).
\end{aligned}$$
\[Alemma\] For any $\alpha\in{\mathbb{Q}}^\times$, $$\frac{(2\pi)^s}{\Gamma(s)}\int_0^\infty A(\alpha+iy)y^s\frac{dy}y$$ continues to an entire function of $s$.
Set $\Phi(s)=\psi'(s)+\psi'(s+1-k)$. From the identity $\psi'(s)=\int_1^\infty\frac{\log{x}}{x-1}x^{-s}\,dx$, we get the integral representation $\Phi(s)=\int_1^\infty\phi(x)x^{-s}\,dx$ for $\Re(s)>k-1$, where $\phi(x)=\frac{(x^{k-1}+1)\log{x}}{x-1}$. Hence $$\begin{aligned}
\Phi(s)\Gamma(s)
&=\int_1^\infty\phi(x)\int_0^\infty e^{-y}(y/x)^s\frac{dy}y\,dx
=\int_1^\infty\phi(x)\int_0^\infty e^{-xy}y^s\frac{dy}y\,dx\\
&=\int_0^\infty\int_1^\infty\phi(x)e^{-xy}\,dx\,y^s\frac{dy}y.\end{aligned}$$ Therefore, by Mellin inversion, $$A(z)=\frac1{2\pi i}\int_{\Re(s)=k+1}\Phi(s)\Gamma(s)
\sum_{n=1}^\infty a_f(n)(-2\pi inz)^{-s}\,ds
=\sum_{n=1}^\infty a_f(n)\int_1^\infty\phi(x)e(nxz)\,dx.$$ Specializing to $z=\alpha+iy$, we get $$A(\alpha+iy)=\sum_{n=1}^\infty a_f(n)\int_1^\infty\phi(x)e(\alpha nx)
e^{-2\pi nxy}\,dx,$$ so that $$\begin{aligned}
\int_0^\infty A(\alpha+iy)y^s\frac{dy}y
&=\sum_{n=1}^\infty a_f(n)\int_1^\infty\phi(x)e(\alpha nx)
\int_0^\infty e^{-2\pi nxy}y^s\frac{dy}y\,dx\\
&=\sum_{n=1}^\infty a_f(n)(2\pi n)^{-s}\Gamma(s)
\int_1^\infty\phi(x)e(\alpha nx)x^{-s}\,dx.\end{aligned}$$
For $j=0,1,2,\ldots$, define functions $\phi_j=\phi_j(x,s)$ recursively by $$\phi_0=\phi,
\quad\phi_{j+1}=x\frac{\partial\phi_j}{\partial x}-(s+j)\phi_j.$$ Then, by integration by parts, $$\int_1^\infty\phi_j(x,s)e(\alpha nx)x^{-s-j}\,dx
=-\frac{e(\alpha n)\phi_j(1,s)}{2\pi i\alpha n}
-\frac1{2\pi i\alpha n}\int_1^\infty\phi_{j+1}(x,s)e(\alpha nx)x^{-s-j-1}\,dx.$$ Applying this iteratively $m$ times, we find $$\begin{aligned}
\int_1^\infty\phi(x)e(\alpha nx)x^{-s}\,dx
&=e(\alpha n)\sum_{j=0}^{m-1}\frac{\phi_j(1,s)}{(-2\pi i\alpha n)^{j+1}}\\
&+(-2\pi i\alpha n)^{-m}\int_1^{\infty}\phi_m(x,s)e(\alpha nx)x^{-s-m}\,dx.\end{aligned}$$ Substituting this back into the above, we have $$\begin{aligned}
\frac{(2\pi)^s}{\Gamma(s)}\int_0^\infty A(\alpha+iy)y^s\frac{dy}y
&=\sum_{j=0}^{m-1}\frac{\phi_j(1,s)}{(-2\pi i\alpha)^{j+1}}
L_f(s+j+1,\alpha)\\
&+(-2\pi i\alpha)^{-m}
\sum_{n=1}^{\infty}\frac{a_f(n)}{n^{s+m}}
\int_1^\infty\phi_m(x,s)e(\alpha nx)x^{-s-m}\,dx.\end{aligned}$$ Each of the terms in the sum over $j$ continues to an entire function of $s$. On the other hand, it is straightforward to prove that $\phi_m(x,s)\ll_m(1+|s|)^mx^{k-1}$. Thus, the final sum over $n$ is holomorphic for $\Re(s)>k-m$. Taking $m$ arbitrarily large establishes the lemma.
\[Fbarlemma\] Let $\alpha\in{\mathbb{Q}}^\times$ and $z=\alpha+iy$ for some $y\in\bigl(0,\frac{|\alpha|}4\bigr]$. Then $$\label{fbartaylor}
\begin{aligned}
\epsilon&(-i\sqrt{N}z)^{-k}\overline{F}\!\left(-\frac1{Nz}\right)\\
&=O_{\alpha,M}(y^{M-\lfloor\frac{k+3}2\rfloor})
+\epsilon N^{-\frac{k}2}\sum_{m=0}^{M-1}
\frac{(-i\alpha)^{-m-k}}{2\pi i}\int_{\Re(s)=\frac{k}2+1}
{{s+m-k}\choose{m}}
\left(\frac{N\alpha^2}{2\pi}\right)^{s+m}\\
&\hspace{8cm}\cdot
\Gamma(s+m)
D_{\bar{f}}\left(s+m,-\frac1{N\alpha}\right)y^{-s}\,ds
\end{aligned}$$ for every $M\in{\mathbb{Z}}_{\ge0}$.
This was essentially done in [@bk-converse §2]; we reproduce the argument here for the sake of completeness. Let $z=\alpha+iy$, $\beta=-1/N\alpha$ and $u=y/\alpha$. Then $$-\frac1{Nz}=\beta+i|\beta u|-\frac{\beta u^2}{1+iu},$$ so that $$\epsilon(-i\sqrt{N}z)^{-k}\overline{F}\!\left(-\frac1{Nz}\right)
=\epsilon (-i\sqrt{N}\alpha)^{-k}
\sum_{n=1}^\infty c_{\bar{f}}(n)e(\beta n)
e^{-2\pi n|\beta u|}
(1+iu)^{-k}e\!\left(-\frac{n\beta u^2}{1+iu}\right).$$ Next, $$\begin{aligned}
(1+iu)^{-k}e\!\left(-\frac{n\beta u^2}{1+iu}\right)
&=\sum_{j=0}^\infty(-iu)^j(1+iu)^{-j-k}\frac{(-2\pi n|\beta u|)^j}{j!}\\
&=\sum_{j=0}^\infty\sum_{\ell=0}^\infty
{{j+k+\ell-1}\choose\ell}(-iu)^{j+\ell}\frac{(-2\pi n|\beta u|)^j}{j!}\\
&=\sum_{m=0}^\infty(-iu)^m\sum_{j=0}^m
{{m+k-1}\choose{m-j}}\frac{(-2\pi n|\beta u|)^j}{j!}.\end{aligned}$$ Note further that for any $M, K\in{\mathbb{Z}}_{\ge0}$ we have $$\begin{aligned}
&\hspace{-1cm}\left|\sum_{m=M}^{\infty}(-iu)^m
\sum_{j=0}^m{{m+k-1}\choose{m-j}}
\frac{(-2\pi n|\beta u|)^j}{j!}\right|\\
&\le (2\pi n|\beta u|)^{-K}K!
\sum_{m=M}^{\infty}|u|^m\sum_{j=0}^m{{m+k-1}\choose{m-j}}{{j+K}\choose{j}}
\frac{(2\pi n|\beta u|)^{j+K}}{(j+K)!}\\
&\le (\pi n|\beta u|)^{-K}K!(3/2)^{k-1}
\sum_{m=M}^{\infty}(3|u|)^me^{2\pi n|\beta u|}\\
&\ll_{\alpha,M,K}|u|^{M-K}n^{-K}e^{2\pi n|\beta u|},
\end{aligned}$$ since $|u|\le1/4$. Hence, substituting the definition of $u$, we have $$\begin{aligned}
&\epsilon(-i\sqrt{N}z)^{-k}\overline{F}\!\left(-\frac1{Nz}\right)
=O_{\alpha,M,K}\Bigl(y^{M-K}\sum_{n=1}^\infty|c_f(n)|n^{-K}\Bigr)\\
&+\epsilon(-i\sqrt{N}\alpha)^{-k}
\sum_{m=0}^{M-1}\left(\frac{-iy}{\alpha}\right)^m
\sum_{j=0}^m{{m+k-1}\choose{m-j}}
\sum_{n=1}^\infty c_{\bar{f}}(n)e(\beta n)
\frac1{j!}\left(-\frac{2\pi ny}{N\alpha^2}\right)^j
e^{-\frac{2\pi ny}{N\alpha^2}}.
\end{aligned}$$ Choosing $K=\lfloor\frac{k-1}2\rfloor+2$, the error term converges and gives the estimate $O_{\alpha,M}(y^{M-K})$.
As for the other terms, we have $$\begin{aligned}
y^m\sum_{n=1}^\infty&c_{\bar{f}}(n)e(\beta n)
\frac1{j!}\left(-\frac{2\pi ny}{N\alpha^2}\right)^j
e^{-\frac{2\pi ny}{N\alpha^2}}
=\frac{y^{j+m}}{j!}\frac{d^j}{dy^j}
\sum_{n=1}^\infty c_{\bar{f}}(n)e(\beta n)
e^{-\frac{2\pi ny}{N\alpha^2}}\\
&=\frac{y^{j+m}}{j!}\frac{d^j}{dy^j}
\frac1{2\pi i}\int_{\Re(s)=m+\frac{k}2+1}
\left(\frac{N\alpha^2}{2\pi}\right)^s\Gamma(s)
D_{\bar{f}}(s,\beta)y^{-s}\,ds\\
&=\frac1{2\pi i}\int_{\Re(s)=\frac{k}2+1}
{{-s-m}\choose{j}}
\left(\frac{N\alpha^2}{2\pi}\right)^{s+m}\Gamma(s+m)
D_{\bar{f}}(s+m,\beta)y^{-s}\,ds.\end{aligned}$$ Moreover, by the Chu–Vandermonde identity we have $$\sum_{j=0}^m{{m+k-1}\choose{m-j}}{{-s-m}\choose{j}}
={{-s+k-1}\choose{m}}=(-1)^m{{s+m-k}\choose{m}}.$$ Collecting these strands together, we arrive at .
\[Blemma\] For any $\alpha\in{\mathbb{Q}}^\times$ there are numbers $P_j(\alpha)$, $j=0,1,2,\ldots$, such that $$B(\alpha+iy)=\sum_{j=0}^{M-1}P_j(\alpha)y^j+O_{\alpha,M}(y^M)$$ for all $M\in{\mathbb{Z}}_{\ge0}$ and $y\in\bigl(0,\frac{|\alpha|}4\bigr]$.
For $z=\alpha+iy$, we have $$\label{izstaylor}
(-iz)^{-s}=e^{i\frac{\pi}2\operatorname{sgn}(\alpha)s}|\alpha|^{-s}
\left(1+\frac{iy}{\alpha}\right)^{-s}
=e^{i\frac{\pi}2\operatorname{sgn}(\alpha)s}|\alpha|^{-s}
\sum_{j=0}^\infty {{-s}\choose{j}}
\left(\frac{iy}{\alpha}\right)^j.$$ Since $y\le\frac{|\alpha|}4$, the crude bound $$\left|{{-s}\choose{j}}\right|
=\left|{{s+j-1}\choose{j}}\right|\le 2^{|s|+j}$$ yields $$\sum_{j=M}^{\infty}
{{-s}\choose{j}}\left(\frac{iy}{\alpha}\right)^j
\ll_{\alpha,M} 2^{|s|}y^M.$$ Hence, if we truncate the sum in at $M$ and substitute it for $(-iz)^{-s}$ in the definition of $B(z)$, then since the contour $\mathcal{C}$ is compact, the first integral of the error term converges to give an $O_{\alpha,M}(y^M)$ error overall. Similarly, by standard estimates, along the line $\Re(s)=k-\frac12$ the function $e^{i\frac{\pi}2\operatorname{sgn}(\alpha)s}|\alpha|^{-s}\Lambda_f(s)$ has at most polynomial growth, and $\frac{\pi^2}{\sin^2(\pi s)}\ll e^{-2\pi|s|}$. Since $e^{2\pi}>2$, the second integral of the error term converges as well, and the lemma follows with $$\begin{aligned}
P_j(\alpha)&=
\frac1{2\pi i}\int_{\mathcal{C}}
(-i\alpha)^{-j}e^{i\frac{\pi}2\operatorname{sgn}(\alpha)s}|\alpha|^{-s}{{-s}\choose j}
\Delta_f(s)\,ds\\
&+\frac1{2\pi i}\int_{\Re(s)=k-\frac12}
(-i\alpha)^{-j}e^{i\frac{\pi}2\operatorname{sgn}(\alpha)s}|\alpha|^{-s}{{-s}\choose j}
\Lambda_f(s)\frac{\pi^2}{\sin^2(\pi s)}\,ds.\end{aligned}$$
Now, to conclude the proof, let us define $$\begin{aligned}
g(y)&=F(\alpha+iy)+A(\alpha+iy)
-\sum_{j=0}^{M-1}P_j(\alpha)y^j\chi_{(0,|\alpha|/4]}(y)\\
&-\epsilon N^{-\frac{k}2}\sum_{m=0}^{M-1}
\frac{(-i\alpha)^{-m-k}}{2\pi i}\int_{\Re(s)=\frac{k}2+1}
{{s+m-k}\choose{m}}
\left(\frac{N\alpha^2}{2\pi}\right)^{s+m}\\
&\hspace{6cm}\cdot\Gamma(s+m)
D_{\bar{f}}\left(s+m,-\frac1{N\alpha}\right)y^{-s}\,ds,\end{aligned}$$ where $\chi_{(0,|\alpha|/4]}(y)=1$ if $y\in\bigl(0,\frac{|\alpha|}4\bigr]$ and $0$ otherwise. Combining Lemmas \[mainidlemma\], \[Fbarlemma\] and \[Blemma\], we have that $g(y)=O_{\alpha,M}(y^{M-\lfloor\frac{k+3}2\rfloor})$ for $y\in\bigl(0,\frac{|\alpha|}4\bigr]$. On the other hand, it is easy to see that $g(y)$ decays rapidly as $y\to\infty$. Thus, $\frac{(2\pi)^s}{\Gamma(s)}\int_0^\infty g(y)y^{s-1}\,dy$ defines a holomorphic function for $\Re(s)>\lfloor\frac{k+3}2\rfloor-M$.
Note that $$\frac{(2\pi)^s}{\Gamma(s)}\int_0^{\infty}\sum_{j=0}^{M-1}P_j(\alpha)
y^j\chi_{(0,|\alpha|/4]}(y)y^s\frac{dy}y
=\frac{(2\pi)^s}{\Gamma(s)}\sum_{j=0}^{M-1}P_j(\alpha)
\frac{(|\alpha|/4)^{s+j}}{s+j}$$ extends to an entire function of $s$. Together with and Lemma \[Alemma\], this shows that is holomorphic for $\Re(s)>\lfloor\frac{k+3}2\rfloor-M$. Finally, we replace $M$ by $M+1$ and discard the final term of the sum over $m$ to see that is in fact holomorphic for $\Re(s)>\frac{k+1}2-M$.
[1]{}
Andrew R. Booker and M. Krishnamurthy, *A strengthening of the [${\rm
GL}(2)$]{} converse theorem*, Compos. Math. **147** (2011), no. 3, 669–715. [MR ]{}[2801397 (2012f:11090)]{}
[to3em]{}, *[W]{}eil’s converse theorem with poles*, preprint, 2011.
Peter Jaehyun Cho, *Simple zeros of [M]{}aass [$L$]{}-functions*, Int. J. Number Theory **9** (2013), no. 1, 167–178.
J. B. Conrey and A. Ghosh, *Simple zeros of the [R]{}amanujan [$\tau$]{}-[D]{}irichlet series*, Invent. Math. **94** (1988), no. 2, 403–419. [MR ]{}[958837 (89k:11078)]{}
J. B. Conrey, A. Ghosh, and S. M. Gonek, *Simple zeros of the zeta function of a quadratic number field. [I]{}*, Invent. Math. **86** (1986), no. 3, 563–576. [MR ]{}[860683 (87m:11114)]{}
Herv[é]{} Jacquet and Joseph A. Shalika, *A non-vanishing theorem for zeta functions of [${\rm GL}_{n}$]{}*, Invent. Math. **38** (1976/77), no. 1, 1–16. [MR ]{}[0432596 (55 \#5583)]{}
Jerzy Kaczorowski and Alberto Perelli, *On the structure of the [S]{}elberg class, [VII]{}: [$1<d<2$]{}*, Ann. of Math. (2) **173** (2011), no. 3, 1397–1441. [MR ]{}[2800717]{}
[^1]: The author was supported by EPSRC Fellowship EP/H005188/1.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: '[[ We study the property of some discount functions known as Decreasing Impatience (DI). As proved by Prelec [@prelec2004decreasing], one discount function exhibits more DI than another, if the logarithm of the former is more convex than the logarithm of the latter. Using this characteristic, we show that the aggregation of distinct discount functions from comparable DI classes will result in the aggregate discount function being strictly more DI than the least DI of its components.]{}]{}'
author:
- Nina Anchugina
- Matthew Ryan
- Arkadii Slinko
bibliography:
- 'Biblio.bib'
date: April 2016
title: 'Aggregating time preferences with decreasing impatience$^*$'
---
[**Abstract.**]{} It is well-known that for a group of time-consistent decision makers their collective time preferences may become time-inconsistent. Jackson and Yariv [@jackson2014present] demonstrated that the result of aggregation of exponential discount functions always exhibits present bias. We show that when preferences satisfy the axioms of Fishburn and Rubinstein [@fishburn1982time], present bias is equivalent to decreasing impatience (DI). Applying the notion of comparative DI introduced by Prelec [@prelec2004decreasing], we generalize the result of Jackson and Yariv [@jackson2014present]. We prove that the aggregation of distinct discount functions from comparable DI classes results in the collective discount function which is strictly more DI than the least DI of the functions being aggregated.
We also prove an analogue of Weitzman’s [@weitzman1998far] result, for hyperbolic rather than exponential discount functions. We show that if a decision maker is uncertain about her hyperbolic discount rate, then long-term costs and benefits will be discounted at a rate which is the probability-weighted harmonic mean of the possible hyperbolic discount rates.
[**Keywords:**]{} Discounting, hyperbolic discounting, decreasing impatience, aggregation.\
[**JEL Classification:**]{} D71, D90.
$^*$ We thank Matthew Jackson, Simon Grant and several seminar audiences for comments and suggestions. Arkadii Slinko was supported by the Marsden Fund grant UOA 1420, and Nina Anchugina gratefully acknowledges financial support from the University of Auckland.
****On Decreasing Impatience****\
Nina Anchugina\
Department of Mathematics, The University of Auckland, n.anchugina@auckland.ac.nz\
Introduction
============
Sometimes decisions about timed outcomes have to be made by a group of individuals, such as boards, committees or households. It is natural to think that individuals may differ in the discounting procedure that they use. If the decision is to be made by a group of individuals it is desirable to have an aggregating procedure that suitably reflects the time preferences of all members. The natural option is to average the discount functions across individuals, which is equivalent to averaging the discounted utilities in the case when all agents have identical utility functions. This approach has been widely used in the existing literature on time preferences. It is known that such collective discount functions need not share properties that are common to the individual discount functions being aggregated. As Jackson and Yariv demonstrate [@jackson2014present], if individuals discount the future exponentially and there is a heterogeneity in discount factors, then their aggregate discount function exhibits present bias, which means that delaying two different dated-outcomes by the same amount of time can reverse the ranking of these outcomes. Moreover, when the number of individuals grows, in the limit the group discount function becomes hyperbolic [@jackson2014present].
Jackson and Yariv [@jackson2014present] give the following example of present-biased group preferences for a household with two time-consistent individuals, Constantine and Patience. Both have identical instantaneous utility functions, and discount the future exponentially, but Constantine has a discount factor of 0.5, whereas Patience has a discount factor of 0.8. Suppose that they need to choose between 10 utiles for each today or 15 utiles for each tomorrow. They calculate the aggregate discounted utility for each option: $10+10=20$ and $15(0.8+0.5)=19.5$. Therefore, 10 utiles today is chosen. Now suppose that they must choose between 10 utiles at time $t\geq 1$ and 15 utiles at $t+1$. The aggregate discounted utilities in this case are $10(0.8^t+0.5^t)$ and $15(0.8^{t+1}+0.5^{t+1})$, respectively. For any $t\geq 1$ the 15 utiles at $t+1$ is preferable to the 10 utiles at $t$, which reverses the initial preference for 10 utiles at $t=0$ over 15 utiles at $t=1$. The behaviour of the household is present-biased.
Another scenario in which the aggregation of time preferences may be required is when a single decision maker is uncertain about the appropriate discount function to apply. For example, discounting may be affected by a survival function with a constant but uncertain hazard rate. Such scenarios are considered by Weitzman [@weitzman1998far] and Sozou [@sozou1998hyperbolic]. If the decision-maker maximizes expected discounted utility, then she maximizes discounted utility for a certainty equivalent discount function, calculated as the probability-weighted average of the different possible discount functions that may apply. Weitzman [@weitzman1998far] shows that if each of the possible rates of time preference converges to some non-negative value (as time goes to infinity), then the certainty equivalent time preference function converges to the lowest of these limits. Similarly, Sozou [@sozou1998hyperbolic] considers a decision maker whose discounting reflects a survival function with a constant, but uncertain, hazard rate. If this hazard rate is exponentially distributed, Sozou shows that the decision-maker’s expected discount function is hyperbolic. Of course, present bias is not limited to aggregate or expected discount functions. It is often observed in experiments that individual decision makers become decreasingly impatient (increasingly patient) as rewards are shifted further into the future. If a decision-maker is indifferent between an early outcome and a larger, later outcome, then delaying both outcomes by the same amount of time will often result in the larger, later outcome being preferred. Such subjects exhibit present bias, or strictly decreasing impatience (DI). Exponential discounting implies constant impatience, so it cannot explain *strictly* decreasing impatience, either globally or locally. The necessity of accommodating DI in individual time preference has made hyperbolic discounting a significant tool in behavioural economics. Several types of hyperbolic discount functions have been introduced, including quasi-hyperbolic discounting [@phelps1968second; @laibson1997golden], discounting for delay [@ainslie1975specious], proportional hyperbolic discounting [@harvey1995proportional; @mazur2001hyperbolic], and generalized hyperbolic discounting [@loewenstein1992anomalies; @al2006note]. Given the widespread use of hyperbolic discount functions to describe individual time preferences, it is important to understand the behaviour of aggregated, or averaged, hyperbolic functions.
The goal of this paper is twofold. Firstly, we seek to extend Jackson and Yariv’s result on the aggregation of exponential discount functions. Two individuals may differ in the rate at which their impatience decreases, but their respective levels of DI may be comparable – the preferences of the one may exhibit unambiguously more DI than the preferences of the other. As Prelec [@prelec2004decreasing] proved, one individual exhibits more DI than another if the logarithm of the discount function of the former is more convex than that of the latter. Can we say anything about the level of DI of the weighted average of individual discount functions that can be (weakly) ordered by DI? Theorem \[main\] establishes that the weighted average always exhibits *strictly* more DI than the component with the *least* DI. This generalizes Jackson and Yariv’s result. Proposition 1 in [@jackson2014present] shows that the weighted average of exponential discount functions with different discount factors exhibits present bias. We show that when preferences satisfy the axioms of Fishburn and Rubinstein [@fishburn1982time], Jackson and Yariv’s definition of present bias is equivalent to strictly decreasing impatience. Since all exponential discount functions exhibit constant impatience – they all exhibit the same degree of DI – Proposition 1 of Jackson and Yariv is a special case of our Theorem \[main\].
Our second goal is to prove an analogue of Weitzman’s [@weitzman1998far] result: one in which discounting is hyperbolic but there is an uncertainty about the hyperbolic discount factor. The answer, given in Theorem \[main2\], is very different to Weitzman’s answer for the case of exponential discounting. We show that the certainty equivalent hyperbolic discount factor converges, not to the lowest individual hyperbolic discount factor, but to the probability-weighted harmonic mean of the individual hyperbolic discount factors.
Preliminaries
=============
In this section we introduce the framework for our investigation and define the two key concepts used in this paper: present bias and strictly decreasing impatience of preferences. We prove that these two concepts coincide when the Fishburn-Rubinstein axioms for a discounted utility representation are satisfied. Taking our lead from Pratt [@pratt1964risk] and Arrow [@arrow1965aspects], these concepts are discussed in terms of log-convexity of discount functions, hence we introduce necessary results and definitions in this regard. Most results are known but included to keep the paper self-contained.
Convexity and log-convexity
---------------------------
Convexity and log-convexity play an important role in the theory of discounting. Let $I$ be an interval (finite or infinite) of real numbers. A function $f \colon I \to \mathbb{R}$ is [*convex*]{} if for any two points $x, y \in I$ and any $\lambda \in [0, 1]$ it holds that: $$f\left( \lambda x+\left( 1-\lambda\right) y\right)\leq \lambda f(x)+(1-\lambda) f(y).$$ A function $f$ is [*strictly convex*]{} if $$f(\lambda x+(1-\lambda) y)< \lambda f(x)+(1-\lambda) f(y)$$ for any $x, y \in I$ such that $x\neq y$ and any $\lambda \in (0, 1)$. If $f$ is twice differentiable convexity is equivalent to $f''\geq 0$, and strict convexity is equivalent to two conditions: the function $f''$ is nonnegative on $I$ and the set $ \{x \in I \ \vline \ f''(x)=0 \} $ contains no non-trivial interval [@stein2012twice].
The following equivalent definition of a (strictly) convex function is well known. A function $f\colon I \to \mathbb{R}$ is (strictly) convex if for every $x, y, v, z \in I$ such that $x-y=v-z>0$ and $y>z$ we have $$f(x)-f(y)\leq [<] f(v)-f(z).$$
Convexity is preserved under composition of functions, as shown in the following lemma, whose straightforward proof is omitted:
\[conv\] Let $f_1\colon I \to \mathbb{R}$ be a non-decreasing and convex function and $f_2\colon I \to \mathbb{R}$ be a convex function, such that the range of $f_2$ is contained in the domain of $f_1$. Then the composition $f = f_1 \circ f_2$ is a convex function. If, in addition, $f_1$ is strictly increasing, and either $f_1$ or $f_2$ is strictly convex, then $f$ is also strictly convex.
A function $f\colon I \to \mathbb{R}$ is called [*log-convex*]{} if $f(x)>0$ for all $x\in I$ and $\ln(f)$ is convex. It is called [*strictly log-convex*]{} if $\ln(f)$ is strictly convex. If follows that if $f$ is a (strictly positive) twice differentiable function, then log-convexity of $f$ is equivalent to the condition $f''f - (f')^2\geq 0$, while strict log-convexity of $f$ requires, in addition, that the set $$\{ \: x \in I \ \vline \ f''(x)f(x) - [f'(x)]^2=0 \: \}$$ contains no non-trivial interval. Log-convexity can also be expressed without using logarithms [@boyd2004convex]. A function $f \colon I \to \mathbb{R}$ is log-convex if and only if $f(x)>0$ for all $x \in I$ and for all $x, y \in I$ and $\lambda \in [0, 1]$ we have: $$\label{ineqlc}
f(\lambda x+(1-\lambda)y) \leq f(x)^{\lambda }f(y)^{1-\lambda}.$$ The function $f$ is strictly log-convex if inequality is strict when $x\neq y$ and $\lambda\in (0,1)$.
The following result appears to be well known, but a formal reference is elusive so we have included a proof here for completeness.
\[log-convex\] Let $f, g \colon I \to \mathbb{R}$ be functions with $f$ strictly log-convex and $g$ log-convex. Then the sum $f+g$ is strictly log-convex.
Since $f(x)>0$ and $g(x)>0$ for all $x\in I$, we have $(f+g)(x)>0$ for all $x\in I$. Let $x, y \in I$ such that $x\neq y$ and let $\lambda \in (0, 1)$. We must show that $$\label{eq}
f(\lambda x+(1-\lambda)y)+g(\lambda x+(1-\lambda)y) < (f(x)+g(x))^{\lambda }(f(y)+g(y))^{1-\lambda}.$$ Since $f$ is strictly log-convex, we have $$\label{eq1}
f(\lambda x+(1-\lambda)y) < f(x)^{\lambda}f(y)^{1-\lambda}.$$ Analogously, since $g(x)$ is log-convex: $$\label{eq2}
g(\lambda x+(1-\lambda)y) \leq g(x)^{\lambda}g(y)^{1-\lambda}.$$ Summing and we obtain: $$\label{eqsum}
f(\lambda x+(1-\lambda)y)+g(\lambda x+(1-\lambda)y) < f(x)^{\lambda }f(y)^{1-\lambda}+g(x)^{\lambda }g(y)^{1-\lambda} .$$ Denote $a=f(x), b=f(y), c=g(x), d=g(y)$. Note that $a, b, c, d >0$. To prove the claim of the lemma, it is sufficient to show that: $$\label{sumnot}
a^{\lambda }b^{1-\lambda}+c^{\lambda }d^{1-\lambda} \leq (a+c)^{\lambda }(b+d)^{1-\lambda}.$$ Since $(a+c)^{\lambda }(b+d)^{1-\lambda}>0$ we can divide both parts of by this expression to get $$\left (\frac{a}{a+c} \right)^{\lambda} \left (\frac{b}{b+d} \right )^{1-\lambda} +\left (\frac{c}{a+c}\right)^{\lambda}\left(\frac{d}{b+d}\right)^{1-\lambda} \ \leq \ 1.$$ By the Weighted AM-GM inequality [@cvetkovski2012inequalities Theorem 7.6, p. 74]: $$\left (\frac{a}{a+c} \right)^{\lambda} \left (\frac{b}{b+d} \right )^{1-\lambda} \ \leq\ \lambda \frac{a}{a+c}+(1-\lambda)\frac{b}{b+d}$$ and $$\left (\frac{c}{a+c} \right)^{\lambda} \left (\frac{d}{b+d} \right )^{1-\lambda} \ \leq\ \lambda \frac{c}{a+c}+(1-\lambda)\frac{d}{b+d}.$$ Hence, $$\left (\frac{a}{a+c} \right)^{\lambda} \left (\frac{b}{b+d} \right )^{1-\lambda} +\left (\frac{c}{a+c}\right)^{\lambda}\left(\frac{d}{b+d}\right)^{1-\lambda}\ \leq\ \lambda+(1-\lambda)=1,$$ which proves the statement in the lemma.
One of the important definitions which will be frequently used throughout the paper is that of a convex transformation. We say that $f_1$ is a [*(strictly) convex transformation*]{} of $f_2$ if there exists a (strictly) convex function $f$ such that $f_1 (x)= (f \circ f_2) (x) = f (f_2 (x))$.
Let $f_1, f_2 \colon I \to \mathbb{R}$ such that $f_2^{-1}$ exists. Then $f_1$ is a (strictly) convex transformation of $f_2$ if and only if the composition $f_1 \circ f_2^{-1}$ is (strictly) convex.
See [@pratt1964risk].
Recall also that a function $f\colon I \to \mathbb{R}$ is called [*concave*]{} if and only if $-f$ is convex. Thus a function $f\colon I \to \mathbb{R}$ is [*log-concave*]{} if and only if $1/f$ is log-convex. Therefore, the definitions and results stated in this section can be easily adapted for (log-)concavity.
Preferences
-----------
Let $X \subset \mathbb{R_+}$ be the set of outcomes. We will assume that $X$ is an interval of non-negative real numbers containing $0$. The natural interpretation is that outcomes are monetary (for an infinitely divisible currency) but this is not essential. Let $T=[0, \infty)$ be a set of points in time where $0$ corresponds to the present moment. The Cartesian product $X \times T$ will be identified with the set of timed outcomes, i.e., a pair $(x, t) \in X \times T$ is understood as a dated outcome, when a decision-maker receives $x$ at time $t$ and nothing at all other time periods in $T\setminus t$.
Suppose that a decision-maker has a preference order $\succcurlyeq$ on the set of timed outcomes with $\succ $ expressing strict preference and $\sim$ indifference. We say that a utility function $U \colon X\times T \to \mathbb{R}$ [*represents*]{} the preference order $\succcurlyeq$, if for all $x, y \in X$ and all $t, s \in T$ we have $ (x, t) \succcurlyeq (y, s)$ if and only if $U(x, t) \geq U(y, s)$. This is a [*discounted utility (DU) representation*]{} if $$\label{eq:DUR}
U(x, t)=D(t)u(x),$$ where $u \colon X \to \mathbb{R}$ is a continuous and strictly increasing function with $u(0)=0$, and $D\colon T \to (0, 1]$ is continuous and strictly decreasing such that $D(0)=1$ and $\displaystyle {\lim _{t \to \infty} D(t)=0}$.
The function $u$ is called the [*instantaneous utility function*]{}, and $D$ is called the [*discount function*]{} associated with $\succcurlyeq$. We say that the pair $(u, D)$ provides a [*discounted utility representation*]{} for $\succcurlyeq$. Fishburn and Rubinstein [@fishburn1982time] provide an axiomatic foundation for a discounted utility representation. A list of their axioms is given in the Appendix. We assume that $\succcurlyeq$ has a discounted utility representation throughout the paper.
As $D$ is strictly decreasing, our decision maker is always impatient. However, as time goes by, her impatience may increase or decrease.
\[DI\] The preference order $\succcurlyeq$ exhibits (strictly) decreasing impatience (DI) if for all $\sigma>0$, all $0\leq t<s$ and all outcomes $y>x>0$, the equivalence $(x, t)\sim (y, s)$ implies $(x, t+\sigma)\preccurlyeq [\prec] \: (y, s+\sigma)$.
Increasing impatience (II) can be defined by reversing the final preference ranking in Definition \[DI\]. However, we focus on DI preferences in the present paper, since this appears to be the empirically relevant case. As in the previous sentence, we also use the acronym “DI” interchangeably as a noun (“decreasing impatience”) and an adjective (“decreasingly impatient”), relying on context to indicate the intended meaning.
In case the preference order $\succcurlyeq$ has a discounted utility representation, the characterization of DI in terms of the discount function is well-known.[^1]
\[DI-conv\] \[strDI\] Let $\succcurlyeq$ be a preference order having discounted utility representation with the discount function $D$. The following conditions are equivalent:
- The preference order $\succcurlyeq$ exhibits (strictly) DI ;
- $D$ is (strictly) log-convex on $[0,\infty)$.
We say that discount function $D$ is (strictly) DI if the preference order $\succcurlyeq$ exhibits (strictly) DI and has a discounted utility representation with discount function $D$.
We next show that a preference order $\succcurlyeq$ with a discounted utility representation exhibits strictly DI if and only if it exhibits present bias in the sense of Jackson and Yariv [@jackson2014present p. 4190]. It is important to note, however, that Jackson and Yariv assume a discrete time setting, whereas we allow time to be continuous. The following Definition \[pb\] is, therefore, the continuous-time analogue of their present bias definition.[^2]
\[pb\] The preference order is present-biased if
(i) $(x,t) \preccurlyeq (y,s)$ implies $(x,t+\sigma) \preccurlyeq (y,s+\sigma)$ for every $x, y$, every $\sigma>0$ and every $s,t\in T$ such that $s>t\geq0$; and
(ii) for every $s,t\in T$ with $s>t\geq 0$ and every $\sigma>0$ there exist $x^*$ and $y^*$ such that $(x^*,t+\sigma) \prec (y^*,s+\sigma)$ and $(x^*,t) \succ (y^*,s)$.
Proposition \[pbeq\] gives conditions which are equivalent to present bias for preferences with a discounted utility representation:
\[pbeq\] Suppose that $\succcurlyeq$ has a discounted utility representation. Then the first condition of Definition \[pb\] is equivalent to convexity of $\ln D(t)$; while the second condition of Definition \[pb\] is equivalent to strict convexity of $\ln D(t)$.
We start by proving the first equivalence. Since a discounted utility representation exists, the first condition is equivalent to: $$u(x)D(t)\leq u(y)D(s) \ \ \text{ implies }\ \ u(x)D(t+\sigma)\leq u(y)D(s+\sigma)$$ for every $x, y$, every $\sigma>0$ and every $s,t\in T$ with $s>t\geq0$. This may be rewritten as follows: $$u(x)\leq \frac{D(s)}{D(t)}u(y)\ \ \text{ implies }\ \ u(x)\leq \frac{D(s+\sigma)}{D(t+\sigma)}u(y).$$ Since $(y, s) \succcurlyeq (x, t)$, $s>t$ and $D$ is strictly decreasing it follows that $u(y)>u(x)$. As $u(0)=0$ and $u$ is strictly increasing we deduce that $u(y)>0$. Since $u$ is continuous, $x$ and $y$ can be chosen so that $u(x)/u(y)$ takes any value in $[0, 1)$. We therefore have: $$\frac{D(s)}{D(t)}\leq \frac{D(s+\sigma)}{D(t+\sigma)}.$$ Alternatively, $$\label{eq:log}
\ln D(s)+\ln D(t+\sigma) \leq \ln D(s+\sigma)+\ln D(t)$$ for every $\sigma>0$ and $s,t\in T$ with $s>t\geq 0$. Inequality is equivalent to convexity of $\ln D(t)$.
The second part is proved analogously. Under a discounted utility representation the second condition is equivalent to the following: for every $s,t\in T$ with $s>t\geq 0$ and every $\sigma>0$ there exist $x^*$ and $y^*$ such that: $$u(x^*)D(t+\sigma)< u(y^*)D(s+\sigma) \ \ \text{ but }\ \ u(x^*)D(t)> u(y^*)D(s).$$ Equivalently, $$\frac{D(s)}{D(t)}u(y^*)< u(x^*)<\frac{D(s+\sigma)}{D(t+\sigma)}u(y^*).$$ From the fact that $(y^*, s+\sigma)$ is preferred to $(x^*, t+\sigma)$ with $s>t$ we deduce that $u(y^*)>0$. Hence, $$\frac{D(s)}{D(t)}<\frac{D(s+\sigma)}{D(t+\sigma)}.$$ This inequality is equivalent to: $$\label{eq:stlog}
\ln D(s)+\ln D(t+\sigma) < \ln D(s+\sigma)+\ln D(t)$$ for every $s,t\in T$ with $s>t\geq 0$ and every $\sigma>0$. Inequality holds if and only if $\ln D(t)$ is [*strictly*]{} convex.
Proposition \[pbeq\] implies that when a discounted utility representation exists the first condition of Definition \[pb\] follows from the second one, since [*strict*]{} convexity of $\ln D(t)$ implies convexity of $\ln D(t)$. An immediate consequence is that present bias is equivalent to strictly DI, as stated below:
\[pb-di\] Suppose the preference order $\succcurlyeq$ admits a discounted utility representation. Then it exhibits present bias if and only if $\succcurlyeq$ exhibits strictly DI.
Comparative DI
==============
More DI and log-convexity
-------------------------
Assume now that there are two decision makers and they are both decreasingly impatient. What does it mean to say that one of them is more decreasingly impatient than the other? The answer to this question is in the following definition:[^3]
\[moreDI\] We say that $\succcurlyeq_1$ exhibits \[strictly\] more DI than $\succcurlyeq_2$, if for every $\sigma>0$, every $\rho$, every $s,t\in T$ with $0\leq t<s$ and every $x,x',y,y'\in X$ with $y>x>0$ and $y'>x'>0$, the conditions $(x', t)\sim_2 (y', s)$, $(x', t+\sigma)\sim_2 (y', s+\sigma+\rho)$ and $(x, t)\sim_1 (y, s)$ imply $(x, t+\sigma) \preccurlyeq_1 [\prec_1] \: (y, s+\sigma+\rho)$.
Not surprisingly, the (strictly) more DI relation may be expressed in terms of the comparative convexity of the logarithms of the respective discount functions, for cases in which both preference relations have discounted utility representations.
\[key\] Let $\succcurlyeq_1$ and $\succcurlyeq_2$ be two preference orders with discounted utility representation by $(u_1, D_1)$ and $(u_2, D_2)$, respectively. The following conditions are equivalent:
(i) The preference order $\succcurlyeq_1$ exhibits (strictly) more DI than $\succcurlyeq_2$;
(ii) $\ln D_1(D_2^{-1}(e^z))$ is (strictly) convex in $z$ on $(-\infty, 0]$.
See the Appendix. We follow Prelec’s argument for his Proposition 1 in [@prelec2004decreasing]. The additional adjustment is the necessity to replace convexity of the log-transformed discount function with strict convexity for the strictly more DI case. The required adjustments are not substantial but we have included a detailed proof as it clarifies some details omitted from Prelec’s original version [@prelec2004decreasing].
Note that the form of the utility functions $u_1$ and $u_2$ does not influence the comparative DI properties of preference relations.
\[exp\] Let $\succcurlyeq_1$ and $\succcurlyeq_2$ be two preference relations with discounted utility representations $(u_1, D_1)$ and $(u_2, D_2)$, respectively, where $D_2(t)=\delta^t$ and $\delta \in (0, 1)$. The preference order $\succcurlyeq_1$ exhibits (strictly) DI if and only if it exhibits (strictly) more DI than $\succcurlyeq_2$.
Prelec [@prelec2004decreasing] proves that a preference relation is DI if and only if it is more DI than an exponential discount function. We prove the “strict” part of the claim.
Since $D_2(t) = \delta^t$ and $\delta \in (0, 1)$ we have $$D_2^{-1}(e^z)\ =\ \frac{z}{\ln \delta}\ \ge\ 0.$$ By Proposition \[key\], for $\succcurlyeq_1$ to exhibit strictly more DI than $\succcurlyeq_2$ it is necessary and sufficient that $\ln D_1(D_2^{-1}(e^z))$ is strictly convex in $z$ on $(-\infty, 0]$. However, $$\ln D_1(D_2^{-1}(e^z))\ =\ \ln D_1 \left( \frac{z}{\ln\delta}\right)\ =\ \ln D_1(t),$$ where $$t = \frac{z}{\ln \delta}\in [0,\infty)$$ when $z$ takes arbitrary values in $(-\infty, 0]$. Therefore, strict convexity of $\ln D_1(D_2^{-1}(e^z))$ in $z$ on $(-\infty,0]$ is equivalent to strict convexity of $\ln D_1(t)$ in $t$ on $[0, \infty)$. By Proposition \[strDI\], strict convexity of $\ln D_1(t)$ in $t$ on $[0, \infty)$ is equivalent to $\succcurlyeq_1$ exhibiting strictly DI.
The following notations will be used below:
- If $D_1$ and $D_2$ represent equally DI preferences, we write $D_1 \sim_{DI} D_2$;
- If $D_1$ represents more DI preferences than $D_2$, we write $D_1 \succcurlyeq_{DI} D_2$;
- If $D_1$ represents strictly more DI preferences than $D_2$, we write $D_1 \succ_{DI} D_2$.
The following corollary, due to Prelec [@prelec2004decreasing], characterizes the relation between any two discount functions from the same DI class.
\[eqDI\] For any two discount functions $D_1$ and $D_2$, we have $D_1 \sim_{DI} D_2$ if and only if $D_1(t) = D_2(t)^c$, where $c>0$ is a constant not depending on $t$.
The $\succcurlyeq_{DI}$ relation is a partial order. In fact, the “more DI” and “strictly more DI” relations are both transitive This is established in the following proposition.
\[trans\] If $D_1 \succcurlyeq_{DI} D_2$ and $D_2 \succcurlyeq_{DI} D_3$, then $D_1\succcurlyeq_{DI}D_3$. If at least one of the relations $D_1 \succcurlyeq_{DI} D_2$ or $D_2 \succcurlyeq_{DI} D_3$ is strict, then $D_1 \succ_{DI} D_3$.
Suppose $D_1 \succcurlyeq_{DI} D_2$ and $D_2 \succcurlyeq_{DI} D_3$. By Proposition \[key\], we know that both $\ln D_1(D_2^{-1}(e^z))$ and $\ln D_2(D_3^{-1}(e^z))$ are convex in $z$ on $(-\infty, 0]$. Defining $h_i=\ln D_i$ for $i\in\{1,2,3\}$, we can equivalently state that $h_1\circ h_2^{-1}$ and $h_2\circ h_3^{-1}$ are convex on $(-\infty, 0]$. To prove transitivity it is sufficient to show that $\ln D_1(D_3^{-1}(e^z))$ is convex in $z$ on $(-\infty, 0]$, or equivalently that $h_1\circ h_3^{-1}$ is convex on $(-\infty, 0]$.
Let $f_1=h_1\circ h_2^{-1}$ and $f_2=h_2\circ h_3^{-1}$. Then $$\ln D_1\left( D_3^{-1}\left( e^z\right)\right)=h_1\left(h_3^{-1}(z)\right) = h_1 h_2^{-1} \left(h_2 h_3^{-1}(z)\right)=f_1\circ f_2(z) = f(z).$$ By the assumption, $f_1$ and $f_2$ are convex functions. Note that $f_1$ is increasing, as the composition of two decreasing functions $h_1$ and $h_2^{-1}$. Indeed, $h_1 = \ln D_1$ is a strictly decreasing function as $D_1$ is strictly decreasing, and $h_2^{-1}$ is a decreasing function as the inverse of the decreasing function $h_2$. Lemma \[conv\] then implies that $f(z)=f_1 \circ f_2(z) = \ln D_1\left( D_2^{-1}(e^z)\right)$ is convex, and that $f$ is strictly convex if $f_i$ is strictly convex for some $i\in\{1,2\}$.
Time Preference Rate and Index of DI {#IndexDI}
------------------------------------
In this section we assume that $D$ is twice continuously differentiable. The [*rate of time preference*]{}, $r(t)$, is defined as follows: $$r(t)\ =\ -\frac{D'(t)}{D(t)}.$$ The following lemma relates the DI property to the behaviour of $r(t)$.[^4]
\[rate\] Let $\succcurlyeq$ be a preference relation with DU representation $(u,D)$ in which $D$ is twice continuously differentiable. Then the following conditions are equivalent:
(i) The preference relation exhibits (strictly) DI;
(ii) The time preference rate $r(t)$ is (strictly) decreasing on $[0, \infty)$.
Suppose that $r(t)$ is decreasing on $[0, \infty)$. This is equivalent to $$r'(t)\ =\ -\frac{D''(t)D(t)-\left( D'(t)\right)^2}{D(t)^2)}\ =\ \frac{\left(D'(t)\right)^2-D''(t)D(t)}{D(t)^2} \ \leq\ 0.$$ Or, alternatively, $D''D-\left( D'\right)^2 \geq 0$. This inequality is equivalent to log-convexity of $D$, which, by Proposition \[DI-conv\], means that the preference order exhibits DI.
To prove the equivalence of strictly DI preferences and a strictly decreasing rate of time preference, recall that a continuously differentiable function $r \colon \mathbb{R}_+ \to \mathbb{R}$ is strictly decreasing if and only if $r'(t) \leq 0$ for all $t$ and the set $\left\{ \: t \; \vline \; r'(t)=0 \: \right\}$ contains no non-trivial interval [@stein2012twice; @rockafellar1998variational]. If a function $v$ is differentiable on an open interval $I \subset \mathbb{R}$, then $v$ is strictly convex on $I$ if and only if $v'$ is strictly increasing on $I$ [@rockafellar1998variational]. Assume that $r(t)$ is strictly decreasing on $[0, \infty)$. Let $M \subseteq \mathbb{R}_+$ be the set of $t$ values such that $r'(t)<0$. Then $D''(t)D(t)-\left[ D'(t)\right]^2 > 0$ for all $t \in M$. Since $\mathbb{R}_+ \setminus M$ contains no non-trivial interval, $r'(t)$ being strictly decreasing is equivalent to $D$ being strictly log-convex.
One way to measure the level of DI for suitably differentiable discount functions was suggested by Prelec [@prelec2004decreasing]. Since more DI preferences have discount functions which are more log-convex, the natural criterion would be some measure of convexity of the log of the discount function. The Arrow-Pratt coefficient, which is a measure of the concavity of a function, can be adapted to this purpose. Indeed, a non-increasing rate of time preference, $r'(t)\leq 0$, is precisely analogous to the notion of decreasing risk aversion in Pratt [@pratt1964risk].
Recall that $D$ is a twice continuously differentiable function. The associated [*rate of impatience, $IR(D)$,*]{} is defined as follows: $$IR(D)\ =\ -\frac{D''}{D'}.$$ The [*index of DI*]{} of $D$, denoted $I_{DI}(D)$, is the difference between the rate of impatience and the rate of time preference: $$I_{DI}(D)\ =\ IR(D)-r(D)\ =\ \left(-\frac{D''}{D'}\right) - \left(-\frac{D'}{D}\right).$$ Note that $$\label{eq:indexf}
I_{DI}(D)(t)\ =\ -\frac{r'(t)}{r(t)}\ =\ -\frac{d}{dt} \ln \left[r(t)\right].$$ Prelec [@prelec2004decreasing] proved that if $\succcurlyeq_1$ and $\succcurlyeq_2$ both have DU representations with twice continuously differentiable discount functions, $D_1$ and $D_2$ respectively, then $\succcurlyeq_1$ exhibits more DI than $\succcurlyeq_2$ if and only if $I_{DI}(D_1) \geq I_{DI}(D_2)$ on the interval $[0, \infty)$. The following proposition strengthens this result.
\[compare\_index\] Let $\succcurlyeq_1$ and $\succcurlyeq_2$ have DU representations with discount functions $D_1$ and $D_2$, respectively, where $D_1$ and $D_2$ are twice continuously differentiable. Then the preference order $\succcurlyeq_1$ exhibits strictly more DI than $\succcurlyeq_2$ if and only if $I_{DI}(D_1) \geq I_{DI}(D_2)$ on the interval $[0, \infty)$ and $\left\{ \: t \ \lvert \ I_{DI}(D_1)(t) = I_{DI}(D_2)(t) \: \right\}$ contains no non-trivial interval.
Prelec’s [@prelec2004decreasing Proposition 2] proof applies the Arrow-Pratt coefficient [@pratt1964risk], which is used to compare the concavity of functions. There is no straightforward adaptation of Prelec’s argument to the case of [*strict*]{} concavity. We therefore adapt Pratt’s [@pratt1964risk] original argument directly.
Recall that $D_1$ is strictly more DI than $D_2$ if and only if $\ln (D_1)$ is strictly more convex than $\ln (D_2)$ on $[0, \infty)$. Let $h_1=\ln (D_1)$ and $h_2=\ln (D_2)$, so $h_1$ and $h_2$ are strictly decreasing functions. The function $h_1$ is strictly more convex than $h_2$ on $(-\infty,0]$ if and only if there exists a strictly convex transformation $f$ such that $h_1=f(h_2)$, or, equivalently, $h_1\left(h^{-1}_2(z)\right)$ is strictly convex on $(-\infty,0]$.
The first derivative of $h_1\left(h^{-1}_2(z)\right)$ is: $$\label{der}
\frac{d h_1\left( h^{-1}_2(z)\right)}{dz}\ =\ \frac{h_1'\left( h^{-1}_2(z)\right)}{h_2'\left(h^{-1}_2(z)\right)}.$$ We need to show that expression is strictly increasing. Note that $h_2^{-1}(z)$ is strictly decreasing since $h_2$ is strictly decreasing. Therefore, is strictly increasing if and only if $h'_1(x)\diagup h'_2(x)$ is strictly decreasing. The latter is satisfied if and only if $$\label{ten}
\log \left[\frac{h_1'(x)}{h_2'(x)}\right]$$ strictly decreases (since $\log(x)$ is strictly increasing). The first derivative of (\[ten\]) is: $$\frac{h_2'(x)}{h_1'(x)}\cdot\frac{h''_1(x)h'_2(x)-h'_1(x)h''_2(x)}{[h_2'(x)]^2}\ =\ \frac{h''_1(x)}{h'_1(x)}-\frac{h''_2(x)}{h'_2(x)}$$ Therefore (\[ten\]) is strictly decreasing if and only if $$\frac{h''_1(x)}{h'_1(x)}-\frac{h''_2(x)}{h'_2(x)} \ \leq\ 0$$ and the set $$\left\{ \: x \ \left| \ \frac{h''_1(x)}{h'_1(x)}-\frac{h''_2(x)}{h'_2(x)} = 0 \:\right. \right\}$$ contains no non-trivial interval.
Note that: $$\frac{h''_i}{h'_i}\ =\ \frac{D_i''}{D'_i}-\frac{D_i'}{D_i}.$$ Therefore, $$\frac{h''_1(x)}{h'_1(x)}-\frac{h''_2(x)}{h'_2(x)} \ \leq \ 0$$ is equivalent to $$-\frac{D_1''}{D'_1}-\left(-\frac{D_1'}{D_1}\right) \ \geq\ -\frac{D_2''}{D'_2}-\left (-\frac{D_2'}{D_2}\right).$$ This means that $D_1 \succ_{DI} D_2$ if and only if $I_{DI}(D_1) \geq I_{DI}(D_2)$ on $[0, \infty)$, and $\left\{ \: t \ \lvert \ I_{DI}(D_1)(t) = I_{DI}(D_2)(t) \: \right\}$ contains no non-trivial interval.
From Proposition \[compare\_index\], Lemma \[rate\] and it follows that $\succcurlyeq$ is DI if and only if $I_{DI}(D)\geq 0$ on $[0, \infty)$, and $\succcurlyeq$ is strictly DI if and only if $I_{DI}(D)\geq 0$ on $[0, \infty)$ and $\left\{ \: t \ \lvert \ I_{DI}(D)(t) =0 \: \right\}$ contains no non-trivial interval. Note that the index of DI equals zero for an exponential discount function.[^5]
The following example illustrates the index of DI for a generalized hyperbolic discount function. We will make use of this information later.
\[fex\]The function $D(t) = (1+ht)^{-\alpha/h}$, with $h>0$ and $\alpha>0$, is called the [*generalized hyperbolic*]{} discount function. For this function we have: $$r(t)=\alpha(1+ht)^{-1} \ \ \text{and} \ \ IR(D)(t)=(\alpha+h)(1+ht)^{-1}.$$ Therefore, $I_{DI}(D)(t)=h(1+ht)^{-1}$. If $D_1(t) = (1+h_1t)^{-\alpha/h_1}$ and $D_2(t) = (1+h_2t)^{-\alpha/h_2}$ are two generalized hyperbolic discount functions then $D_1 \succcurlyeq_{DI} [\succ_{DI}] \: D_2$, if and only if $h_1\geq [>] \: h_2$.
Thus the parameter $h$ may be used as a measure of the degree of DI of a generalized hyperbolic discount function, while the parameter $\alpha$ has no influence on $I_{DI}(D)$. We call parameter $h$ the [*hyperbolic discount rate*]{}. The special case of a generalized hyperbolic discount function with $\alpha=h>0$ is called the [*proportional*]{} hyperbolic discount function.
Mixtures of Discount Functions
==============================
As described in the introduction, there are some situations in which the necessity arises to calculate a convex combination (mixture) of discount functions.
Two scenarios
-------------
The first situation where a convex combination of discount functions may be used is when there is a group of decision makers with different discount functions and a social discount function needs to be constructed. The natural option is averaging the discount functions among individuals, which is equivalent to averaging the discounted utilities when all agents have identical utility functions. This approach has been widely used in the existing literature on time preferences ([@jackson2014present]).
Indeed, suppose that we have a set of agents $M=\{1,\ldots,m\}$. Assume that agent $i$ has time preferences with DU representation $(u,D_i)$. Thus, all agents have the same instantaneous utility function. Then we define the [*collective (utilitarian) utility function*]{} as $\hat{u}(x)=mu(x)$ and the [*collective total utility*]{} at time $t$ is $$\hat{U}(x,t)=\sum_{i=1}^m D_i(t) u(x) = \left(\frac{1}{m} \sum_{i=1}^m D_i(t) \right) \hat{u}(x).$$ Thus, we obtain the [*collective discount function*]{} $D=\frac{1}{m} \sum_{i=1}^m D_i$.
In the second possible scenario, discussed by Sozou [@sozou1998hyperbolic] and Weitzman [@weitzman1998far], there is a single decision maker with some uncertainty about her discount function. For example, there may be some possibility of not surviving to any given period, $t$, described by a survival function with and uncertain (constant) hazard rate [@sozou1998hyperbolic]. Then the expected discount function of this decision maker can be calculated as a weighted average of the distinct discount functions that may eventuate.
If the discount function $D_i$ eventuates with probability $p_i$, then the expected utility of the decision maker is $$\hat{U}(x,t)=\sum_{i=1}^m p_iD_i(t) u(x) = \left(\sum_{i=1}^m p_iD_i(t) \right) u(x).$$ and the [*certainty equivalent discount function*]{} will be $D=\sum_{i=1}^m p_iD_i$.
The same question arises in both cases: Is it possible to make some conclusion about the behaviour of the convex combination of distinct discount functions in comparison with its components, if all the component discount functions exhibit DI?
Mixtures of discount functions with decreasing impatience {#mix}
---------------------------------------------------------
Given a set of discount functions $\{\row Dn\}$, we define a [*mixture*]{} of them as $$D=\sum_{i=1}^n \lambda_i D_i,$$ where $0<\lambda_i<1$ for all $i$ and $\sum_{i=1}^n \lambda_i=1$. Note that we define a mixture such that each $D_i$ has a [*strictly*]{} positive weight.
We first discuss some known results related to the mixture of discount functions.
One of the most recent results was obtained by Jackson and Yariv [@jackson2014present], who demonstrated that if all decision makers in a group have exponential discount functions, but they do not all have the same discount factor, then their collective discount function must be present biased.
It has also been noted by several authors (for example, [@prelec1991decision] and [@quiggin1995time]), that time preferences have strong similarities with risk preferences and that some results from risk theory are relevant in the context of intertemporal choice. Pratt [@pratt1964risk] showed that decreasing risk aversion is preserved under linear combinations. As was observed in Section \[IndexDI\], decreasing risk aversion is analogous to non-decreasing time preference rate, or DI of the discount function. Therefore, Pratt’s result can be translated into our time preference framework as follows:
\[Pratt\] \[pratheorem\] Let $\succcurlyeq_1,\succcurlyeq_2, \ldots, \succcurlyeq_n$ have DU representations with twice continuously differentiable discount functions $D_1, \ldots, D_n$, respectively. Assume that $\succcurlyeq_1,\succcurlyeq_2, \ldots, \succcurlyeq_n$ all exhibit DI. Let $$D=\sum_{i=1}^n \lambda_iD_i,$$ be a mixture of $D_1, \ldots, D_n$. Then $D$ is DI. It is strictly DI if and only if $$\left\{ \: t \ \lvert \ r_1(t) = r_2(t) =\ldots =r_n(t) \ \text{and } r'_1(t) = r'_2(t) = \ldots = r'_n(t) = 0 \: \right\}$$ contains no non-trivial interval.
From the definition of time preference rate it follows that $D'_i=-r_iD_i$ for all $i=1, \ldots, n$. The time preference rate for $D$ is: $$r\ =\ -\frac{D'}{D}\ =\ -\frac{\sum_{i=1}^n\lambda_i D'_i}{D}\ =\
\sum_{i=1}^n \frac{\lambda_i D_i}{D} r_i.$$ By Lemma \[rate\], to prove that $D$ exhibits DI we must show that $r'(t) \leq 0$: $$\begin{aligned}
r'\ =\ &\sum_{j=1}^n \frac{\lambda_j D'_j\sum_{i=1}^n\lambda_i D_i-\lambda_j D_j \sum_{i=1}^n\lambda_i D'_i}{ D^2}r_j+\sum_{j=1}^n\frac{\lambda_j D_j}{D}r'_j.\end{aligned}$$
Rearranging and substituting $D'_i=-r_iD_i$ we obtain: $$r'\ =\ \sum_{j=1}^n\frac{\lambda_j D_j}{D}r'_j+\frac{Q}{D^2},$$ where $$\begin{aligned}
Q=-\sum_{j=1}^n \left[ \lambda_j r_jD_j\sum_{i=1}^n\lambda_i D_i-\lambda_j D_j \sum_{i=1}^n\lambda_i r_iD_i \right]r_j\end{aligned}$$ This is a quadratic form in $\row Dn$ with the coefficient on $D_iD_j$ being $$\lambda_i\lambda_j\left(r_ir_j-r_j^2\right)+\lambda_i\lambda_j\left(r_ir_j-r_i^2\right)=-\lambda_1\lambda_2(r_i-r_j)^2.$$ Hence $$Q=-\sum_{i< j}\lambda_i \lambda_j D_i D_j(r_i-r_j)^2.$$ Since $D_i \in (0, 1]$, $\lambda_i>0$ and $r'_i \leq 0$ for all $i=1, \ldots n$, we have $r'(t)\leq 0$. Therefore, $\succcurlyeq$ is DI. The preference relation $\succcurlyeq$ is strictly DI if and only if $r'(t)$ is strictly decreasing. We see that $r'(t)$ is strictly decreasing iff $$\{ \: t \ \lvert \ r_1(t) = r_2(t)=\ldots = r_n(t)\ \ \text{and}\ \ r'_1(t) = r'_2(t)=\ldots=r'_n(t) = 0 \: \}$$ contains no non-trivial interval.
The following corollary describes an important special case of Proposition \[Pratt\]:
\[JYC\] Mixtures of non-identical exponential discount functions are strictly DI.
Corollary \[JYC\] is therefore a continuous-time version of Jackson and Yariv (2014, Proposition 1). Prelec [@prelec2004decreasing Corollary 4] considers the mixture of two discount functions only, but does not require differentiability. He proves that the mixture of two equally DI discount functions is more DI than its components. Prelec [@prelec2004decreasing Corollary 4] implies the special case of Jackson and Yariv’s [@jackson2014present] result when $n=2$.
Our objective is to establish such a result which is more general than both Prelec [@prelec2004decreasing] and Jackson and Yariv [@jackson2014present]. The result we obtain is stated in the following theorem:
\[main\] Let $n\geq 2$ and $D_1,\ldots, D_n$ be distinct discount functions such that $$D_1 \succcurlyeq_{DI} D_2 \succcurlyeq_{DI} \ldots \succcurlyeq_{DI} D_n.$$ If $D$ is a mixture of $D_1, \ldots, D_n$, then $D \succ_{DI} D_n$.
To construct the proof of this theorem, two preliminary results will be useful. The first is a strengtheneing of a result in Prelec [@prelec2004decreasing].
\[mixture\] Let $\lambda \in (0, 1)$. If two distinct discount functions $D_1$ and $D_2$ satisfy $D_1 \sim_{DI} D_2$, then their mixture, $D=\lambda D_1+(1-\lambda) D_2$, represents strictly more DI preferences than each $D_i$. That is, $D \succ_{DI} D_1$ and $D \succ_{DI} D_2$.
As $D_1 \sim_{DI} D_2$, then, by Corollary \[eqDI\], $D_1(t) = D_2(t)^c$, where $c \neq 1, c>0$. By Proposition \[key\], it is necessary to demonstrate strict convexity of $f(z)=\ln D\left(D_1^{-1}(e^z)\right)$ for $z\leq 0$. By Proposition \[trans\] it is also sufficient. We note that: $$\begin{aligned}
f(z)&=\ln D\left(D_1^{-1}(e^z)\right)= \ln \left(\lambda D_1\left(D_1^{-1}(e^z)\right)+(1-\lambda) D_2 \left(D_1^{-1}(e^z)\right)\right) \\
&= \ln \left(\lambda e^z+(1-\lambda)e^{z/c}\right).\end{aligned}$$ The first-order derivative of $f(z)$ is: $$f'(z)\ =\ \frac{\lambda e^z+\frac{1}{c}(1-\lambda)e^{z/c}}{\lambda e^z+(1-\lambda)e^{z/c}}.$$ The second-order derivative is: $$f''(z)\ =\ \frac{\left(\lambda e^z+\frac{1}{c^2}(1-\lambda)e^{z/c}\right)\left(\lambda e^z+(1-\lambda)e^{z/c}\right)-\left(\lambda e^z+\frac{1}{c}(1-\lambda)e^{z/c}\right)^2}{\left(\lambda e^z+(1-\lambda)e^{z/c}\right)^2}\ =\ \frac{p(z)}{q(z)}.$$ Both the denominator $q(z)$ and the numerator $p(z)$ of this fraction are strictly positive. The former is obvious. To see the latter, note that the numerator $p(z)$ can be simplified as follows: $$\begin{aligned}
p(z) & = \left(\lambda e^z+\frac{1}{c^2}(1-\lambda)e^{z/c}\right)\left(\lambda e^z+(1-\lambda)e^{z/c}\right)-\left(\lambda e^z+\frac{1}{c}(1-\lambda)e^{z/c}\right)^2 \\
& = \lambda (1-\lambda) e^{\left(1+\frac{1}{c}\right)z}-\frac{2}{c} \lambda (1-\lambda)e^{\left(1+\frac{1}{c}\right)z} + \frac{1}{c^2}\lambda (1-\lambda) e^{\left(1+\frac{1}{c}\right)z} \\
& = e^{\left(1+\frac{1}{c}\right)z} \lambda (1-\lambda) \left (1-\frac{1}{c} \right )^2.\end{aligned}$$ Therefore, $f$ is a strictly convex function.
Proposition \[mixture\] is a stronger version of Corollary 4 in [@prelec2004decreasing], as we show that the mixture of two discount functions that are equally DI represents *strictly* more DI preferences, rather than just more DI preferences. It is important to note that Proposition \[mixture\] cannot be directly generalized to $n$ discount functions by induction. To obtain the path to such generalization, we will need the following lemma.
\[new1\] Let $\lambda \in (0, 1)$. If two distinct discount functions $D_1$ and $D_2$ satisfy $D_1 \succcurlyeq_{DI} D_2$, then their mixture $D=\lambda D_1+(1-\lambda) D_2$ represents strictly more DI preferences than $D_2$. That is, $D \succ_{DI} D_2$.
If $D_1 \sim_{DI} D_2$ then the conclusion follows from Proposition \[mixture\]. Suppose that $D_1$ represents strictly more DI preferences than $D_2$. Then, by Proposition \[key\], $\ln D_1 \left(D_2^{-1}(e^z)\right)$ is strictly convex on $(\infty,0]$. We need to demonstrate that $\ln D \left(D_2^{-1}(e^z)\right)$ is also strictly convex on $(\infty,0]$, or, equivalently, that the following function is strictly log-convex: $$D \left(D_2^{-1}\left(e^z\right)\right)=\lambda D_1D_2^{-1} (e^z)+(1-\lambda)D_2 D_2^{-1} (e^z) = \lambda D_1D_2^{-1} (e^z)+(1-\lambda) e^z.$$ Denote $\lambda D_1D_2^{-1} (e^z) = f$ and $(1-\lambda) e^z = g$. Then we have: $$D \left(D_2^{-1}(e^z)\right)= f+g,$$ where $f$ is strictly log-convex by assumption and $g = (1-\lambda) e^z$ is log-convex. By Lemma \[log-convex\], the sum of a strictly log-convex function and a log-convex function is strictly log-convex, hence $D \left(D_2^{-1}(e^z)\right)$ is strictly log-convex.
We are now ready to provide the proof of Theorem \[main\], since Lemma \[new1\] can be straightforwardly generalized to the case of $n$ distinct discount functions.
We prove this statement by induction on $n$. By Lemma \[new1\] the result holds for $n=2$. Suppose that the statement of the theorem is true for $n = k$. Let $$D^{(k+1)}=\eta_1 D_1+ \ldots + \eta_{k+1} D_{k+1}$$ where $\sum_{i=1}^{k+1} \eta_i=1$ and each $\eta_i \in (0, 1)$. Then $$\begin{aligned}
D^{(k+1)}\ =\ &\eta_1 D_1+ \ldots + \eta_{k+1} D_{k+1}\\
= \ &(1-\eta_{k+1}) \left ( \frac{\eta_1}{1-\eta_{k+1}} D_1 +\ldots + \frac{\eta_k}{1-\eta_{k+1}} D_k \right) + \eta_{k+1} D_{k+1}.\end{aligned}$$ Let $$\displaystyle D^{(k)}=\frac{\eta_1}{1-\eta_{k+1}} D_1 +\ldots + \frac{\eta_k}{1-\eta_{k+1}} D_k.$$ By the induction hypothesis, $D^{(k)} \succ_{DI} D_k$. It is also known that $D_k\succcurlyeq_{DI} D_{k+1}$, and hence, by Proposition \[trans\], $D^{(k)} \succ_{DI} D_{k+1}$. However, the mixture of these two functions is exactly $$D^{(k+1)} = (1-\eta_{k+1}) D^{(k)} + \eta_{k+1} D_{k+1}.$$ Then, by Proposition \[new1\], $D^{(k+1)}\succ_{DI} D_{k+1}$, which completes the induction step.
Mixtures of twice continuously differentiable\
discount functions {#mixarb}
----------------------------------------------
Note that when discount functions are suitably differentiable, Theorem \[main\] and Proposition \[compare\_index\] imply that $$\label{indexineq}
I_{DI}(D) \geq \min_i \{I_{DI}(D_i)\} \ \ \text{on} \ \ [0, \infty)$$ and the set of $t$ values at which equality holds does not include any non-trivial interval.
Consider the following example:
\[orderimp\] Let $D_1(t)=(1+ht)^{-2}$ be a zero-speed hyperbolic discount function [@jamison2011characterizing] and $D_2(t)=\exp(-\alpha t^{1/2})$ be a slow Weibull discount function [@jamison2011characterizing]. As shown in Example \[fex\], $I_{DI}(D_1)(t)=h(1+ht)^{-1}>0$ for all $t$. We also have $I_{DI}(D_2)(t)= (2t)^{-1}>0$ for all $t$ since $$r_2(t)=\frac{\alpha}{2}t^{-1/2}\ \ \text{and}\ \ r'_2(t)=-\frac{\alpha^2}{4}t^{-3/2}.$$ Therefore, both $D_1$ and $D_2$ exhibit strict DI. Assume that $h=0.1$. Then $$I_{DI}(D_1)(t)-I_{DI}(D_2)(t)\ =\ \frac{0.1}{1+0.1t}-0.5\frac{1}{t}\ =\ 0.05 \frac{t-10}{1+0.1t}.$$ Obviously, $I_{DI}(D_1)(t) \leq I_{DI}(D_2)(t)$ if and only if $t \leq 10$ and $I_{DI}(D_1)(t) > I_{DI}(D_2)(t)$ if and only if $t > 10$. It follows that $D_1$ and $D_2$ both are from incomparable DI classes. Since $D_1$ and $D_2$ both exhibit strictly DI, Proposition \[pratheorem\] implies that their mixture $D$ also exhibits strictly DI.
![Index of DI for the Mixture of $D_1$ and $D_2$[]{data-label="graph"}](Fig1)
By direct calculation we obtain the following expression: $$\begin{aligned}
I_{DI}(D)(t)\ =\ &\frac{6\lambda_1 h^2(1+ht)^{-4}+1/4\lambda_2 \alpha \exp{-\alpha t^{0.5}}t^{-1}(\alpha+t^{-0.5})}{2\lambda_1 h(1+ht)^{-3}+1/2\lambda_2 \alpha \exp{-\alpha t^{0.5}}t^{-0.5}}\\
\ &\ \\
\ &-\ \frac{2\lambda_1 h(1+ht)^{-3}+1/2\lambda_2 \alpha \exp{-\alpha t^{0.5}}t^{-0.5}}{\lambda_1 (1+ht)^{-2}+\lambda_2 \exp{-\alpha t^{0.5}}}.\end{aligned}$$ The behaviour of $I_{DI}(D)$ with parameters $\lambda_1=\lambda_2=0.5$, $h=0.1$ and $\alpha=0.12$ is illustrated in Figure \[graph\]. It can be clearly seen from Figure \[graph\] that neither $D \succ_{DI} D_1$ nor $D \succ_{DI} D_2$. However, $I_{DI}(D) \geq \min \{I_{DI}(D_1), I_{DI}(D_2)\}$ on $[0, \infty)$.
Example \[orderimp\] suggests that the inequality may continue to hold even if the discount functions are not DI-comparable. Theorem \[main3\] verifies this conjecture.
\[main3\] Let $\succcurlyeq_1, \succcurlyeq_2, \ldots, \succcurlyeq_n$ have DU representations with twice continuously differentiable discount functions $D_1, D_2, \ldots, D_n$, respectively. Let $D=\sum_{i=1}^n \lambda_i D_i$ be a mixture of $D_1, D_2, \ldots, D_n$. Then $I_{DI}(D) \geq \ds \min_{i}\{I_{DI}(D_i)\}$ on $[0, \infty)$, and $$I_{DI}(D)(t) > \ds \min_{i}\{I_{DI}(D_i)(\hat{t})$$ if $r_j(\hat{t}) \neq r_k(\hat{t})$ for some $j\neq k$.
Let $I_i=I_{DI}(D_i)$ for all $i\in\{ 1, \ldots, n\}$ and let $I=I_{DI}(D)$. Recall that $D'=-rD$ and hence $D''=Dr^2-Dr'=Dr(r+I)$. Recall also that $$I\ =\ -\frac{D''}{D'}+\frac{D'}{D}.$$ Therefore, $$I\ =\ \frac{-\sum_{i=1}^n \lambda_i D''_i}{\sum_{i=1}^n \lambda_i D'_i}+\frac{\sum_{i=1}^n \lambda_i D'_i}{\sum_{i=1}^n \lambda_i D_i}
\ =\ \frac{\sum_{i=1}^n \lambda_i D_i r_i(r_i+I_i)}{\sum_{i=1}^n \lambda_i D_i r_i}-\frac{\sum_{i=1}^n \lambda_i D_i r_i}{\sum_{i=1}^n \lambda_i D_i}.$$ This expression can be rearranged as follows: $$I\ =\ \frac{\sum_{i=1}^n \lambda_i D_i r_iI_i}{\sum_{i=1}^n \lambda_i D_i r_i}+\frac{\sum_{i=1}^n \lambda_i D_i r_i^2}{\sum_{i=1}^n \lambda_i D_i r_i}-\frac{\sum_{i=1}^n \lambda_i D_i r_i}{\sum_{i=1}^n \lambda_i D_i}\ =\ \sum_{i=1}^n \alpha_i(t) I_i+Q,$$ where $$Q\ =\ \frac{\sum_{i=1}^n \lambda_i D_i r_i^2}{\sum_{i=1}^n \lambda_i D_i r_i}-\frac{\sum_{i=1}^n \lambda_i D_i r_i}{\sum_{i=1}^n \lambda_i D_i}
\ \ \text{and}\ \
\alpha_i\ =\ \frac{\lambda_i D_i r_i}{\sum_{i=1}^n \lambda_i D_i r_i}$$ with $\sum_{i=1}^n \alpha_i=1$ and $\alpha_i\geq 0$. Note that $$\min_i\{I_i\} \ \leq\ \sum_{i=1}^n \alpha_i I_i \ \leq \ \max_i\{I_i\} \ \ \text{for all } t\in [0,\infty).$$
The expression $Q$ can be rewritten as: $$Q\ =\ \frac{\Big[\sum_{i=1}^n \lambda_i D_i r_i^2\Big]\cdot\Big[\sum_{i=1}^n \lambda_i D_i\Big] - \Big[\sum_{i=1}^n \lambda_i D_i r_i\Big]^2}{\Big[\sum_{i=1}^n \lambda_i D_i r_i\Big]\cdot\Big[\sum_{i=1}^n \lambda_i D_i\Big]}.$$ The denominator of $Q$ is strictly positive, so the sign of $Q$ depends on the sign of the numerator. Let $N$ be the numerator of $Q$: $$N=\Big[\sum_{i=1}^n \lambda_i D_i r_i^2\Big]\cdot\Big[\sum_{i=1}^n \lambda_i D_i\Big] - \Big[\sum_{i=1}^n \lambda_i D_i r_i\Big]^2.$$ We can simplify $N$ as follows: $$N=\sum_{i=1}^n \sum_{j=1}^n \lambda_i \lambda_j D_i D_j r_i^2 - \sum_{i=1}^n \sum_{j=1}^n \lambda_i \lambda_j D_i D_j r_i r_j.$$ Therefore, we have: $$N\ =\ \sum_{i=1}^n\sum_{j=1}^n\theta_{ij}r_{i}\left(r_{i}-r_{j}\right)$$ where $\theta_{ij}=\lambda_{i}\lambda_{j}D_{i}D_{j}$. Since $\theta
_{ij}=\theta_{ji}>0$ for all $i$ and $j$ we see that$$N\ =\ \sum_{i<j}\theta_{ij}\left[ r_{i}\left( r_{i}-r_{j}\right)
+r_{j}\left( r_{j}-r_{i}\right) \right] \ =\ \sum_{i<j}\theta_{ij}\left(
r_{i}-r_{j}\right) ^{2}\text{.}$$
Hence, $N \geq 0$ and $N > 0$ if $r_j \neq r_k$ for some $j \neq k$. It follows that $Q \geq 0$ and $Q > 0$ if $r_j \neq r_k$ for some $j \neq k$. Therefore, since $I=\sum_{i=1}^n \alpha_i I_i+Q$ and $$\min_i\{I_i\} \ \leq\ \sum_{i=1}^n \alpha_i I_i \ \leq\ \max_i\{I_i\} \ \text{for all } t\in [0,\infty),$$ we have: $$\min_i\{I_i\}\ \leq\ \min_i\{I_i\}+Q \ \leq\ \sum_{i=1}^n \alpha_i I_i+Q=I.$$ In other words, $I\geq\ds\min_i\{I_i\}$ on $[0, \infty)$, and $I(\hat{t})>\ds\min_i\{I_i(\hat{t})\}$ if $r_j(\hat{t}) \neq r_k(\hat{t})$ for some $j \neq k$.
Observe that this result does not require discount functions to exhibit decreasing impatience. Therefore, Theorem \[main3\] makes less restrictive assumptions than Proposition \[Pratt\] – it allows the discount functions to exhibit increasing impatience.
Mixtures of proportional hyperbolic discount functions {#prob}
------------------------------------------------------
Weitzman [@weitzman1998far] shows that if different discount functions may eventuate with certain probabilities, then future costs and benefits must eventually be discounted at the lowest possible limiting time preference rate. This result is particularly salient when the possible discount functions are all exponential, with constant time preference rates. The purpose of this section is to give an analogous result for proportional hyperbolic discount functions, with constant hyperbolic discount rates (Example \[fex\]). The result in this case is very different to Weitzman’s. Long-term future benefits and costs are discounted, not at the lowest hyperbolic discount rate, but at the probability-weighted harmonic mean of the individual hyperbolic discount rates.
Suppose that there is some uncertainty about the rate of time preference, and we have a set of possible scenarios $N=\{1,\ldots,n\}$ where each time preference rate $r_i(t)$ may eventuate with probability $p_i\geq 0$, such that $\sum_{t=1}^n p_i=1$. Since for each $i$ $$r_i(t)=-\frac{D_i'(t)}{D_i(t)},$$ the corresponding discount function can be expressed in terms of the rate of time preference as follows $$\label{discthroughrate}
D_i(t)=\exp{\left(-\int_0^t{r_i(\tau)d\tau}\right)} \ \text{for each}\ i \in N.$$ The certainty equivalent discount function will be: $$D=\displaystyle\sum_{i=1}^n p_i D_i, \ \text{where} \ p_i\geq 0 \ \text{and} \ \sum_{t=1}^n p_i=1.$$ Then the certainty equivalent time preference rate is $r=-\frac{D'}{D}$. Weitzman [@weitzman1998far] proved that if each rate of time preference converges to a non-negative value as time goes to infinity, then the certainty equivalent rate of time preference converges to the lowest of these values. In other words, if $\lim_{t\to \infty} r_i(t)=r_i^*$ with $r_i^*\geq 0$ and $r_1^*<r_{i}^*$, where $i\neq 1$, then $\lim_{t\to \infty} r(t) = r^*_1$.
Note that $r_i(t)$ in is constant if and only if $D_i$ is exponential. In this case we have: $$D_i(t)=\exp{\left(-r_it\right)} \ \text{for each}\ i \in N,$$ where $r_i=const$. Therefore, Weitzman’s result implies that $\lim_{t\to \infty} r(t) = \min_i{r_i}$. Figure \[graphMixExp\] illustrates for the case $n=3$, $r_1=0.01$, $r_2=0.02$, $r_3=0.03$ and $p_1=p_2=p_3=1/3$. We also observe that the certainty equivalent rate of time preference $r(t)$ decreases monotonically towards $r_1$. This is a consequence of Corollaries \[exp\] and \[JYC\] and the fact that $I_{DI}(D)=-r'/r$.
![Mixture of Exponential Discount Functions[]{data-label="graphMixExp"}](Fig2)
However, Weitzman’s result [@weitzman1998far] does not provide much insight in the special case when each possible time preference has a DU representation with a proportional hyperbolic discount function. Suppose $$D_i(t)\ \ = \frac{1}{1+h_i t}$$ for each $i\in N$, where $h_i>0$ is the hyperbolic discount rate. Without loss of generality we assume that $h_1 > h_2 > \ldots > h_n$. Suppose that $D_i$ eventuates with probability $p_i$ where $p_i\geq 0$ and $\sum_{i=1}^n p_i=1$. Then the certainty equivalent discount function would be $$D(t)\ =\ \frac{p_1}{1+h_1 t}+\ldots+\frac{p_n}{1+h_n t}.$$ The rate of time preference is $$r_i(t)\ =\ \frac{h_i}{1+h_it}$$ for all $i$. It is obvious that $r^*_i=r^*_j=0$ for all $i\neq j$ and $ \lim_{t \to \infty} r(t)=0$, which, indeed, corresponds to Weitzman’s result. However, this conclusion does not give much information about the asymptotic behavior of the certainty equivalent discount function. Given that each possible discount function comes from a different DI class (unlike in the case of heterogeneous exponential discount functions) we would like to know which (if any) most closely characterizes the asymptotic behaviour of the certain equivalent function.
To answer this question we need to modify the analysis of Weitzman. Note that the certainty equivalent discount function can be written as $$D(t)\ =\ \frac{1}{1+h(t)t},$$ where $h(t)$ is the [*certainty equivalent hyperbolic discount rate*]{}. In particular, $$h(t)=(\frac{1}{D(t)}-1)\frac{1}{t},$$ so $h(t)$ is well-defined for $t\in (0, \infty)$. We ask: [*How does $h(t)$ behave as $t\to \infty$?*]{}
We remind the reader that the [*weighted harmonic mean*]{} of non-negative values $x_1, x_2, \ldots, x_n$ with non-negative weights $\row an$ satisfying $a_1+\ldots+a_n=1$ is $$H(x_1, a_1;\ldots;x_n, a_n)=\left( \sum_{i=1}^n \frac{a_i}{x_i}\right)^{-1}.$$ It is well-known that the weighted harmonic mean is smaller than the corresponding expected value (weighted arithmetic mean).
\[main2\] Suppose that each $D_i$ ($i\in N$) is a proportional hyperbolic discount function, with associated hyperbolic discount rate $h_i$. Discount function $D_i$ will eventuate with probability $p_i$. Then the long-term certainty equivalent hyperbolic discount rate is the probability-weighted harmonic mean of the individual hyperbolic discount rates, $H(h_1, p_1;\ldots;h_n, p_n)$.
We note that $$\frac{p_i}{1+h_it} \ =\ \frac{p_i}{h_it}+\epsilon_i(t),$$ where $\epsilon_i(t)/t^2 \to 0$ when $t\to\infty$. Let $\epsilon(t)=\epsilon_1(t)+\ldots+\epsilon_n(t)$. Hence it follows that: $$\begin{aligned}
\frac{1}{1+h(t)t} \ =\ \sum_{i=1}^n p_iD_i(t) & \ =\ \frac{p_1}{1+h_1 t}+\ldots+\frac{p_n}{1+h_n t}\\ & \ =\ \frac{p_1}{h_1 t}+\ldots+\frac{p_n}{h_n t} +\epsilon(t)
\\ &\ =\ \biggl( \frac{p_1}{h_1}+\ldots+\frac{p_n}{h_n}\biggr ) \frac{1}{t} +\epsilon(t)\\ &\ =\ \frac{1}{H(h_1, p_1;\ldots;h_n, p_n) t}+\epsilon(t)\\
&\ =\ \frac{1}{1+H(h_1, p_1;\ldots;h_n, p_n) t}+\hat{\epsilon}(t),
\end{aligned}$$ where $\hat{\epsilon}(t)/t^2 \to 0$ as $t\to\infty$. This implies that $h(t) \to H(h_1, p_1;\ldots;h_n, p_n)$ as $t\to\infty$.
![Mixture of Hyperbolic Discount Functions[]{data-label="gMH"}](Fig3)
Figure \[gMH\] illustrates Theorem \[main2\] for the case $n=3$, when hyperbolic rates $h_1=0.01$, $h_2=0.02$ and $h_3=0.03$ eventuate with equal probabilities. Note that $h_2=0.02$ corresponds to the arithmetic mean of $h_1$, $h_2$ and $h_3$. Figure \[gMH\] displays the convergence of the certainty equivalent hyperbolic discount rate to the weighted harmonic mean $H(h_1, p_1;h_2, p_;h_3, p_3)$. It also shows the certainty equivalent hyperbolic discount rate decreasing monotonically. The following proposition proves that this is always the case.
\[hyp2\] Suppose that each $D_i$ ($i\in N$) is a proportional hyperbolic discount function, with associated hyperbolic discount rate $h_i$. Discount function $D_i$ will eventuate with probability $p_i$. Then the certainty equivalent hyperbolic discount rate is strictly decreasing on $(0, \infty)$.
We prove this statement by induction on $n$. First we need to prove that the statement holds for $n=2$. The respective certainty equivalent hyperbolic discount rate is: $$\begin{aligned}
h(t) =\left[ \frac{1}{p_1(1+h_1t)^{-1}+p_2(1+h_2t)^{-1}} -1 \right] \frac{1}{t}\end{aligned}$$ for each $t>0$. Rearranging: $$\begin{aligned}
h(t)=\left[ \frac{(1+h_1t)(1+h_2t)}{p_1(1+h_2t)+p_2(1+h_1t)}-1 \right]\frac{1}{t}\ =\
\left[ \frac{1+(h_1+h_2)t+h_1h_2t^2} {p_1 +p_2+(p_1h_2+p_2h_1)t}-1\right]\frac{1}{t}.\end{aligned}$$ Since $p_1+p_2=1$ we obtain: $$\begin{aligned}
h(t)&=\left[ \frac{1+\left(h_1+h_2\right)t+h_1h_2t^2} {1+\left( p_1h_2+p_2h_1\right)t}-1\right]\frac{1}{t}\ =\
\frac{\left( h_1+h_2-p_1h_2-p_2h_1\right)t+h_1h_2t^2} {1+\left( p_1h_2+p_2h_1\right)t} \cdot\frac{1}{t} \\
&\ =\ \frac{h_1+h_2-p_1h_2-p_2h_1+h_1h_2t} {1+\left( p_1h_2+p_2h_1\right)t}\ =\ \frac{p_1h_1+p_2h_2+h_1h_2t} {1+\left( p_1h_2+p_2h_1\right)t}.\end{aligned}$$ By differentiating $h(t)$: $$\label{deriv}
h'(t)=\frac{h_1h_2\left( 1+\left( p_1h_2+p_2h_1 \right)t \right)-\left( p_1h_1+p_2h_2+h_1h_2t\right) \left( p_1h_2+p_2h_1\right)}{ \left[ 1+(p_1h_2+p_2h_1)t \right]^2}$$ We need to show that $h'(t) < 0$. Since the denominator of is positive, the sign of $h'(t)$ depends on the sign of the numerator. Therefore, we denote the numerator of by $Q$ and analyse it separately: $$\begin{aligned}
Q(t)&\ =\ h_1h_2\left[1+\left(p_1h_2+p_2h_1\right)t\right]-\left(p_1h_1+p_2h_2+h_1h_2t\right)\left(p_1h_2+p_2h_1\right)\\
&\ =\ h_1h_2+h_1h_2(p_1h_2+p_2h_1)t-(p_1h_1+p_2h_2)(p_1h_2+p_2h_1)-h_1h_2(p_1h_2+p_2h_1)t\\
&\ =\ h_1h_2-\left(p_1h_1+p_2h_2\right)\left(p_1h_2+p_2h_1\right).\end{aligned}$$ By expanding the brackets and using the fact that $p_1+p_2=1$ implies $1-p_1^2-p_2^2=2p_1p_2$ expression $Q$ can be simplified further: $$\begin{aligned}
Q(t)&\ =\ h_1h_2-p_1^2h_1h_2-p_1p_2h_1^2-p_1p_2h_2^2-p_2^2h_1h_2 \\&\ =\ h_1h_2(1-p_1^2-p_2^2)-p_1p_2(h_1^2+h_2^2)\\
&\ =\ 2p_1p_2h_1h_2-p_1p_2(h_1^2+h_2^2)\\&\ =\ -p_1p_2(h_1-h_2)^2.\end{aligned}$$ Therefore, since $h_1\neq h_2$ we have $Q<0$. Hence it follows that $h'(t)<0$ and $h(t)$ is strictly decreasing.
Suppose that the proposition holds for $n=k$. We need to show that it also holds for $n=k+1$. When $n=k+1$ the certainty equivalent hyperbolic discount rate is: $$h_{k+1}(t) = \left[ \frac{1}{D^{(k+1)}}-1 \right] \frac{1}{t},$$ where $$D^{(k+1)}\ =\ \sum_{i=1}^{k+1}p_i D_i\ =\ \left( 1-p_{k+1}\right)\left(\sum_{i=1}^{k}\frac{p_i}{1-p_{k+1}} D_i\right)+p_{k+1}D_{k+1}.$$ Since $$\sum_{i=1}^{k}\frac{p_i}{1-p_{k+1}}=1,$$ we have $$D^{(k+1)}\ =\ \left( 1-p_{k+1}\right)D^{(k)}+p_{k+1}D_{k+1}.$$ where $$D^{(k)}\ =\ \sum_{i=1}^{k}\frac{p_i}{1-p_{k+1}} D_i.$$ By the induction hypothesis it follows that $$D^{(k)}=\frac{1}{1+h_k(t)t},$$ where $h_k$ is strictly decreasing. Therefore, $$\begin{aligned}
h^{(k+1)}(t)&=\left[\frac{1}{(1-p_{k+1})D^{(k)}+p_{k+1}D_{k+1}}-1\right]\frac{1}{t}\\
&=\left[\frac{1}{\left( 1-p_{k+1}\right)\left( 1+h_k(t)t\right)^{-1}+p_{k+1}\left(1+h_{k+1}t\right)^{-1}}-1\right]\frac{1}{t}.\end{aligned}$$
Let $\hat{p_1}=1-p_{k+1}$, $\hat{p_2}=p_{k+1}$, $\hat{h_1}(t)=h_k(t)$ and $\hat{h_2}=h_{k+1}=const$. Then we have $$h^{(k+1)}(t)=\left[ \frac{1}{\hat{p_1}(1+\hat{h_1}(t)t)^{-1}+\hat{p_2}(1+\hat{h_2}t)^{-1}}-1\right]\frac{1}{t}.$$ Analogously to the case $n=2$, this expression can be rearranged to give: $$h^{(k+1)}(t)=\frac{\hat{p_1}\hat{h_1}+\hat{p_2}\hat{h_2}+\hat{h_1}\hat{h_2}t}{1+\hat{p_1}\hat{h_2}t+\hat{p_2}\hat{h_1}t}.$$ However, by contrast to the case $n=2$, $\hat{h_1}$ is now a function of $t$. The derivative of $h^{(k+1)}$ is: $$\begin{aligned}
&\frac{dh^{(k+1)}(t)}{dt}=\\
&\frac{\left( \hat{p_1}\hat{h}'_1+\hat{h_1}\hat{h_2}+\hat{h}'_1\hat{h_2}t\right) \left( 1+\hat{p_1}\hat{h_2}t+\hat{p_2}\hat{h_1}t\right)-
\left(\hat{p_1}\hat{h_1}+\hat{p_2}\hat{h_2}+\hat{h_1}\hat{h_2}t\right)\left(\hat{p_1}\hat{h_2}+\hat{p_2}\hat{h_1}+\hat{p_2}\hat{h}'_1 t\right)
}{\left[1+\hat{p_1}\hat{h_2}t+\hat{p_2}\hat{h_1}t\right]^2}.\end{aligned}$$ The denominator of this fraction is strictly positive, so the sign of the derivative depends on the numerator only. Denote the numerator by $N$: $$\begin{aligned}
N=&\left( \hat{p_1}\hat{h}'_1+\hat{h_1}\hat{h_2}+\hat{h}'_1\hat{h_2}t\right)\left( 1+\hat{p_1}\hat{h_2}t+\hat{p_2}\hat{h_1}t \right)\\
&-\left( \hat{p_1}\hat{h_1}+\hat{p_2}\hat{h_2}+\hat{h_1}\hat{h_2}t\right)\left( \hat{p_1}\hat{h_2}+\hat{p_2}\hat{h_1}+\hat{p_2}\hat{h}'_1 t\right).\end{aligned}$$ Note that$$N\ =\ \hat{Q}\left( t\right) +\hat{h}_{1}^{\prime}\left[ \left( \hat
{p}_{1}+\hat{h}_{2}t\right)\left(1+\hat{p}_{1}\hat{h}_{2}t+\hat{p}_{2}\hat{h}_{1}t\right)-\hat{p}_{2}t\left( \hat{p}_{1}\hat{h}_{1}+\hat{p}_{2}\hat{h}_{2}+\hat{h}_{1}\hat{h}_{2}t\right) \right],$$ where $\hat{Q}\left( t\right) $ is defined as in the proof of Proposition 1, but with $h_{1}=\hat{h}_{1}\left( t\right) $ and $h_{2}=\hat{h}_{2}$. Since Proposition 1 establishes that $\hat{Q}\left( t\right) \leq0$ (with equality if and only if $\hat{h}\left( t\right) =h_{2}$) and $\hat{h}_{1}^{\prime}<0$, it suffices to show that $$\label{eqn:ineq}
\left( \hat{p}_{1}+\hat{h}_{2}t\right) \left( 1+\hat{p}_{1}\hat{h}_{2}t+\hat{p}_{2}\hat{h}_{1}t\right) -\hat{p}_{2}t\left( \hat{p}_{1}\hat
{h}_{1}+\hat{p}_{2}\hat{h}_{2}+\hat{h}_{1}\hat{h}_{2}t\right) >0$$
Cancelling terms on the left-hand side of leaves us with:$$\hat{p}_{1}\left( 1+\hat{p}_{1}\hat{h}_{2}t\right) +\hat{h}_{2}t\left(
1+\hat{p}_{1}\hat{h}_{2}t\right) -\left( \hat{p}_{2}\right) ^{2}\hat{h}
_{2}t.$$ We now use the fact that $\left( \hat{p}_{2}\right) ^{2}=\left( 1-\hat
{p}_{1}\right) ^{2}=1-2\hat{p}_{1}+\left( \hat{p}_{1}\right) ^{2}$ to get $$\begin{aligned}
& \hat{p}_{1}\left( 1+\hat{p}_{1}\hat{h}_{2}t\right) +\hat{h}_{2}t\left(
1+\hat{p}_{1}\hat{h}_{2}t\right) -\left[ 1-2\hat{p}_{1}+\left( \hat{p}
_{1}\right) ^{2}\right] \hat{h}_{2}t\\
& =\ \hat{p}_{1}+\left( t\hat{h}_{2}\right) ^{2}\hat{p}_{1}+2\hat{p}_{1}
\hat{h}_{2}t,\end{aligned}$$ which is strictly positive as required. Therefore, $h^{(k+1)}(t)$ is strictly decreasing.
Discussion
==========
We generalized Jackson and Yariv’s result [@jackson2014present] by proving that whenever we aggregate different discount functions from comparable DI classes, the weighted average function is always *strictly more DI* than the least DI of its constituents. This also strengthens the conclusion of the theorem of Prelec [@prelec2004decreasing] who demonstrates that the mixture of two different discount functions from the same DI class represents more DI preferences.
When a decision maker is uncertain about her hyperbolic discount rate, we showed that long-term costs and benefits must be discounted at the probability-weighted harmonic mean of the hyperbolic discount rates that might eventuate. This complements the well-known result of Weitzman [@weitzman1998far].
One natural question that arises is whether it is possible to prove a result analogous to Proposition \[Pratt\] when all preference orders exhibit increasing impatience (II). Will the mixture of II discount functions be (strictly) II? Perhaps surprisingly, the answer to this question is negative in general.
This follows from results in the literature on survival analysis and reliability theory. The similarity between reliability theory and temporal discounting is discussed in [@sozou1998hyperbolic]. Takeuchi [@takeuchi2011non] also notes that a discount function is analogous to a survival function, $S(t)$. The failure rate associated with $S(t)$ is $$g(t)=-\frac{S'(t)}{S(t)},$$ which behaves as a time preference rate. For twice continuously differentiable survival functions, a decreasing failure rate (DFR) corresponds to a decreasing time preference rate, and hence to DI, whereas an increasing failure rate (IFR) corresponds to II. Mixtures of probability distributions are a common topic in survival and reliability analysis. Proschan [@proschan1963theoretical] established that mixtures of distributions with DFR always exhibit DFR.[^6] However, Gurland and Sethuraman [@gurland1994shorter; @gurland1995pooling] provide striking examples of mixtures of very quickly increasing failure rates that are eventually decreasing.
Appendix
========
Fishburn and Rubinstein’s axioms for a discounted utility representation
------------------------------------------------------------------------
After Fishburn and Rubinstein [@fishburn1982time], we assume that:
Axiom 1. (Weak Order)
: The preference order $\succcurlyeq$ is a weak order, i.e., it is complete and transitive.
Axiom 2. (Monotonicity)
: For every $x, y \in X$, if $x<y$, then $(x, t) \prec (y, t)$ for every $t \in T$.
Axiom 3. (Continuity)
: For every $(y, s) \in X \times T$ the sets $\{ (x, t)\in X \times T \colon (x, t) \succcurlyeq (y, s)\}$ and $\{ (x, t)\in X \times T \colon (x, t) \preccurlyeq (y, s)\}$ are closed.
Axiom 4. (Impatience)
: For all $t, s\in T$ and every $x>0$, if $t<s$, then $(x, t) \succ (x, s)$. If $t<s$ and $x=0$, then $(x, t) \sim (x, s)$ for every $t, s \in T$, that is, $0$ is a time-neutral outcome.
Axiom 5. (Separability)
: For every $x, y, z \in X$ and every $r, s, t \in T$ if $(x, t) \sim (y, s)$ and $(y, r) \sim (z, t)$ then $(x, r) \sim (z, s)$.
Fishburn and Rubinstein [@fishburn1982time] proved the following result:
The preferences $\succcurlyeq$ on $X \times T$ satisfy Axioms 1-5 if and only if there exists a discounted utility representation for $\succcurlyeq$ on $X \times T$. If $(u, D)$ and $(u_0, D_0)$ both provide discounted utility representations for $\succcurlyeq$ on $X \times T$, then $u = \alpha u_0$ for some $\alpha > 0$, and $D = \beta D_0$ for some $\beta > 0$.
Proof of Proposition \[key\]
----------------------------
We need to prove the following lemma first:
\[hh\] Suppose that $h_1$ and $h_2$ are strictly decreasing functions. Then $h_1$ is a (strictly) convex transformation of $h_2$ if and only if $h_2(s)-h_2(t)=h_2(s+\sigma+\rho)-h_2(t+\sigma)$ implies that $h_1(s)-h_1(t)\leq [<] h_1(s+\sigma+\rho)-h_1(t+\sigma)$ for every $s$. $t$, $\sigma$ and $\rho$ satisfying $0<t<s\leq t+\sigma<s+\sigma+\rho$.
We prove necessity first. Suppose that $h_1$ is a (strictly) convex transformation of $h_2$; that is, there exists a (strictly) convex function $f$ such that $h_1=f(h_2)$. Assume also that $0<t<s\leq t+\sigma<s+\sigma+\rho$ and $$\label{eq:h2}
h_2(s)-h_2(t)=h_2(s+\sigma+\rho)-h_2(t+\sigma).$$ We need to show that $$h_1(s)-h_1(t)\leq [<]\: h_1(s+\sigma+\rho)-h_1(t+\sigma)$$ whenever $0<t<s\leq t+\sigma<s+\sigma+\rho$. Since $h_2$ is strictly decreasing, it follows that $$h_2(s+\sigma+\rho)<h_2(t+\sigma)\leq h_2(s)<h_2(t).$$ Recall that $f$ is a (strictly) convex function. Therefore, as equality holds, it implies that $$f(h_2(t+\sigma))-f(h_2(s+\sigma+\rho))\leq [<] \: f(h_2(t))-f(h_2(s)).$$ Since $h_1=f(h_2)$, this inequality is equivalent to $$h_1(t+\sigma)-h_1(s+\sigma+\rho) \leq [<] \: h_1(t)-h_1(s).$$ Rewriting: $$\label{eq:h1}
h_1(s)-h_1(t)\leq [<] \: h_1(s+\sigma+\rho)-h_1(t+\sigma),$$ whenever $0<t<s\leq t+\sigma<s+\sigma+\rho$.
To show the sufficiency, suppose that implies for every $s$, $t$, $\sigma$ and $\rho$ satisfying $0<t<s\leq t+\sigma<s+\sigma+\rho$. Define $f$ such that $f=h_1\circ h_2^{-1}$. Note that we can do so because $h_2^{-1}$ exists (since $h_2$ is a strictly decreasing function). Then if $$h_2(s+\sigma+\rho)<h_2(t+\sigma)\leq h_2(s)<h_2(t)$$ and equation holds, we have $$f(h_2(t+\sigma))-f(h_2(s+\sigma+\rho)) \leq [<] \: f(h_2(t))-f(h_2(s)).$$ Therefore, $f$ is a (strictly) convex function, which means that $h_1$ is a (strictly) convex transformation of $h_2$.
We can now prove Proposition \[key\].
Observe that $D_i\colon [0,\infty) \to (0,1]$ is one-to-one and onto, so $D_i^{-1}\colon (0,1] \to [0, \infty)$.
Let us first prove that condition (i) follows from condition (ii). The proof is by contraposition. We show that not (i) implies not (ii). Assume that (i) fails; that is, there exist $s$ and $t$ with $0<t<s$, $\rho>0$, $\sigma>0$ and $x,y,x',y'\in X$ with $0<x<y$ and $0<x'<y'$ such that $(x', t) \sim_2 (y', s)$, $(x', t+\sigma) \sim_2 (y', s+\sigma+\rho)$, $(x, t) \sim_1 (y, s)$ and $$(x, t+\sigma) \succ_1 [\succcurlyeq_1] \: (y, s+\sigma+\rho).$$ Since $u_1(y)>0$ and $u_2(y')>0$ by assumption, this implies $$\frac{u_2(x')}{u_2(y')}=\frac{D_2(s)}{D_2(t)}=\frac{D_2(s+\sigma+\rho)}{D_2(t+\sigma)}$$ and $$\frac{u_1(x)}{u_1(y)}=\frac{D_1(s)}{D_1(t)}> [\geq] \:\frac{D_1(s+\sigma+\rho)}{D_1(t+\sigma)}.$$
Let $h_1=\ln D_1$ and $h_2=\ln D_2$. Note that $h_1$ and $h_2$ are both strictly decreasing functions. Observe also that $h_i \colon [0,\infty) \to (-\infty, 0]$ is one-to-one and onto. Thus $h_i^{-1} \colon (-\infty,0] \to [0,\infty)$, where $h_i^{-1}(z)=D_i^{-1}(e^z)$. Rewriting these expressions we get $D_i(t)=e^{h_i(t)}$ for each $i\in\{1,2\}$. Thus: $$\frac{e^{h_2(s)}}{e^{h_2(t)}}=\frac{e^{h_2(s+\sigma+\rho)}}{e^{h_2(t+\sigma)}}$$ and $$\frac{e^{h_1(s)}}{e^{h_1(t)}}> [\geq]\:\frac{e^{h_1(s+\sigma+\rho)}}{e^{h_1(t+\sigma)}}.$$ Equivalently, $$\label{eq:contr1}
h_2(s)-h_2(t)=h_2(s+\rho+\sigma)-h_2(t+\sigma)$$ and $$\label{eq:contr2}
h_1(s)-h_1(t)> [\geq]\:h_1(s+\rho+\sigma)-h_1(t+\sigma).$$
Note that $\ln D_1(D_2^{-1}(e^z))$ (strictly) convex in $z$ on $(-\infty, 0]$ is equivalent to $h_1 \circ h_2^{-1}$ (strictly) convex in $z$ on $(-\infty, 0]$. In other words, $h_1$ is a (strictly) convex transformation of $h_2$. By Lemma \[hh\] this conclusion contradicts equation and inequality . Therefore, not (i) implies not (ii).
Secondly, we need to demonstrate that (i) implies (ii). Using the previously introduced notation, we show that for every for every $s$, $t$, $\sigma$ and $\rho$ satisfying $$0<t<s\leq t+\sigma<s+\sigma+\rho$$ the equation $$h_2(s)-h_2(t)=h_2(s+\sigma+\rho)-h_2(t+\sigma)$$ implies $$h_1(s)-h_1(t)\leq [<]\:h_1(s+\sigma+\rho)-h_1(t+\sigma).$$ As $h_1$ and $h_2$ are decreasing functions, this proves that $h_1$ is a (strictly) convex transformation of $h_2$. Assume that $0\leq t<s\leq t+\sigma<s+\sigma+\rho$ such that $$h_2(s)-h_2(t)=h_2(s+\sigma+\rho)-h_2(t+\sigma).$$ By definition of $h_i=\ln D_i$ this expression is equivalent to $$\frac{D_2(s)}{D_2(t)}=\frac{D_2(s+\sigma+\rho)}{D_2(t+\sigma)} \ \in (0, 1).$$ As $u_2$ is continuous, we can choose $0<x'<y'$ such that: $$\frac{D_2(s)}{D_2(t)}\ =\ \frac{D_2(s+\sigma+\rho)}{D_2(t+\sigma)}\ =\ \frac{u_2(x')}{u_2(y')}.$$ Therefore, $D_2(t)u_2(x')=D_2(s)u_2(y')$ and $D_2(t+\sigma)u_2(x')=D_2(s+\sigma+\rho)u_2(y')$. This means that $(x', t) \sim_2 (y', s)$ and $(x', t+\sigma) \sim_2 (y', s+\sigma+\rho)$.
Analogously, because $u_1$ is continuous, we can choose $x, y$ such that: $$\frac{D_1(s)}{D_1(t)}=\frac{u_1(x)}{u_1(y)} \ \in (0, 1).$$ Hence, $(x, t) \sim_1 (y, s)$.
But according to (i), if $(x', t) \sim_2 (y', s)$, $(x', t+\sigma) \sim_2 (y', s+\sigma+\rho)$ and $(x, t) \sim_1 (y, s)$ then $(x, t+\sigma) \preccurlyeq_1 [\prec_1]\: (y, s+\sigma+\rho)$. The latter is equivalent to: $$\frac{D_1(s+\sigma+\rho)}{D_1(t+\sigma)} \geq [>] \:\frac{u_1(x)}{u_1(y)}.$$ It follows that $$\frac{D_1(s)}{D_1(t)}\leq [<]\: \frac{D_1(s+\sigma+\rho)}{D_1(t+\sigma)},$$ which is equivalent to $$\ln{D_1(s)}-\ln{D_1(t)}\leq [<] \ln{D_1(s+\sigma+\rho)}-\ln{D_1(t+\sigma)}$$ or $$h_1(s)-h_1(t)\leq [<]\:h_1(s+\sigma+\rho)-h_1(t+\sigma).$$ Therefore, $$h_2(s)-h_2(t)=h_2(s+\sigma+\rho)-h_2(t+\sigma)$$ implies $$h_1(s)-h_1(t)\leq [<]\:h_1(s+\sigma+\rho)-h_1(t+\sigma)$$ whenever $0\leq t<s\leq t+\sigma< s+\sigma+\rho$. Hence, by Lemma \[hh\], $h_1$ is a (strictly) convex transformation of $h_2$.
[^1]: The proof in [@harvey1995proportional Theorem 3.3] can be easily adapted to demonstrate an analogous result for increasing impatience: the preference order $\succcurlyeq$ exhibits (strictly) II if and only if $D$ is (strictly) log-concave on $[0,\infty).$
[^2]: There is an inconsistency between Jackson and Yariv’s Present Bias definition in their 2014 paper (referenced here) and their 2015 paper [@jackson2015collective]. We adhere to the former definition.
[^3]: Since the sign of $\rho$ is not restricted in Definition \[moreDI\], it actually applies to preferences that exhibit decreasing or increasing impatience.
[^4]: Takeuchi [@takeuchi2011non] contains a related result. His Corollary 1 says that the [*hazard function*]{} is decreasing (increasing) if and only if preferences exhibit decreasing (increasing) impatience. Takeuchi’s hazard function $h(t)$ corresponds to our time preference rate $r(t)$. However, Takeuchi does not analyse the case of strictly decreasing impatience.
[^5]: Similarly, $\succcurlyeq$ is II if and only if $I_{DI}(D)\leq 0$ on $[0, \infty)$ and strictly II if and only if $I_{DI}(D)\leq 0$ on $[0, \infty)$ and $\left\{ \: t \ \lvert \ I_{DI}(D)(t) =0 \: \right\}$ contains no non-trivial interval.
[^6]: This result is comparable to the “non-strict” part of our Proposition \[Pratt\].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We use microlocal arguments to suggest that Lorentz symmetry breaking must occur in a reasonably behaved tachyonic quantum field theory that permits renormalizability. In view of this, we present a scalar tachyonic quantum field model with manifestly broken Lorentz symmetry and without exponentially growing/decaying modes. A notion of causality, in which anti-telephones are excluded, and which is viewed as a form of chronology protection, is obeyed. The field theory is constructed in a preferred tachyon frame in terms of commuting creation/annihilation operators. We calculate some sample (renormalized) operators in this preferred frame, argue that the Hadamard condition is satisfied, and discuss the PCT and spin-statistics theorems for this model.'
author:
- |
Marek J. Radzikowski$^*$ ()\
\
\
bibliography:
- 'big.bib'
title: 'Stable, Renormalizable, Scalar Tachyonic Quantum Field Theory with Chronology Protection'
---
ł[[**l**]{}]{} [**u**]{} ß
\#1[\#1]{} \#1[0= 0=0 1= 1=1 0>1 \#1 ]{} \#1[0=\#1-0to0[*/*]{}]{} \#1[\#1]{}
2em
Introduction
============
Ever since the notion of [*tachyons*]{}, i.e., particles which always travel faster than light, was conceived as a possibility within the basic framework of special relativity [@BilDesSud62], a quantum field theory describing such particles has been sought. Despite numerous attempts to formulate such a theory [@Tan60; @Fei67; @AroSud68; @DhaSud68; @Sch71], it appears that a model which is consistent with the desiderata of conventional quantum field theory has been wanting [@KamKam71; @KamKam78; @HugSte90]. However, due to the interest in the possibility that the neutrino may turn out to be a tachyon [@CHK85], which has been supported, at least naïvely, by tritium beta decay experiments during the 90’s [@PDG96], one perhaps should be motivated to clarify the issue either by presenting a viable model, or by showing that no viable tachyonic quantum field model, according to some reasonable definition of “viable”, can exist.
It is apparent from the early literature cited above, that various authors have attempted to hold tightly to the assumption of [*Poincaré invariance*]{}, no matter what the consequences for the theory. The usual way of requiring such invariance in the context of a scalar quantum field theory is by demanding that the [*two-point distribution*]{} $\left<0\left|\phi(x)\phi(y)\right|0\right>$ be left unchanged in value when the spacetime points $x,y$ are simultaneously replaced by $\Lambda x+a,\Lambda y+a$ respectively. (Here, $\left|0\right>$ is some notion of “vacuum” or “ground” state, $\phi(x)$ is an operator-valued distribution called the [*field operator*]{}, $\Lambda$ is a proper, orthochronous Lorentz transformation, and $a$ is a constant spacetime vector.) However, such a restriction, under further suitable physical assumptions, as will be explained in Section \[SymmBreak\], must necessarily lead to a non-renormalizable theory, i.e., one which could not be incorporated into any renormalizable interacting (or self-interacting) theory, and for which the renormalized stress-energy tensor of the free field would make no sense. The argument used in Section \[SymmBreak\] comes from the arena of [*microlocal quantum field theory*]{} [@Rad92; @Rad96a], which has proven to be reasonably successful, for example, in clarifying renormalization in [*quantum field theory on curved spacetime*]{} [@BruFre00; @HolWal02].
From the standpoint of quantum field theory on curved spacetime, it is clear that, applying the usual methods to construct a scalar model of tachyons on flat [*Minkowski*]{} spacetime, one obtains a quantum field theory satisfying the usual axioms of [*Wightman positivity*]{} (namely the condition on the two-point distribution corresponding to the positive definiteness of the Hilbert space inner product), local commutativity, and the [*Hadamard condition*]{}, with, however, no guarantee of [*Poincaré invariance*]{}. (Note that this does not nullify the existence of Poincaré [*symmetry*]{}, which is always a symmetry of the underlying Minkowski spacetime, but that one allows for the possibility that the construction of the various quantities needed in the quantum field theory may manifest [*spontaneous Poincaré symmetry breaking*]{}.) That one is permitted to effect this construction, follows from the quite general existence (and uniqueness) theorems for the typical quantities needed for a quantum field theory, such as the advanced and retarded fundamental solutions $\Delta_A,\Delta_R$ to the inhomogeneous wave equation on a globally hyperbolic spacetime [@Lic61; @Ler63]. Indeed, there is also an existence theorem (and, up to smooth function, a uniqueness theorem) for a Feynman propagator satisfying the Hadamard condition on such a spacetime [@DuiHoer72; @Rad92; @Rad96a]. Furthermore, the smooth part of the Feynman propagator may be chosen so that the corresponding two-point distribution satisfies Wightman positivity [@DuiHoer72]. Note that it would be at the point of introducing this smooth function that the Poincaré symmetry may need to be broken.
Carrying out this construction for the [*tachyonic*]{} Klein-Gordon equation (i.e., with the opposite sign in the mass squared term) leads, however, to a theory which yields exponentially growing renormalized expectation values of typical observables, since in order to ensure that the strict local commutativity condition is satisfied, the mode solutions of the wave equation which are exponentially growing and decaying in time must be included in the mode expansion of the field operator. This situation is unpalatable from a physical point of view, since we would hope that a valid quantum field theory describing tachyonic neutrinos would at least preserve [*stability*]{}. To support this, one may simply observe that neutrinos from SN1987A, which were detected after travelling a distance of 150,000 light years at very close to the speed of light, have evidently manifested this property. A further aspect of the theory which may at first glance appear slightly disconcerting is that the two-point distribution does indeed manifest explicit breaking of Lorentz invariance [*and*]{} time translation invariance (but space translation invariance is maintained).
There is at our disposal, however, the simple procedure of deleting the undesirable exponentially growing/decaying modes from the theory. Since these modes contribute only a smooth function to the two-point distribution, the resulting theory still satisfies the Hadamard condition, and so retains the possibility of being incorporated into a renormalizable interacting or self-interacting quantum field theory according to the criteria of Weinberg’s theorem [@BruFre00]. The theory also maintains Wightman positivity since it is explicitly obtained from a mode sum. Two new features (from the point of view of conventional QFT) remain or emerge (from the construction described in the previous paragraph): the theory still breaks Lorentz invariance (but now preserves both time and space translation invariance), and the usual notion of local commutativity (commuting fields at spacelike separation) is not satisfied.
However, neither of these two features appears insurmountable. Indeed, the notion of a preferred [*tachyon frame*]{}, “seen” only by the tachyons in the theory, would constitute one of the predictions of the theory. (Note particularly that the preferred tachyon frame is not one in which the speed of light takes on a special value $c$, different from what would be measured in other frames; rather $c$ is the same in every inertial frame, as in the usual formulation of special relativity.) Furthermore, the local commutativity axiom is violated so weakly that physical signals, made up of tachyons described by this QFT, can be sent backward in time, but only to points which are spacelike separated from the sender. This is true even for devices which use relays to attempt to send messages backward in time to the sender (called [ *anti-telephones*]{}). Such devices cannot be constructed, according to this QFT, since the (tachyonic) particles required for such a device cannot be simultaneously created from any of the vacuum states allowed in the theory. We consider this property to be a manifestation of [*chronology protection*]{}, which has otherwise appeared in the quite different context of QFT on curved spacetime [@Haw92; @KayRadWal97].
Working in the preferred tachyon frame, we begin the construction of the scalar model from scratch, presenting some of the Green’s functions, in Section \[ScaModel\]. Continuing in the preferred frame, we then utilize a Lagrangian approach to determine the (renormalized) Hamiltonian and momentum operators for a Hermitian scalar tachyon, as well as the charge operator for a charged scalar tachyon, in Section \[Ops\]. Further elaboration on the Hadamard and chronology protection properties, as well as some remarks on the $PCT$ and spin-statistics theorems for this particular model are given in a final Discussion section (Section \[Disc\]).
Note that a different approach (in appearance) is adopted by [@CibRem96], who also obtain a stable, causal QFT based on a preferred frame and a non-standard synchronization scheme, without, however, a discussion of the renormalizability of their theory. Also, they obtain some first results for the beta decay spectrum near the endpoint (for tachyonic neutrinos), as well as suggest an alternative mechanism to neutrino oscillations involving 3-body (tachyonic) decay channels. We conjecture that their QFT approach and ours can be mapped to each other by a suitable (general linear) coordinate transformation. It is hoped that, if these two approaches are indeed found to be compatible, that the present formulation in terms of the more familiar Minkowski coordinates would bring further clarity to the approach of [@CibRem96].
Necessity of Lorentz symmetry breaking {#SymmBreak}
======================================
Here we draw upon tools from the microlocal approach to quantum field theory. This approach grew out of the necessity of dealing with the singularities inherent in quantum field theory on curved spacetime [@BirDav82; @Ful89; @Wal94] in a general and coherent manner. Specifically, one applies techniques and theorems from microlocal analysis [@Hoer71; @DuiHoer72] to the Green’s functions (or, rather, distributions) of the quantum field theory. Such distributions include the advanced and retarded fundamental solutions $\Delta_A,\Delta_R$ to the inhomogeneous wave equation, the Feynman propagator $\Delta_F$ (and its complex conjugate), the two-point distribution $\Delta^{(+)}(x_1,x_2)=
\left<0\left|\phi(x_1)\phi(x_2)\right|0\right>$, and, more generally, the $n$-point or Wightman distributions, which are the vacuum expectation values of the $n$-fold products of the field operator $\phi(x)$ at the $n$ points $x_1,\ldots,x_n$. We remark that this incorporation of microlocal techniques into quantum field theory on curved spacetime has led to, among other results, a characterization of the [*global Hadamard condition*]{} [@KayWal91] in terms of a restriction on the [*wave front set*]{} of the two-point distribution [@Rad96a], a resolution of Kay’s singularity conjecture [@Rad92; @Rad96b], and the development of an Epstein-Glaser-like approach to renormalization on curved spacetime [@BruFre00; @HolWal02]. Note that, as input to the renormalization programme, the two-point distribution must be kept globally Hadamard. Since the characterization of this condition in terms of wave front sets is important in the present context, we start with an introduction to the main concepts involved in this characterization in the following paragraph.
Recall that the [*singular support*]{} of a distribution $F(x)$ consists of all the points $x$ at which $F$ is not smooth. (In order for $F$ to be smooth at $x$, an open neighbourhood of this point must exist, on which $F$ and all its derivatives exist and are finite.) The [*wave front set*]{} of $F$ is an extension of the singular support of $F$ consisting of the pairs $(x,k)$ where $x$ is in the singular support, and $k\ne 0$ is a direction in the cotangent space of the spacetime at $x$. The wave front set is [*conic*]{} in the sense that if $k$ is the second component of a point in the wave front set, then so is $\alpha k$, where $\alpha>0$ is a real number. The directions $k$ indicate, heuristically speaking, the directions in the “local Fourier space” at $x$ in which the “local Fourier transform” of $F$ near $x$, along with all its derivatives, does [*not*]{} decay more rapidly than any polynomial in the Euclidean distance as this distance tends to infinity. The wave front set is a general enough construct that it can be defined on a curved spacetime as readily as on a flat spacetime, and, on flat spacetime, it is defined for distributions more general than the tempered distributions (whose Fourier transforms exist). Furthermore, some general results for performing operations with such distributions on manifolds (such as multiplication, convolution, and restriction to a submanifold) have simple statements in terms of their wave front sets. For more precise definitions, please see [@Hoer90].[^1]
We now describe the wave front set of a two-point distribution $F^{(+)}(x_1,x_2)$ (on a globally hyperbolic curved spacetime $(M,g)$) satisfying the Hadamard condition [@Rad92; @Rad96a].[^2] The quadruple $(x_1,k_1,x_2,k_2)$ is in the wave front set precisely when $x_1$ and $x_2$ either coincide or are on the same null geodesic. If $x_1=x_2$, the covectors $k_1,k_2$ are both null and $k_1=-k_2$. If $x_1\ne x_2$ (but $x_1$ and $x_2$ are still on the same null geodesic) then $k_1,k_2$ are null and tangent to this null geodesic at the points $x_1,x_2$ respectively. Furthermore, $k_1$ is minus the parallel transport of $k_2$ from $x_2$ to $x_1$ along this null geodesic. An additional restriction is that $k_1$ is always pointing in the [*future*]{} time direction (for both $x_1=x_2$ and $x_1\ne x_2$). Thus, $k_2$ is always pointing in the [*past*]{} time direction. There are no other restrictions on the covectors; hence all covectors satisfying the above criteria are included in the wave front set. If, in addition, the spacetime is Minkowski, and translation invariance holds, the two-point distribution is a distribution of the difference variable $x=x_1-x_2$, and the wave front set of the distribution in one variable $f(x)$ consists of pairs of points $(x,k)$ where $x$ and $k\ne 0$ are null, $k$ (as a vector) is parallel (or anti-parallel) to $x$, if $x\ne 0$, and $k$ is future pointing.
We sketch the steps of the proof that for a tempered, Poincaré invariant two-point distribution satisfying the tachyonic Klein-Gordon equation $(\dal -m^2)\phi=0$, the Hadamard condition cannot be satisfied. (Temperedness is assumed here in order to guarantee good behaviour of renormalized observables, e.g., to avoid exponential growth of observables in time.) Translation invariance implies that we can write the two-point distribution as a distribution of the difference variable $x=x_1-x_2$, i.e., as $u(x)$. Now $\hat u (k)$ is also a tempered distribution, and the wave equation implies that $\hat u (k)$ is nonzero only for $k^2=-m^2$, which is a one-sheeted hyperboloid. Furthermore, Lorentz invariance of the distribution $u(x)$ implies Lorentz invariance of $\hat u(k)$. The above considerations imply that $\hat u (k)$ may be written as $c(k)\delta(k^2+m^2)$, where $c(k)$ is a smooth function on the one-sheeted hyperboloid. Since a spacelike $k$ may be mapped by a proper, orthochronous Lorentz transformation to $-k$, we obtain $c(k)=c(-k)$ (in fact $c$ is a constant), and therefore $\hat u (-k)=\hat u (k)$. This implies $u(-x)=u(x)$, which, as is readily verified, leads to the result that $(x,k)\in\WF(u)$ if and only if $(-x,-k)\in\WF(u)$. Thus the wave front set covectors are not restricted to be future pointing; past pointing ones are also required. (Note that the wave front set of the two-point distribution must contain [*some*]{} pairs, if this theory is to even vaguely resemble a typical nontrivial quantum field theory!) This finishes the sketch that the Hadamard condition cannot be satisfied for a Poincaré invariant, tempered two-point distribution of a tachyonic quantum field theory.
We note that a similar “no-go” result, in which the requirement of temperedness is dropped, appears to have the following counterexample: First note that the symmetric part of the two point distribution (without exponentially growing/decaying modes) is Poincaré invariant, while the same is true of the commutator distribution (determined through the Leray-Lichnerowicz uniqueness theorem, which [*requires*]{} the presence of the growing and decaying modes). Thus, we may combine these terms to form a “hybrid” two-point distribution which satisfies the Hadamard condition. However, Wightman positivity evidently fails for this two-point distribution (at least it apparently cannot be constructed from a mode sum). Hence it seems reasonable to conjecture that a Poincaré invariant two-point distribution, satisfying Wightman positivity and the tachyonic Klein-Gordon equation, [*cannot*]{} satisfy the Hadamard condition. The author is aware only of arguments in favour of this conjecture being true [@Fre98]. In any case, the constructed examples lend support to this conjecture’s validity.
A further observation is that, with the insertion of the extra assumption of Wightman positivity for the two-point distribution in our no-go “lemma” for the tempered case, we can assert more strongly that the two-point distribution becomes [*real*]{} $u(x)^*=u(x)$, in addition to being even $u(-x)=u(x)$. This is so because Wightman positivity implies that $\hat u(k)$ is real and positive-valued (in the appropriate sense of distribution theory), besides being even $\hat u (-k) = \hat u (k)$. Hence the antisymmetric (and imaginary) part of $u(x)$ is zero. This appears to go far astray from describing a QFT with which we are presently familiar. In any case, the Hadamard condition still fails (this is now readily seen in the fact that the usual leading order Hadamard singularity must have a non-zero imaginary part). As before, the breakdown of this important input to renormalization forces us to consider such a model as [*unphysical*]{} for our purposes.
Finally, we note that non-Hadamard two-point distributions render the existence of the renormalized stress-energy tensor problematic, since the conditions on the wave front set under which one can construct the two-point distribution of the Wick-ordered polynomials (of which the stress-energy tensor is one) are not satisfied. See [@BruFreKoeh96; @Wal77; @Wal78] for more discussion on the stress-energy tensor and the need for the Hadamard condition to be satisfied in order that the stress-energy tensor be defined.
Construction of the scalar model {#ScaModel}
================================
Having argued that one should expect Poincaré symmetry to be broken in a reasonable quantum field model satisfying the tachyonic Klein-Gordon equation, we now present such a model. We shall, in this and the next sections, perform the constructions in just the preferred frame. Note that values of the scalar Green’s functions in a boosted frame are readily obtained from these by the obvious change of coordinates. E.g., if a Green’s function in the preferred frame is $G(x',y')$, then the same function in the frame obtained by the boost $\Lambda\colon x'\to x$ is the pullback by the inverse Lorentz transformation, namely $G(\Lambda^{-1}x,\Lambda^{-1}y)$.
The first part of the construction in any frame is a listing of all the [*oscillatory*]{} mode solutions of the tachyonic Klein-Gordon equation (using $\dal=\partial_t^2 - \nabla^2$) $$\label{eq:TKG}
(\dal - m^2)u=0\;.$$ In our choice of preferred frame, one has a natural choice of [*inner product*]{} for such modes, namely $$\label{eq:InnPro}
(u,v) = i\int_{t=a} u^*(x) \overleftrightarrow{\partial_t} v(x)\,d^3\x$$ where $$u^*(x)\overleftrightarrow{\partial_t} v(x) = u^*(x)(\partial_t v(x))
- (\partial_t{u^*(x)}) v(x)\;.$$ This inner product turns out to be independent of the choice of the constant $a$ (as will be evident for the orthonormal basis we shall construct). However, the inner product [*does*]{} in general depend on the choice of space-like hypersurface over which the integral is defined. This is in constrast to the case of massive or light-like particles.
We choose (in this reference frame) a basis for the space of [*positive energy*]{} or [*positive frequency*]{} (oscillatory) solutions as follows. We consider the solutions $u_\k(t,\x) = M_\k\E^{-i(\wk t-\k\cdot\x)}$, where $\wk = \sqrt{\k^2 - m^2}$ and $|\k|>m$. In order to obtain the orthonormal relations $$\begin{aligned}
\label{eq:RESoln}
(u_\k,u_\l) &=& \delta^{(3)}(\k-\l) \\
(u^*_\k,u^*_\l) &=& -\delta^{(3)}(\k-\l) \\
\label{eq:RESoln2}
(u_\k,u^*_\l) &=& (u^*_\k,u_\l) = 0\;,
\end{aligned}$$ we choose the normalization factors to be $M_\k = ((2\pi)^3\cdot
2\wk)^{-\half}$, whence $$\label{eq:PosFreMod}
u_\k(t,\x) = {\frac{1}{\sqrt{(2\pi)^3\cdot 2\wk}}}\E^{-i(\wk
t-\k\cdot\x)}\;.$$
For mode solutions with $|\k|=m$, which are also oscillatory, the frequency is $0$, and thus time derivatives of the modes give zero. Thus the modes here are of the form $$\label{eq:ZeroFre}
v_\k(t,\x) = \E^{i\k\cdot\x}$$ with $$\label{eq:NormZerFreq}
(v_\k,v_\l) = 0\;.$$ Hence any normalization constant will do here, and we just leave the $v_\k$ as they are. Note that inner products with the $u_\k, |\k|>m$ are zero. It will turn out that incorporating these modes into the theory does not ultimately affect the two-point distribution, since the integrations involving only these “zero modes” are over a set of measure zero. Hence we may safely ignore them in the further development of the theory.
Note that the [*negative energy*]{} or [*negative frequency*]{} solutions are simply taken to be the complex conjugates of the positive energy modes. Thus they are $$\label{eq:NegFreMod}
u^*_\k(t,\x) = {\frac{1}{\sqrt{(2\pi)^3\cdot 2\wk}}}\E^{i(\wk
t-\k\cdot\x)}\;,$$ where the parameter space labelled by $\k$ is restricted to $|\k|>m$ as above.
As stated in the Introduction, the exponentially growing and decaying modes, which would be labelled by $|\k|<m$, are omitted from the model in order to avoid exponential blow-up of renormalized observables, and thus they shall be ignored henceforth in this paper. A model in which these are incorporated (as partly described in the Introduction) was given in [@Sch71], and also was considered in the context of quantum field theory on curved spacetime by [@Rad98b; @Fre98].
In order to quantize Eq.(\[eq:TKG\]) without the exponentially growing/decaying modes, we seek a field operator $\phi(x)$ of the following form $$\phi(x) = \int_{|\k|>m} (a_\k u_\k(x) + a^\dag_\k u^*_\k(x))\,d^3\k \;,$$ which, besides satisfying Eq.(\[eq:TKG\]), also satisfies the [*equal time commutation relations*]{}, modified so as not to include frequencies $\k$ for which $|\k|<m$: $$\begin{aligned}
\label{eq:ETCR}
\left[\phi(t,\x),\phi(t,\y)\right] &=& 0 \\
\left[\phi(t,\x),\partial_t\phi(t,\y)\right] &=& i\delta_m(\x-\y)\;.\end{aligned}$$ Here, $\delta_m$ is a modification of the Dirac delta distribution obtained by deleting modes $\E^{i\k\cdot\x}$ with frequencies $|\k|<m$ from the latter, namely, $$\delta_m(\x)=\frac{1}{(2\pi)^3}\int_{|\k|>m}\E^{i\k\cdot\x}\,d^3\k\;.$$
We find the following relations: $$\label{eq:Rel1}
(u_\k,\phi)=a_\k, \ (u^*_\k,\phi) = -a^\dag_\k\;,$$ and $$\label{eq:Rel2}
\left[a_\k,\phi(x)\right] = u^*_\k(x)\;,$$ where we have employed $$\int u^*_\k(t,\y)\delta_m(\x-\y)\,d^3\y = u^*_\k(t,\x)$$ for $|\k|>m$. These relations lead to the commutators $$\label{eq:Comm}
\left[a_\k,a_\l\right] = \left[a^\dag_\k,a^\dag_\l\right] = 0, \
\left[a_\k,a^\dag_\l\right]=\delta^{(3)}(\k-\l)\;.$$
The [*vacuum*]{} or [*ground state*]{} $\gs$ associated with this particular choice of preferred tachyon frame is then defined by $a_\k\gs = 0$. The [*two-point distribution*]{} is therefore $$\begin{aligned}
\Delta^{(+)}(x,y) &=& \gsb\phi(x)\phi(y)\gs \\
&=& \int_{|\k|>m} u_\k(x) u^*_\k(y)\,d^3\k \\
&=& \frac{1}{2}\Delta^{(1)}(x,y) + i\frac{1}{2}\Delta(x,y)\;.\end{aligned}$$ Taking the anti-symmetric part of the two-point distribution, namely, $$\gsb[\phi(x),\phi(y)]\gs =i\Delta(x,y) = 2i{\rm Im}\gsb\phi(x)\phi(y)\gs\;,$$ we obtain the commutator distribution $$\Delta(x,y) = -{\frac{1}{(2\pi)^3}}\int_{|\k|>m}d^3\k\,\E^{i\k\cdot(\x-\y)}
{\frac{\sin[\sqrt{\k^2 -m^2}(t-s)]}{\sqrt{\k^2 - m^2}}}\;.$$ This is seen to be the unique distributional bisolution of the Cauchy problem $$\begin{aligned}
\label{eq:CauPro}
\Delta(t,\x,t,\y) &=& 0 \\
\partial_t\Delta(t,\x,s,\y)|_{s=t} &=& -\delta_m(\x-\y)\;.\end{aligned}$$ The symmetric part of the two-point distribution, namely $$\Delta^{(1)}(x,y)= \gsb\{\phi(x),\phi(y)\}\gs =
2{\rm Re}\gsb\phi(x)\phi(y)\gs\;,$$ is $$\Delta^{(1)}(x,y) = {\frac{1}{(2\pi)^3}}\int_{|\k|>m}d^3\k\,
\E^{i\k\cdot(\x-\y)}
{\frac{\cos[\sqrt{\k^2 -m^2}(t-s)]}{\sqrt{\k^2 - m^2}}} \;.$$ Note that this is a Lorentz invariant distribution.
Renormalized operators {#Ops}
======================
Taking the (classical) Lagrangian density of our field theory to be $${\cal L} = \frac{1}{2}\left(\partial^\mu\phi\partial_\mu\phi + m^2\phi^2\right)\;,$$ where the field is real-valued, we obtain, via the usual method, the Hamiltonian $$H=\frac{1}{2}\int \left(\pi^2 + (\nabla\phi)^2 - m^2 \phi^2\right)\,d^3\x\;,$$ where $\pi(x) = \partial_t\phi(x)$. Note that, because of the minus sign in the third term above, the Hamiltonian is generally not positive. However, we shall shortly show that the quantized version is indeed a positive operator in the preferred frame. Similarly, the momentum of the field is $${\bf P} =-\int \partial_t\phi\nabla\phi\,d^3\x\;.$$
The quantization of the above expressions proceeds along the familiar lines: classical fields are replaced by the quantized versions, and products of field operators are normal ordered. We express the final result in terms of creation/annihilation operators, recalling that normal ordering places creation operators before annihilation operators in the expressions. Thus, we obtain $$\begin{aligned}
\pi &=& \partial_t\phi = -i\int_{|\k|>m} \wk (a_\k u_\k - a^\dag_\k u^*_\k)\,d^3\k \\
\nabla\phi &=& i\int_{|\k|>m} \k(a_\k u_\k - a^\dag_\k u^*_\k )\,d^3\k\;,\end{aligned}$$ and, using the orthonormality property of the exponential functions $\E^{i\k\cdot\x}$, $$\begin{aligned}
H &=& \int_{|\k|>m} \wk a^\dag_\k a_\k\,d^3\k \\
{\bf P} & = & \int_{|\k|>m} \k a^\dag_\k a_\k\,d^3\k\;.\end{aligned}$$ Note that these expressions are as expected, in analogy with the usual creation/annihilation operator formalism in conventional QFT. We anticipate that these operators will transform in the usual manner when we calculate them in a boosted frame. Hence the Hamiltonian, although it will remain Hermitian, will not remain positive in any frame boosted with respect to the preferred one.
For the case of a charged scalar tachyon, the Lagrangian density becomes $${\cal L} = \partial^\mu\phi^*\partial_\mu\phi + m^2\phi^*\phi\;,$$ where the field is now complex-valued. The conserved quantity derived from global phase invariance (i.e., invariance under $\phi \to \E^{i\alpha}\phi$) is $$Q = -iq\int\phi^*\overleftrightarrow{\partial_t} \phi\,d^3\x\;.$$ The ansatz for the field operator is now $$\phi = \int_{|\k|>m} (a_\k u_\k + b^\dag_\k u^*_\k)\,d^3\k\;,$$ while the quantized version of the above conserved quantity evaluates to $$\begin{aligned}
Q&=&-iq\int\colon\phi^\dag\overleftrightarrow{\partial_t}\phi\colon\,d^3\x \\
&=& q\int_{|\k|>m}(b^\dag_\k b_\k - a^\dag_\k a_\k)\,d^3\k\;.\end{aligned}$$ Thus the operator $b^\dag_\k$ may be interpreted as creating a particle of charge $q$ from the vacuum, while $a^\dag_\k$ may be interpreted as creating an anti-particle of charge $-q$ from the vacuum.
Discussion {#Disc}
==========
A sketch that the two-point distribution constructed in Section \[ScaModel\] is of the Hadamard form is now presented. One notes that the two-point distribution (of the difference variable) for our tachyonic theory may be written as $$\label{tpd}
\Delta^{(+)}(x) = \frac{1}{(2\pi)^3}\int_{|\k|>m} \frac{1}{2\wk}\E^{-i(\wk t-\k\cdot\x)}\,d^3\k \;.$$ It is simple to show that the [*scaling limit*]{} of our two-point distribution is $$\lim_{\lambda\to 0}\lambda^2\Delta^{(+)}(\lambda x) = \Delta^{(+)}_0 (x)\;,$$ which is the two-point distribution for the massless theory. Thus the leading order singularity of $\Delta^{(+)}(x)$ is the usual Hadamard one. Conceivably, $\Delta^{(+)}(x)$ may have lower order singularities different from the ones emanating along the light cone from the origin, or the lower order singularities may produce extra covectors for the singularities emanating from the origin. However, noting that Eq.(\[tpd\]) describes a [*Fourier integral operator*]{} [@Hoer71], with [*homogeneous phase function*]{} $\phi(x,\k)=|\k|t-\k\cdot\x$, we may avail ourselves of Theorem 8.1.9 of [@Hoer90],[^3] which restricts the pairs $(x,k)$ in the wave front set to be precisely those in the leading order singularity. Thus the wave front set of $\Delta^{(+)}$ is of the Hadamard type. Since the anti-symmetric part of $\Delta^{(+)}$ is the same as $i$ times the advanced minus retarded fundamental solutions, up to smooth function, the equivalence theorem of [@Rad96a] tells us that $\Delta^{(+)}$ satisfies the global Hadamard condition. Thus, we expect that the free field theory presented in Section \[ScaModel\] may be used as input to a renormalizable (self-) interacting theory, satisfying the criteria of Weinberg’s theorem [@BruFre00]. That it leads to sensible renormalized expectation values of observables quadratic in the fields has already been demonstrated for some typical cases in Section \[Ops\].
We now explain the “no anti-telephones” property of this model, which we consider to be a version of [*chronology protection*]{}, as mentioned in the Introduction. An [*anti-telephone*]{} is a device in which a relay at $B$ is set up at a spacelike separation to a message sender at $A$. The relay is designed to receive a message sent via tachyons from $A$ and immediately to resend the message, again using tachyons, back to a receiver at $A'$ (the same individual as the initial sender, but at a different time). The idea of the anti-telephone is to attempt to violate causality by sending the relayed message from $B$ to $A'$ [*more backward in time*]{} (resp. less forward in time) than the starting message was sent from $A$ to $B$ forward in time (resp. backward in time). Then one would have used tachyons to effect what should be considered a highly egregious violation of causality. (If this could be done, then a signal could be sent whose effect would be to prevent the message from being sent in the first place. However, if the message had not been sent, then nothing would have hindered it from being sent.) However, it is apparent that, due to the cutoff in the spectrum of the one-particle states of the model presented in Section \[ScaModel\], there is a global directional dependence in the lower bound of the allowed energies, such that, if the tachyons managed to travel backward in time in moving from $A$ to $B$, they would certainly be forced to travel as much forward in time, or more so, in travelling back from $B$ to $A'$. Only particles with such a directionally dependent lower bound in the energy may be created out of the vacuum either in the preferred frame, or in any other frame, boosted with respect to the preferred one. Similarly, no sequence of relays could be constructed to guide the tachyons along some path so that they arrive back to $A'$ at a previous time to $A$, since it is clear that such a path of tachyonic world-lines, any one of which goes backward in time, is not constructible in the preferred frame. If it is impossible in one inertial frame, then it must be so for all inertial frames. Note that we should postulate that the same preferred frame must be universal to all tachyonic particles in the same interacting theory, since otherwise, severe violations of causality, in principle, could be brought about.
It is of interest to determine whether any of the “big theorems” of axiomatic QFT [@StreWig64; @Jos65] remain in our model, or in models constructed in a similar approach (by restricting $4$-momenta to lie in the upper half of a single-sheeted hyperboloid cut through by a spacelike hyperplane through the origin). As a first step, we consider the PCT theorem for the Hermitian scalar model. The relevant property to be proved for the two-point distribution is $$\Delta^{(+)}(x)^* = \Delta^{(+)}(-x)\;.$$ This is equivalent to the statement that the Fourier transform of $\Delta^{(+)}$ is real-valued, which is evidently true for our model in any inertial frame. (The Fourier transform of the two-point distribution is a positive multiple of $\theta(k^0+\beta k^z)\delta(k^2+m^2)$ in a frame boosted by a speed $\beta$ in the $z$ direction relative to the preferred frame.) Hence a PCT theorem holds for this model. (The above property, appropriately reformulated, extends to all the Wightman distributions, since the theory we have constructed is [*quasi-free*]{}, and the Wightman distributions determine the full theory by the Wightman reconstruction theorem [@Wig56].)
Next, we touch upon the spin-statistics theorem. We would expect that a well-behaved tachyonic model would reduce to a physically well-behaved massless theory as the tachyonic parameter $m$ tends to $0$. That would mean that the Gårding-Wightman axioms (Wightman positivity, Poincaré invariance, spectral condition, local commutativity) should hold for the scaling limit two-point distribution. However, if the wrong connection between spin and statistics is assumed in this case, then one would necessarily obtain a non-Lorentz invariant theory in the scaling limit, since all the other properties would presumably be satisfied for this limit. (If all the axioms hold in the limit, the limit two-point distribution must be zero, by the usual spin-statistics theorem. However, this would contradict the definition of the scaling limit as the leading order \[non-zero\] behaviour of the two-point distribution as the difference variable $x$ tends to $0$, unless, of course, the two-point distribution of the original tachyonic theory is itself $0$.) Thus, to avoid this undesirable failure of Lorentz invariance in the scaling limit, we must retain the usual connection between spin and statistics. Note that this model then stands in constrast to the one suggested by Feinberg [@Fei67], who assumed the wrong connection of spin with statistics, e.g., anti-commutation relations in the scalar theory.
Finally, we observe that the spacelike hypersurface through the origin (in Fourier space), which bounds the upper half of the single-sheeted mass hyperboloid from below (i.e., the one-particle spectrum of the model described here), may be regarded as defining a [*frame-dependent interpretation rule*]{} for the allowed $4$-momenta of particles and anti-particles in the QFT. This is, in effect, a use of the “Re-interpretation Principle” of [@BilDesSud62], which proposes to regard a negative energy, backward-in-time-moving particle/anti-particle of momentum $\k$ as a positive energy, forward-in-time-moving anti-particle/particle of momentum $-\k$. This would at first seem to suggest an identification of the $4$-momentum $k$ with $-k$ on the full single-sheeted hyperboloid. However, we find it more appropriate to pick a [*single*]{} description from each pair $(k,-k)$, to describe [*both*]{} a particle and anti-particle, and to do so in each frame in a way that preserves chronology protection (a plane must be used to cut the hyperboloid) and the Hadamard condition (the half containing arbitrarily large [*positive*]{} energies must be chosen), and is consistent with Lorentz covariance. (The fact that, in the quantum field model, we choose a single description from among two descriptions which are equally valid from the classical viewpoint, suggests an [*economy of description*]{} principle inherent in the quantum field theory of tachyons. This, of course, is satisfied in the usual choice of the upper mass hyperboloid \[out of two sheets\], as is made in the regular massive Klein-Gordon theory.) Such a “halving” of the single sheeted mass hyperboloid is, up to a boost, unique. We find it rather remarkable that such a simple interpretaion rule leads to [*both*]{} chronology protection [*and*]{} the Hadamard condition (i.e., renormalizability) being satisfied. This points to a deep unity among the “axioms” which we adopt as physical.
In conclusion, we expect that future work in this subject will be done to further develop and clarify the quantum field theoretic aspects of tachyons, especially those involving interactions. A basic step in this direction has been to clarify the calculation of the phase space factor that appears in two-body decay in which one of the products is a tachyon [@Rad08a]. Note that, in that paper, the underlying quantum field theory is implicitly assumed to be in accord with the model presented in this paper (in Section \[ScaModel\]). We also foresee the development started here as extending consistently to Dirac-like tachyonic ([*Dirachyonic*]{}) quantum field theory, whose ramifications (especially the inherent maximal parity breaking that arises in such a model) would tend to support the possibility that the neutrino may be a tachyon.
[**Acknowledgements:**]{} I wish to thank A.S. Wightman and K. Fredenhagen for first asking me questions about tachyonic QFT which led (indirectly or directly) to the present attempt at formulating such a theory consistently. Also, G. Heinzelmann, B.S. Kay, R. Atkins, A. Chodos, W.G. Unruh, and the UBC Dept. of Physics and Astronomy are acknowledged for helpful discussions, encouragement, and hospitality.
[^1]: \[fn1\]Note that we have implicitly used a definition of Fourier transform that differs from that of [@Hoer90] by an extra minus sign in the argument of the exponential function. This accounts for the apparent discrepancy in signs in some formulae, e.g., the exponent in our formula Eq.(\[tpd\]) is minus that in the formula for a Fourier integral operator used in Theorem [8.1.9]{} of [@Hoer90].
[^2]: For corrections to the original proofs given in these references (which contain gaps), please consult [@Koeh95a; @SahVer01].
[^3]: See Footnote \[fn1\].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have investigated, using both a theoretical and an empirical approach, the frequency of low redshift galaxy-galaxy lensing systems in which the signature of weak lensing might be directly detectable. We find good agreement between these two approaches. In order to make a theoretical estimate of the weak lensing shear, $\gamma$, for each galaxy in a catalogue, we have made an estimate of the asymptotic circular velocity from the stellar mass using three different approaches: from a simulation based relation, from an empirically-derived relation, and using the baryonic Tully-Fisher relation. Using data from the Galaxy and Mass Assembly redshift survey we estimate the frequency of detectable weak lensing at low redshift. We find that to a redshift of $z\sim 0.6$, the probability of a galaxy being weakly lensed by at least $\gamma = 0.02$ is $\sim 0.01$. A scatter in the $M_*-M_h$ relation results in a shift towards higher measured shears for a given population of galaxies. Given this, and the good probability of weak lensing at low redshifts, we have investigated the feasibility of measuring the scatter in the $M_*-M_h$ relation using shear statistics. This is a novel measurement, and is made possible because DSM is able to make individual *direct shear measurements, in contrast to traditional weak lensing techniques which can only make statistical measurements. We estimate that for a shear measurement error of $\Delta\gamma = 0.02$ (consistent with the sensitivity of DSM), a sample of $\sim$50,000 spatially and spectrally resolved galaxies would allow a measurement of the scatter in the $M_*-M_h$ relation to be made. While there are no currently existing IFU surveys of this size, there are upcoming surveys which will provide this data (e.g The Hobby-Eberly Telescope Dark Energy Experiment (HETDEX), surveys with Hector, and the Square Kilometre Array (SKA)).*'
author:
- 'C. O. de Burgh-Day$^{1,2,3}$[^1], E. N. Taylor$^{1,3}$, R. L. Webster$^{1,3}$and A. M. Hopkins$^{2,3}$\'
bibliography:
- 'bibliography.bib'
title: 'Direct Shear Mapping: Prospects for weak lensing studies of individual galaxy-galaxy lensing systems'
---
=5
gravitational lensing – weak lensing
Introduction {#sec:Introduction}
============
Weak gravitational lensing is a powerful probe of dark matter in the universe (eg. @1993ApJ...404..441K [-@1993ApJ...404..441K]). Following initial investigations by @2002ApJ...570L..51B and @2006ApJ...650L..21M, @2015MNRAS.451.2161D have developed a new method to measure the weak lensing signal in individual galaxies called Direct Shear Mapping (DSM). The primary scientific application considered by @2015MNRAS.451.2161D is the measurement of mass and mass distribution in dark matter halos around individual low-redshift galaxies. In particular, since the dark matter halo properties can be measured for individual galaxies, DSM will enable the measurement of the dispersion in the galaxy luminous matter to dark matter ratio, as a function of other galaxy observables.\
The possibility of measuring individual galaxy dark matter halo masses through DSM is an exciting prospect, however the measurement itself is challenging, and potentially observationally expensive. We have consequently developed the approach presented here for identifying the most robust candidates for such a measurement. We have also used this approach to investigate the possibility of measuring the scatter in the $M_*-M_h$ relation using shear statistics, and we estimate the size of the statistical sample that would be required to make this measurement.\
In this paper the probability of weak lensing shear has been estimated as a function of the redshifts of the source and lensing galaxies, and a catalogue of candidate galaxy pairs is selected from the Galaxy and Mass Assembly Phase 1 Survey (GAMA I) Data Release 2 (DR2) catalogue (@2011MNRAS.413..971D [-@2011MNRAS.413..971D], Liske et al. in prep). We also find that the distribution of shears in a galaxy sample is influenced by the relationship between stellar mass and halo circular velocity, and the scatter in this relation. With enough shear measurements it may be possible to constrain this relationship, and to measure the scatter.\
DSM uses spatially resolved velocity field information for an object to obtain a shear measurement from the velocity map. DSM assumes intrinsic rotational symmetry in the the velocity map, and searches for departures from this symmetry. This requires either radio data cubes or spatially resolved optical spectroscopy. To identify prospective targets, it is desirable to first obtain an estimate of the shear signal present in a galaxy.\
While galaxy-galaxy lensing has been used to measure halo masses in the past, those studies stack many galaxy-galaxy pairs statistically, to obtain average halo masses [@1996ApJ...466..623B; @1998ApJ...503..531H; @2001ApJ...555..572W; @2006MNRAS.368..715M]. In addition, measurements of galaxy halo shapes have been made from stacked galaxy-galaxy weak lensing measurements .\
To test our target selection algorithm, the weak lensing statistics of a sample of galaxies in the Galaxy and Mass Assembly Data Release 2 (GAMA-DR2) catalogue were investigated, using the stellar mass estimates from @2011MNRAS.418.1587T. The purpose of our lensing frequency algorithm is to estimate of the distribution of shear signals present in a dataset. This algorithm enables novel measurements to be made, and will improve the success rate of any survey intended to measure weak lensing via the DSM method.\
The rest of this paper is organised as follows. In Section \[sec:WeakLensing\] relevant weak lensing theory is described. In Section \[sec:LensingProb\] a theoretical estimate of the probability of weak lensing at low redshift is made, following @2000MNRAS.319..860M. In Section \[sec:Lensingfreqalg\] the inputs, structure and outputs of the lensing frequency algorithm are outlined. In Section \[sec:ApplicationtotheGAMAsurveycatalogue\] results of the application of the algorithm to a dataset obtained from GAMA-DR2 are presented, along with an investigation of the possibility of using shear statistics to measure the scatter in the $M_*-M_h$ relation. Conclusions and a summary are presented in Section \[sec:Conclusions\]. Throughout the paper we assume a flat Concordance cosmology, with $\Omega_m = 0.3$, $\Omega_\Lambda = 0.7$, $H_0 = 70$km$\,$s$^{-1}\,$Mpc$^{-1}$, and $h_{70} = H_0 / 100 = 0.7$.
Weak lensing {#sec:WeakLensing}
============
In this section we will outline the relevant weak lensing theory. Gravitational lensing is the deflection of light from some source on its path to the observer by an intervening mass. The deflection angle is given by $$\boldsymbol{\alpha}(\boldsymbol{\xi}) = \int \mathrm{d}^2\boldsymbol{\xi}^\prime\,\frac{4\,G\,\Sigma(\boldsymbol{\xi}^\prime)}{c^2}\,\frac{\boldsymbol{\xi}-\boldsymbol{\xi}^\prime}{|\boldsymbol{\xi}-\boldsymbol{\xi}^\prime|^2},$$ where $G$ is the gravitational constant, $\Sigma$ is the projected mass distribution of the deflector, $c$ is the speed of light and $\xi$ is the distance from the deflector (i.e the impact parameter). Since the angle of incidence of the light has been altered, the source will appear to be in a different location to its true position. The true position of the source can be found by solving the lens equation $$\label{eq:lens_eq}
\boldsymbol{\beta} = \boldsymbol{\theta} - \frac{D_\mathrm{ds}}{D_\mathrm{s}}\boldsymbol{\alpha},$$ where $\theta$ is the apparent angular separation of the deflector and source, $\beta$ is the true angular separation of the deflector and source, $D_\mathrm{ds}$ is the angular diameter distance between the deflector and source and $D_\mathrm{s}$ is the angular diameter distance between the observer and source. The angular coordinates $\boldsymbol{\beta}$ and $\boldsymbol{\theta}$ can be related to the corresponding physical coordinates in the source and lens planes as $$\begin{aligned}
\boldsymbol{\eta} &= D_\mathrm{s}\boldsymbol{\beta}\\
\boldsymbol{\xi} &= D_\mathrm{d}\boldsymbol{\theta}\\\end{aligned}$$ In the case of a circularly symmetric lens, and perfect alignment between the observer, lens and source, the lens equation can be solved to obtain the Einstein radius, a characteristic length scale $$\theta_\mathrm{E}^2 = \frac{4\pi GM(<\theta_\mathrm{E}D_\mathrm{d})}{c^{2}}\frac{D_\mathrm{ds}}{D_\mathrm{s}D_\mathrm{d}},\label{eq:bend_angle}$$ where $M(<\theta_\mathrm{E}D_\mathrm{d})$ is the mass enclosed within the Einstein radius, and $D_\mathrm{d}$ is the observer-deflector angular diameter distance.\
In the weak lensing regime the light from the source passes well outside the Einstein radius, and the source is singly-imaged. In this case one can assume that so long as the length scale of the source is much less than that of the deflector, then the lensing will be linear. Thus it can be represented by a first-order Taylor expansion, allowing Equation (\[eq:lens\_eq\]) to be re-expressed as a linear coordinate mapping between the lensed and un-lensed coordinate systems $$\boldsymbol{\beta} = \mathbf{A}\,\boldsymbol{\theta}\label{eq:lensmapping},$$ where $\mathbf{A}$ is the Jacobian of transformation; $$\mathbf{A} = \frac{\partial\boldsymbol{\beta}}{\partial\boldsymbol{\theta}} = \left(\begin{array}{cc}
1-\kappa-\gamma_1 & -\gamma_2\\
-\gamma_2 & 1-\kappa+\gamma_1\\
\end{array}\right)\label{eq:A_ij}.$$ Here $\kappa$ is the convergence, and $\gamma_1$ and $\gamma_2$ are the two components of the shear vector $\boldsymbol{\gamma} = (\gamma_1,\gamma_2)$. Equation (\[eq:A\_ij\]) can be inverted to obtain $$\begin{aligned}
\kappa &= 1-(A_{11}+A_{22})/2 \label{eq:kappa}\\
\gamma_1 &= -(A_{11}-A_{22})/2 \label{eq:gamma1}\\
\gamma_2 &= -A_{21} = -A_{12} \label{eq:gamma2}.\end{aligned}$$ The shear vector $\boldsymbol{\gamma}$ can be rewritten in polar co-ordinates as a function of a shear magnitude, $\gamma$, and an angle, $\phi$ $$\gamma_1 = \gamma\cos\phi; \gamma_2 = \gamma\sin\phi,$$ where $\phi$ is the angle of the shear vector and $\gamma = \sqrt{\gamma_1^2+\gamma_2^2}$. In this work we are interested in the value of $\gamma$ for a given lens-source system, which is a function of the lens projected mass density and the lens-source angular separation.\
We assume a Singular Isothermal Sphere (SIS) lens profile throughout this paper, whose projected surface density has the form $$\Sigma(\boldsymbol{\xi}) = \frac{\sigma_v^2}{2G}\frac{1}{|\boldsymbol{\xi}|},$$ where $\sigma_v$ is the halo velocity dispersion. Although the SIS profile is a primitive lens model, it is sufficient for this work and allows for a simple shear estimation. The Einstein radius for an SIS lens is given by $$\theta_\mathrm{E}^2 = \frac{4\pi\sigma_v^2}{c^2}\frac{D_\mathrm{ds}}{D_\mathrm{s}}\label{eq:theta_E_SIS},$$ and from e.g. @2009MNRAS.396.2257L, the shear components for an SIS are $$\gamma_1 = \frac{D_\mathrm{d}\theta_\mathrm{E}}{2|\boldsymbol{\xi}|^3}(\xi_2^2-\xi_1^2);\,\,\gamma_2 = -\frac{D_\mathrm{d}\theta_\mathrm{E}}{|\boldsymbol{\xi}|^3}\xi_1\xi_2,$$ so that $$\gamma = \frac{D_\mathrm{d}\theta_\mathrm{E}}{2|\boldsymbol{\xi}|}.$$ Thus we have an expression for the shear magnitude as a function of lens-source projected separation and lens velocity dispersion, the latter of which can be related to the halo mass.
An estimation of the probability of lensing {#sec:LensingProb}
===========================================
In this section we describe the process by which a theoretical estimate may be made of the probability of a given source being weakly lensed, as a function of the source redshift. We begin by introducing a similar calculation for strong lensing in analytically solvable cosmologies from @2000MNRAS.319..860M. We then adapt this work to the weak lensing case, and for a more realistic Concordance cosmology. In the following sections, we will assume two different lens populations: a population of halos housing elliptical galaxies, as in @2000MNRAS.319..860M, and a Press-Schechter [@1974ApJ...187..425P] population of halos. In both cases we assume an SIS halo. We will discuss the general steps for obtaining the expression for the weak lensing optical depth, and will then discuss the elliptical galaxy and Press-Schechter halo cases individually.\
Given an estimate for the spatial distribution and mass of lensing galaxies in a given volume, it is possible to make an estimate of the distribution of shears in the volume. This can in turn be used to make an estimate of the probability distribution of weak shears across the sky, assuming the distribution of lensing galaxies is isotropic. @2000MNRAS.319..860M have used these arguments to make an estimate of the probability of lensing of quasars by elliptical galaxies for three simplified cosmologies. They define the lensing optical depth, $\tau$, as the fraction of the source plane within which the lens equation has multiple solutions. It can be used as an estimator for strong lensing probability. The contribution to the total optical depth by one lensing galaxy is $$\tau_{\mathrm{g}} = \frac{\pi\beta_\mathrm{crit}^{2}}{4\pi},$$ i.e. the fraction of the sky covered by its lensing cross-section. Here $\beta_\mathrm{crit}$ is the angular distance from the deflector where a background source transitions between being singly or multiply imaged.\
In the case of weak lensing, rather than being interested in the region where the source is multiply imaged, we are interested in the region where the source is singly imaged, yet sheared sufficiently such that it is still a measurable effect. This region takes the form of an annulus about the lens (assuming a spherically symmetric lens), the inner bound of which is the Einstein radius, $\theta_\mathrm{E}$, and the outer bound of which is a function of some limiting shear value, $\gamma_\mathrm{lim}$. Thus, $\tau$ is re-written as a function of a new area, $a(\gamma_\mathrm{lim})$, defined as the area covered by this annulus: $$\tau_{\mathrm{g}}(\gamma_\mathrm{lim}) = \frac{a(\gamma_\mathrm{lim})}{4\pi},$$ and $$a(\gamma_\mathrm{lim}) = \pi \left[\left(\frac{1}{2\gamma_\mathrm{lim}} - 1\right)\theta_\mathrm{E}\right] ^{2},$$ and substituting in equation (\[eq:theta\_E\_SIS\]) this becomes $$\tau_{\mathrm{g}}(\gamma_\mathrm{lim}) = \pi^{2} \left(\frac{1}{2\gamma_\mathrm{lim}} - 1\right)^{2}\left(\frac{\sigma}{c}\right)^{4}\left(\frac{D_\mathrm{ds}}{D_\mathrm{s}}\right)^{2},$$ where $\sigma$ is the velocity dispersion of the deflector, $c$ is the speed of light, $D_\mathrm{ds}$ is the deflector-source distance, and $D_\mathrm{s}$ is the observer-source distance.\
If it is assumed that the population of lensing objects are non-evolving and have uniform volume density at all redshifts (which is reasonable at low redshift), then the differential number of objects at redshift $z$ with velocity dispersion $\sigma$ is $$\frac{\mathrm{d}^{2}N_\mathrm{d}}{\mathrm{d}z_\mathrm{d}\mathrm{d}\sigma} = \frac{\mathrm{d}V_{0}}{\mathrm{d}z_\mathrm{d}}\frac{\mathrm{d}n_\mathrm{d}}{\mathrm{d}\sigma}.$$ Here $\mathrm{d}V_{0}/\mathrm{d}z$ is the comoving volume element at redshift $z$, $$\frac{\mathrm{d}V_{0}}{\mathrm{d}z} = \frac{c}{H_{0}}\frac{(1+z)^{2}D_{A}(z)^{2}}{E(z)}\mathrm{d}\Omega,$$ where $c$ is the speed of light, $D_A(z)$ is the angular diameter distance at redshift $z$, and $$E(z) = \sqrt{(1+z)^{3}\Omega_{\mathrm{M}}+\Omega_{\Lambda}},$$ assuming a flat universe.\
The optical depth to a redshift $z_\mathrm{s}$ is then obtained by integrating over the optical depths of the entire population of deflectors up to $z_\mathrm{s}$ $$\label{eq:tau_zs}
\tau(z_{\mathrm{s}},\gamma_\mathrm{lim}) = \int^{z_\mathrm{s}}_0\int^{\infty}_0 \frac{\mathrm{d}^{2}N_\mathrm{d}}{\mathrm{d}z_\mathrm{d}\mathrm{d}\sigma} \tau_\mathrm{g}(\gamma_\mathrm{lim}) \mathrm{d}\sigma\mathrm{d}z_\mathrm{d}.$$
We will now discuss using this relation to estimate the weak lensing optical depth as a function of source redshift for a population of elliptical galaxies, $n_\mathrm{g}$, and a population of dark matter halos, $n_\mathrm{h}$.
A population of halos housing elliptical galaxies
-------------------------------------------------
@2000MNRAS.319..860M gives the local comoving number density of elliptical galaxies as $$\frac{\mathrm{d} n_\mathrm{g}}{\mathrm{d} \sigma_{\parallel}} = \frac{\delta n_{*}}{\sigma_{*}} \left(\frac{\sigma_{\parallel}}{\sigma_{*}}\right)^{\delta(1+\alpha)-1}\exp\left[-\left(\frac{\sigma_{\parallel}}{\sigma_{*}}\right)^\delta\right],$$ where $\sigma_{\parallel}$ is the observed line-of-sight velocity dispersion, $\alpha=-1.07\pm0.05$, $n_{*} = (0.0019\pm0.003)h_{70}^{3}\,$Mpc$^{-3}$, $\sigma_{*} = 225\pm 20\,$km$\,$s$^{-1}$ and $\delta = 3.7\pm 1$. The values of $\alpha$ and $n_{*}$ are drawn from @1988MNRAS.232..431E, and the values of $\sigma_{*}$ and $\delta$ from @1982ApJ...256..346D. Substituting this into equation (\[eq:tau\_zs\]), the inner integral can be solved analytically, to give $$\label{equation:tau}
\tau(z_\mathrm{s},\gamma_\mathrm{lim}) = C_\mathrm{g}
\int^{z_\mathrm{s}}_{0} \frac{(1+z_\mathrm{g})^{2}}{E(z_\mathrm{g})} \left(\frac{D_\mathrm{d}D_\mathrm{ds}}{D_\mathrm{s}}\right)^{2}\mathrm{d}z_\mathrm{g},$$ where $$C_\mathrm{g} = \pi^{2}n_{*}\left(\frac{\sigma_\parallel}{c}\right)^{4} \left(\frac{1}{2\gamma_\mathrm{lim}}-1\right)^2 \Gamma(1+\alpha+\frac{4}{\delta}).$$ For a given limiting shear value, $\gamma_\mathrm{lim}$, to obtain the probability of measuring a shear of *[at least]{} the limiting value, one must solve equation (\[equation:tau\]) numerically as a function of the source redshift, $z_\mathrm{s}$.*
The dashed grey line in Figure \[fig:Probability\_of\_lensing\_afo\_zs\] shows the lensing optical depth as a function of source redshift for a limiting shear of $\gamma_\mathrm{lim} = 0.02$ based on a population of elliptical galaxies. The coloured lines in Figure \[fig:Probability\_of\_lensing\_afo\_zs\] are discussed in the next section.
A Press-Schechter halo population
---------------------------------
The Press-Schechter mass function is given by $$\frac{\mathrm{d}n}{\mathrm{d}M} = \sqrt{\frac{2}{\pi}}\frac{\rho_m}{M}\times-\frac{\mathrm{d}\ln\sigma(M)}{\mathrm{d}M}\frac{\delta_{crit}}{\sigma(M)}\exp^{-\frac{\delta^2_{crit}}{2\sigma^2(M)}}$$ and gives the co-moving number density of dark matter halos as a function of halo mass [@1974ApJ...187..425P].\
Substituting this into equation (\[eq:tau\_zs\]), $$\label{eq:tau_zs_halo}
\tau(z_{\mathrm{s}},\gamma_\mathrm{lim}) = \int^{z_\mathrm{s}}_0\int^{\infty}_{M_\mathrm{min}} \frac{\mathrm{d}^{2}n_\mathrm{d}}{\mathrm{d}z_\mathrm{d}\mathrm{d}M} \tau_\mathrm{d}(\gamma_\mathrm{lim}) \mathrm{d}M\mathrm{d}z_\mathrm{d},$$ where $M_\mathrm{min}$ is the mass of the smallest halo capable of containing a galaxy, and $$\tau_{\mathrm{g}}(\gamma_\mathrm{lim}) = \pi^{2} \eta \left(\frac{1}{2\gamma_\mathrm{lim}} - 1\right)^{2}\left(\frac{(M\,h_{70}^{-1})^{0.316}}{c}\right)^{4}\left(\frac{D_\mathrm{ds}}{D_\mathrm{s}}\right)^{2},$$ where $\eta = 6.14656\times 10^{-7}$ and we have used the relation $$V_\mathrm{c} = 2.8 \times 10^{-2} \left(M\,h_{70}^{-1}\right) ^{0.316},$$ from @2011ApJ...740..102K. $V_\mathrm{c}$ is the halo asymptotic circular velocity, and we have assumed $V_\mathrm{c} \simeq \sigma$. The lower bound of the integral over $M$ in equation (\[eq:tau\_zs\_halo\]) has been truncated at $M_\mathrm{min}$ because we are only interested in halos large enough to contain at least one galaxy. The value of $M_\mathrm{min}$ at $z\simeq0$ is usually taken to be $\log_{10}(M_\mathrm{min}/M_\odot) \simeq 10$ [@2001PhR...349..125B].\
In this case, the integral over mass is not analytically solvable, however the two integrals are still separable, giving the following expression: $$\begin{split}
\label{eq:tau_zs_halo_2}
\tau(z_{\mathrm{s}},\gamma_\mathrm{lim}) & = C_h \int^{\infty}_{M_\mathrm{min}} \frac{\mathrm{d}n_h}{\mathrm{d}M} \left(M\,h_{70}^{-1}\right) ^{1.264} \mathrm{d}M\\
& \times \int^{z_\mathrm{s}}_0 \frac{(1+z_h)^{2}}{E(z_h)} \left(\frac{D_\mathrm{d}D_\mathrm{ds}}{D_\mathrm{s}}\right)^{2}\mathrm{d}z_h,
\end{split}$$ where $$C_h = \pi^2\eta^2 \left(\frac{1}{2\gamma_\mathrm{lim}} - 1\right)^{2}\left(\frac{\sigma^*}{c}\right)^{4},$$ and $\sigma^* \simeq 321$ km$\,$s$^{-1}$ is a halo characteristic velocity dispersion, corresponding to a characteristic halo mass $M^*$ defined as where $\sigma(M^*) = \delta_c(z)$, where $\delta_c(z)$ is the critical density and $\delta_c(z=0) \simeq 1.686$. At $z=0$, $\log_{10}(M^*/M_\odot) \simeq 13$ [@2001PhR...349..125B].\
To obtain the probability of a shear of at least $\gamma_\mathrm{lim}$ from halos, as a function of source redshift, the mass and redshift integrals in equation (\[eq:tau\_zs\_halo\_2\]) must be solved numerically. In order to illustrate the effect of varying the minimum halo mass, the mass integral in equation (\[eq:tau\_zs\_halo\_2\]) has been solved for three values of the minimum halo mass: $\log_{10}( M_\mathrm{min} / M_\odot )= 9.5,10,$ and $10.5$. The results of this are shown by the three coloured curves in Figure \[fig:Probability\_of\_lensing\_afo\_zs\]. As expected, the probability curve which best matches that of an elliptical population is for the case $\log_{10}( M_\mathrm{min} / M_\odot ) = 10$. Not surprisingly, the higher the minimum halo mass the lower the probability of lensing, since the majority of halos are of lower mass.\
![The lensing optical depth as a function of source redshift and a limiting shear of $\gamma_\mathrm{lim} = 0.02$. The coloured solid lines show the optical depths obtained when using a population of lenses drawn from a Press-Schechter halo mass function. The three lines show the effect of different minimum halo masses. The minimum halo mass is usually taken to be $10^{10}$M$_\odot$ (e.g. in the Millennium simulation). The dashed grey line shows the optical depth obtained when using a population of lenses drawn from an elliptical galaxy population. It is reassuring to see that the Press-Schechter curve which best matches the dashed curve is that which uses the commonly used minimum halo mass of $10^{10}$M$_\odot$.[]{data-label="fig:Probability_of_lensing_afo_zs"}](Probability_of_lensing_afo_zs_HALOES.pdf){width="8.5cm"}
From Figure \[fig:Probability\_of\_lensing\_afo\_zs\], roughly 1 in 1,000 sources at $z \sim 0.2$ will be sheared by at least $\gamma = 0.02$. In the calculation of halo lensing probability, four major assumptions have been made:
- The lensing cross-sections of each galaxy in the population do not overlap. This assumption breaks down if: 1) The redshift is high, since the number of lenses contributing to the optical depth increases with increasing source redshift. 2) $M_\mathrm{min}$ is small, since this leads to more lenses and an increasing cross-section. 3) $\gamma_\mathrm{lim}$ is small, since the cross-section of each lens is larger for a smaller limiting shear.
- Both assumed halo populations do not evolve with redshift. This assumption is only correct over small redshift ranges. The elliptical and halo populations used in Figure \[fig:Probability\_of\_lensing\_afo\_zs\] are for $z = 0$.
- The Press-Schechter halo population is dependent on the minimum halo mass, i.e. the smallest halo which will house a galaxy. This is discussed in detail in Section \[subsubsec:GAMADiscussion\].
- For this calculation we have assumed a uniform distribution of lens redshifts, which is not the case in real data. If the total lens mass distribution is specified however, any inhomogeneity in the galaxy distribution will average out over a large sample of source galaxies. Thus when comparing the results of our theoretical calculation to lensing probabilities in real data, it is acceptable to use the redshifts of the source galaxies in the sample and treat them as homogenous. Such a comparison is discussed in greater detail in Section \[subsubsec:GAMADiscussion\].
In spite of minor limitations, the results shown in Figure \[fig:Probability\_of\_lensing\_afo\_zs\] are strong motivation for a more through calculation and estimation of the probability of detecting weak lensing, since we are interested primarily in relatively low redshift weak lensing, where the DSM technique will be useful.
The lensing frequency algorithm {#sec:Lensingfreqalg}
===============================
We now describe an algorithm for estimating the lensing signal per galaxy in a catalogue. We can use this algorithm to estimate the probability of any given object in the catalogue being lensed by some value, and to identify potentially suitable targets for lensing studies.\
There are three factors which will determine the strength of the shear signal a lens imposes on a source: the mass of the lens, the angular separation of the lens and source, and the redshifts of the lens and source. Different combinations of these can lead to systems in which the lens appears large, and almost obscuring the source, but has a shear of the same magnitude as well separated systems with a small but dense lens. The simplest approach to estimating the strength of the shear present in a potential lens-source system is to simply inspect their separations and redshifts. Clearly this is prone to misidentifications, since systems may be falsely rejected using this method if the lens is particularly dense, and systems may be falsely selected if the lens appears large and close to a background source, but is of a very low density.\
In the lensing frequency algorithm the shear probability is estimated utilising information about the stellar mass of the lens, and the redshifts and angular separations of each lens-source pair. There are a number of steps involved in this process, beginning with the stellar masses of the galaxies in the catalogue, assuming a Singular Isothermal Sphere density distribution for each galaxy, estimating the halo circular velocities of each galaxy, and ultimately returning an estimate of the shear each object imposes on its nearest neighbours. These measurements can then be binned in shear to obtain an estimate of the probability distribution function for shear, and to flag particularly promising targets for follow up and direct measurement of the shear. In this investigation we focus on galaxy-galaxy lensing only, however this process is equally applicable to galaxy-group lensing, provided good group masses are available.
Estimation of the circular velocity, $V_\mathrm{c}$ {#sec:Calc_of_V_circ}
---------------------------------------------------
The most uncertain step in the estimation of the estimated shears for each object is the calculation of the asymptotic circular velocity of the lens, $V_c$. This step is important as it takes us from the stellar mass to the total mass of each lens. Three approaches to this calculation were used:
1. Use a power law relation between halo mass and halo circular velocity obtained from the Bolshoi simulations [@2011ApJ...740..102K]. In this case it is necessary to compute the halo masses from the stellar masses, which is done using the relation derived by @2010ApJ...710..903M.
2. Assume the circular velocity at the outermost regions of the disk is approximately equal to the circular velocity in the halo. In this case, the baryonic Tully-Fisher relation [@2000ApJ...533L..99M] can be used to obtain $V_c$ from the known stellar masses of the objects.
3. Use an empirically derived relation between $V_\mathrm{c}$ and $\sigma_0$ from @2007ApJ...655L..21C. This relation has a significant scatter.
There are significant uncertainties in all of the methods discussed here, and the discrepancies in the results obtained are at times significant (in particular for very high and very low mass objects). However, there is no single best method, and each method approaches the problem from a different starting point. These three methods may bracket the reality, and so by utilising all three methods simultaneously a good representation of possible values is obtained. Therefore the shear resulting from *all three methods are computed.\
The steps taken by the algorithm are as follows:*
1. The objects are sorted by redshift and the angular size distance for each object is computed by integrating over the Friedman equation: $$D = \frac{c}{H_0} \int [\Omega_\mathrm{m} (1+z)^{3} + \Omega_\Lambda ]^{-1} \mathrm{d}z.$$
2. Beginning with the lowest redshift galaxy, assume it is a lens, and perform the following steps:
1. Compute the angular separation between the object and its neighbours out to a specified projected radius $\theta_\mathrm{max}$, excluding those at a lower redshift: $$\theta_\mathrm{sep} = \sqrt{\Delta[R\cos D]^2 + \Delta [D]^2},$$ where $\Delta[R\cos D] = (R_1 - R_2)\cos (D_1 + D_2)$, $\Delta D = D_1-D_2$ and $R_i$, $D_i$ are an objects right ascension and declination[^2] . \[it:N\_neighbours\]
2. Sort the neighbours by angular distance from the object;
3. Compute the halo mass from the stellar mass [@2010ApJ...710..903M] $$\hspace*{-4pt}
\frac{m(M_h)}{M_h} = 2 \left(\frac{m}{M_h}\right)_0\left[ \left(\frac{M_h}{M_1}\right)^{-\phi} + \left(\frac{M_h}{M_1}\right)^\eta \right]^{-1}\label{eq:M*_Mh_rel}$$ where $M_h$ is the halo mass, m is the stellar mass, $(m/M_h)_{0}$ is a normalisation, $M_1$ is a characteristic mass where $m(M_h)/M_h$ is equal to $(m/M_h)_{0}$, and $\phi$ and $\eta$ are two slopes which indicate the behaviour of the relation at the low and high mass ends. The values for the free parameters used in this work are the best-fit values from @2010ApJ...710..903M: $$\begin{split}
\log{(M_1)} & = 11.884^{+0.030}_{-0.023} \\
\left(\frac{m}{M_h}\right)_0 & = 0.02820^{+0.00061}_{-0.00053} \\
\phi & = 1.057^{+0.054}_{-0.046} \\
\eta & = 0.556^{+0.010}_{-0.004} ; \\
\end{split}$$
4. Compute the circular velocity three different ways:
1. Using a power law relation derived from the Bolshoi simulation [@2011ApJ...740..102K], which uses a $\Lambda$CDM cosmology: $$V_\mathrm{c} = 2.8 \times 10^{-2} M_\mathrm{vir} ^{0.316}\label{eq:LCDM_meth},$$ where $M_\mathrm{vir}$ is the virial mass and it is assumed that $M_\mathrm{vir} = M_h$.
2. From the stellar mass using the baryonic Tully-Fisher relation [@2000ApJ...533L..99M] $$V_\mathrm{c} = \left(\frac{M_*}{{A}}\right)^{0.25}\label{eq:bTF_meth},$$ where $A = 26.25\,h_{70}^{-2}\,M_{\odot}\,$km$^{-4}\,$s$^{4}$.
3. Using an empirically derived relation between $V_\mathrm{c}$ and $\sigma_0$ [@2007ApJ...655L..21C]: $$V_\mathrm{c} = \sqrt{2}\sigma_0\label{eq:sig_0_meth}.$$
These will henceforth be referred to as the $\Lambda$CDM, bTF and $\sigma_0$ methods respectively.
5. The Einstein radius of the object is computed for each of the three $V_\mathrm{c}$ values (where it is assumed that due to the virial theorem one can write $V_c \sim \sigma_v$), and the shear for each $V_\mathrm{c}$ value is computed;
3. Move to the next highest redshift galaxy, assume it is a lens, and repeat the above steps.
4. Continue iterating, increasing in redshift until the second-highest redshift has been reached.
Application to the GAMA survey data release 2 catalogue {#sec:ApplicationtotheGAMAsurveycatalogue}
=======================================================
The target selection procedure outlined in Section \[sec:Lensingfreqalg\] has been applied to the Galaxy and Mass Assembly Phase 1 Survey (GAMA I) Data Release 2 (DR2) catalogue (@2011MNRAS.413..971D [-@2011MNRAS.413..971D], Liske et al. in prep). The GAMA Survey is part of a larger project aiming to exploit the latest generation of ground-based and space-borne survey facilities to study cosmology and galaxy formation and evolution . Phase I of the GAMA Survey is a magnitude limited spectroscopic survey measuring galaxy spectra and redshifts in three equatorial regions centred at $9^{h}$, $12^{h}$ and $14.5^{h}$ (called G09, G12 and G15 respectively), each with an area of $12\times 4$ deg$^2$ [@2010MNRAS.404...86B] . The fields were observed to a limiting r-band apparent magnitude of $r_\mathrm{app} < 19.4$, $r_\mathrm{app} < 19.8$ and $r_\mathrm{app} < 19.4$ mag respectively. The target galaxies are distributed over a redshift range $0 < z \lesssim 0.5$ with a median redshift of $z \simeq 0.17$. The GAMA DR2 catalogue contains all GAMA I main survey objects down to $r<19.0$ mag (for G09 and G12) and $r<19.4$ mag (for G15) including spectral redshifts [@2014MNRAS.441.2440B]. The catalogue contains a total of 72,225 objects, of which 71,599 have derived stellar masses [@2011MNRAS.418.1587T]. GAMA survey data is available on the GAMA website[^3].\
The lensing frequency algorithm has been applied to the GAMA I DR2 catalogue in two ways:
1. As the entire catalogue (with some minor cuts detailed below) to investigate the probability of any object being sheared by at least some value, and to identify conceivably suitable targets for potential follow up observation and shear measurement with DSM;
2. As a smaller sample to investigate the possibility of measuring the scatter in the $M_*-M_h$ relation, and to determine what sample size of galaxies is required to make this measurement.
These two analyses are described in detail in the following subsections.
Identifying weak lensing candidates in the GAMA survey Catalogue {#subsec:targets_in_GAMA}
----------------------------------------------------------------
In this section we look at the entire GAMA 1 DR2 catalogue (with minor cuts described in the next paragraph), to investigate the probability of detecting shear of at least $\gamma$ for any given object in the catalogue, and identify candidate targets for follow up observation and shear measurement with DSM.
After removing objects with undefined or uncertain redshift, and selecting for stellar mass in the range $8<\log_{10}(M_*/M_\odot)<12$, a dataset of 69,434 objects was obtained. The data was passed to the lensing frequency algorithm in two ways;
- As the original set of galaxies (i.e. 69,434 objects). The shears present in the dataset were calculated via the three halo circular velocity estimates ($\Lambda$CDM, bTF and $\sigma_0$). The three $M_*-V_\mathrm{c}$ relations are shown in Figure \[fig:Vcirc\_vs\_M\_star\] as dotted ($\Lambda$CDM), dashed (bTF) and solid ($\sigma_0$) lines. A histogram of shears present in the GAMA DR2 sample is shown in Figure \[fig:raw\_LCDMdist\_bTFdist\_sigma\_0\_together\] as solid lines. The probability of a given galaxy being sheared by a particular value is obtained by normalising by the total number of galaxies in the sample.
- As a larger (synthetic) population of galaxies, to account for the scatter in the $M_*-V_\mathrm{c}$ relation. The synthetic population is produced by making 100 realisations of each input galaxy, with a scatter introduced to the realisations (giving a total dataset of 7,029,800 objects). The scatter in the population arises from the intrinsic scatter in each $M_*-V_\mathrm{c}$ relation and so is different for each of the three relations used. A scatter of $\sigma_{\Lambda\mathrm{CDM}} = 0.15$ dex was introduced into Equation (\[eq:M\*\_Mh\_rel\]) for the $\Lambda$CDM method [@2010ApJ...710..903M]; a scatter of $\sigma_\mathrm{bTF} = 0.14\,h_{70}^{-2}$ dex was introduced into Equation (\[eq:bTF\_meth\]) for the bTF method [@2000ApJ...533L..99M]; and a scatter of $\sigma_{\sigma_0} = 0.08$ dex was introduced into Equation (\[eq:sig\_0\_meth\]) for the $\sigma_0$ method. The $\sigma_0$ values and associated scatter used in the $\sigma_0$ method were obtained by fitting for the stellar masses using velocity dispersions obtained from the Sloan Digital Sky Survey Data Release 10 (SDSSDR10; @2014ApJS..211...17A [-@2014ApJS..211...17A]). The $M_*-V_\mathrm{c}$ relation of the synthetic population of galaxies is shown in Figure \[fig:Vcirc\_vs\_M\_star\] as blue ($\Lambda$CDM), maroon (bTF) and green ($\sigma_0)$ points. One can see that in the range $8<\log_{10}(M_*/M_\odot) <12$, where most galaxies are situated, the three methods agree well. The shears present in the synthetic dataset were computed for every point, and the resulting number of sheared objects, normalised by 100, are presented in Figure \[fig:raw\_LCDMdist\_bTFdist\_sigma\_0\_together\], as dotted lines. The synthetic data is normalised by 100 to allow for better comparison to the real data, since for every galaxy in the real dataset there are 100 galaxies in the synthetic dataset. One can see that the three methods agree well *within within each of the real and synthetic datasets, with the synthetic datasets sitting slightly above the real datasets. The implications of this upward shift are discussed in Section \[subsec:Scatter\_in\_M\*\_Mh\_rel\]. The probability of a given galaxy being sheared by a particular value is obtained by normalising by the total number of galaxies in the sample.*
![ The $M_*-V_\mathrm{c}$ relation for three methods of calculation. ‘K’ denotes the $\Lambda$CDM method [@2011ApJ...740..102K], ‘M’ the bTF method [@2000ApJ...533L..99M] and ‘C’ the $\sigma_0$ method [@2007ApJ...655L..21C]. The dotted, dashed and solid lines show the relation with no scatter introduced. The blue, maroon and green points show a synthetic dataset obtained by making 100 realisations of the original dataset, and introducing a scatter of $\sigma_{\Lambda\mathrm{CDM}} = 0.15$ dex, $\sigma_\mathrm{bTF} = 0.14\,h_{70}^{-2}$ dex, and $\sigma_{\sigma_0} = 0.08$ dex for the $\Lambda$CDM [@2010ApJ...710..903M], bTF [@2000ApJ...533L..99M] and $\sigma_0$ methods respectively. The three methods agree well in the range $8<\log_{10}(M_*/M_\odot) <12$ (to within $\sim 15\%$), where most galaxies are situated. []{data-label="fig:Vcirc_vs_M_star"}](V_circ_vs_M_star.pdf){width="8.5cm"}
![ Histogram of estimated shears for galaxies in the GAMA survey. The solid, coarsely-binned lines correspond to the shears present in the original GAMA DR2 sample. The dotted lines correspond to the shears present in a synthetic dataset, obtained by producing 100 relisations of the original dataset, and introducing a scatter in the $M_*-V_\mathrm{c}$ relation. For every galaxy in the GAMA survey, there are 100 galaxies in the synthetic dataset. Therefore for ease of comparison of the histograms of the synthetic and real datasets, the number of sheared objects in the synthetic dataset have been divided by 100. One can see that the three methods agree well *within each of the real and synthetic datasets, with the synthetic datasets sitting slightly above the real datasets.* []{data-label="fig:raw_LCDMdist_bTFdist_sigma_0_together"}](raw_LCDMdist_LCDM_bTFdist_bTF_sigma_0dist_sigma0_together.pdf){width="8.5cm"}
### Discussion {#subsubsec:GAMADiscussion}
The number of galaxies with shears above the cutoff value of 0.02 varied between the three methods of calculating $V_\mathrm{c}$, with the $\Lambda$CDM method giving the largest estimate of measurable shears in the data, and the bTF method giving the smallest estimate. The probability of a shear of at least $\gamma = 0.02$ was $\mathrm{P}(\geq\gamma) = 0.018, 0.005$ and $0.007$ for the $\Lambda$CDM, bTF and $\sigma_0$ methods respectively. The bTF method was chosen for selecting candidate targets because it gives the smallest shear probability, and so will be the least likely to overestimate the shears present in a sample. Using the bTF method, the number of objects in GAMA with an estimated shear of $\gamma\geq 0.02$ is $393$. The bTF method always estimates a smaller shear than the $\Lambda$CDM and $\sigma_0$ methods for any given object, so any of the $393$ objects selected by the bTF method would also have been selected by the other two methods (although of course the converse is not true).\
These objects were extracted from the sample, and matched to objects in the SDSS DR10 catalogue. The lens-source pairs in the sample were ranked by each of the authors by eye, based on their morphology, surface brightness, inclination angle, separation and environment. The motivation for selecting on this criterion stems from the requirements of the DSM algorithm; that the source galaxy be undisturbed and stably-rotating, neither face-on nor edge-on and bright enough for observation with an IFU. For a more detailed discussion of the requirements of the DSM algorithm, see @2015MNRAS.451.2161D. Two highly-ranked example lens-source pairs are shown in Figure \[fig:twopairs\].\
![ Two example galaxy pairs from the GAMA DR2 Sample identified with the target selection algorithm. In both cases the galaxy in the crosshairs is the source galaxy (i.e. the galaxy being lensed). The left-hand images show thumbnails of the galaxy pairs from the SDSS DR10 Finding Chart Tool, while the middle and right-hand images show the J-band images from the UKIDSS survey, with the residuals from 2D Sérsic fits, taken from GAMA’s online Single Object Viewer tool. The top pair is at RA $= 213.705$ deg, DEC $= 1.623$ deg, and has an estimated shear of $\gamma = 0.023$. The lens and source redshifts are $z = 0.128$ and $z = 0.186$ respectively. The bottom pair is at RA $= 213.705$ deg, DEC $= 1.623$ deg, and has an estimated shear of $\gamma = 0.053$. The lens and source redshifts are $z = 0.088$ and $z = 0.190$ respectively. []{data-label="fig:twopairs"}](all.pdf){width="8.5cm"}
![Lens stellar mass as a function of lens redshift for the GAMA DR2 sample, showing the distribution of shears with these parameters. Grey contours show the density of all galaxies in the sample. Blue squares show galaxies with a shear in the range $0.01<\gamma<0.05$, and red circles show those with a shear in the range $\gamma>0.5$. The shears were estimated using the bTF method.[]{data-label="fig:lens_mstar_v_z_w_gamma"}](M_star_vs_z_w_gammas.pdf){width="8.5cm"}
Figure \[fig:lens\_mstar\_v\_z\_w\_gamma\] shows the distribution of galaxies with significant estimated shear, with lens stellar mass plotted as a function of lens redshift. As might be expected, the selection favours more massive lensing galaxies, with significant numbers of candidates at all redshifts. The quality of the shear measurements however depends on the characteristics of the source galaxy.\
The normalised cumulative sum was taken of the data shown in Figure \[fig:raw\_LCDMdist\_bTFdist\_sigma\_0\_together\], the results of which are shown in Figure \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\], giving the probability of measuring a shear of *at least $\gamma_\mathrm{lim}$. One can again see the upward shift in the distributions for synthetic data, relative to those for the real data. Also shown are two dashed lines corresponding to our theoretical estimate of the probability of lensing from Section \[sec:LensingProb\]. The probability of at least a given shear being observed in a galaxy increases as the minimum shear decreases. Since the lensing cross-section of each galaxy goes as $\gamma_\mathrm{lim}^{-2}$, this is not surprising, as smaller minimum shears rapidly increase the fraction of the sky in which measurable lensing will occur. The method which gives the lowest overall probability is the bTF method. Since we expect to be able to measure shears as small as $\gamma = 0.02$ with DSM, we can expect to have *[at least]{} a one in two hundred chance of measuring shear in a randomly chosen galaxy (from the bTF method), and up to a roughly one in fifty chance (from the $\Lambda$CDM method).\
It is obvious that the $\Lambda$CDM method results in much higher probabilities than the bTF and $\sigma_0$ methods. It is interesting to note that the $\Lambda$CDM method is the one which uses only simulation and theoretically-derived relations to obtain values of $V_\mathrm{c}$ from $M_*$. In contrast, the bTF and $\sigma_0$ methods utilise empirically derived relationships between $V_\mathrm{c}$ and $M_*$. One can easily identify the origin of the higher numbers of lensed galaxies from the $\Lambda$CDM method; the break in the $M_* - V_\mathrm{c}$ relation for this method. While giving good agreement in the intermediate mass range, this results in a much larger corresponding halo circular velocity. These higher $V_\mathrm{c}$ galaxies will have a much larger lensing cross section, resulting in a larger number of lensed objects. Hence, while there are relatively few galaxies with $\log_{10}(M_*/M_\odot)>12$, they contribute strongly to the total lensing probability. This raises the question: how does this over-abundance of higher-$V_\mathrm{c}$ galaxies affect the results of $\Lambda$CDM-based cosmological simulations? Further consideration of this topic is beyond the scope of this paper, and is left to future work.\
**
![ Estimated probability of measuring a shear of *[at least]{} $\gamma$ for galaxies in the GAMA survey. The dotted lines are the probabilities with scatter introduced into the $M_*-V_\mathrm{c}$ relation (i.e. the synthetic dataset), and the solid lines are the probabilities with no scatter introduced. Again one can see that the synthetic datasets sit higher than the real datasets. The dashed lines are included for comparison purposes, and correspond to the probabilities as a function of $\gamma$ derived from the theoretical calculation of probability in Section \[sec:LensingProb\], using the redshifts of the sheared objects in the GAMA survey.* []{data-label="fig:cumPDF_LCDMdist_bTFdist_sigma_0_together_AT_LEAST_x"}](cumPDF_LCDM_LCDMdist_bTF_bTFdist_sigma_0_sigma_0dist_together_AT_LEAST_x_OVERPLOT.pdf){width="8.5cm"}
![ Distribution of redshifts in the GAMA sample, after a redshift quality cut has been made. The mean and median of this distribution are $0.18$ (solid white line) and $0.17$ (dashed white line) respectively, and the range of the distribution is $0\lesssim z\lesssim 0.65$. []{data-label="fig:hist_of_z"}](hist_of_z.pdf){width="8.5cm"}
As mentioned already, in Figure \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\] we have included two dashed lines corresponding to our theoretical estimate of the probability of lensing from Section \[sec:LensingProb\]. The probability calculated in Section \[sec:LensingProb\] is the probability that a source at redshift $z_s$ is lensed by at least some value $\gamma_\mathrm{lim}$. The probability obtained from the lensing frequency algorithm however is the probability that a single galaxy from the GAMA DR2 sample, chosen at random, is lensed by at least $\gamma_\mathrm{lim}$. In order to compare the two, we need to calculate the theoretical probability that one galaxy, chosen at random from a GAMA-like population of galaxies, is lensed by at least $\gamma_\mathrm{lim}$. From Section \[sec:LensingProb\], the probability that some object in the GAMA DR2 sample is lensed by at least $\gamma_\mathrm{lim}$ will be dependent on its redshift only (once a lens mass distribution has been assumed). Thus the probability that *[any]{} galaxy in the sample has been lensed by at least $\gamma_\mathrm{lim}$ will be the sum of the individual probabilities of each one having been lensed. Then the probability that one galaxy chosen at random has been lensed by at least $\gamma_\mathrm{lim}$ will be the probability of any having been lensed, divided by the total number of galaxies in the sample. Figure \[fig:hist\_of\_z\] shows the histogram of the redshifts of the GAMA DR2 sample after a redshift quality cut has been made. While the distribution of redshifts in the GAMA sample is not uniform, it is sufficiently smoothly varying that the assumption made in Section \[sec:LensingProb\] (that the lenses are uniformly distributed in redshift) will suffice. Thus we have arrived at a method by which we can compare the results of Section \[sec:LensingProb\] and Section \[subsec:targets\_in\_GAMA\].\
From Section \[sec:LensingProb\], the theoretical probability of measuring a shear of at least $\gamma = 0.02$ in any one galaxy in the GAMA survey is $\mathrm{P}(\geq\gamma) \simeq 0.002$ and $0.003$ for the halos housing elliptical galaxies and Press-Schechter halo populations respectively. Comparing to the probabilities obtained from the GAMA DR2 sample ($\mathrm{P}(\geq\gamma) = 0.018, 0.005$ and $0.007$ for the $\Lambda$CDM, bTF and $\sigma_0$ methods respectively), we see the difference is a factor of $\sim 6$ for the $\Lambda$CDM method, and a factor of $\sim 2$ for the bTF and $\sigma_0$ methods. This is an acceptable level of agreement, given the assumptions made in the calculations in Section \[sec:Lensingfreqalg\]. The level of agreement between the two approaches can be seen in Figure \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\].\
The theoretical estimate of $P(>\gamma_\mathrm{lim})$ drawn from the Press-Schechter halo population is dependent on the value of the minimum halo mass in which a galaxy will form. As can be seen in Figure \[fig:Probability\_of\_lensing\_afo\_zs\], a smaller (larger) value of $M_\mathrm{min}$ will result in a higher (lower) probability of measuring a shear of at least $\gamma_\mathrm{lim}$ as a function of redshift. It can be seen in Figure \[fig:GAMA\_cf\_varying\_M\_star\] that adjusting $M_\mathrm{min}$ will alter the level of agreement between the theoretical estimate and the probabilities derived from the GAMA DR2 sample. Each dashed grey line represents a minimum halo mass incremented by $0.5\log_{10}(M_\odot)$. The line with the highest corresponding probability is that with $\log_{10}(M_\mathrm{min}/M_\odot) = 8$, and that with the lowest corresponding probability $\log_{10}(M_\mathrm{min}/M_\odot) = 12$. The probabilities from the lensing frequency algorithm are as in Figure \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\]. It is interesting to note that for the bTF and $\sigma_0$ methods, the best agreement with the theoretical approach is obtained for a minimum halo mass in the range $9\lesssim \log_{10}(M_\mathrm{min}/M_\odot)\lesssim 10$, whereas the currently accepted value is $\log_{10}(M_\mathrm{min}/M_\odot) = 10$.*
![Comparison of the results from the lensing frequency algorithm with the theoretically derived Press-Schechter probabilities, with varying $M_\mathrm{min}$. The highest theoretically-derived probability corresponds to a minimum halo mass of $\log_{10}(M_\mathrm{min}/M_\odot) = 8$, while the lowest corresponds to $\log_{10}(M_\mathrm{min}/M_\odot) = 12$. The results of the lensing frequency algorithm are plotted as in Figure \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\]. The bTF and $\sigma_0$ methods best agree with a Press-Schechter halo population with a minimum halo mass in the range $9\lesssim \log_{10}(M_\mathrm{min}/M_\odot)\lesssim 10$. []{data-label="fig:GAMA_cf_varying_M_star"}](GAMA_cf_varying_M_star.pdf){width="8.5cm"}
Measuring the scatter in the $M_*-M_h$ relation {#subsec:Scatter_in_M*_Mh_rel}
-----------------------------------------------
Since halos of given mass can have different halo concentrations, spin parameters and merger histories, we expect them to house galaxies with a range of masses. This manifests as a scatter in the stellar mass to halo mass relation. The value of the scatter in the $M_*-M_h$ relation is not well constrained, as directly measuring the masses of galaxies *and their host halos is not trivial. Abundance matching techniques can be useful for describing the relationship between stellar mass and halo mass, however they cannot constrain the scatter in the $M_*-M_h$ relation [@2009ApJ...696..620C]. @2012ApJ...744..159L used traditional weak lensing techniques to constrain the $M_*-M_h$ relation, however as with abundance matching they are unable to place any meaningful constraints on the scatter in the relation. An alternative approach has been to utilise satellite galaxies to constrain the galaxy luminosity-halo mass relation. Early work involved stacking the central galaxies to obtain a statistical measure of the kinematics of the satellite galaxies [@1987Natur.325..779E; @1993ApJ...405..464Z; @1994ApJ...435..599Z; @1997ApJ...478...39Z]. However, recent work has avoided the need for stacking, largely by utilising the larger datasets offered by the 2dF Galaxy Redshift Survey (2dFGRS; @2001MNRAS.328.1039C [-@2001MNRAS.328.1039C]) and SDSS [@2002ApJ...571L..85M; @2003ApJ...593L...7B; @2003ApJ...598..260P; @2004MNRAS.352.1302V; @2004bdmh.confE..41V; @2005ApJ...635..982C; @2007ApJ...654..153C; @2009MNRAS.392..801M]. These techniques to not utilise weak lensing, and do not attempt to measure the scatter in the $M_*-M_h$ relation.\
In this section we test whether DSM measurements of a population of galaxies could be used to measure the scatter in the $M_*-M_h$ relation. This is a novel measurement which is difficult with traditional weak lensing techniques, but is made possible with DSM because it can measure individual shears with far greater accuracy around individual galaxies. We then briefly describe how to fit for the scatter in the $M_*-M_h$ relation, and measure the scatter in a set of simulated shear datasets with a known scatter incorporated.\
As was noted in Section \[subsec:targets\_in\_GAMA\], and can be seen in Figures \[fig:raw\_LCDMdist\_bTFdist\_sigma\_0\_together\] and \[fig:cumPDF\_LCDMdist\_bTFdist\_sigma\_0\_together\_AT\_LEAST\_x\], scatter in the $M_*-M_h$ relation results in a shift in the distribution of shears present in a population of galaxies towards larger shears. If the measurement error in the shear is sufficiently small, or the population of galaxies with known stellar masses is large enough, it is possible to measure the scatter in the $M_*-M_h$ relation by comparing the distribution of shears to those obtained from a $M_*-M_h$ relation with zero scatter.\
To perform this measurement with DSM, velocity maps from from an intermediate redshift survey would be required. The velocity maps can be obtained from several components of the galaxy, such as HI in radio wavelengths, or Integral Field Unit (IFU) observations of stellar velocities and gas emission in optical wavelengths. IFU maps obtained from bright, H$\alpha$ emitting galaxies are the most practical and easily obtainable in the immediate term however, and so we will focus on these galaxies in this analysis.\
We assume that the scatter in the $M_*-M_h$ relation is lognormal, which is the standard form assumed in the literature (e.g. @2010ApJ...717..379B [-@2010ApJ...717..379B], @2010ApJ...710..903M [-@2010ApJ...710..903M], @2013ApJ...770...57B [-@2013ApJ...770...57B]). The mean of a lognormal distribution is a function of the size of the scatter in the distribution, and is given by $\exp(\mu+\sigma^2/2)$, where $\sigma$ is the scatter in the distribution, and $\mu$ is the natural logarithm of the mean of the underlying normal distribution. It can be seen that the mean increases with increasing scatter. Thus if a lognormal scatter is introduced into the $M_*-M_h$ relation for a population of galaxies, the result is a larger number of galaxies with higher mass halos. It is this property of the lognormal scatter in the $M_*-M_h$ relation which leads to an increased probability of larger shears.\
To compare the shear distributions obtained from populations with and without scatter in the $M_*-M_h$ relation, the lensing frequency algorithm was applied to a selection of mock catalogues generated from a subsample of the complete GAMA DR2 sample. This subsample was obtained by performing the same cuts on the GAMA DR2 as in the previous section, along with the following additional cuts: $0.1<z<0.15$, $r<17.5$, and keeping only galaxies with H-$\alpha$ emission. The resulting dataset contained 2,861 galaxies. When computing shears in the mock catalogues, a maximum lens-source separation of $D_\mathrm{max} = 0.2$ Mpc was used for step (\[it:N\_neighbours\]) of the lensing frequency algorithm.\
A successful measurement of the scatter in the $M_*-M_h$ relation will be limited by survey sample size and the size of the scatter. To investigate the range in which these parameters would enable a measurement of the scatter, a set of simulated datasets of varying size was created, with a range of scatters in the $M_*-M_h$ relation.\
To produce the mock catalogues, larger populations of galaxies were generated from the subsample of GAMA DR2 galaxies by producing Monte-Carlo realisations of each galaxy. The resulting datasets contained $N$= 1,000, 15,000, 50,000 and 150,000 galaxies. The distribution of shears for a tight $M_*-M_h$ relation was computed for the simulated datasets. A scatter, $\sigma_{M_*-M_h}$, was then included in the $M_*-M_h$ relation and the shears were computed again. This process was repeated for a selection of values of $\sigma_{M_*-M_h}$ between $0.1$ and $0.5$. This range of scatter is chosen to bracket the current literature values (e.g.. @2009ApJ...693..830Y [-@2009ApJ...693..830Y], @2009MNRAS.392..801M [-@2009MNRAS.392..801M], @2010ApJ...717..379B [-@2010ApJ...717..379B], @2010ApJ...710..903M [-@2010ApJ...710..903M]). The resulting ‘true’ shears for each dataset were then ‘observed’ with a range of measurement errors, so that the ‘observed’ shear in the tight datasets contained a shear measurement error only, while the ‘observed’ shear in the scattered datasets contained a shear measurement error, and additional scatter from the $M_*-M_h$ relation.\
*
{width="17cm"}
{width="17cm"}
{width="17cm"}
The key questions are: 1) Can we identify the scatter from the $M_*-M_h$ relation over the scatter from measurement error? 2) How many objects do we need to do so? 3) What is the uncertainty on the measurement? To answer these questions, one first needs to establish that for a given number of galaxies, the distribution of shears arising from a tight $M_*-M_h$ relation can be distinguished from the distribution with scatter in the $M_*-M_h$ relation. The distributions of shears with and without scatter in the $M_*-M_h$ relation are shown in Figures \[fig:fourfigs\_w\_n\]-\[fig:fourfigs\_w\_obs\]. Figure \[fig:fourfigs\_w\_n\] shows the observed shear as a function of the predicted shear for $N =$ 1,000, 15,000, 50,000 and 150,000, for a shear measurement error of $\sigma_{\gamma_\mathrm{obs}}=0.02$ and $M_*-M_h$ relation scatter of $\sigma_{M_*-M_h} =0.3$ dex. The solid blue and red lines show the mean in the tight and scattered distributions respectively, and the dashed blue and red lines show $\pm1\sigma$ from the mean in the tight and scattered distributions respectively. The shaded background shows the relative excess (red) or shortfall (blue) of galaxies with scatter relative to galaxies without scatter, in 2D bins of size $(\Delta\gamma_\mathrm{obs}\times\Delta\gamma_\mathrm{pred}) = (0.02\times 0.002)$. Figure \[fig:twofigs\_w\_scatt\] shows the observed shear as a function of the predicted shear for $N =$ 50,000, shear measurement error of $\sigma_{\gamma_\mathrm{obs}}=0.02$, and scatters of $\sigma_{M_*-M_h} =0.1$ and $0.5$ dex. Lines and shading are as in Figure \[fig:fourfigs\_w\_n\]. Figure \[fig:fourfigs\_w\_obs\] shows the observed shear for $\sigma_{\gamma_\mathrm{obs}}$ = 0.005, 0.01, 0.03, and 0.1 for $N =$ 50,000 and $\sigma_{M_*-M_h}= 0.3$dex. Lines and shading are as in Figure \[fig:fourfigs\_w\_n\].\
The distributions of observed shears with and without scatter in the $M_*-M_h$ relation were compared using the Two Sample Kolmogorov-Smirnov (2SKS) Test. For it to be possible for the scatter in the $M_*-M_h$ relation to be measurable from comparing the distribution of observed to predicted shears with and without scatter, we require the distributions to fail the hypothesis that they are drawn from the same distribution, under the 2SKS Test. That is, we require the p-value of the test to be small. The resulting 2SKS scores and p-values (denoted $S_\mathrm{2SKS}$ and $P_\mathrm{2SKS}$) for the combinations of $n$ and $\sigma_{M_*-M_h}$ considered are presented in Table \[table:2SKSreults\].\
\[table:2SKSreults\]
[c c c c c ]{}\
$\mathbf{N}$ & $\boldsymbol{\sigma_{\gamma_\mathrm{obs}}}$ & $\boldsymbol{\sigma_{M_*-M_h}}$ & $\boldsymbol{S_\mathrm{2SKS}}$ & $\boldsymbol{P_\mathrm{2SKS}}$\
\[0.5ex\]\
1,000 & 0.02 & $ 0.3 $ & $2.5\times 10^{-2}$ & 0.91\
15,000 & 0.02 & $ 0.3 $ & $8.7\times 10^{-3}$ & 0.61\
50,000 & 0.02 & $ 0.3 $ & $1.1\times 10^{-2}$ & $4.4\times 10^{-3}$\
150,000 & 0.02 & $ 0.3 $ & $7.9\times 10^{-3}$ & $1.9\times 10^{-3}$\
\
50,000 & 0.005 & $ 0.3 $ & $1.6\times 10^{-2}$ & $6.2\times 10^{-6}$\
50,000 & 0.01 & $ 0.3 $ & $1.4\times 10^{-2}$ & $1.4\times 10^{-4}$\
50,000 & 0.03 & $ 0.3 $ & $1.0\times 10^{-2}$ & $9.5\times 10^{-2}$\
50,000 & 0.1 & $ 0.3 $ & $8.7\times 10^{-3}$ & $4.7\times 10^{-2}$\
\
50,000 & 0.02 & $ 0.1 $ & $7.6\times 10^{-3}$ & 0.11\
50,000 & 0.02 & $ 0.5 $ & $1.9\times 10^{-2}$ & $1.6\times 10^{-8}$\
The next step is to fit for the scatter in $M_*-M_h$ relation in simulated datasets. To do this, we assume that the observed shears are distributed according to both the scatter in the $M_*-M_h$ relation, and an observation error, so that the likelihood of an observed shear given the true underlying shear is given by $$\begin{split}
L(\gamma_\mathrm{obs}&|\gamma_\mathrm{true}, \sigma_{M_*-M_h}, \sigma_{\gamma_\mathrm{obs}})\\
&= \int N(\tau - \gamma_\mathrm{obs},0,\sigma_{\gamma_\mathrm{obs}}) \\
&\times M(\tau,\gamma_\mathrm{true},\sigma_{M_*-M_h})\mathrm{d}\tau,
\end{split}$$ where $N$ is a Gaussian distribution representing the measurement error in the DSM method, and $M$ is a lognormal distribution representing the scatter in the $M_*-M_h$ relation. $N$ is given by $$N(x,\mu,\sigma) = \frac{1}{\sqrt{2\pi}\sigma}\exp{\left[-\frac{1}{2}\left( \frac{x-\mu}{\sigma} \right)^2\right]}$$ where $\mu$ is the mean of the distribution and $\sigma$ is the standard deviation. The lognormal, $M$, is a distribution whose logarithm is normally distributed. It is given by $$M(x,\mu,\sigma) = \frac{1}{\sqrt{2\pi}\sigma x}\exp{\left[-\frac{1}{2}\left( \frac{\log (x/\mu)}{\sigma} \right)^2\right]}\label{eq:lognormal}$$ where $\mu$ and $\sigma$ are the mean and standard deviation of the underlying normal distribution. The total likelihood of a given value of $\sigma_{M_*-M_h}$ for fixed $\sigma_{\gamma_\mathrm{obs}}$ is then given by $$\mathcal{L}(\sigma_{M_*-M_h}) = \sum_i L(\gamma_{\mathrm{obs},i}|\gamma_{\mathrm{true},i}, \sigma_{M_*-M_h} ,\sigma_{\gamma_\mathrm{obs}}).$$ We have fitted for the scatter in the $M_*-M_h$ relation for an assumed true scatter of $\sigma_{M_*-M_h}=0.3$ by maximising $\log[\mathcal{L}(\sigma_{M_*-M_h})]$ for a range of values of $N$ and $\sigma_{\gamma_\mathrm{obs}}$ to investigate the behaviour of the uncertainty in the fit with these parameters. A Fisher Matrix analysis was used to estimate the standard error in the maximum likelihood for each $\sigma_{\gamma_\mathrm{obs}}$ and $N$. Figure \[fig:stand\_err\_plot\] shows the standard error, $\Delta\sigma_{M_*-M_h}$, as a function of these parameters. We find that for an assumed scatter of $\sigma_{M_*-M_h}=0.3$, to obtain a robust fit with a measurement error of $\sigma_{\gamma_\mathrm{obs}}=0.02$ a dataset of $N\sim$ 50,000 DSM measurements is required, a result which is consistent with the results of the 2SKS test.\
![The standard error in the maximum value of $\log[\mathcal{L}(\sigma_{M_*-M_h})]$ as a function of the shear measurement error, $\sigma_{\gamma_\mathrm{obs}}$ for values of $N$ in the range 1,000$<N<$150,000. []{data-label="fig:stand_err_plot"}](theplot.pdf){width="8.5cm"}
### Discussion {#subsubsec:Scatt_in_M*_Mh_rel}
For fixed $\sigma_{\gamma_\mathrm{obs}}$ and $\sigma_{M_*-M_h}$, the p-value for $N =$ 15,000 is too large to rule out the hypothesis that the two samples of data come from the same distribution, while the p-value for $N =$ 50,000 is sufficient to rule out this hypothesis, and the p-value for $N =$ 150,000 can easily do so. Similarly, for fixed $N$ and $\sigma_{M_*-M_h}$, the p-value for $\sigma_{\gamma_\mathrm{obs}}$ = 0.03 is too large to confidently rule out the hypothesis that the two samples of data come from the same distribution, while the p-value for $\sigma_{\gamma_\mathrm{obs}}$ = 0.02 is sufficient to rule out this hypothesis, and the p-value for $\sigma_{\gamma_\mathrm{obs}}$ = 0.01 can easily do so. Not surprisingly, a larger scatter in the $M_*-M_h$ relation results in a smaller p-value, since a larger scatter directly increases the difference between the datasets with and without scatter in them. From the above considerations we conclude that in order to measure a scatter of $\sigma_{M_*-M_h} = 0.3$, a sample of $\gtrsim$ 50,000 galaxies with a shear measurement error of $\sigma_{\gamma_\mathrm{obs}} \lesssim$ 0.02 would be required. For larger values of the scatter, larger values of $\sigma_{\gamma_\mathrm{obs}}$ and smaller values of $N$ would be sufficient.\
This result is confirmed in our fits for a scatter in the $M_*-M_h$ relation of $0.3$ for a range of values of $N$ and $\sigma_{\gamma_\mathrm{obs}}$, presented in Figure \[fig:stand\_err\_plot\]. For a true $M_*-M_h$ scatter of $0.3$, a shear measurement error of $\sigma_{\gamma_\mathrm{obs}} = 0.02$, and a sample of $N=$50,000 shear measurements, we recover a fitted scatter of 0.308$\pm$0.02.\
DSM is expected to achieve measurement errors of $\sigma_{\gamma_\mathrm{obs}} \sim$ 0.02, and so shear measurement error is not seen to be a limiting factor in measuring scatter in the $M_*-M_h$ relation. While a sample of $\gtrsim 50,000$ galaxies with spatial and spectral resolution does not yet exist, there are several surveys beginning in the near to intermediate future which will provide datasets of a sufficient size to perform this experiment, for example surveys with the Hector instrument [@2012SPIE.8446E..53L] on the Anglo-Australian Telescope (AAT), and The Hobby-Eberly Telescope Dark Energy Experiment (HETDEX; @2008ASPC..399..115H, [-@2008ASPC..399..115H]), or surveys on the Square Kilometre Array (SKA). We conclude that while it would not be possible to utilise DSM to measure the scatter in the $M_*-M_h$ relation with existing IFU survey data, it will be possible with data from upcoming surveys.
Conclusions {#sec:Conclusions}
===========
We have made an analytical estimate of the frequency of a source being weakly lensed given a uniformly distributed population of lenses, following @2000MNRAS.319..860M. We have adapted their work for the weak lensing case, in which we consider the probability of the source being lensed by *[at least]{} some limiting value $\gamma_\mathrm{lim}$. The results of this analysis suggest the probability of detecting weak lensing greater than a limiting value of $\gamma_\mathrm{lim} = 0.02$ in a realistically observable redshift range ($z\lesssim 1$) is non-negligible. Given this, we have created a lensing frequency algorithm which searches an input dataset for all lens-source pairs with an estimated shear greater than a limiting value of $\gamma_\mathrm{lim}$. Our algorithm has been applied to a dataset extracted from the GAMA survey catalogue, and the number of objects with an estimated shear of at least $\gamma = 0.02$ in the sample was found to be $\sim 393$. These targets can be matched to objects in the SDSS DR10 Catalogue, and a subsample of good targets can be chosen from this selection for follow up observations.\
A scatter in the $M_*-M_h$ relation results in a shift towards higher measured shears for a given population of galaxies. Given this, we have investigated the feasibility of measuring the scatter in the $M_*-M_h$ relation using shear statistics. We find that for a given shear measurement error, our ability to differentiate between a distribution of shears from a ‘tight’ $M_*-M_h$ relation, and one with scatter, is dependent on the size of the scatter, the number of objects in the sample, and the shear measurement error. For a scatter of $0.3$ dex in the $M_*-M_h$ relation, we find that a sample size of $\sim$50,000 galaxies would be needed to measure the scatter, for a measurement error on the shear of $0.02$ (a value consistent with the shear measurement accuracy achievable with DSM). We attempt to fit for the scatter in the $M_*-M_h$ relation for a set of simulated datasets. The result of this is shown in Figure \[fig:stand\_err\_plot\]. For a true scatter of $0.3$, a shear measurement error of $0.02$ and $\sim$50,000 shear measurements, we recover a scatter of 0.308$\pm$0.02. It should be noted that the technique we have demonstrated here is based on a relatively untested algorithm, however it is adequate as an illustrative example and our results are promising for future measurements. While there are no existing IFU survey catalogues of a sufficient size to apply this technique, there are several surveys beginning in the near to intermediate future which will provide datasets of a sufficient size to perform this experiment, for example HETDEX which aims to observe $\sim 10^6$ galaxies, or surveys on the SKA, such as the ‘billion galaxy survey’ which aims to observe $\sim 10^9$ galaxies.*
We thank the referee for their constructive feedback. Their input into this publication has greatly improved it. This research was conducted as part of the Australian Research Council Centre for Excellence for All-Sky Astrophysics (CAASTRO), through project number CE110001020. We acknowledge financial support from The University of Melbourne, and the Australian Astronomical Observatory (AAO). Special thanks goes to the members of the GAMA survey team for their advice, input and helpful discussions. GAMA is a joint European-Australasian project based around a spectroscopic campaign using the Anglo-Australian Telescope. The GAMA input catalogue is based on data taken from the Sloan Digital Sky Survey and the UKIRT Infrared Deep Sky Survey. Complementary imaging of the GAMA regions is being obtained by a number of independent survey programs including GALEX MIS, VST KiDS, VISTA VIKING, WISE, Herschel-ATLAS, GMRT and ASKAP providing UV to radio coverage. GAMA is funded by the STFC (UK), the ARC (Australia), the AAO, and the participating institutions. The GAMA website is <http://www.gama-survey.org/>. Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is <http://www.sdss3.org/>.
\[lastpage\]
[^1]: cdbd@student.unimelb.edu.au$\,$(CDBD); ent@ph.unimelb.edu.au$\,$(ENT); r.webster@unimelb.edu.au$\,$(RLW); ahopkins@aao.gov.au$\,$(AMH)
[^2]: The purpose of only including galaxies with $\theta_\mathrm{sep} < \theta_\mathrm{max}$ in this step is to improve computation time, otherwise every object would be compared to every other higher redshift object. Background objects with a large separation from the source are likely to have negligible shears, and so can be safely excluded.
[^3]: http://www.gama-survey.org
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'R. K. Saito'
- 'D. Minniti'
- 'R. Angeloni'
- 'M. Catelan'
- 'J. C. Beamin'
- 'J. Borissova'
- 'I. Dékány'
- 'E. Kerins'
- 'R. Kurtev'
- 'R. E. Mennickent'
date: 'Received ; Accepted '
title: 'A near-infrared catalogue of the Galactic novae in the VVV survey area[^1]'
---
[Near-infrared data of Classical Novae contain useful information about the ejected gas mass and the thermal emission by dust formed during eruption, and provide independent methods to classify the objects according to the colour of their progenitors, and the fading rate and features seen after eruption. The VISTA Variables in the Vía Láctea survey (VVV) is a near-IR ESO Public Survey mapping the Milky Way bulge and southern plane. Data taken during $2010 -
2011$ covered the entire area in the $JHK_{\rm s}$ bands plus some epochs in $K_{\rm s}$-band of the ongoing VVV variability campaign.]{} [We used the VVV data to create a near-IR catalogue of the known Galactic novae in the 562 sq. deg. area covered by VVV. We also compiled the information about novae from the variability tables of the VVV variability campaign.]{} [We used the novae list provided by VSX/AAVSO catalogue to search for all objects within the VVV area. From the 140 novae, we were able to retrieve the $JHK_{\rm s}$ colours of 93 objects. We also checked in the ongoing VVV variability campaign for the light-curves of novae that erupted in the last years.]{} [ The VVV near-IR catalogue of novae contains $JHK_{\rm s}$ photometry of 93 objects completed as of December 2012. VVV allows to monitor objects within up to $\Delta K_{\rm s}\sim10$ mag range. VVV images can also be used to discover and study novae by searching for the expanding shell. Since objects are seen at different distances and reddening levels, the colour-magnitude and colour-colour diagrams show the novae spread in magnitude as well as in colour. Dereddened colours and reddening-free indices were used with caution and cannot be a good approach in all cases since the distance and spectral features prevent more conclusive results for some extreme objects. Light-curves for some recent novae are presented.]{} [Thanks to its high spatial resolution in the near-IR, and large $K_{\rm s}$-range, the VVV survey can be a major contributor for the search and study of novae in the most crowded and high-extinction regions of the Milky Way. The VVV survey area contains $\sim35\%$ of all known novae in the Galaxy.]{}
Introduction {#intro}
============
Cataclysmic variable stars (CVs) are close binary systems composed by a late-type star which transfers material to a more massive white dwarf (WD) companion. The typical mass transfer rates in CVs are $10^{-8} - 10^{-11}$ M$_{\odot}$ yr$^{-1}$, and the progressive accretion of hydrogen-rich material from the secondary star onto the hot surface of the WD can lead to a thermonuclear runaway, in which the accreted material is expelled from the system. This phenomenon is a so called nova eruption [e.g., @2002apa..book.....F; @2003cvs..book.....W].
During a nova eruption the object suddenly rises in brightness, becoming typically 8-15 magnitudes brighter than its progenitor. The recurrence between nova outbursts is expected to be $10^3 - 10^6$ years, and depends on the mass of the white dwarf and the mass transfer rate. The systems are classified as Classical Novae when just one nova eruption was recorded. However, this is strongly biased by ancient records, which are limited to a few thousand years. Systems with two or more registered nova eruptions are classified as Recurrent Novae, in general associated with a high-mass transfer rate system onto a primary star with the mass close to the Chandrasekhar limit [@2003cvs..book.....W].
Since novae are usually found in eruption, the novae classification is in general based on the features seen during this phase, such as the fading (speed class) rate after the nova eruption [e.g., @1957gano.book.....G] and the spectral differences [e.g., @1992AJ....104..725W].
More recently, alternative classifications are proposed based either on the Galactic component, with disk and bulge novae [e.g., @1998ApJ...506..818D], or using automated selection and classification based on optical photometry [@2004MNRAS.353..571D].
The infrared light can also contain the thermal emission by dust formed during the shell ejection. Models for thermonuclear runway predict the presence of material from the white dwarf in the ejecta, thus its spectroscopic analysis allows to distinguish between CO and ONe white dwarfs [@2011ApJ...727...50S]. The importance of observing Galactic novae in the IR spectral region was fully understood when it was realized that they can represent a significant contributor to the interestellar medium (at least on local scales) of highly processed material, thus playing an active role in the chemical evolution of the Galaxy. Furthermore, IR observations provide independent methods for determining the ejected gas mass, a key parameter of every thermonuclear runaway model [@1998ApJ...494..783M].
IR data can also help in the determination both of the global Galactic nova rate and the separate Galactic bulge and disk nova rates. Current Galactic nova samples, obtained mostly from optical surveys, are biased towards regions of lower optical extinction and therefore are incomplete within both the bulge and disk towards the Galactic plane. Previous determinations of the Galactic nova rate therefore vary by more than an order of magnitude, from 11 to 260 yr$^{-1}$ [@1997ApJ...487..226S]. The nova rate in M31 is better constrained from optical studies [e.g., @2006MNRAS.369..257D] but it remains unclear whether the rates within both galaxies are consistent with fixed bulge and disk nova rates per unit galactic luminosity. Again this uncertainty is due mostly to the poorly constrained knowledge of the Galactic rate. A large scale near-infrared time domain survey can allow novae to be identified throughout the Galactic bulge and disk and therefore facilitate a more complete determination of the nova rate within each stellar population.
The VISTA Variables in the Vía Láctea is a ESO Public near-IR survey scanning the Milky Way (MW) bulge and southern plane, in five near-IR bands ($ZYJHK_{\rm s}$), plus a variability campaign of about 100 epochs in the $K_{\rm s}$-band spanning over many years [@2010NewA...15..433M]. VVV is about four magnitudes deeper than previous IR surveys, and thanks to its higher spatial resolution in the near-IR, enables deep observations in the most crowded and high-extinction regions of the Milky Way .
Here we present a catalogue with the $JHK_{\rm s}$ colours of 93 novae in the VVV area. We discuss the colour properties and features seen in the near-IR. We also present the first results from the VVV variability campaign, including light-curves for some novae. VVV can be a major contributor for the discovery and study of novae in the innermost regions of the Galaxy, objects that are beyond detection in the current novae searches.
Observations {#obs}
============
VVV is an ESO Public Survey observing the MW bulge and inner southern plane in the near-IR with the VISTA Telescope located at Cerro Paranal Observatory in Chile [@2010NewA...15..433M and references therein]. VVV observes an area of about 562 deg$^2$, within $-10.0^\circ \lesssim l \lesssim +10.5^\circ$ and $-10.3^\circ \lesssim b \lesssim +5.1^\circ$ in the bulge, and within $294.7^\circ \lesssim l \lesssim 350.0^\circ$ and $-2.25^\circ
\lesssim b \lesssim +2.25^\circ$ in the plane (see Fig. \[fig:area\]). The whole bulge and disk areas were fully observed in five near-IR passbands ($ZYJHK_{\rm s}$) during 2010–2011, while a variability campaign in the $K_{\rm s}-$band is ongoing, with $\sim$100 pointings planned over many years .
Photometric catalogues for the VVV observations are provided by the Cambridge Astronomical Survey Unit (CASU)[^2]. The catalogues contain the positions, fluxes, and some shape measurements obtained from different apertures, with a flag indicating the most probable morphological classification. The VVV data are in the natural VISTA Vegamag system, with the photometric calibration in $JHK_{\rm s}$ performed using the VISTA magnitudes of unsaturated 2MASS stars present in the images. The single-band CASU catalogues for each tile were matched, creating a $JHK_{\rm s}$ catalogue of more than 173 million sources in the VVV bulge area , and about 148 million sources in the disk. A similar procedure was adopted to match the $K_{\rm
s}$-band catalogues from the variability campaign in order to create variability tables for selected fields.
Different observing strategies between the bulge and disk area, and effects caused by crowding and extinction cause both the limiting magnitude and saturation to vary along the VVV area . For instance, the $K_{\rm s}$-band limiting magnitude on the outer bulge is about 18.5 mag, while in the Galactic center the limit is $K_{\rm
s}\sim16.5$ mag (see Fig. \[fig:maglim\]). The saturation limit varies in a shorter range. The $JHK_{\rm s}$ observations for a given field were taken in the same “observational block” (OB) which guarantees quasi-simultaneous observations in the three bands (there is a time gap of only $\sim190$ s between each band). On the other hand, due to multiple scheduling constraints on the side of the VISTA telescope operations, there is a little control on the cadence of the $K_{\rm s}$-band variability data within a given season, with the number of epochs planned in a given year being observed together with the other VISTA surveys according with the surveys requirements on the weather and visibility, for instance.
The VVV Novae Catalogue {#sec:cat}
=======================
Our novae list was taken from The International Variable Star Index (VSX), provided by the American Association of Variable Star Observers[^3] (AAVSO). In this catalogue there are about 400 known novae in the Galaxy, 140 of them lie within the VVV area. Fig. \[fig:area\] shows their spatial distribution, with most objects concentrated towards the Galactic bulge. We note that the information about some novae are from historical records or amateur astronomers, with the coordinates in general taken from observations while the objects were in eruption. Thus, in order to certify the coordinates with enough accuracy during the quiescent state, we checked for all objects in the SIMBAD database[^4]. Finally, we defined as our coordinates the ICRS J2000 RA and DEC taken from SIMBAD, which were used as entries in our search for the VVV $JHK_{\rm
s}$ photometry and composite colour images.
We defined as a valid match all cases where a source was found within 1 arcsec from the position given by the catalogue. This value is agreement with the median values of the VVV image quality in the first season ($0\farcs89-0\farcs87$, respectively to $J$ and $K_{\rm s}$ observations) and the typical astrometric accuracy of the photometric catalogues . A complementary visual inspection of the images was performed for all novae candidates, in order to certify the presence (or absence) of a target in the given position. In some cases the source was rejected because of the presence of a faint object in a closer position to the coordinate, seen during the visual inspection of the images, but beyond the detection limit of the CASU aperture photometry, used to build the VVV photometric catalogues (see Section \[obs\]). In other cases multiple sources are seen in a similar close distance from the entry coordinates. In this procedure we secured the information for 55 novae within $\leq1~$arcsec from the catalogue position, being 27 of them within the VIRCAM pixel scale of 0$\farcs$34 pixel$^{-1}$. A complementary search included also objects matching the position given by the catalogue within 2 arcsec. By this procedure we retrieved the photometry of another 37 objects. These are all fainter than $K_{\rm
s}\sim13$ mag and were classified as doubtful sources. We note that due to the source confusion in the most crowded regions of Galactic plane and bulge, a search based only on the coordinates does not guarantee that in all cases the source found in a closer position to the coordinates is in fact the nova remnant. Thus these should be seen as candidate near-IR counterparts for the novae remnant.
Therefore, from the 140 novae within the VVV area, we are able to provide the $JHK_{\rm s}$ colours of 93 objects: 55 classified as valid matches (27 within the pixel scale), and 37 as doubtful sources. The other objects are beyond our detection limit (progenitors in quiescence with $K_{\rm s}\gtrsim18$ mag), or with coordinates not sufficiently accurate to allow us to identify the target in the field. Table \[table:novae\] presents the $JHK_{\rm s}$ photometry of all novae detected in the VVV area while in the Appendix we provide the composite $JHK_{\rm s}$ finding charts for all objects.
Since the VVV $JHK_{\rm s}$ observations were performed during 2010–2011, we are also able to provide the information about the progenitors of the most recent novae, as well as the fading curve after the nova eruption (see Section \[sec:var\]). Interestingly, the observations of Nova Sgr 2010b were secured while the object was in eruption [@2012ATel.4353....1S]. Complementary information, including the timings of the VVV colour and variability campaign observations for all novae erupted in this century, is presented in Table \[table:time\].
[l l l l l l]{} Nova & Other & Eruption & VVV $JHK_{\rm s}$ & First $K_{\rm s}$-band epoch & Note\
& Designation & Date & Observation & (variability campaign) &\
Sco 2001 & V1178 Sco & 2001 & 2010 08 10 & 2010 09 12 & Remnant\
Sgr 2001 & V4643 Sgr & 2001 & 2010 04 09 & 2011 08 06 & Remnant\
Sgr 2001b & V4739 Sgr & 2001 & 2010 08 14 & 2010 10 15 & Remnant\
Sgr 2001c & V4740 Sgr & 2001 & 2010 08 15 & 2011 08 05 & Remnant\
Sgr 2002 & V4741 Sgr & 2002 & 2010 08 10 & 2010 09 12 & Remnant\
Sgr 2002b & V4742 Sgr & 2002 & 2010 03 28 & 2010 03 30 & Remnant\
Sgr 2002d & V4744 Sgr & 2002 & 2010 03 18 & 2010 09 29 & Remnant\
Sgr 2003b & V5113 Sgr & 2003 & 2010 03 28 & 2010 08 26 & Remnant\
Oph 2004 & V2574 Oph & 2004 & 2010 03 30 & 2010 04 10 & Remnant\
Sco 2004 & V1186 Sco & 2004 & 2010 08 03 & 2010 08 26 & Remnant\
Sco 2004b & V1187 Sco & 2004 & 2010 08 03 & 2010 08 26 & Remnant\
Sgr 2004 & V5114 Sgr & 2004 & 2010 08 15 & 2010 10 24 & Remnant\
Cen 2005 & V1047 Cen & 2005 & 2010 03 07 & 2010 03 29 & Remnant\
Nor 2005 & V0382 Nor & 2005 & 2010 03 06 & 2010 03 06 & Remnant\
Sco 2005 & V1188 Sco & 2005 & 2010 10 01 & 2011 07 27 & Remnant\
Sgr 2005a & V5115 Sgr & 2005 & 2010 03 28 & 2010 03 30 & Remnant\
Sgr 2005b & V5116 Sgr & 2005 & 2010 08 15 & 2010 10 15 & Remnant\
Oph 2006 & V2575 Oph & 2006 & 2010 03 30 & 2010 04 09 & Remnant\
Sgr 2006 & V5117 Sgr & 2006 & 2010 04 23 & 2010 10 25 & Remnant\
Oph 2007 & V2615 Oph & 2007 & 2010 03 17 & 2011 06 12 & Remnant\
Nor 2007 & V0390 Nor & 2007 & 2010 03 16 & 2010 08 15 & Remnant\
Oph 2008a & V2670 Oph & 2008 & 2010 03 17 & 2011 06 12 & Remnant\
Oph 2008b & V2671 Oph & 2008 & 2011 05 07 & 2011 06 13 & Remnant\
Sgr 2008 & V5579 Sgr & 2008 & 2010 04 12 & 2011 08 05 & Remnant\
Sgr 2008b & V5580 Sgr & 2008 & 2010 04 08 & 2010 08 27 & Remnant\
Oph 2009 & V2672 Oph & 2009 & 2010 04 13 & 2011 06 13 & Remnant\
Sgr 2009a & V5581 Sgr & 2009 & 2010 03 30 & 2010 03 30 & Remnant\
Sgr 2009b & V5582 Sgr & 2009 & 2010 04 10 & 2011 06 14 & Remnant\
Sgr 2009c & V5583 Sgr & 2009 & 2010 04 23 & 2010 08 18 & Remnant\
Oph 2010 & V2673 Oph & 2010 01 15 $^{a}$ & 2010 04 10 & 2011 06 14 & Remnant\
Sgr 2010 & V5585 Sgr & 2010 01 20 $^{b}$ & 2010 04 23 & 2010 08 26 & Remnant\
Oph 2010b & V2674 Oph & 2010 02 18 $^{c}$ & 2011 05 09 & 2011 07 26 & Remnant\
Sgr 2010b & V5586 Sgr & 2010 04 23.782 $^{d}$ & 2010 04 23.2457 ($H$) & 2010 09 12 & During eruption\
& & & 2010 04 23.2479 ($K_{\rm s}$) & &\
& & & 2010 04 23.2501 ($J$) & &\
Sgr 2011 & V5587 Sgr & 2011 01 25 $^{e}$ & 2010 03 18 & 2010 09 29 & Progenitor\
Sgr 2011b & V5588 Sgr & 2011 04 07 $^{f}$ & 2010 04 21 & 2010 10 06 & Progenitor\
Cen 2012b & TCP J14250600-5845360 & 2012 04 05 $^{g}$ & 2010 03 27 & 2010 03 09 & Progenitor\
Sgr 2012 & PNV J17452791-2305213 & 2012 04 21 $^{h}$ & 2010 03 18 & 2010 09 29 & Progenitor\
Oph 2012b & PNV J17395600-2447420 & 2012 05 19 $^{i}$ & 2010 04 13 & 2011 06 12 & Progenitor\
Sco 2012 & MOA 2012 BLG-320 & 2012 05 22 $^{j}$ & 2010 08 29 & 2010 09 25 & Progenitor\
Sgr 2012c & PNV J17522579-2126215 & 2012 06 26 $^{k}$ & 2010 03 25 & 2010 09 29 & Progenitor\
Sgr 2012d & PNV J18202726-2744263 & 2012 07 12 $^{l}$ & 2010 04 08 & 2010 08 27 & Progenitor\
& OGLE-2012-NOVA-01 & 2012 05 02 $^{m}$ & 2010 03 28 & 2010 08 26 & Progenitor\
\[table:time\]
The near-IR Colours {#sec:col}
===================
Figure \[fig:cmd\] shows in the top panels the $K_{\rm
s}~vs.~(J-K_{\rm s})$ colour-magnitude diagram (CMD) and the $(J-H)~vs.~(H-K_{\rm s})$ colour-colour diagram (CCD) for all novae in the VVV area. We used different symbols to denote the objects classified as valid matches and the doubtful ones, as described in the previous Section. The reddening vector associated with an extinction of $E(B-V)=1$, based on the relative extinctions of the VISTA filters, and assuming the [@1989ApJ...345..245C] extinction law, is also shown.
The Galactic novae range in magnitude and colour, with all objects appearing red, with $(J-K_{\rm s})>0$ (the only exception is Nova Cir 1906, with $(J-K_{\rm s})=-0.228$). Interestingly, a large fraction of novae are beyond detection in previous near-IR surveys , thus we are now reporting their IR-colours for the first time in an homogeneous data set. We note that the novae are spread across the whole VVV area (see Fig. \[fig:area\]), and therefore are affected by different extinction. For instance, in the bulge area where most novae are located, reddening spans from $E(B-V)\lesssim0.2$ at $b\sim-10^{\circ}$ to $E(B-V)\sim10$ closer to the Galactic centre .
In the $(J-H)~vs.~(H-K_{\rm s})$ CCD the objects are seen more concentrated at low values of $(H-K_{\rm s})$, but spread along $(J-H)$ by more than 1 mag. The colours of main sequence stars [with spectral type from B0 to M4, adapted from @2001ApJ...558..309D] are presented in the CMD in the case of $A_{Ks}=0$ mag (no extinction) an applying the median value of extinction among our targets of $A_{Ks}=0.243$ mag (corresponding to $E(B-V)\sim0.67$ mag). The comparison reveals that part of our sources coincide in colour with the main sequence stars and therefore suggests that some of our novae candidates can be actually old field stars, which match a position close to the novae coordinates, and suffer from reddening at different levels. Individual values of $A_{Ks}$ for all bulge novae are listed in Table 1. Although the [@2001ApJ...558..309D] data are in the $JHK$ system, it is sufficiently close to the VISTA system that is useful for the comparison. Four objects are seen with the most extreme colours: Nova Sgr 2001 and Nova Sgr 2005b, with the largest $(J-H)$ but low $(H-K_{\rm s})$; and Nova Sgr 2008 and Nova Sgr 2009b, showing the largest $(H-K_{\rm s})$ values at relatively low values of $(J-H)$. The complementary $JHK_{\rm s}$ spectral energy distributions (SED) for these selected objects are shown in Fig. \[fig:sed\]. While Nova Sgr 2001 and Nova Sgr 2005b show a maximum close to the $H$-band, the SED of Nova Sgr 2008 and Nova Sgr 2009b reveal a monotonic increase in flux towards longer wavelengths.
In order to minimize the reddening effects we calculated the dereddened $J_{\rm 0}$, $H_{\rm 0}$, and $K_{\rm s\,0}$, colours for the bulge novae, using the maps of , the relative extinctions of the VISTA filters, and assuming the [@1989ApJ...345..245C] extinction law. A dereddened $K_{\rm
s\,0}~vs.~(J-K_{\rm s})_0$ CMD is presented in the bottom-left panel of Fig. \[fig:cmd\]. Since the maps do not cover the disk area we prefer to exclude from the plot the few novae located in this region, instead of using a different prescription for the extinction in the disk, which does not guarantee a good agreement with the procedure used in the bulge area . We also note that the reddening law is known to change close to in the Galactic center [e.g., @2009ApJ...696.1407N], but in our sample just one object (Nova Sgr 2001) belongs to this region.
A set of reddening-free indices provided by [@2011rrls.conf..145C], based on the extinction law of [@1985ApJ...288..618R] for the VISTA filters was also computed. The $m_4$ pseudo-magnitude and the $c_3$ pseudo-colour are defined as
$$m_4\equiv K_{\rm s}-1.22(J-H),$$
$$c_3\equiv (J-H)-1.47(H-K_{\rm s}).$$
The $m_4~vs.~c_3$ CMD is shown in the bottom-right panel of Fig. \[fig:cmd\]. We note the indices shown in the equations depend on the effective wavelengths of each filter, assumed to be a flat distribution here, which is certainly different from the near-IR spectra of novae. Refined indices based on actual near-IR spectra covering different luminosity classes and spectral types are currently being computed, and quantitative tests in the case of actual near-IR spectra of novae will also be carried out in the future.
One important aspect seen in all panels of Fig. \[fig:cmd\] is the absence of a clear-cut correlations or trends, even using dereddened colours and reddening-free indices, the resulting distributions do not appear to have straightforward interpretations. The differences in colour arise mostly because ($i$) the objects are seen at different distances and positions in the Galaxy, and thus can suffer from extinction and reddening at different levels, compared to the 2D reddening maps of and ($ii$) the spectral energy distribution in the near-IR can be affected by emission lines and thermal emission by a dust shell produced during eruption. The latter are time-dependent and valid mostly for recent novae. Depending on the evolution of the dust shells created in the outbursts, a variety of emission lines are seen in early post-outburst near-IR spectra. Moreover, for older novae the peak emission from the dust shells is shifted to longer wavelengths [e.g., @2002AJ....124.3009V; @2012BASI...40..243B].
Four objects seen with the extreme colours on the plots of Fig. \[fig:cmd\] are discussed in more detail in the following, namely Nova Sgr 2001, Nova Sgr 2005b, Nova Sgr 2008 and Nova Sgr 2009b.
[@2006MNRAS.368..592A] classified Nova Sgr 2001 as a “NA” fast novae, with absence of dust formation after the eruption. Its near-IR spectrum evolves significantly from the early decline stage (March 2001) to the coronal phase [August 2001, @2006MNRAS.368..592A]. Nova Sgr 2001 is the innermost known nova in the Galaxy, with coordinates $(l,b)=(3.345, -0.337)$ deg, with an extinction of $A_{Ks}=1.88$ mag according to the maps of using [@2009ApJ...696.1407N] extinction law, equivalent to $A_{V}\simeq15.9$ mag.
Our data taken during the quiescent phase (on April 2010) show the remnant of Nova Sgr 2001 as a relatively blue object, with the maximum intensity in the near-IR around the$H$-band (see Fig. \[fig:sed\]). However, data taken with UKIRT on March 2012 show Nova Sgr 2001 with a flatter SED in the near-IR, with the maximum intensity towards longer wavelengths [@2012ATel.4405....1V].
[@2006MNRAS.368..592A] estimated a distance to the object of $d=3.3$ kpc. Thus, since our calculations using maps assume the total extinction in the line of sight, it could not be a good approach to estimate the reddening for nearby objects, overestimating the corrections. In this case, three-dimensional extinction maps are necessary, in order to estimate the relative extinction at given distance . On the other hand, the $c_3$ colour seems to be dominated by the $H$-band (by construction, see Equation 2), thus the presence of features such as emission lines or simply the slope of the SED contribute to shift the value of the $c_3$ pseudo-color from a relatively small to a relatively large value, depending on whether the emission peaks around the $H$-band or not. This explains why Nova Sgr 2001 and Nova Sgr 2005b, whose spectra do peak around $H$ (Fig. 4), have large [*positive*]{} $c_3$ values, whereas the remaining novae, whose spectra monotonically increase towards longer wavelengths, have – on the contrary – large [*negative*]{} $c_3$ values.
Nova Sgr 2005b is similar in color with Nova Sgr 2001. Its remnant is a Supersoft X-ray variable source, presenting an orbital period of $P_{orb}=2.97$ hr , which puts the object on the edge of the $2-3$ hr CVs period gap [e.g., @2003cvs..book.....W]. The object is located beyond the bulge, with $(l,b)=(2.136, -6.832)$ deg and a distance of $d=11\pm3$ kpc estimated by [@2008ApJ...675L..93S]. Nova Sgr 2005b lies in the lower part of the CMDs, with $K_{\rm s}=16.163$ and $K_{\rm s,0}=16.079$. The presence of such distant objects in our data demonstrates our capability to monitor even the most distant novae in the Galaxy.
In the other extreme are Nova Sgr 2008 and 2009b. Nova Sgr 2008 was discovered in eruption on 2008 April 18 [@2008CBET.1342....1N]. It is a nearby fast Fe II nova, located at $(l,b)=(3.734,-3.022)$ deg at distance of $d=4.4\pm0.2$ kpc [@2008IAUC.8948....1R; @2011MNRAS.415.3455R]. The very red colors even in $(J-K_{\rm s})_0$ are related to the presence of dust. Spectra taken in the late stages after the eruption indicate dust formation during the nova remnant’s development [@2011MNRAS.415.3455R]. Nova Sgr 2009b [@2009IAUC.9049....1S] is located in a low-extinction region with $(l,b)=(7.531,4.719)$ deg and $A_{K_s}=0.1845$ mag. There are no entries in the literature estimating its distance or classifying the nova remnant. However, by comparing their colours, one can infer that Nova Sgr 2009b was similar to the Nova Sgr 2008 remnant. In this case, Nova Sgr 2009b could be surrounded by dust, expelled from the system after the eruption.
The VVV variability campaign on Novae {#sec:var}
=====================================
The IR photometric monitoring of several novae outbursts has allowed already in the 1990’s to distinguish between two fundamentally different types of objects, usually referred to as CO and ONe novae. While the first seems to arise from thermonuclear runaway on a relatively low-mass WD and tends to form considerable amounts of dust at later stages, the latter turn out to be associated with massive ($\geq$1.2 $M_{\sun}$) WD that have outbursts which exhibit a coronal line emission phase with little or no dust production.
Particularly useful in identifying extreme CO and ONe novae are the $K_{\rm s}$-band light curves: because of the thermal emission from dust during the condensation phase, CO Novae reach an emission peak within a few months from the optical maximum, declining exponentially thereafter as consequence of the decreasing density of the shell owing to expansion [emblematic is the case of NQ Vul; @1978PASJ...30..419S]. ONe novae, on the contrary, decline rapidly shortly after the optical maximum and then have a positive change in the slope rate [e.g., V1974 Cyg; @1997ApJ...477..817W] because of the increasing contribution of several IR forbidden emission lines (e.g., \[Ca VIII\] 2.32 $\mu$m, \[Si VII\] 2.47 $\mu$m), whose intensity strengthens during the coronal phase acting thus as effective coolant of the ejecta [e.g., as in QU Vul; @1988AJ.....95..172G].
The VVV Survey, for the first time, will provide well-sampled light curves in the $K_{\rm s}$-band for about $100$ Galactic novae that exploded in the last $\sim$150 years in the Galactic bulge and in the southern part of the Galactic disk. At the time of writing, the VVV variability campaign is ongoing, and in the next few years until the completion of the survey (2016 or later), we would be able to finally build a coherent picture of nova IR behavior, at quiescence, as well as to discover new novae in highly-extincted regions that are missed in the optical.
There are about 400 known novae in the Milky Way, but the comparison with the nova rate in nearby galaxies shows that a number of novae are lost every year in the Galaxy. While the historical records show a rate below a dozen nova discovered per year in the MW (see Fig. \[fig:year\]), several results point to a nova rate spanning between $\sim30-40$ yr$^{-1}$, or even higher reaching up to $\sim260$ yr$^{-1}$ .
Not coincidentally, the spatial distribution of novae in the bulge and southern plane shows a “zone of avoidance”(see Fig. \[fig:area\]), with just a few objects belonging to the high-extinction regions of the Galaxy. This arises because the current searches for novae are performed with small-aperture telescopes in the optical. The VVV Survey combines deep and high-spatial resolution observations in the near-infrared, allowing one to search for novae even in the most crowded and high extinction regions of our Galaxy. Very illustrative is the recent discovery of VVV-NOV-001, the first Nova candidate from the VVV data, located in a high extinction region in the inner bulge with $(l,b)=(8.897,-0.158)$ deg and $A_V\sim 11$ mag [@2012ATel.4426....1S].
Fig. \[fig:maglim\] shows the magnitude range covered by VVV at different levels of crowding in the bulge. Even in the most problematic case which is close to the Galactic center, the VVV observations allow us to monitor over a range in magnitudes spanning from $\Delta K_{\rm s}>7$ mag, reaching even $\Delta K_{\rm s}\sim10$ mag when using data taken below 5-$\sigma$ accuracy and saturated sources. As of this writing, we are able to provide light curves with a few points for some objects, which demonstrates our capability to monitor novae at all phases depending on the magnitude in each case. Figs. \[fig:n2010\], \[fig:n2011\] and \[fig:n2012\] show VVV light curves of Nova Sgr 2010b [@2012ATel.4353....1S], 2011 and 2012 [@2012ATel.4372....1S], respectively.
These $K_{\rm s}$-band light curves at quiescence will allow us to study any kind of orbital effects which could be related to the influence on the secondary star by the accreting WD, and when combined with the IR colors will also help quantify (via modeling of ellipsoidal variations of the secondary) fundamental parameters of the nova system, such as spectral type of the secondary, mass ratios, orbital periods and inclinations [@1998ApJ...494..783M].
Perspectives on the search for novae using the VVV images {#sec:ima}
=========================================================
There is a well-known discrepancy on the actual amount of mass ejected in a CN outburst as predicted by theory and numerical simulations on one side, and as derived from infrared and radio observations on the other side - being the latter systematically higher by up to a factor $\sim$10 [@2002AIPC..637..259S; @2012MmSAI..83..792S].
Even if there are strong caveats on the methods and assumptions used for determining the mass of the ejecta [e.g., @2010ApJ...712L.143S], the analysis of the most recent observational material seems in fact to suggest for the ejecta of classical novae a lower mass and higher velocities [$10^{-4}-10^{-5}$M$_{\sun}$ and up to 6000 km s$^{-1}$, respectively, as in the extreme case of LMC1991 – @2001MNRAS.320..103S] than thought before. Identifying where such a discrepancy stems from would have dramatic consequences on our understanding of the nova phenomenon, e.g., by clarifying its role in the chemical evolution of the galaxy.
By supposing expansion velocities of the nova remnants larger than 1000 km s$^{-1}$, it is reasonable to expect that at least the closest/fastest novae produce nebular remnants that eventually may be spatially resolved even from ground-based telescopes. Imaging such remnants allows the investigation of fundamental correlations between the properties of the central binary, the evolution of the outburst and the ejecta shaping mechanism, and also provides a direct measure of the distance to the nova by combining the angular expansion rate with spectroscopically derived expansion velocities, i.e., the so-called *expansion parallax method* [@1940PASP...52..386B; @1989clno.conf...73M].
While the infrared imagery has several advantages over the optical one (e.g., the extinction – which is often poorly known for novae – is conveniently reduced), up to now only a handful of nova remnants have been searched (and even fewer spatially resolved) in this spectral region, compared to more than 40 nova remnants resolved in the optical and $\sim$10 in the radio.
Thanks to the large sky area covered and to the higher spatial resolution than previous IR surveys (such as 2MASS and DENIS), VVV is in unique position to perform for the first time a systematic search for remnants of classical novae that exploded in the last decades within the inner, highly-extincted regions of the MW. Even assuming a representative distance to a nova of 4 kpc (about half-way to the Galactic center) and a very conservative expansion velocity of 1000 km s$^{-1}$, VVV is in fact potentially able to spatially resolve the corresponding nebular remnants after $\sim$40 years from the outburst, taking into account the VIRCAM pixel scale (0$\farcs$34 pixel$^{-1}$) and a nebula angular size of 2". No resolved novae shells have yet been detected in the VVV imaging data. In a forthcoming phase of the survey a search for novae using imaging will be conducted on the deep, stacked $K_{\rm s}$-band images, using more focused image analysis such as point spread function (PSF) subtraction or difference image analysis .
Conclusions {#conclu}
===========
We presented a near-IR catalogue of novae within the VVV area covering the MW bulge and southern Galactic plane area. From the 140 objects found in the VSX/AAVSO catalogue, we reported the $JHK_{\rm s}$ colours of 93 objects. The rest were beyond detection or heavily blended sources. Colour-magnitude and colour-colour diagrams were presented, as well as CMDs using dereddened colours and reddening-free indices. These should be used with caution since in the case of nearby objects the dereddened colours can overestimate the corrections, by assuming the total extinction in the line of sight. Likewise, reddening-free near-IR indices specifically devised for the study of novae would also be of considerable interest, given their different spectral shapes compared with those typically assumed in the computation of these indices.
Thanks to its higher spatial resolution in the near-IR, and large $K_{\rm s}$-range covered by the observations, the VVV survey can be a major contributor for the search and study of novae in the Galaxy, mainly in the most crowded and high-extinction regions of the MW, beyond the capabilities of the current searches for novae in optical wavelengths.
VVV can produce well-sampled light curves covering many years for Galactic novae belonging both to the bulge and the southern part of the disk, even for objects during or fading after eruption. The recent report of VVV-NOV-001, the first nova candidate from the VVV data discovered in the inner bulge, is an illustrative example. Surely, the first of many more to come. The possibility to search for novae using the VVV imaging is accordingly a very promising path to unveiling the heretofore hidden population of heavily obscured novae.
We gratefully acknowledge use of data from the ESO Public Survey programme ID 179.B-2002 taken with the VISTA telescope, data products from the Cambridge Astronomical Survey Unit, and funding from the FONDAP Center for Astrophysics 15010003, the BASAL CATA Center for Astrophysics and Associated Technologies PFB-06, the FONDECYT from CONICYT, and the Ministry for the Economy, Development, and Tourism’s Programa Iniciativa Científica Milenio through grant P07-021-F, awarded to The Milky Way Millennium Nucleus. Support for R.A. is provided by Proyecto GEMINI CONICYT 32100022 and via a Postdoctoral Fellowship by the School of Engineering at Pontificia Universidad Católica de Chile. M.C. and I.D. acknowledge funding from Proyecto FONDECYT Regular 1110326. J.B. acknowledges funding from Proyecto FONDECYT Regular 1120601. R.K. acknowledges partial support from FONDECYT through grant n. 1130140. This research has made use of the International Variable Star Index (VSX) database, operated at AAVSO, Cambridge, Massachusetts, USA.
[aa]{}
Ashok, N. M., Banerjee, D. P. K., Varricatt, W. P., & Kamath, U. S. 2006, , 368, 592
Baade, W. 1940, PASP, 52, 386
Banerjee, D. P. K., & Ashok, N. M. 2012, Bulletin of the Astronomical Society of India, 40, 243
Cardelli, J. A., Clayton, G. C., & Mathis, J. S. 1989, , 345, 245
Catelan, M., Minniti, D., Lucas, P. W., et al. 2011, RR Lyrae Stars, Metal-Poor Stars, and the Galaxy, 145
Chen, B. Q., Schultheis, M., Jiang, B. W., et al. 2013, , 550, A42
Darnley, M. J., Bode, M. F., Kerins, E., et al. 2004, , 353, 571
Darnley, M. J., Bode, M. F., Kerins, E., et al. 2006, , 369, 257
della Valle, M., & Livio, M. 1994, , 286, 786
della Valle, M., & Livio, M. 1998, , 506, 818
Dobrotka, A., Retter, A., & Liu, A. 2008, , 478, 815
Ducati, J. R., Bevilacqua, C. M., Rembold, S. B., & Ribeiro, D. 2001, , 558, 309
Epchtein, N., de Batz, B., Copet, E., et al. 1994, , 217, 3
Frank, J., King, A., & Raine, D. J. 2002, Accretion Power in Astrophysics, by Juhan Frank and Andrew King and Derek Raine, pp. 398. ISBN 0521620538. Cambridge, UK: Cambridge University Press, February 2002.,
Gaposchkin, C. H. P. 1957, Amsterdam, North-Holland Pub. Co.; New York, Interscience Publishers, 1957.,
Green, D. W. E. 2012, Central Bureau Electronic Telegrams, 3073, 1
Green, D. W. E. 2012, Central Bureau Electronic Telegrams, 3089, 1
Greenhouse, M. A., Grasdalen, G. L., Hayward, T. L., Gehrz, R. D., & Jones, T. J. 1988, AJ, 95, 172
Gonzalez, O. A., Rejkuba, M., Zoccali, M., et al. 2012, , 543, A13
Liller, W., & Mayer, B. 1987, , 99, 606
Liller, W. 1993, , 26, 41
Martin, P. G. 1989, Classical Novae, 73
Mason, C. G., Gehrz, R. D., Woodward, C. E., et al. 1998, ApJ, 494, 783
Minniti, D., Lucas, P. W., Emerson, J. P., et al. 2010, , 15, 433
Nakano, S., Nishiyama, K., & Kabashima, F. 2008, Central Bureau Electronic Telegrams, 1342, 1
Nakano, S., Nishimura, H., Itagaki, K., et al. 2010, , 9111, 2
Nakano, S., Nishimura, H., Kiyota, S., et al. 2010, , 9119, 1
Nakano, S., Nishimura, H., Kiyota, S., & Yusa, T. 2011, , 9196, 1
Nishimura, H., & Nakano, S. 2012, Central Bureau Electronic Telegrams, 3166, 1
Nishiyama, S., Tamura, M., Hatano, H., et al. 2009, , 696, 1407
Nishiyama, K., Kabashima, F., Liller, W., Yusa, T., & Maehara, H. 2010, , 9140, 1
Nishiyama, K., Kabashima, F., Maehara, H., & Kiyota, S. 2011, , 9203, 1
Kiyota, S., Vollmann, W., Koberger, H., et al. 2010, , 9112, 1
Kozlowski, S., Poleski, R., Udalski, A., et al. 2012, The Astronomer’s Telegram, 4323, 1
Raj, A., Ashok, N. M., & Banerjee, D. P. K. 2011, , 415, 3455
Rieke, G. H., & Lebofsky, M. J. 1985, , 288, 618
Russell, R. W., Rudy, R. J., Lynch, D. K., et al. 2008, , 8948, 1
Saito, R. K., Hempel, M., Minniti, D., et al. 2012, , 537, A107
Saito, R. K., Minniti, D., Dias, B., et al. 2012, , 544, A147
Saito, R. K., Minniti, D., Angeloni, R., & Catelan, M. 2012, The Astronomer’s Telegram, 4426, 1
Saito, R. K., & Minniti, D. 2012, The Astronomer’s Telegram, 4353, 1
Saito, R. K., & Minniti, D. 2012, The Astronomer’s Telegram, 4372, 1
Sala, G., Hernanz, M., Ferri, C., & Greiner, J. 2008, , 675, L93
Sala, G., Hernanz, M., Ferri, C., & Greiner, J. 2010, Astronomische Nachrichten, 331, 201
Sato, S., Kawara, K., Kobayashi, Y., et al. 1978, PASJ, 30, 419
Schwarz, G. J., Shore, S. N., Starrfield, S., et al. 2001, , 320, 103
Shafter, A. W. 1997, , 487, 226
Shafter, A. W., & Irby, B. K. 2001, , 563, 749
Shafter, A. W., Bode, M. F., Darnley, M. J., et al. 2011, , 727, 50
Shara, M. M., Yaron, O., Prialnik, D., & Kovetz, A. 2010, , 712, L143
Shaviv, N. J. 2002, Classical Nova Explosions, 637, 259
Shaviv, N. J., & Dotan, C. 2012, , 83, 792
Skrutskie, M. F., et al. 2006, , 131, 1163
Sun, G., & Gao, X. 2009, , 9049, 1
Udalski, A., Szymanski, M., Kaluzny, J., Kubiak, M., & Mateo, M. 1993, , 43, 69
Varricatt, W. P., Banerjee, D. P. K., & Ashok, N. M. 2012, The Astronomer’s Telegram, 4405, 1
Venturini, C. C., Rudy, R. J., Lynch, D. K., Mazuk, S., & Puetter, R. C. 2002, , 124, 3009
Waagen, E. O. 2012, AAVSO Alert Notice, 461, 1
Walter, F. M., & Buil, C. 2012, Central Bureau Electronic Telegrams, 3124, 1
Warner, B. 2003, Cataclysmic Variable Stars, by Brian Warner, pp. 592. ISBN 052154209X. Cambridge, UK: Cambridge University Press, September 2003.,
Williams, R. E. 1992, , 104, 725
Woodward, C. E., Gehrz, R. D., Jones, T. J., Lawrence, G. F., & Skrutskie, M. F. 1997, ApJ, 477, 817
Yamaoka, H., & Itagaki, K. 2012, Central Bureau Electronic Telegrams, 3156, 1
VVV Novae finding charts {#charts}
========================
In this section we present the finding charts for all Novae in the VVV Survey area, listed in Table \[table:novae\]. In most cases the charts are $JHK_{\rm s}$ composite images, unless when explicitly marked as a $K_{\rm s}$-band image. All charts are $30\arcsec \times
30\arcsec$ in size, oriented in Galactic coordinates, with the positive Galactic longitude pointing up. A cross marks the coordinates given by the catalogue and helps one to check the notes we adopted in Table \[table:novae\].
[^1]: Based on observations taken within the ESO VISTA Public Survey VVV, Programme ID 179.B-2002
[^2]: [ http://casu.ast.cam.ac.uk/vistasp/]{}
[^3]: [http://www.aavso.org/vsx/index.php]{}
[^4]: [simbad.u-strasbg.fr/simbad/]{}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We report results of a deep all-sky search for periodic gravitational waves from isolated neutron stars in data from the S6 LIGO science run. The search was possible thanks to the computing power provided by the volunteers of the distributed computing project. We find no significant signal candidate and set the most stringent upper limits to date on the amplitude of gravitational wave signals from the target population. At the frequency of best strain sensitivity, between $170.5$ and $171$ Hz we set a 90% confidence upper limit of $\sci{5.5}{-25}$, while at the high end of our frequency range, around 505 Hz, we achieve upper limits $\simeq {10}^{-24}$. At $230$ Hz we can exclude sources with ellipticities greater than $10^{-6}$ within 100 pc of Earth with fiducial value of the principal moment of inertia of $10^{38} \textrm{kg m}^2$. If we assume a higher (lower) gravitational wave spindown we constrain farther (closer) objects to higher (lower) ellipticities.\
author:
- The LIGO Scientific Collaboration and the Virgo Collaboration
title: ' Results of the deepest all-sky survey for continuous gravitational waves on LIGO S6 data running on the volunteer distributed computing project '
---
[^1]
Introduction {#sec:introduction}
============
In this paper we report the results of a deep all-sky [@EaHweb] search for continuous, nearly monochromatic gravitational waves (GWs) in data from LIGO’s sixth science (S6) run. A number of all-sky searches have been carried out on LIGO data, [@S6Powerflux; @Aasi:2015rar; @S5GC1HF; @FullS5EH; @FullS5Semicoherent; @S5EH; @EarlyS5Paper; @S4IncoherentPaper; @S4EH; @S2FstatPaper], of which [@S4EH; @S5EH; @FullS5EH] also ran on . The search presented here covers frequencies from 50 Hz through 510 Hz and frequency derivatives from [[$\sci{3.39}{-10}~\mathrm{Hz/s}$]{}]{} through [$-\sci{2.67}{-9}~\mathrm{Hz/s}$]{}. In this range we establish the most constraining gravitational wave amplitude upper limits to date for the target signal population.
LIGO interferometers and the data used {#sec:S6intro}
======================================
The LIGO gravitational wave network consists of two observatories, one in Hanford (WA) and the other in Livingston (LA) separated by a 3000-km baseline [@LIGO_detector]. The last science run (S6) [@LIGO:2012aa] of this network before the upgrade towards the advanced LIGO configuration [@TheLIGOScientific:2014jea] took place between July 2009 and October 2010. The analysis in this paper uses a subset of this data: from GPS 949469977 (2010 Feb 6 05:39:22 UTC) through GPS 971529850 (2010 Oct 19 13:23:55 UTC), selected for good strain sensitivity [@Shaltev]. Since interferometers sporadically fall out of operation (“lose lock”) due to environmental or instrumental disturbances or for scheduled maintenance periods, the data set is not contiguous and each detector has a duty factor of about 50% [@detchar2]. As done in [@S5EH], frequency bands known to contain spectral disturbances have been removed from the analysis. Actually, the data has been substituted with Gaussian noise with the same average power as that in the neighbouring and undisturbed bands. Table \[A:cleanedBands\] identifies these bands.
The Search
==========
The search described in this paper targets nearly monochromatic gravitational wave signals as described for example by Eqs. 1-4 of [@S5EH]. Various emission mechanisms could generate such a signal as reviewed in Section IIA of [@S2FstatPaper]. In interpreting our results we will consider a spinning compact object with a fixed, non-axisymmetric mass quadrupole, described by an ellipticity $\epsilon$.
We perform a stack-slide type of search using the GCT (Global correlation transform) method [@Pletsch:2008; @Pletsch:2010]. In a stack-slide search the data is partitioned in segments and each segment is searched with a matched-filter method [@cutler]. The results from these coherent searches are combined by summing (stacking) the detection statistic values from the segments (sliding), one per segment (${\mathcal{F}}_i$), and this determines the value of the core detection statistic: $$\label{eq:avF}
{{\overline{{\mathcal{F}}}}}:={1\over{{N_{\mathrm{seg}}}}} \sum_{i=1}^{{{N_{\mathrm{seg}}}}} {\mathcal{F}}_i.$$ There are different ways to combine the single-segment ${\mathcal{F}}_i$ values, but independently of the way that this is done, this type of search is usually referred to as a “semi-coherent search”. So stack-slide searches are a type of semi-coherent search. Important variables for this type of search are: the coherent time baseline of the segments $\Tcoh$, the number of segments used ${{N_{\mathrm{seg}}}}$, the total time spanned by the data $\Tobs$, the grids in parameter space and the detection statistic used to rank the parameter space cells. For a stack-slide search in Gaussian noise, ${{N_{\mathrm{seg}}}}\times 2{{\overline{{\mathcal{F}}}}}$ follows a $\chi^2_{4{{N_{\mathrm{seg}}}}}$ chi-squared distribution with $4{{N_{\mathrm{seg}}}}$ degrees of freedom.
![Polar plots ($r,\theta$ plots with $\theta=\alpha$ and $r=\cos\delta$) of the grid points in the northern equatorial hemisphere sky for the band 50-60Hz (left panel) and for the band 110-120Hz (right panel). $\alpha$ is the right ascension coordinate and $\delta$ the declination coordinate. One can clearly see the higher density in the $-0.5\leq \delta \leq 0.5$ equatorial region and the higher density ($\propto {{f^2}}$) of grid points at higher frequencies. The southern hemispheres looks practically identical to the respective northern ones.[]{data-label="fig:SkyGrids"}](SkyGrids2.pdf){width="1.0\columnwidth"}
These parameters are summarised in Table \[tab:GridSpacings\]. The grids in frequency and spindown are each described by a single parameter, the grid spacing, which is constant over the search range. The same frequency grid spacings are used for the coherent searches over the segments and for the incoherent summing. The spindown spacing for the incoherent summing, $\delta{\dot{f}}$, is finer than that used for the coherent searches, $\delta{\dot{f_c}}$, by a factor $\gamma$. The notation used here is consistent with that used in previous observational papers [@GalacticCenterSearch] and in the GCT methods papers cited above.
The sky grid is the union of two grids: one is uniform over the projection of the celestial sphere onto the equatorial plane, and the tiling (in the equatorial plane) is approximately square with sides of length $$d(m_{\text{sky}})={1\over f}
{
{\sqrt{ m_{\text{sky}} } }
\over {\pi \tau_{E}}
},
\label{eq:skyGridSpacing}$$ with $m_{\text{sky}}=0.3$ and $\tau_{E}\simeq0.021$ s being half of the light travel time across the Earth. As was done in [@S5EH], the sky-grids are constant over 10 Hz bands and the spacings are the ones associated through Eq. \[eq:skyGridSpacing\] to the highest frequency $f$ in the range. The other grid is limited to the equatorial region ($0\leq \alpha\leq 2\pi$ and $-0.5\leq \delta \leq 0.5$), with constant right ascension $\alpha$ and declination $\delta$ spacings equal to $d(0.3)$ – see Fig.\[fig:SkyGrids\]. The reason for the equatorial “patching” with a denser sky grid is to improve the sensitivity of the search: the sky resolution actually depends on the ecliptic latitude and the uniform equatorial grid under-resolves particularly in the equatorial region. The resulting number of templates used to search 50-mHz bands as a function of frequency is shown in Fig. \[fig:NumberOfTemplatesIn50mHz\].
![Number of searched templates in 50-mHz bands. The variation with frequency is due to the increasing sky resolution. $N_f \times N_{\dot{f}} \sim 3.7\times 10^{8}$, where $N_f$ and $N_{\dot{f}}$ are the number of $f$ and $\dot{f}$ templates searched in 50-mHz bands. The total number of templates searched between 50 and 510 Hz is [[$\sci{6.3}{16}$]{}]{}. []{data-label="fig:NumberOfTemplatesIn50mHz"}](NumberOfTemplatesIn50mHzLR.jpg){width="\columnwidth"}
$\Tcoh$ 60 hrs
------------------------ ----------------------------
$\Tref$ 960499913.5 GPS sec
${{N_{\mathrm{seg}}}}$ 90
$\delta f$ $1.6\times 10^{-6}$ Hz
$\delta {\dot{f_c}}$ $5.8 \times 10^{-11}$ Hz/s
$\gamma$ 230
$m_{\text{sky}}$ 0.3 + equatorial patch
The search is split into work-units (WUs) sized to keep the average volunteer computer busy for about [[6]{}]{} hours. Each WU searches a $50$ mHz band, the entire spindown range and [[13]{}]{} points in the sky, corresponding to [[$\sci{4.9}{9}$]{}]{} templates out of which it returns only the top 3000. A total of [[12.7]{}]{} million WUs are necessary to cover the entire parameter space. The total number of templates searched is [$\sci{6.3}{16}$]{}.
### The ranking statistic
The search was actually carried out in separate runs that used different ranking statistics to define the top-candidate-list, reflecting different stages in the development of a detection statistic robust with respect to spectral lines in the data [@Keitel:2013]. In particular, three ranking statistics were used: the average $2{\mathcal{F}}$ statistic over the segments, $2{{\overline{{\mathcal{F}}}}}$, which in essence at every template point is the likelihood of having a signal with the shape given by the template versus having Gaussian noise; the line-veto statistic ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{\mathrm{\MakeUppercase{L}}}}}}}$ which is the odds ratio of having a signal versus having a spectral line; and a general line-robust statistic, ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$, that tests the signal hypothesis against a Gaussian noise + spectral line noise model. Such a statistic can match the performance of both the standard average $2{{\overline{{\mathcal{F}}}}}$ statistic in Gaussian noise and the line-veto statistic in presence of single-detector spectral disturbances and statistically outperforms them when the noise is a mixture of both [@Keitel:2013].
We combine the $2{{\overline{{\mathcal{F}}}}}$-ranked results with the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{\mathrm{\MakeUppercase{L}}}}}}}$-ranked results to produce a single list of candidates ranked according to the general line-robust statistic ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$. We now explain how this is achieved. Alongside the detection statistic value and the parameter space cell coordinates of each candidate, the application also returns the single-detector $2{{{\overline{{\mathcal{F}}}}}^X}$ values (“$X$” indicates the detector). These are used to compute, for every candidate of any run, the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ through Eq. 61 of [@Keitel:2013]: $$\begin{split}
\label{eq:logOSNsc}
\ln &{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}= \ln {{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}}+ {{\widehat{{\mathcal{F}}}}}- {{{\widehat{{\mathcal{F}}}}}''_{\mathrm{max}}}\\
&- \ln \left( {\ensuremath{\mathrm{e}^}}{{{{\widehat{{\mathcal{F}}}}}_*}-{{{\widehat{{\mathcal{F}}}}}''_{\mathrm{max}}}} + {\left\langle {{\widehat{r}}}^X {\ensuremath{\mathrm{e}^}}{{{{\widehat{{\mathcal{F}}}}}^X}-{{{\widehat{{\mathcal{F}}}}}''_{\mathrm{max}}}} \right\rangle} \right)\,,
\end{split}$$ with the angle-brackets indicating the average with respect to detectors ($X$) and $$\begin{aligned}
\label{eq:DefsforLogOSGL1}
&{{\widehat{{\mathcal{F}}}}}={{N_{\mathrm{seg}}}}{{\overline{{\mathcal{F}}}}}\\
&{{{\widehat{{\mathcal{F}}}}}^X}={{N_{\mathrm{seg}}}}{{{\overline{{\mathcal{F}}}}}^X}\\
&{{{\widehat{{\mathcal{F}}}}}''_{\mathrm{max}}}\equiv \text{max}\left( {{{\widehat{{\mathcal{F}}}}}_*},\,{{{\widehat{{\mathcal{F}}}}}^X}+ \ln {{\widehat{r}}^X}\right)\\
&{{{\widehat{{\mathcal{F}}}}}_*}\equiv {{{\widehat{{\mathcal{F}}}}}_*^{(0)}}- \ln {{{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}}\\
\label{eq:cstar}
&{{{\widehat{{\mathcal{F}}}}}_*^{(0)}}\equiv \ln {c_*}^{{{N_{\mathrm{seg}}}}} ~~{\text{with}}~{c_*}~{\text{set to }} 20.64\\
&{{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}= \sum_X {{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}^X \\
&{{\widehat{r}}}^X \equiv \frac{{{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}^X}{{{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}/{{N_{\mathrm{det}}}}}\\
&{{\widehat{p}}_{{{{\mathrm{\MakeUppercase{L}}}}}}}\equiv \frac{{{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}}{1 + {{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}}\end{aligned}$$ where ${{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}^X$ is the assumed prior probability of a spectral line occuring in any frequency bin of detector X, ${{\widehat{p}}_{{{{\mathrm{\MakeUppercase{L}}}}}}}$ is the line prior estimated from the data, ${{N_{\mathrm{det}}}}= 2$ is the number of detectors, and $ {{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}}$ is an assumed prior probability of a line being a signal (set arbitrarily to 1; its specific value does not affect the ranking statistic). Following the reasoning of Eq. 67 of [@Keitel:2013], with ${{N_{\mathrm{seg}}}}=90$ we set ${c_*}=20.64$ corresponding to a Gaussian false-alarm probability of $10^{-9}$ and an average $2{{\overline{{\mathcal{F}}}}}$ transition scale of $\sim 6$ (${{{\mathcal{F}}_*}^{(0)}}\sim 3$). The $ {{\MakeLowercase{{{\widehat{O}}_{{{{\mathrm{\MakeUppercase{L}}}}{\mathrm{\MakeUppercase{G}}}}}}}}}^X$ values are estimated from the data as described in Section VI.A of [@Keitel:2013] in 50-mHz bands with a normalized-SFT-power threshold ${{\mathcal{P}}_{\mathrm{thr}}}^X={{\mathcal{P}}_{\mathrm{thr}}}({p_{\mathrm{FA}}}=10^{-9},{N_{{\mathrm{SFT}}}}^X\sim 6000)\approx 1.08$. For every 50-mHz band the list of candidates from the $2{{\overline{{\mathcal{F}}}}}$-ranked run is merged with the list from the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{\mathrm{\MakeUppercase{L}}}}}}}$-ranked run and duplicate candidates are considered only once. The resulting list is ranked by the newly-computed ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ and the top 3000 candidates are kept. This is our result-set and it is treated in a manner that is very similar to [@S5GC1HF].
### Identification of undisturbed bands {#sec:visualInspection}
Even after the removal of disturbed data caused by spectral artefacts of known origin, the statistical properties of the results are not uniform across the search band. In what follows we concentrate on the subset of the signal-frequency bands having reasonably uniform statistical properties. This still leaves us with the majority of the search parameter space while allowing us to use methods that rely on theoretical modelling of the significance in the statistical analysis of the results. Our classification of “clean" vs. “disturbed" bands has no pretence of being strictly rigorous, because strict rigour here is neither useful nor practical. The classification serves the practical purpose of discarding from the analysis regions in parameter space with evident disturbances and must not dismiss real signals. The classification is carried out in two steps: a visual inspection and a refinement on the visual inspection.
The visual inspection is performed by three scientists who each look at various distributions of the detection statistics over the entire sky and spindown parameter space in 50-mHz bands. They rank each band with an integer score 0,1,2,3 ranging from “undisturbed" (0) to “disturbed" (3) . A band is considered “undisturbed" if all three rankings are 0. The criteria agreed upon for ranking are that the distribution of detection statistic values should not show a visible trend affecting a large portion of the $f-\dot{f}$ plane and, if outliers exist in a small region, outside this region the detection statistic values should be within the expected ranges. Fig. \[fig:VIGquietdisturbed\] shows the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ for three bands: two were marked as undisturbed and the other as disturbed. One of the bands contains the $f-\dot{f}$ parameter space that harbours a fake signal injected in the data to verify the detection pipelines. The detection statistic is elevated in a small region around the signal parameters. The visual inspection procedure does not mark as disturbed bands with such features.
![On the z-axis and color-coded is the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ in three 50-mHz bands. The top band was marked as “undisturbed". The middle band is an example of a “disturbed band". The bottom band is an example of an “undisturbed band" but containing a signal, a fake one, in this case.[]{data-label="fig:VIGquietdisturbed"}](VIG1.pdf "fig:"){width="\columnwidth"} ![On the z-axis and color-coded is the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ in three 50-mHz bands. The top band was marked as “undisturbed". The middle band is an example of a “disturbed band". The bottom band is an example of an “undisturbed band" but containing a signal, a fake one, in this case.[]{data-label="fig:VIGquietdisturbed"}](VIG2.jpg "fig:"){width="\columnwidth"} ![On the z-axis and color-coded is the ${{\widehat{O}}_{{{{\mathrm{\MakeUppercase{S}}}}{{{\mathrm{\MakeUppercase{G}}}{{\mathrm{\MakeUppercase{L}}}}}}}}}$ in three 50-mHz bands. The top band was marked as “undisturbed". The middle band is an example of a “disturbed band". The bottom band is an example of an “undisturbed band" but containing a signal, a fake one, in this case.[]{data-label="fig:VIGquietdisturbed"}](VIG3.jpg "fig:"){width="\columnwidth"}
Based on this visual inspection 13% of the bands between 50 and 510 Hz are marked as “disturbed". Of these, 34% were given by all visual inspectors rankings smaller than 3, i.e. they were only marginally disturbed. Further inspection “rehabilitated” 42% of these. As a result of this refinement in the selection procedure we exclude from the current analysis 11% of the searched frequencies.
![Highest values of $2{{\overline{{\mathcal{F}}}}}$ in every half-Hz band as a function of band frequency. Since the number of templates increases with frequency so does the loudest $2{{\overline{{\mathcal{F}}}}}$.[]{data-label="fig:loudestHalfHzVersusFreq"}](loudestHalfHzVersusFreqWithOutDisturbed50mHzLR.jpg){width="\columnwidth"}
![Highest values of the significance (CR) in every half-Hz band as a function of band frequency. Since the significance folds in the expected value for the loudest $2{{\overline{{\mathcal{F}}}}}$ and its standard deviation, the significance of the loudest in noise does not increase with frequency. Our results are consistent with this expectation.[]{data-label="fig:loudestSigmaHalfHzVersusFreq"}](loudestSigmaHalfHzVersusFreqWithOutDisturbedLR.jpg){width="\columnwidth"}
![Histogram of the highest values of the significance CR in every half-Hz band.[]{data-label="fig:loudestSigmaHistHalfHzVersusFreq"}](sigmaHistLoudestInHalfHzWithOutDisturbedLR.jpg){width="\columnwidth"}
![The fraction of 50mHz bands (in signal frequency) which contribute to the results in every half-Hz band. As explained in the text, some bands are excluded because they are all from fake data or because they are marked as disturbed by the visual inspection. The list of excluded bands is given in Table \[A:excluded50mHzBands\].[]{data-label="fig:fill-level"}](HalfHzbandsFillLevelDistrLR.jpg){width="\columnwidth"}
Fig. \[fig:loudestHalfHzVersusFreq\] shows the highest values of the detection statistic in half-Hz signal-frequency bands compared to the expectations. The set of candidates that the highest detection statistic values are picked from, does not include the 50mHz signal-frequency bands that stem entirely from fake data, from the cleaning procedure, or that were marked as disturbed. In this paper we refer to the candidates with the highest value of the detection statistic as the [*[loudest]{}*]{} candidates. The loudest expected value over $N_{\text{trials}}$ independent trials of $2{{\overline{{\mathcal{F}}}}}$ is determined[^2] by numerical integration of the probability density function given, for example, by Eq. 7 of [@GalacticCenterSearch]. For this search, we estimate that $N_{\text{trials}}\simeq 0.87 \,N_{\text{templ}}$, with $N_{\text{templ}}$ being the number of templates searched. As a uniform measure of significance of the highest $2{{\overline{{\mathcal{F}}}}}$ value across bands that were searched with different values of $N_{\text{trials}}$ we introduce the critical ratio CR defined as the deviation of the measured highest $2{{\overline{{\mathcal{F}}}}}$ from the expected value, measured in units of the standard deviation: $$\label{eq:CR}
{\text{CR}}:={{2{{\overline{{\mathcal{F}}}}}_{\text{meas}}- 2{{\overline{{\mathcal{F}}}}}_{\text{exped}}}\over{\sigma_{\text{exped}}}}.$$
The highest and most significant detection statistic value from our search is $2{{\overline{{\mathcal{F}}}}}=8.6$ at a frequency of about 52.76 Hz with a CR=29. This is due to a fake signal. The second highest value of the detection statistic is 7.04 at a frequency of about 329.01 Hz corresponding to a CR of 4.6. The second highest-CR candidate has a $2{{\overline{{\mathcal{F}}}}}$ of 6.99, is at 192.16 Hz and has a CR=4.8.
Sorting loudest candidates from half-Hz bands according to detection statistic values is not the same as sorting them according to CR. The reason for this is that the number of templates is not the same for all half-Hz bands. This is due to the grid spacings decreasing with frequency (Eq. \[eq:skyGridSpacing\]) and to the fact that, as previously explained, some 50-mHz bands have been excluded from the current analysis and hence some half-Hz bands comprise results from fewer than ten 50mHz bands. Fig.\[fig:fill-level\] gives the fill-level of each half-Hz band, i.e. how many 50mHz bands have contributed candidates to the analysis out of ten. We use the CR as a measure of the significance because it folds in correctly the effect of varying number of templates in the half-Hz bands.
![p-values for the loudest in half-Hz bands of our data (histogram bars) and expected distribution of pure noise data for reference (black markers).[]{data-label="fig:pvalues"}](pvaluesLR.jpg){width="\columnwidth"}
After excluding the candidate due to the fake signal, in this data we see no evidence of a signal: the distribution of p-values associated with every measured half-Hz band loudest is consistent with what we expect from noise-only across the measured range (Fig.\[fig:pvalues\]). In particular we note two things: 1) the two candidates at CR=4.6 and CR=4.8 are not significant when we consider how many half-Hz bands we have searched and 2) there is no population of low significance candidates deviating from the expectation of the noise-only case. The p-value for the loudest measured in any half-Hz band searched with an effective number of independent trials $N_{\text{trials}}=0.87 ~N_{\text{trials}}$ is obtained by integrating Eq. 6 of [@GalacticCenterSearch] between the observed value and infinity.
Upper limits {#sec:upper limits}
============
{width="80.00000%"}
{width="80.00000%"}
The search did not reveal any continuous gravitational wave signal in the parameter volume that was searched. We hence set frequentist upper limits on the maximum gravitational wave amplitude consistent with this null result in half-Hz bands : $h_0^{90\%}$(f). $h_0^{90\%}$(f) is the GW amplitude such that 90% of a population of signals with parameter values in our search range would have produced a candidate louder than what was observed by our search. This is the criterion hereafter referred to as “detection".
Evaluating these upper limits with injection-and-recovery Monte Carlo simulations in every half-Hz band is too computationally intensive. So we perform them in a subset of 50 bands and infer the upper limit values in the other bands from these. The 50 bands are evenly spaced in the search frequency range. For each band $j=1\dots 50$, we measure the 90% upper limit value corresponding to different detection criteria. The different detection criteria are defined by different CR values for the assumed measured loudest. The first CR bin, $\CR_0$, is for CR values equal to or smaller than $0$, the next bins are for $i < \CR_i \leq (i+1)$ with $i=1 \ldots 5$. Correspondingly we have ${h_{0,\CR_i}^{90\%,j}}$ for each band. For every detection criteria and every band we determine the sensitivity depth [@GalacticCenterMethod], and by averaging these sensitivity depths over the bands we derive a sensitivity depth for every detection criteria: ${\mathcal{D}}_{\CR_i}^{90\%}={1/50}\sum_j {\mathcal{D}}_{\CR_i}^{90\%,j}$. We use these to set upper limits in the bands $k$ where we have not performed injection-and-recovery simulations as $${ h{_{0}^{90\%}}(f_k) } =
{
{\sqrt{S_h(f_k)}} \over {
{{\mathcal{D}}}_{\CR_i(k)}^{90\%}
}
},
\label{eq:ULfromSensDepth}$$ where $\CR_i(k)$ is the significance bin of the loudest candidate of the $k^{\text{th}}$ band and $S_h(f_k)$ the power spectral density of the data (measured in $1/\sqrt{\text{Hz}}$). The values of the sensitivity depths range between ${{\mathcal{D}}}_{\CR_6}^{90\%} \simeq 33 ~(1/\sqrt{\text{Hz}})$ and ${{\mathcal{D}}}_{\CR_0}^{90\%}\simeq 37 ~(1/\sqrt{\text{Hz}})$. The uncertainties on the upper limit values introduced by this procedure are $\simeq 10\%$ of the nominal upper limit value. We represent this uncertainty as a shaded region around the upper limit values in Fig.\[fig:ULs\]. The upper limit values are also provided in tabular form in the Appendix in Table IV. We do not set upper limits in half-Hz bands where the results are entirely produced with fake data inserted by the cleaning procedure described in Section \[sec:S6intro\]. Upper limits for such bands will not appear in Table \[A:ULs\] nor in Fig. \[fig:ULs\]. There also exist 50-mHz bands that include contributions from fake data as a result of the cleaning procedure or that have been excluded from the analysis because they were marked as disturbed by the visual inspection procedure described in Section \[sec:visualInspection\]. We mark the half-Hz bands which host these 50mHz bands with a different colour (light red) in Fig.\[fig:ULs\]. In Table \[A:excluded50mHzBands\] in the Appendix we provide a complete list of such 50-mHz bands because the upper limit values do not apply to those 50-mHz bands. Finally we note that, due to the cleaning procedure, there exist signal frequency bands where the search results [*[might]{}*]{} have contributions from fake data. We list these signal-frequency ranges in Table \[A:suspectFreqRanges\]. For completeness this table also contains the cleaned bands of Table \[A:excluded50mHzBands\], under the column header “all fake data”.
Conclusions
===========
Our upper limits are the tightest ever placed for this set of target signals. The smallest value of the GW amplitude upper limit is $5.5\times 10^{-25}$ in the band 170.5-171 Hz. Fig. \[fig:ULs\] shows the upper limit values as a function of search frequency. We also show the upper limits from [@S6Powerflux], another all-sky search on S6 data, rescaled according to [@WetteSensEst:2012] to enable a direct comparison with ours. Under the assumption that the sources are uniformly distributed in space, our search probes a volume in space a few times larger than that of [@S6Powerflux]. It should however be noted that [@S6Powerflux] examines a much broader parameter space than the one presented here. The Virgo VSR2 and VSR4 science runs were contemporary to the S6 run and more sensitive at low frequency with respect to LIGO. The Virgo data were analysed in search of continuous signals from the whole sky in the frequency range 20Hz - 128Hz and a narrower spindown range than that covered here, with $|{\dot{f}}| \leqslant 10^{-10}$ Hz/s [@Aasi:2015rar]. Our sensitivity is comparable to that achieved by that search and improves on it above 80 Hz.
Following [@Ming:2015jla], we define the fraction $x$ of the spindown rotational energy emitted in gravitational waves. The star’s ellipticity necessary to sustain such emission is $$\label{eq:GWspindown}
\epsilon(f,x{\dot{f}})=\sqrt{{5c^5\over 32\pi^4 G}{x{\dot{f}}\over If^5}},$$ where $c$ is the speed of light, $G$ is the gravitational constant, $f$ is the GW frequency and $I$ the principal moment of inertia of the star. Correspondingly, $x{\dot{f}}$ is the spindown rate that accounts for the emission of GWs and this is why we refer to it as the GW spindown. The gravitational wave amplitude $h_0$ at the detector coming from a GW source like that of Eq.\[eq:GWspindown\], at a distance $D$ from Earth is $$\label{eq:h0}
h_0(f,x\dot{f},D)={1\over D}\sqrt{{5GI\over 2c^3}{x{\dot{f}}\over f}}.$$ Based on this last equation, we can use our GW amplitude upper limits to bound the minimum distance for compact objects emitting continuous gravitational waves under different assumptions on the object’s ellipticity (i.e. gravitational wave spindown). This is shown in Fig. \[fig:reach\]. We find that for most frequencies above 230 Hz our upper limits exclude compact objects with ellipticities of $10^{-6} {\sqrt{10^{38} \textrm{kg\,m}^2\over I}}$ (corresponding to GW spindowns between $10^{-12}$ Hz/s and $10^{-11}$ Hz/s) within 100 pc of Earth. Both the ellipticity and the distance ranges span absolutely plausible values and could not have been excluded with other measurements.
We expect the methodology used in this search to serve as a template for the assessment of run results in the future, for example the next run, using advanced LIGO data that is being processed as this paper is written. Results of searches for continuous wave signals could also be mined further, probing sub-threshold candidates with a hierarchical series of follow-up searches. This is not the topic of this paper and might be pursued in a forthcoming publication.
Acknowledgments
===============
The authors gratefully acknowledge the support of the Einstein@Home volunteers, of the United States National Science Foundation for the construction and operation of the LIGO Laboratory, the Science and Technology Facilities Council of the United Kingdom, the Max-Planck-Society, and the State of Niedersachsen/Germany for support of the construction and operation of the GEO600 detector, and the Italian Istituto Nazionale di Fisica Nucleare and the French Centre National de la Recherche Scientifique for the construction and operation of the Virgo detector. The authors also gratefully acknowledge the support of the research by these agencies and by the Australian Research Council, the International Science Linkages program of the Commonwealth of Australia, the Council of Scientific and Industrial Research of India, the Istituto Nazionale di Fisica Nucleare of Italy, the Spanish Ministerio de Educación y Ciencia, the Conselleria d’Economia Hisenda i Innovació of the Govern de les Illes Balears, the Foundation for Fundamental Research on Matter supported by the Netherlands Organisation for Scientific Research, the Polish Ministry of Science and Higher Education, the FOCUS Programme of Foundation for Polish Science, the Royal Society, the Scottish Funding Council, the Scottish Universities Physics Alliance, The National Aeronautics and Space Administration, the Carnegie Trust, the Leverhulme Trust, the David and Lucile Packard Foundation, the Research Corporation, and the Alfred P. Sloan Foundation.
This document has been assigned LIGO Laboratory document number `LIGO-P1600156-v22`.
[99]{}
Tabular data
============
Upper limit values {#A:ULs}
------------------
Cleaned-out frequency bins {#A:cleanedBands}
--------------------------
50mHz signal-frequency bands that did not contribute to results {#A:excluded50mHzBands}
---------------------------------------------------------------
Signal-frequency ranges where the results might have contributions from fake data {#A:suspectFreqRanges}
---------------------------------------------------------------------------------
[^1]: Full author list given at the end of the article.
[^2]: After a simple change of variable from $2{{\overline{{\mathcal{F}}}}}$ to ${{N_{\mathrm{seg}}}}\times 2{{\overline{{\mathcal{F}}}}}$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate spectral evolution in 37 bright, long gamma-ray bursts observed with the BATSE Spectroscopy Detectors. High resolution spectra are characterized by the energy of the peak of and the evolution of this quantity is examined relative to the emission intensity. In most cases it is found that this peak energy either rises with or slightly precedes major intensity increases and softens for the remainder of the pulse. Inter-pulse emission is generally harder early in the burst. For bursts with multiple intensity pulses, later spikes tend to be softer than earlier ones indicating that the energy of the peak of is bounded by an envelope which decays with time. Evidence is found that bursts in which the bulk of the flux comes well after the event which triggers the instrument tend to show less peak energy variability and are not as hard as several bursts in which the emission occurs promptly after the trigger. Several recently proposed burst models are examined in light of these results and no qualitative conflicts with the observations presented here are found.'
author:
- 'L. A. Ford, D. L. Band, J. L. Matteson'
- 'M. S. Briggs, G. N. Pendleton, R. D. Preece, W. S. Paciesas'
- 'B. J. Teegarden, D. M. Palmer, B. E. Schaefer, T. L. Cline'
- 'G. J. Fishman, C. Kouveliotou, C. A. Meegan, R. B. Wilson'
- 'J. P. Lestrade'
title: |
BATSE Observations of Gamma-Ray Burst Spectra.\
II. Peak Energy Evolution in Bright, Long Bursts
---
=-0.25truein
$\underline{~}$
Introduction
============
Gamma-ray burst continuum spectra provide the most direct information about the emission processes involved in these violent events. Unfortunately, the continuum generating process is unknown, making physical interpretation of spectral observations difficult. A striking feature of gamma-ray bursts is the temporal variability of spectra both between and within bursts. Therefore, an empirical study of the dynamics of burst spectra may provide vital clues for resolving the mystery surrounding these puzzling events.
This is the second in a series of reports describing spectral observations of gamma-ray bursts as seen by the Burst and Transient Source Experiment (BATSE) on the [*Compton Gamma-Ray Observatory*]{} ([*CGRO*]{}). The goal of this series is to discover global properties of burst continuum spectra which may shed light on the gamma-ray burst problem. Gamma-ray bursts remain one of the least understood phenomena in astrophysics despite years of intense study. Their heterogenous nature makes classification difficult and few uniform burst properties are known. Bursts have only been observed above $\gtrsim$2 keV and no emission has been seen in quiescence ([@sch94]). In the first paper of this series ([@band93], hereafter Paper I), spectral properties of time integrated spectra from 54 bright bursts were studied. While spectra are quite diverse, it was shown that a simple and flexible empirical model successfully described all spectra. This model is used in the present study to describe burst spectra and monitor the evolution of the energy at which the energy flux per logarithmic energy band peaks (, energy of the peak of ). In future work, similar evolution studies will be performed with other spectral characteristics, such as continuum shape or spectral bandwidth parameters. The data used here are high energy resolution spectra from the Spectroscopy Detectors (SDs) with moderate time resolution ($\geq 0.128$s). Subsequent work will use medium energy resolution data with fine time resolution from the BATSE Large Area Detectors (LADs) to examine weaker bursts and explore techniques using time-tagged counts from the SDs to improve time resolution.
The study of burst spectral evolution has a long history. Golenetskii et al. (1983, hereafter G83) examined two-channel data covering $\sim 40$-700 keV with $\sim$0.5 s time resolution from five bursts observed by the Konus experiment on [*Venera*]{} 13 and 14. Spectra were described by an optically thin thermal bremsstrahlung (OTTB) model, $N_E(E) \propto E^{-1}\exp(-E/kT)$ . A correlation between luminosity and temperature parameter $T$ was discovered ($L\propto T^\gamma,~\gamma\approx$1.5-1.7), implying spectral hardness at a given time is related to the intensity of the burst in a simple way. Laros et al. (1985) performed a similar analysis using five bursts observed by the [*Pioneer Venus Orbiter*]{} ($\sim0.1$-2 MeV, $\sim$0.2 s time resolution) but found no correlation and speculated that the G83 results were an artifact of the way temperature was inferred.
Norris et al. (1986, hereafter N86) investigated ten bursts seen by instruments on the [*Solar Maximum Mission*]{} satellite using hardness ratios (the ratio of observed flux in two energy bands). The energy bands used in the hardness ratio were 52-182 keV and 300-350 keV and the time resolution ranged over 0.128-1s, depending on the data available. They found that individual intensity pulses evolved from hard-to-soft with the hardness peaking before intensity. This implies a more complex relationship between hardness and intensity than proposed by G83.
Recent studies of spectral evolution have tended to follow the lead of G83 or N86, obtaining similar results. Kargatis et al. (1994, hereafter K94) fit OTTB and thermal synchrotron models ($N_E(E) \propto \exp[-(E/E_c)^{1/3}]$ ) to sixteen bursts from the SIGNE experiment, covering $\sim$50-700 keV with 0.5 second resolution. A luminosity-temperature corelation was found in seven of the bursts ($\gamma\approx$ 1.4-3.0) but two clearly did not show this correlation (the remainder were questionable) indicating that while many bursts may have a luminosity-temperature correlation, the trend is not universal. On the other hand, Band et al. (1992a) analyzed nine bursts observed by the BATSE SDs ($\sim$25-1000 keV, $\geq$0.128s) using models of the form $N_E(E) \propto E^\alpha \exp[-(E/kT)]$ , confirming the results of N86 including $T$ leading the intensity when fine time resolution was available. Bhat et al. (1994) used hardness ratios with BATSE LAD data to study single pulse bursts which had a fast rise followed by a gradual decay. The energy bands used for the ratio were $E$=25-100 keV and $E>$100 keV. The time resolution of the sample was 64 ms. Hard-to-soft spectral evolution was found with hardness leading the intensity. The time lag between the peak hardness ratio and peak count rate was found to be directly correlated with the rise time of the counting rate.
Although the analyses exemplified by G83 and N86 give apparently inconsistent results, they can be reconciled by considering the time resolution of the data. The analysis of G83 and K94 had a time resolution of $\sim 0.5$ s while the resolution of N86 was as small as $\sim 0.128$ s and Bhat et al. (1994) had 64 ms resolution. Therefore, slightly asynchronous behavior could be masked by the poor time resolution in the G83 and K94 samples. This conclusion is supported by the results of Band et al. (1992a), who performed an analysis similar to that of G83 and K94, but found that the spectral hardness (parameterized by $T$) led the intensity on short timescales. The need for fine time resolution is emphasized by Kouveliotou et al. (1992), who examined 22 bursts observed with the LADs and performed Fourier transforms on intensity profiles in different energy bands. The hard band was found to lead the soft by $\sim$0.1 second. In addition, Kouveliotou et al. (1994a) reported significant variations in the hardness ratio at the 2 ms level in an extremely intense burst observed with BATSE.
The spectral analyses of G83 and N86 characterize the spectrum in very different ways. Both G83 and K94 use spectral models which are assumed to describe the gamma-ray burst continuum and whose parameters are interpreted as physically meaningful. Most often, the model chosen is one which cuts off rapidly at high energies (e.g., OTTB). It was demonstrated in Paper I that such models do not reproduce spectra observed by the BATSE SDs and conflict with the observation of high energy emission by other experiments ([@matz85]; [@sch92]; [@han94]). Consequently, these models cannot be physically correct and though the parameters are certainly physically linked to the emission process, one must realize that the parameters do not necessarily have their apparent meaning (e.g., the OTTB cutoff energy may not be temperature). On the other hand, hardness ratios (such as used by N86) do not assume any knowledge of the source physics but do depend on instrumental properties which are accounted for in spectral fitting. While this characterization tracks changes in the shape and slope of a spectrum, the numerical values do not have a quantitative physical meaning and do not lend themselves to additional insights as a characteristic temperature or spectral shape can. Therefore it is helpful to use an empirical model which describes the photon spectrum but does not depend strongly on preconceived notions of spectrum formation in bursts.
In this work, a description of the photon spectrum independent of physical models is achieved by fitting the empirical spectral model from Paper I to BATSE SD data. This model was shown to adequately describe spectra in the BATSE energy range and incorporates many simple physical models as special cases, accommodating uncertainty about the actual continuum emission process. Although the model is phenomenological, physically meaningful parameters can be derived from it. In this work, (the energy of the peak of ) is used to quantify fitted spectra, a parameter which indicates the energy of maximum radiated power. This allows a study of spectral evolution in gamma-ray bursts based on a well-defined physical measure of spectral hardness which does not depend strongly on assumed emission processes.
In this work, the word hardness is used to describe the energy of . Spectra in which is at high energies are called hard and soft spectra have small . Before discussing the analysis, the SDs are briefly described ($\S$2). In $\S$3 the analysis techniques used are presented in detail along with several consistency tests for both the data and methods. The results are presented in $\S$4 followed by a discussion of their implications for both burst phenomenology and recently proposed theoretical models ($\S$5). The entire work is then briefly summarized ($\S$6).
The Instrument
==============
BATSE is a set of eight detector modules located on the corners of the [*CGRO*]{} spacecraft. Each module contains two detectors: an LAD, optimized for the detection and location of gamma-ray transients, and an SD, optimized for energy resolution. The SDs were designed for spectral studies and are used in this work because of their superior energy resolution to the LADs, which have large area but are thin (20 diameter by 0.5 thick) NaI(Tl) crystals.
The gamma-ray detector for an SD is a 5 diameter by 3 thick NaI(Tl) crystal. Because of their thickness, the response to incident gamma-rays is roughly constant over a large range of energies and viewing angles. The SDs independently cover two energy decades in the range $\sim$10 keV–100 MeV (the exact energy range is determined by a commandable gain setting, with higher gains covering lower energy ranges). The fractional full width at half maximum energy resolution of the SDs at 662 keV is $\sim$7%, with a $\sim E^{-0.4}$ dependence. A 3 beryllium window in the front of the aluminum case containing the NaI allows the response to extend down to $\sim$5 keV for face-on bursts. In general however, the lower level discriminator is set so that the spectrum is cut-off below $\sim$10 keV in the highest gain setting. An electronic artifact near the low end of spectra (the Spectroscopy detector Low Energy Distortion \[SLED\], Band et al. 1992b) exists in several channels above the instrument’s low energy cutoff which can cause problems for spectral analysis. In this work, only data which are uncontaminated by this artifact are used. The energy deposited in the crystal is analyzed into 2782 linear channels which are rebinned on the spacecraft into 256 quasi-logarithmic channels for transmission to Earth. Despite this compression, the width of most of the compressed channels is significantly smaller than the detector resolution and little information is lost in this process.
BATSE enters a 4-10 minute long burst mode when the count rate in at least two LADs exceeds the background count rate by a predetermined significance threshold (usually 5.5$\sigma$) on one of three timescales: 64 ms, 256 ms, and 1024 ms. Several data types are collected in this mode. The type used in this study consists of high energy resolution SD spectra with accumulation time based on a time-to-spill criterion in the LADs (Spectroscopy detector, High Energy Resolution, Burst, called SHERB). When the total number of counts in the LADs exceeds a pre-determined threshold, the current SHERB accumulation ends and the next begins. A total of 192 spectra are recorded from the SDs associated with the four most brightly illuminated LADs. Half of these spectra are from the SD associated with the brightest LAD, one-quarter with the second brightest, and one-eighth each for the other two detectors. The length of the accumulation is a multiple of 64 ms with a minimum accumulation of 128 ms. Spectra are accumulated until either all spectra are exhausted or the burst mode ends. The total time interval covered by each detector is equal: in the time required to accumulate one spectrum in the third and fourth brightest detector, four spectra from the brightest and two spectra from the second brightest detectors are recorded. Additional details about the BATSE detectors can be found in Fishman et al. (1989) and Horack (1991).
Analysis
========
It was shown in Paper I that the following empirical model is sufficient to describe burst spectra: $$N_E(E)~\biggl({{\rm photons}\over\hbox{keV-s-cm}^2}\biggr)=
\cases{A\biggl({E\over{100 {\rm ~keV}}}\biggr)^\alpha
e^{-E/E_0},&$E \le (\alpha-\beta)E_0$\cr
A^\prime\biggl({E\over{100 {\rm ~keV}}}\biggr)^\beta,&$E>(\alpha-\beta)E_0$\cr}$$ where $A,~\alpha,~\beta,$ and $E_0$ are fit to observed spectra and $A^\prime$ is chosen so that the function is continuously differentiable everywhere. The success of this model can be traced to its flexibility. Included in equation (1) as special cases are a power law, photon exponential, OTTB, and a broken power law with continuous transition. Since the emission process is unknown, equation (1) is a logical choice for determining physically meaningful parameters. The quantity of interest in this study is the energy of the peak of () and since $N_E(E)E^2\propto \nu F_\nu$, $=(2+\alpha)E_0$ (assuming $\beta < -2$).
Burst data were selected from detectors in high gain states, most covering $\sim$15-1500 keV with about a third extending to higher energies. This choice was motivated by the results of Paper I which demonstrated that for most bursts the spectral slope changes at energies less than 1 MeV and that the observed signal is greatest around 100 keV (a detector dependent property, not necessarily true for the incident spectrum).
Bursts which had many high quality spectra were required so that could be reliably determined and meaningful trends discovered. However, most bursts were not sufficiently long or intense to provide enough spectra to resolve the evolution. To understand the necessary signal strength, a Monte Carlo simulation of a typical spectrum was performed to investigate the reliability of the determination of as a function of the signal-to-noise ratio (S/N) in the 60-200 keV energy band. The results of this simulation are given in Table 1 which shows that a large S/N level is required to reduce the variance in to reasonable values.
Even for many strong bursts, most individual SHERB spectra do not have large S/N levels. However, the quality of the signal can be improved by averaging spectra together weighted by livetime. An explicit demonstration of this process for burst 2B910807 is given in Figure 1. In this figure, as determined from spectra at the finest time resolution available and from averaged pairs of spectra are plotted together. S/N in the 60-200 keV band ranged from 7.0-17.0 with a median of 11.7 for the fine time resolution data and 9.7-23.1 with median 16.3 for the averaged spectra. It is apparent that not only do the estimated errors in the determination of get smaller as spectra are accumulated, but that the fitted value converges to the average of determined from the fine resolution spectra.
After selecting candidate bursts, background models were created using a channel-by-channel fit of a time dependent polynomial to data before and after the burst. The order of the polynomial varied from burst to burst; typically the lowest order which gave an acceptable fit was used (never higher than fourth order). After subtracting the background, consecutive spectra were averaged together until S/N$\geq$15 in the 60-200 keV energy band. Although Table 1 shows that this choice will result in a sizable dispersion in , a larger S/N level would have essentially destroyed the time resolution in most of the sample. Therefore, some certainty in was sacrificed for time resolution. Exceptions to this S/N criterion were made for a few bursts in which a long period of low activity was bounded by interesting pulse structure. In those cases, all data in between the intense periods were averaged despite a low S/N value. An exception was also made for burst 2B910813 in which the data did not extend below 100 keV. The S/N criterion for this burst was S/N$\geq$12 in the 100-200 keV band.
A burst was included in the sample if it survived the process outlined above with at least six spectra. Since most bursts do not have many spectra with large S/N, only $\sim$4% of all bursts observed by BATSE and 17 of the 54 bursts from Paper I were contained in the sample (this sample includes bursts which occurred after those in Paper I). The empirical model (eqn. \[1\]) was fitted to data using forward folding deconvolution ([@lor89]). This method determines the incident photon spectrum by folding an assumed photon spectrum through a model of the detector response and comparing the result to observed data. The best fit was found using a modified version of the Levenberg-Marquardt iterative $\chi^2$ minimization algorithm ([@bev69], p. 237; [@pre89], p. 521). The modifications were the use of model variances instead of data variances in $\chi^2$ (see the appendix for details) and stopping criteria based on $\chi^2$, $\chi^2$ per degree of freedom in the fit ($\chi^2_\nu$), and a test for false minima. The detector response model included the direct component along with scatter off the spacecraft ([@pen89]) and the Earth’s atmosphere ([@pen92]). The error in was determined through standard error propagation using the error in $\alpha$ and $E_0$ and the correlation between these two parameters ([@bev69], p. 61).
Each spectrum was fit over the entire energy range above the SLED. The data were not rebinned from the telemetered format so that in most cases the average number of counts per bin in the high energy channels was less than a few. In these channels, the Poisson statistical distribution is not well approximated by a normal distribution and the $\chi^2$ statistic is therefore not strictly applicable. This is a common difficulty in high energy astrophysics which can be avoided by binning the data until the Gaussian limit is reached. To determine the importance of this problem, a set of Monte Carlo simulations was performed for several types of binning. The results for four different binning modes in Table 1 show that the fits are not improved by rebinning the data even if most high energy bins have only a few counts. There are two reasons for this. First, the data at higher energies are so sparse that when the Gaussian limit is achieved, the bins are so wide that the upper power law in eqn. (1) cannot be accurately determined. Second, the fit is largely determined by the signal below a few hundred keV where the number of counts per channel is largest (Max(C) in Table 1). The Gaussian limit can also be achieved if there are a large number of background counts, which can be important at low energies.
These simulations demonstrate that for the SHERB data type used in this sample, any binning is valid. To ensure that this was the case for real data, fits using the telemetered binning and eight broad energy bins were made to spectra from burst 3B930916. The comparison of the values of is presented in Figure 2. is consistent between the two fits and determined with roughly equal accuracy. Since the choice of data binning is unimportant, the telemetered channels were not rebinned because the analysis software was optimized for this scheme.
One difficulty with using high gain detectors (where the spectrum extends to low energy) was that occasionally the upper spectral index $\beta$ was larger than $-2$ which implies that was beyond the upper end of the fit range, that the spectrum has more gradual curvature than the empirical model could accomodate, or that the high energy signal was insufficient to fix $\beta$. For this work, $\beta$ was constrained to be less than $-2$ for all fits. In bursts for which this constraint was important (implying a flatter spectrum), it was found that either $\beta$ tended to shift rapidly from the upper constraint to the lower ($\beta=-5$) or that it remained at the upper constraint ($-2.01$) consistently. In the former case, the high energy spectrum was probably not well-determined and should be ignored when calculating . For the latter, should reflect changes in the shape (i.e. hardness) of the spectrum since $E_0$ will move to higher energies to compensate for the inability of $\beta$ to adequately describe this region. No bursts were observed to show convincing evidence for evolution of $\beta$ from above $-2$ to lower values. Therefore, variations in the fitted value of should be consistent with the true variations in spectral hardness if quantitatively inaccurate (with systematically smaller values of ). Omitting the upper power law would force to higher energies as the model accounts for power at the high end. However, this simpler model may overestimate , placing it far above the detector’s energy range where the model is not as sensitive to changes in . As a subset of the empirical four parameter model of equation (1), the model without $\beta$ is not a more accurate description of the true photon spectrum. Since evolution of is the primary focus of this study, equation (1) with $\beta$ constrained is used to model all spectra since it remains sensitive to hardness variations over a broader energy range than a model which cuts off at high energies.
To test whether changes in the fitted values were indeed consistent with real variations in for cases where $\beta>-2$, detectors at lower gain were used to determine for bursts in which the constraint was important. Unfortunately, in all but one case the high energy signal was too weak to reliably fit with any time resolution. Figure 3 shows a comparison of the high and low gain detectors for the exception (2B910503). It is apparent from this figure that not only do the two detectors have similar time histories, but that determined by the high gain detector is not far from the value of measured by the low gain detector (for which $\beta < -2$). Therefore, it is not expected that qualitative statements about the evolution of depend strongly on the accuracy with which $\beta$ is determined.
That the low gain detector measures a consistently larger value of than the high gain detector in the previous example suggests the calculated value of may depend on gain setting even though is well defined in all detectors. To ensure that this was not the case, the value of from two detectors at different gains viewing burst 2B921207 were compared ($\beta<-2$ for both detectors). In this case, spectra accumulated by the high gain detector were averaged so that the time intervals covered by both sets of spectra were identical. The comparison in Figure 4 shows that the determination of was consistent.
The tests outlined above cover the obvious difficulties involved in the analysis. Although solutions to the S/N and $\beta > -2$ problems are not ideal, they do allow the analysis to proceed so that useful information about burst continua can be derived. Also, the inter-detector test indicates that SD observations are consistent with each other and the results should not depend on the variable energy range covered by the SDs.
Results
=======
Of the 862 gamma-ray bursts detected by BATSE until the end of 1993, only the 37 bursts listed in Table 2 met the criteria outlined above. As expected, equation (1) fit the data well, $\chi^2_\nu \lesssim 1$ for nearly all spectra. For these bursts, was compared to the count rate (see Figure 5). Two bursts which clearly demonstrate the general trends observed in the sample as a whole are 2B921207 and 2B920525. In 2B921207, it is apparent that softens over the whole burst except for the increase along with the intensity 8 seconds after the trigger. 2B920525 is a burst in which two strong intensity pulses were resolved. softens after intial hardening within both spikes and the second spike is softer than the first.
The plots of Figure 5 were used to characterize types of evolution in . The classification categories used were motivated by the work of G83 and N86 in which a spectral characteristic (i.e., OTTB temperature and hardness ratio respectively) was tied to changes in burst intensity. These categories describe both the tendency for variations in to reflect changes in intensity (G83) and hard-to-soft evolution (N86). The categories are:
1. An increase in occurring in proximity to a major intensity increase. and intensity need not be morphologically identical and can lead or lag the intensity increase by a small amount ($\lesssim$ 1 s).
2. General softening of in time outside of intensity pulses over the entire burst. Regions which do not have a spike-like intensity profile are compared for this category. Spikes are considered to be short time intervals (less than a few seconds) in which the count rate was much greater than the rate in nearby intervals. For broader pulses, the first few seconds after the maximum count rate was achieved are considered part of the pulse, the remainder was considered inter-pulse emission.
3. Softening of within intensity pulse structures. should be harder early in the intensity pulse for a burst to fall in this category. It can be measured only if the pulse is temporally resolved.
4. Later intensity pulses have a softer peak in than earlier pulses. This can only be measured if the burst has multiple intensity spikes.
Bursts were judged as showing the property in question, showing an opposite trend, or showing neither trend ($+$, $-$, and $0$ respectively in Table 2). To demonstrate how a burst was categorized, consider 2B921207. This burst satisfies the first category because hardens as the intensity increases at 0.5 and 8 seconds after the trigger. The interpulse region lies between 2.5 and 8 seconds, where clearly softens. Both intensity spikes (0.5-2.5 seconds and 8-10 seconds after trigger) are temporally resolved and softens in each. Finally, the peak is softer in the later spike than in the first.
The rating system used here is admittedly simplistic, but is preferable to schemes which allow for finer distinctions within a class (e.g., rating bursts on a scale of 1-10, either discretely or continuously) for two reasons. First, the sample size is small so that the subdivision of bursts into several discrete subcategories reduces the significance of the classification. Second, the mediocre time resolution for some time intervals and sizable errors for some create problems for a continuous classification scheme. Therefore, although the criteria outlined above are subjective, they are the least problematic given the state of the observations.
Individual bursts are detailed in Table 2 and characteristics for the sample as a whole are given in Table 3. Included in Table 2 are the median values of and the range in \[$\equiv$max()-min()\] for each burst. Not all bursts could be classified because of large errors in , unresolved structure, lack of structure, poor time resolution, etc. Bursts in which $\beta$ was fit consistently at the upper constraint are noted in Table 2 as are bursts which were simultaneously observed in the 1-30 MeV range with the COMPTEL instrument on [*CGRO*]{} ([@han94]). In all four cases, the fitted $\beta$ was consistent with power law fits to the COMPTEL data in that for bursts in which COMPTEL reported a power law of index $>-2$, $\beta$ was consistently fit to the upper constraint. It can be seen from Figure 5 that there are several instances in which was poorly fit resulting in a large error ($\sigma_{E_P}$) for this quantity. Spectra for which $\sigma_{E_P}\gtrsim$0.5 were ignored when characterizing bursts and calculating $\Delta E_P$.
Table 2 clearly demonstrates that increases in are associated with rises in intensity and examination of Figure 5 shows that in many cases, the peak in leads the intensity peak. A quantitative analysis of leading intensity was not performed since Kouveliotou et al. (1992) found the typical lead time is shorter then the resolution of this sample. In addition, the time profile is extremely complex on short timescales so that a detailed study would best be performed using data with better time resolution. This data can be obtained from the LADs and the analysis is currently being performed (Kouveliotou et al. 1994b). Table 3 shows that many bursts undergo hard-to-soft evolution both within and outside of intensity spikes as well as between successive pulses. Softening of successive spikes is explicitly demonstrated in Figure 6, where the peak in for early intensity spikes is compared to the same quantity in later spikes in multi-spike bursts. It can be seen that late pulses are more likely to be softer than those which occurred previously, suggesting that the maximum possible value for is bounded by an envelope decaying in time. The trends noted above suggest that spectral hardness leads the intensity modulated by a decay envelope.
It should be mentioned that there are bursts in the sample which do not share the traits outlined above. This shows that while typical burst characteristics may have been identified, they are not universal on the timescale at which this sample was studied. However, it is not unreasonable to expect that leading intensity and the softening trends noted above will be found in most bursts when observed at finer time resolution.
The general softening trends in imply that the time at which emission occurs relative to the beginning of the burst is an important factor in spectral evolution. This raises an interesting question: assuming that the maximum values of are bounded by an envelope decaying in time, then if a burst has a precursor several seconds before most of the emission occurs, does this envelope begin to decay when the precursor happens or is the precursor irrelevant with respect to spectral evolution? To examine this question, the range in and the time at which significant emission occurred relative to the BATSE trigger were compared (Figure 7). Significant emission was defined as emission capable of producing spectra which met the S/N criterion outlined in $\S$3. In Figure 7 the horizontal bar indicates the period of significant emission. It begins at $t_d$, the time delay between the instrument trigger and the beginning of significant emission, and has length equal to the duration of significant emission ($\Delta T$). A large value of $t_d$ results when a long period of time passes between the trigger (a burst precursor) and the brightest emission. The data in Figure 7 occupy a triangular region indicating the sample was devoid of bursts with hard, highly evolving spectra which began emitting significantly well after the BATSE trigger: the few bursts with large $t_d$ have small values of and . To determine the importance of these bursts, the sample was divided into four groups based on the values of $t_d$ and . Since there are no clear groupings in Figure 7, the divisions were $t_d=0$ (significant emission at trigger) and =400 keV, which each divide the sample in half, producing an unbiased division. As shown by Table 4, bursts with $t_d \ne 0$ do tend to evolve less, but the significance of this result is small ($\sim 1\sigma$), implying that the apparent $t_d-$ association may be a statistical aberration which will vanish when more bursts with large $t_d$ are observed.
For several bursts, the nonzero time delay was short compared to the duration of significant emission. This suggests that the measure of time delay should be the ratio of $t_d$ and $\Delta t$. This definition is preferable to the absolute measure since it incorporates the timescale of the event from the trigger to the end of significant emission, allowing comparisons based on the time the burst began its most active stages and the total length of the burst. Bursts were divided into two groups based on the value of $t_d/\Delta t$. Table 5 shows the number of bursts in each group for several values of this ratio as well as the median and extreme values of for each group. The table shows that late emitters tend to show less variability than bursts which emit promptly after the trigger. Although this is very suggestive, there is sizable overlap between the groups. Considering the small size of the sample, this result is not too significant and therefore the possibility of a dependence of on $t_d$ should be treated with some caution.
Discussion
==========
In this work it was shown that for most gamma-ray bursts, the energy of the peak of is associated with and sometimes leads the intensity of the burst modulated by an envelope decaying in time. Indications that bursts whose main emission occurs at relatively long times after the trigger are generally soft and tend to evolve less in were also found. In this section the phenomenological implications of these results are discussed. Some recently proposed burst models are also examined in light of the observations reported here and no qualitative conflicts between the models considered and observations are found.
Phenomenological Implications
-----------------------------
While it has been known for some time that burst spectra tend to evolve from hard to soft (N86), the observations presented here show that the time of emission relative to the beginning of the burst may also determine how the spectral hardness evolves. For optically thin burst models (in which particles radiate on timescales much shorter than the resolution of SHERB spectra), these results imply that as the burst transpires, the emission region ‘remembers’ previous emission suggesting that either emission comes from only one region or that different emission regions must be physically connected. The softening of successive intensity pulses could be caused by significant change in the emitting region during intense episodes (such as an expansion or an increase in the number density of particles). The softness of delayed emission implies that the emission region can evolve without observable radiation. For optically thick models, a burst could be the result of a single energy input since the radiation does not escape the region on short timescales. If the region cools by adiabatic expansion and the surface of last scattering does not propagate inward to higher temperature regions too quickly, then observable radiation would be expected to evolve from hard-to-soft. The softness of delayed emission could also be explained in this way if the region was not able to emit significantly during the initial stages of expansion.
The hard-to-soft spectral evolution within pulse structures indicates that intensity spikes are not symmetric in time. Nemiroff et al. (1994b) reached a similar conclusion by examining the time structure of intensity profiles. These observations clearly show that time symmetric models of spectra formation as well as overall intensity profiles cannot be correct. An example of the kind of model eliminated by these observations is symmetric beam models in which the beam sweeps by much faster than the timescale over which the beam changes.
Another implication of this work relates to burst duration classes. The distribution of gamma-ray burst durations observed by BATSE is bimodal and short bursts (duration $\lesssim 2$ s) tend to have larger hardness ratios than longer ones ([@kou93]; [@lam94]) where the hardness ratios were averaged over the whole burst. Based on the large values of reported here and the fact that the hardest emission tends to come early in a burst, short bursts are expected to be harder than long ones since they do not have softer tails, not necessarily because of an inherent difference in the radiation mechanism. An example of the effect of time averaging is burst 2B920525. If spectra from 4.16-6.72 s (the first spike) are used =650$\pm$70 keV but falls to 421$\pm$25 keV if the entire burst is integrated into one spectrum. As Norris et al. (1994) also suggested, if hardnesses are measured over the same absolute time interval (e.g., compare short bursts to the first few seconds of long bursts), the hardness distribution may be the same for both classes.
In Paper I it was shown that varied considerably from burst to burst indicating that there is no universal characteristic energy of burst emission. One conclusion which was drawn from this was that pair processes could be directly observed only if burst sources had a broad range of redshifts. Such processes would produce spectra with an annihilation feature at 511 keV and would cut off above this energy because of pair opacity. The observed energy of these features depends on the redshift of the source region. The argument against observable pair processes can be extended by noting the large changes in observed within many bursts in this sample (Table 3). Unless the emission process involves large and rapid changes in redshift, pair processes cannot be directly observed in burst spectra.
The data in Figure 7 include several bursts in which the emission was prompt but had small . An interesting possibility is that these prompt emitting, relatively non-evolving bursts were preceded by a precursor too faint to trigger BATSE. In this case, bursts which occupy the lower left-hand corner of the figure are late emitters for which BATSE did not trigger on the initial event. A less exciting possibility is that bursts which emit late but evolve considerably exist and there is no relation between emission time and . Although this type of behavior is not seen, it cannot be ruled out because of the small sample size and rarity of late emitters (about one strong burst per year with $t_d>$60s). Therefore, this may remain an open question even if BATSE detects many bursts over the next few years.
Implications for Burst Models
-----------------------------
A plethora of gamma-ray burst scenarios have been proposed ([@nem94a]). However, most attempt to explain only general spectral characteristics and energetics with little or no mention of how spectra are expected to evolve. Therefore, several models are examined to find those which can accommodate the observations presented here. Because BATSE observations of burst isotropy and inhomogeneity have cast serious doubt on older models, only those proposed after 1991 (the year BATSE was launched) are considered. Discussion is further restricted to models which purport to explain all or most gamma-ray bursts. Although arguments have been made for two population distributions ([@lin92]; [@smi93]), the inability to separate bursts into two uncontroversial classes makes such discussion premature. Jet models are also avoided (c.f. [@bri92]; [@der94]) since the phenomenon is poorly understood and the models are incomplete. ‘Exotic’ sources such as cosmic strings ([@pac88]), strange stars ([@hae91]), or primordial black holes ([@cli92]) are also omitted since the existence of such sources is highly speculative.
The distribution of bursts seen by BATSE has forced bursts out to extended halo or cosmological distances if they arise from a single source population. Since source regions are inferred to be small from the fast rise times of bursts, the energy density at the source must be enormous, particularly if the burst radiates isotropically. Therefore, many models incorporate optically thick relativistic electron-positron fireballs. As originally proposed ([@good86]; [@pac86]), fireballs cannot explain the observed nonthermal emission and will not be relativistic if the fireball is contaminated by baryons ([@cav78]; [@pac90]).
Many different scenarios have been proposed recently (c.f. [@nar92]; [@usov92]; [@woo92]) which can essentially be reduced to a fireball which must either form in a region free of baryons or break through a cloud of baryons to be observed. In these cases, the observed nonthermal spectra are produced either by several thermal blackbodies at different temperatures (corresponding to different fireballs) or through interactions with a strong or turbulent magnetic field. If a number of fireballs occur, then hard-to-soft spectral evolution could be explained by the creation of low temperature fireballs or rapid cooling as the burst transpires. For magnetic interaction models, the general softening of may reflect a gradual decline in available energy as the field evolves into a stable configuration.
Mészáros & Rees (1992) have proposed a model involving a baryon contaminated fireball from any source, galactic or cosmological. The baryons in the fireball interact with the interstellar medium (ISM) to form a relativistic shock where gamma-rays are radiated. Hard-to-soft evolution could be explained either by the most energetic baryons reaching the shock first or the gradual decline of average energy as more particles encounter the shock. The softness of late emission could be a consequence of the baryon front losing energy to the ISM before the shock is fully formed or could arise from the shock losing energy as it propagates through the ISM.
In a different type of model suggested by Melia & Fatuzzo (1992), a gamma-ray burst is generated by sheared Alfvén waves from a radio pulsar. The waves are created by crustal disturbances at the polar cap which flood the magnetosphere with charged particles. These charges upscatter ambient radio photons to gamma-ray energies which are beamed outward. Hard-to-soft evolution may result from the most energetic charges scattering first and successive spikes soften if subsequent disturbances are less energetic than the original disturbance. Since the emission is beamed, the hard emission might occur when the beam was not pointed in our direction. Delayed emission might be soft because the early hard emission was not beamed in our direction.
Summary
=======
In this work, the evolution of the energy of the maximum flux per logarithmic energy band () in long, bright gamma-ray bursts was studied using high energy resolution spectra from the BATSE SDs. It was found that is coupled with the intensity of a burst (either leading or accompanying intensity increases) but is apparently modulated by an envelope decaying in time. These results are consistent with previous studies by N86 and Bhat et al. (1994). Indications were also found that bursts in which significant emission is substantially delayed relative to the instrument trigger tend to be soft and evolve less in although the significance of this result is low. Several burst models were examined in light of these results and it was found that none of those discussed could be ruled out on the basis of these results alone. However, as the models mature to the point of more precise spectral predictions, the observations presented here may prove to be a useful constraint or suggest directions for further effort.
We thank D. Gruber, J. Higdon, R. Lingenfelter, and R. Rothschild for helpful advice and useful information. D. Marsden and D. Potter provided additional assistance. BATSE work at UCSD is supported by NASA contract NAS8-36081.
Model Weights for $\chi^2$
==========================
The Levenberg-Marquardt algorithm (Bevington 1969, p. 237; Press et al. 1989, p. 521) is a prescription to efficiently minimize goodness-of-fit statistics. The most common statistic used is $$\chi^2=\sum^N_{i=1}\biggl[{y_i-y(x_i;{\rm\bf a})\over\sigma_i}\biggr]^2,$$ where: $x$ is the independent variable (here energy); $y$ is an observed variable (count rate in this case); $y(x_i;{\rm\bf a})$ is the value of the model at $x_i$ for parameters [**a**]{}; $N$ is the number of measurements (here, channels in a spectrum); [**a**]{} represents the parameters to be fit; and $\sigma_i$ are the errors associated with $y_i$. $\chi^2$ is minimized using an iterative gradient procedure for which the first and second partial derivatives of $\chi^2$ with respect to [**a**]{} are needed. In its usual implementation, the algorithm assumes that the $\sigma_i$ are taken directly from the data and are constant.
The problem with using data variances ($\sigma_i$) in equation (A1) is that $\sigma_i$ weights downward fluctuations in the data too strongly, a result of the Poisson nature of count data in which $\sigma_i^2\propto n_i$, the number of observed counts. In this case, a small number of counts create a smaller than average $\sigma_i$ which causes the associated data points to be more important in the calculation of $\chi^2$. Wheaton et al. (1994, hereafter W94) examined this problem in detail using a Monte Carlo simulation of 24,000 one-parameter data sets and found that the mean of fitted values deviated from the true value by $-170\sigma$ (note the downward shift). The solution to this problem is to use $\sigma_i$ generated from the model under consideration rather than the data, replacing $\sigma_i^2\propto n_i$ with $\sigma_i^2(x_i;{\rm\bf a})\propto n_i(x_i;{\rm\bf a})$ (model variance). When W94 performed the simulation with this new scheme, the bias was $-0.34\sigma$.
Although W94 demonstrated the importance of using weights derived from the model, they do not mention a necessary change in the standard Levenberg-Marquardt prescription to accommodate the new weighting. The algorithm assumes that $\sigma_i$ is taken from the data so that changes in the $\chi^2$ surface are contained in only the $y(x_i;{\bf a})$ term in equation (A1). However, by making $\sigma_i$ model dependent without changing the partial derivatives, assumed changes in the $\chi^2$ surface caused by variation in [**a**]{} will be incorrect, resulting in slightly off-course iterations and an incorrect minimum since $\partial \chi^2/\partial$[**a**]{}=0 can be at the wrong value of [**a**]{}. It was found that when this effect is not accounted for, the true minimum of the $\chi^2$ surface is not reached and the values of the fitted parameters differ from the true best fit by roughly the quoted error in the fitted parameters though it is not biased toward higher or lower values making it difficult to detect in large collections of fits.
Alterations in the second derivative terms only affect the route used to find the minimum of $\chi^2$ and therefore need not be corrected for the new weighting scheme ([@pre89], p. 523). On the other hand, the first partial derivatives need to reflect any changes which might occur in the $\chi^2$ surface and must be corrected. The corrected version is $${\partial\chi^2\over\partial a_j} = -2\sum^N_{i=1}\biggl[{y_i-
y_i(x_i;{\rm\bf a})\over\sigma_i^2(x_i;{\rm\bf a})}\biggr]
\biggl({\partial y_i(x_i;{\rm\bf a})\over\partial a_j}-{[y_i-
y_i(x_i;{\rm\bf a})]\over\sigma_i(x_i;{\rm\bf a})}{\partial
\sigma_i(x_i;{\rm\bf a})\over\partial a_j}\biggl).$$ The first partial derivative term in this equation is equal to the partial derivative specified by the Levenberg-Marquardt algorithm with model variances replacing data variances but the second term is new. For this work, equation (A2) can be expressed in a simpler form since background subtracted count rates are used ($y_i=n_i/f_i-b_i$ where $f_i$ is a term which converts counts to a count rate and $b_i$ is the background counting rate). In this case, $\sigma_i^2(x_i;{\rm\bf a})$ becomes $$\sigma_i^2(x_i;{\rm\bf a})={y_i(x_i;{\rm\bf a})\over f_i}+\sigma^2_{b_i},$$ where $\sigma^2_{b_i}$ is the error in the determination of the background rate in channel $i$. Differentiating (A3) with respect to [**a**]{} and substituting into equation (A2), it is found that $${\partial\chi^2\over\partial a_j} = -\sum^N_{i=1}[y_i-y_i(x_i;{\rm\bf a})]
{\sigma_i^2+\sigma_i^2(x_i;{\rm\bf a})\over\sigma_i^4(x_i;{\rm\bf a})}
{\partial y_i(x_i;{\rm\bf a})\over\partial a_j}.$$ This expression is equal to the uncorrected expression when $\sigma_i(x_i;{\rm\bf a})$ is replaced by $\sigma_i$ which allows the standard algorithms (e.g. Bevington 1969, p. 237; Press et al. 1989, p. 521) to be modified easily. When this expression is substituted into the algorithm, the true minimum of the $\chi^2$ surface with model variances is found.
[rcccccc]{} nobin & 0.1 & 0.23 & 4.66 & 259 & 286 & 10.0 & 0.25 & 0.99 & 7.78 & 285 & 77 & 15.6 & 0.5 & 2.11 & 16.48 & 265 & 50 & 21.6 & 1.0 & 4.41 & 30.31 & 251 & 36 & 29.6 & 2.0 & 9.70 & 43.72 & 250 & 21 & 39.5 half & 0.1 & 1.18 & 9.52 & 271 & 176 & 10.0 & 0.25 & 2.78 & 26.85 & 259 & 85 & 15.6 & 0.5 & 4.66 & 45.71 & 256 & 56 & 21.6 & 1.0 & 9.05 & 86.85 & 252 & 39 & 29.6 & 2.0 & 21.50 & 219.33 & 251 & 22 & 39.5 full & 0.1 & 2.46 & 15.47 & 252 & 153 & 10.0 & 0.25 & 4.95 & 36.03 & 244 & 68 & 15.6 & 0.5 & 8.55 & 72.05 & 253 & 50 & 21.6 & 1.0 & 19.30 & 141.77 & 251 & 35 & 29.6 & 2.0 & 45.47 & 321.04 & 250 & 21 & 39.5 eight & 0.1 & 3.29 & 10.52 & 220 & 86 & 10.0 & 0.25 & 8.54 & 24.35 & 240 & 63 & 15.6 & 0.5 & 17.67 & 46.76 & 247 & 46 & 21.6 & 1.0 & 34.85 & 102.28 & 250 & 40 & 29.6 & 2.0 & 73.27 & 195.23 & 248 & 25 & 39.5
[ccccccc]{} 2B910503$^{*\dagger}$ & $+$ & $+$ & $+$ & $+$ & 540 & 1350 2B910601$^\dagger$ & $+$ & $0$ & & & 700 & 9502B910807 & $+$ & $0$ & $+$ & $0$ & 170 & 175 2B910814 & & & & & 490 & 900 2B910814C$^\dagger$ & & $+$ & & & 1400 & 1300 2B911031$^*$ & & $+$ & & & 400 & 850 2B911106 & & $0$ & & & 165 & 120 2B911118$^\dagger$ & $+$ & $+$ & $+$ & $+$ & 180 & 440 2B911126$^*$ & $+$ & $+$ & $+$ & $0$ & 250 & 900 2B911127 & & & & & 190 & 250 2B911202$^*$ & $0$ & $+$ & $0$ & & 350 & 1300 2B911209$^*$ & $+$ & $+$ & $+$ & $+$ & 270 & 400 2B920218$^*$ & & $+$ & & & 220 & 150 2B920311 & $+$ & $+$ & $+$ & $+$ & 440 & 1150 2B920406$^*$ & $+$ & $0$ & $+$ & $0$ & 240 & 300 2B920513$^*$ & $+$ & & $+$ & & 220 & 200 2B920517$^*$ & $+$ & & $+$ & & 230 & 150 2B920525$^*$ & $+$ & $+$ & $+$ & $+$ & 380 & 2075 2B920617 & & $+$ & & & 145 & 175 2B920622 & $+$ & $0$ & $+$ & & 440 & 1550 2B920627$^*$ & & & & & 210 & 325 2B920711 & $+$ & $0$ & & $0$ & 510 & 650 2B920720 & $+$ & $-$ & $-$ & $-$ & 270 & 525 2B920902$^*$ & & $+$ & $+$ & & 550 & 1500 2B921009 & $+$ & $0$ & & & 300 & 225 2B921123 & $+$ & $+$ & & & 250 & 350 2B921207 & $+$ & $+$ & $+$ & $+$ & 140 & 650 2B921209 & $+$ & $+$ & $+$ & $+$ & 180 & 375 2B930120 & & & & & 125 & 80 3B930405 & & $-$ & & & 280 & 300 3B930425 & & & & & 250 & 225 3B930506 & $+$ & $+$ & $0$ & & 1030 & 1750 3B930916$^*$ & $0$ & $0$ & $0$ & $0$ & 390 & 350 3B930922 & $+$ & $+$ & $+$ & $0$ & 110 & 130 9 3B931103 & $+$ & $+$ & $+$ & $+$ & 380 & 1300 3B931126 & $+$ & $+$ & $+$ & $+$ & 240 & 300 3B931204 & $+$ & $+$ & $+$ & $0$ & 380 & 775
[rrl]{} 23 & 25 & – intensity association 20 & 30 & softens over whole burst 2 & 30 & hardens over whole burst 18 & 22 & softens within intensity spikes 1 & 22 & hardens within intensity spikes 9 & 17 & Later spikes softer than earlier ones 1 & 17 & Later spikes harder than earlier ones
[lcc]{} $t_d \ne 0$ & 10 & 7 $t_d = 0$ & 8 & 12
[rcccc]{} $>0.25$ & 10 & 80 & 300 & 950 $\leq 0.25$ & 27 & 120 & 525 & 2075 $>0.5$ & 8 & 80 & 300 & 950 $\leq 0.5$ & 29 & 120 & 525 & 2075 $>0.9$ & 7 & 80 & 300 & 950 $\leq 0.9$ & 30 & 120 & 525 & 2075 $>1$ & 5 & 80 & 300 & 950 $\leq 1$ & 32 & 120 & 440 & 2075
Band, D. L., et al. 1992, in [*Gamma Ray Bursts*]{}, eds. W. S. Paciesas and G. J. Fishman (New York: AIP), p. 169 Band, D. L., et al. 1992, Exp. Astron., 2, 307 Band, D. L., et al. 1993, , 413, 281 Bevington, P. R. 1969, Data Reduction and Error Analysis for the Physical Sciences, (New York: McGraw-Hill) Bhat, N. P., Fishman, G. J., Meegan, C. A., Wilson, R. B., Kouveliotou, C., Paciesas, W. S., Pendleton, G. N., & Schaefer, B. E. 1994, , 426, 604 Brainerd, J. J. 1992, , 394, L33 Cavallo, G. & Rees, M. J. 1978, , 183, 359 Cline, D. B., & Hong, W. 1992, , 401, L57 Dermer, C. D., & Schlickeiser, R. 1994, in [*Proc. of the Huntsville Gamma-Ray Burst Workshop*]{}, eds. G. J. Fishman, K. C. Hurley, & J. J. Brainerd (New York: AIP), in press Fishman, G. J., et al., 1989, in [*Proc. Gamma Ray Observatory Science Workshop*]{} (Washington: NASA), 2-39 Fishman, G. J., et al., 1994, , in press Golenetskii, S. V., Mazets, E. P., Aptekar, R. L., & Ilyinskii, V. N. 1983, Nature, 306, 451 Goodman, J. 1986, , 308, L47 Haensel, P., Paczyński, B., & Amsterdamski, P. 1991, , 375, 209 Hanlon, L. O., et al. 1994, , in press Horack, J. M. 1991, [*Development of the Burst and Transient Source Experiment*]{}, NASA reference publication 1268 Kargatis, V. E., Liang, E. P., Hurley, K. C., Barat, C., Eveno, E., & Niel, M., , 422, 260 Kouveliotou, C., Norris, J. P., Fishman, G. J., Meegan, C. A., Wilson, R. B., & Paciesas, W. S. 1992, in [*Gamma Ray Bursts*]{}, eds. W. S. Paciesas and G. J. Fishman (New York: AIP), p. 299 Kouveliotou, C., Meegan, C. A., Fishman, G. J., Bhat, N. P., Briggs, M. S., Koshut, T. M., Paciesas, W. S., & Pendleton, G. N. 1993, , 413, L101 Kouveliotou, C., et al. 1994, , 422, L59 Kouveliotou, C. et al. 1994, in preparation Lamb, D. Q., Graziani, C., & Smith, I. A. 1994, , in press Laros, J. G., et al. 1985, , 290, 728 Lingenfelter, R. E. & Higdon, J. C. 1992, Nature, 356, 132 Loredo, T. J. & Epstein, R. I. 1989, , 336, 896 Matz, S. M., Forrest, D. J., Vestrand, W. T., Chupp, E. L., Share, G. H., & Rieger, E. 1985, , 288, L37 Melia, F., & Fatuzzo, M. 1992, , 398, L85 Mészáros, P. & Rees, M. J. 1992, , 405, 278 Narayan, R., Paczyński, B., & Piran, T. 1992, , 395, L83 Nemiroff, R. J. 1994, Comm. on Astrophys., in press Nemiroff, R. J., Norris, J. P., Kouveliotou, C., Fishman, G. J., Meegan, C. A., & Paciesas, W. S. 1994, , 423, 432 Norris, J. P., Share, G. H., Messina, D. C., Dennis, B. R., Desai, U. I., Cline, T. L., Matz, S. M., & Chupp, E. L. 1986, , 301, 213 Norris, J. P., Nemiroff, R. J., Davis, S. P., Kouveliotou, C., Fishman, G. J., Meegan, C. A., & Paciesas, W. S. 1994, in [*Proc. of the Huntsville Gamma-Ray Burst Workshop*]{}, eds. G. J. Fishman, K. C. Hurley, & J. J. Brainerd (New York: AIP), in press Paczyński, B. 1986, , 308, L43 Paczyński, B. 1988, , 335, 525 Paczyński, B. 1990, , 363, 218 Pendleton, G. N., Paciesas, W. S., Lestrade, J. P., Fishman, G. J., Wilson, R. B., & Meegan, C. A. 1989, in [*Proc. Gamma Ray Observatory Science Workshop*]{} (Washington: NASA), 4-547 Pendleton, G. N., et al. 1992, [*The Compton Observatory Science Workshop*]{} (Washington:NASA), 47 Press, W. H., Flannery, B. P., Teukolsky, S. A., Vetterling, W. T. 1989, Numerical Recipes (New York: Cambridge) Schaefer, B. E., 1994, in [*Proc. of the Huntsville Gamma-Ray Burst Workshop*]{}, eds. G. J. Fishman, K. C. Hurley, & J. J. Brainerd (New York: AIP), in press Schneid, E. J., et al. 1992, , 255, L13 Smith, I. A., & Lamb, D. Q. 1993, in [*Compton Gamma-Ray Observatory*]{}, eds. M. Friedlander, N. Gehrels, & D. J. Macomb (New York: AIP), p. 1020 Usov, V. V. 1992, Nature, 357, 472 Wheaton, W. A., Dunklee, A. L., Jacobson, A. S., Ling, J. C., Mahoney, W. A., & Radocinski, R. G. 1994, , submitted Woosley, S. E. 1992, , 405, 273.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We report a Hitomi observation of IGR J16318–4848, a high-mass X-ray binary system with an extremely strong absorption of $N_{\rm H}$ $\sim$ $10^{24}$ cm$^{-2}$. Previous X-ray studies revealed that its spectrum is dominated by strong fluorescence lines of Fe as well as continuum emission. For physical and geometrical insight into the nature of the reprocessing material, we utilize the high spectroscopic resolving power of the X-ray microcalorimeter (the soft X-ray spectrometer; SXS) and the wide-band sensitivity by the soft and hard X-ray imager (SXI and HXI) aboard Hitomi. Even though photon counts are limited due to unintended off-axis pointing, the SXS spectrum resolves Fe K$\alpha_1$ and K$\alpha_2$ lines and puts strong constraints on the line centroid and width. The line width corresponds to the velocity of 160$^{+300}_{-70}$ km s$^{-1}$. This represents the most accurate, and smallest, width measurement of this line made so far from any X-ray binary, much less than the Doppler broadening and shift expected from speeds which are characteristic of similar systems. Combined with the K-shell edge energy measured by the SXI and HXI spectra, the ionization state of Fe is estimated to be in the range of FeI–IV. Considering the estimated ionization parameter and the distance between the X-ray source and the absorber, the density and thickness of the materials are estimated. The extraordinarily strong absorption and the absence of a Compton shoulder component is confirmed. These characteristics suggest reprocessing materials which are distributed in a narrow solid angle or scattering primarily with warm free electrons or neutral hydrogen. This measurement was achieved using the SXS detection of 19 photons. This provides strong motivation for follow-up observations of this and other X-ray binaries using the X-ray Astrophysics Recovery Mission, and other comparable future instruments.'
bibliography:
- 'mybibfile.bib'
title: 'Glimpse of the highly obscured HMXB IGR J16318–4848 with Hitomi [^1] '
---
Introduction
============
High-mass X-ray binaries (HMXBs) consist of a compact object (neutron star or black hole candidate) and a massive companion star that is typically a Be star or a supergiant O or B type star. HXMBs with Be companions often show periodic variability in X-ray flux when the compact object passes through a circumstellar decretion disk surrounding the star. Supergiant HMXBs exhibit X-ray time variability associated with eclipse, or partial eclipse, of the compact object by the companion star.
In addition to the comprehensive catalog of the galactic HMXBs by , a recent deep survey in the hard X-ray and soft gamma-ray band performed by IBIS/ISGRI onboard International Gamma-Ray Astrophysics Laboratory (INTEGRAL) has discovered a considerable number of HMXBs that are summarized in a catalog by @2017MNRAS.470..512K. More than half exhibit persistent time variability in the hard X-ray band [@2013MNRAS.431..327L]. One of the highlights of the survey is the discovery of a number of HMXBs that exhibit extraordinarily strong absorption with their distribution in the galaxy correlating with that of star forming regions [@2012ApJ...744..108B; @2013ApJ...764..185C]. IGR J16318–4848 (hereafter IGR J16318) was the first discovered and remains the most extreme example of such objects.
IGR J16318 was discovered in the scanning observation of the Galactic plane by the INTEGRAL/IBIS/ISGRI . Examination of archival ASCA data revealed extremely strong X-ray absorption toward the direction of the source [@2003IAUC.8070....3M]. The X-ray spectrum is dominated by Fe K$\alpha$, K$\beta$, and Ni K$\alpha$ fluorescence emission lines and continuum [@2003MNRAS.341L..13M; @2003AstL...29..644R]. The fluorescence lines as well as the continuum vary on time scales of thousands of seconds, corresponding to an upper limit on the emitting region size approximately $10^{13}$ cm .
The optical/near-infrared (NIR) counterpart exhibits less absorption than that measured in the X-ray band, which implies that the absorbing material is concentrated around the compact object . The NIR spectroscopy suggests that the counterpart is a supergiant B\[e\] star [@2004ApJ...616..469F] based on the detection of forbidden lines of Fe. Such stars are also known to contain dust in their envelopes [@2007ApJ...667..497M]; a mid-infrared observation revealed that it is surrounded by dust and cold gas with a heated inner rim [@2012ApJ...751..150C]. The distance to the target was derived by @2004ApJ...616..469F based on fitting of the optical/NIR spectral energy distribution (SED) fitting to be 0.9–6.2 kpc. performed SED fitting from optical to mid-infrared band, and utilizing the known stellar classification of the companion star obtained a distance of 1.6 kpc.
Long term monitoring of the hard X-ray flux with Swift/BAT shows a periodicity of $\sim$ 80 d [@2009RAA.....9.1303J; @Iyer17]. Although the companion star belongs to the spectral type of B\[e\], there is no obvious coincidence between numbers of outbursts and orbital phase [@2009RAA.....9.1303J]. Monitoring in the soft and hard X-ray band shows that the source is always bright with flux dynamic range of a factor 5 and Compton thick ($N_{\rm H}$ $\ge$ 1.1 $\times$ $10^{24}$ cm$^{-2}$) [@2010int..workE.135B]. The statistically best spectrum obtained with Suzaku [@2007PASJ...59S...1M] shows no Compton shoulder, which implies a non-spherical and inhomogeneous absorber . The average X-ray spectrum of the source exhibits a continuum typical for neutron stars [@2004ESASP.552..417W]. Moreover, the source shows disagreement in its X-ray/radio flux relationship with that observed in the low/hard state of black hole binaries [@2004ApJ...616..469F]. Nevertheless, the nature of the compact source (neutron star or black hole candidate) is uncertain because pulsations have not been detected.
Hitomi, the Japan-led X-ray astronomy satellite [@Takahashi17], carried a microcalorimeter array (SXS; soft X-ray spectrometer) [@Kelley17] which had outstanding energy resolution in the energy band containing the Fe K-shell lines. Combined with an X-ray CCD camera (SXI; soft X-ray imager) [@Tanaka17] and a hard X-ray imager (HXI) [@Nakazawa17], it provided unprecedented wide-band imaging spectroscopy. Hitomi was lost due to an accident a month after the launch. The observation of IGR J16318 was performed during the instrument check-out phase to demonstrate the spectroscopic performance of Hitomi. In spite of offset pointing during the observation due to incomplete attitude calibration, it is possible to extract significant scientific results from the limited data.
In the remainder of this paper, we first describe the observation log including some notes on the data reduction in section \[sec:obs\]. The imaging and spectroscopic analyses (section \[sec:ana\]) are followed by the discussion (section \[sec:discussion\]) and summary (section \[sec:summary\]). Measurement errors correspond to the 90 % confidence level, unless otherwise indicated.
Observation and data reduction {#sec:obs}
==============================
Observation {#ssec:obs}
-----------
Pointing toward IGR J16318 started on 22:28 10th March 2016 UT and ended on 16:20 14th March 2016 UT. While the SXS and SXI were already in operation, the HXI was undergoing the startup procedure of one of the two sensors (HXI-1). Because the observation was performed before optimizing the alignment matrices of star trackers (STT1 and STT2), the target was at off-axis positions throughout the observation. The off-axis angle was according to the SXI image after the switch of the STT from STT1 to STT2 on 17:58 13th March, which limit the effective area of all the instruments. The fields of view (FoV) of the SXS and HXI are and square, respectively. Therefore only the SXI caught the target securely within its FoV thanks to its large FoV of square [@Nakajima17].
The microcalorimeter array in the SXS was already in thermal equilibrium at the time of our observation [@2016SPIE.9905E..3SF; @2016SPIE.9905E..3RN]. The energy resolution of the onboard radioactive $^{55}$Fe source was 4.9 eV full width half maximum (FWHM) as reported by @leutenegger17. However, the SXS was not in the normal operation mode in terms of some calibration items as follows. The gate valve was still closed and hence the effective area in the soft energy band was limited. The Modulated X-ray Source (MXS; [@devries17]) was also not yet available for contemporaneous gain measurement, which forces us to estimate the gain uncertainty only by onboard radioactive $^{55}$Fe sources. The SXI was in normal operation with the “Full Window + No Burst” mode [@Tanaka17]. Temperature of the CCDs was already stable at $-110^{\circ}$C at the time of the exposure [@Nakajima17b]. The observation was carried out before optimizing the parameters for the dark level calculation and hence the SXI suffered from a cross-talk issue. That is, an anomalously low dark level can be induced in a pixel by a charged particle event in the adjacent segment. The dark level leads to continuous false events in the pixel and the erroneously higher pulse heights for the normal events around the pixel. To minimize the effect of the cross-talk issue, the lower threshold of the effective energy band was set to be 100 ch, which corresponds to 600 eV.
The HXI-1 completed its startup procedure and started observation on 21:30 12th March UT. The target came at the edge of the HXI-1 FoV after the switch of the STT. Another sensor HXI-2 was still undergoing increasing of the high voltage for the Si/CdTe double-sided strip detectors.
Data reduction {#ssec:reduction}
--------------
Hereafter, we concentrate on the data after the STT switch because event files of all the three instruments are available in the interval. We utilize the data cleaned and processed with a script version 03.01.005.005. All the reduction and analyses below employ the Hitomi software version 5b and the calibration database released on 11th May 2017 [@Angelini17] [^2]. The effective exposure times of the SXI, SXS, and HXI-1 are 39.4, 68.9, and 39.4 ks, respectively, after the data reduction.
### SXS {#sssec:sxsreduction}
Owing to the shape of the point spread function (PSF) of the soft X-ray telescope (SXT-S; [@maeda17]), some photons from the target reached the SXS in spite of the off-axis pointing. Furthermore, there was a wobble of the satellite at the beginning of the observation, so that the optical axis of the SXT-S temporarily approached the target direction. Then a part of the FoV of the SXS overlapped with a photon extracting region for the SXI as shown in figure \[fig:sxs\_expimg\] top panel.
![(Top) SXS exposure map with the designation of the exposure time for each pointing position. The magenta circle corresponds to the source extraction region for the SXI (see figure \[fig:sxiregion\] bottom panel). (Bottom) Spatial event distribution of the SXS microcalorimeter array in the DET coordinate in the energy band from 6.38 to 6.42 keV. Blue, red, yellow and white pixels correspond to detection of one, two, three and four events, respectively. The black pixel at the bottom right is the calibration pixel that is not directly exposed to the sky. []{data-label="fig:sxs_expimg"}](ah100042040sxs_p0px1010_exp4fig_exp_rev.pdf "fig:"){width="\linewidth"}\
![(Top) SXS exposure map with the designation of the exposure time for each pointing position. The magenta circle corresponds to the source extraction region for the SXI (see figure \[fig:sxiregion\] bottom panel). (Bottom) Spatial event distribution of the SXS microcalorimeter array in the DET coordinate in the energy band from 6.38 to 6.42 keV. Blue, red, yellow and white pixels correspond to detection of one, two, three and four events, respectively. The black pixel at the bottom right is the calibration pixel that is not directly exposed to the sky. []{data-label="fig:sxs_expimg"}](ah100042040sxs_6p38-6p42keV_cl2_DET.pdf "fig:"){width="0.75\linewidth"}
To retrieve photons from the target during the wobbling, we relax the standard screening criteria for the angular distance between the actual pointing and the mean pointing position (`ANG_DIST`) from to . Besides the grade filtering in the standard screening, events flagged due to close proximity in time of 0.72 ms to other events are additionally filtered.
![(Top) Event light curve of the SXS full array in the energy band from 6.38 to 6.42 keV binned with 400 s. (Middle) Same as the top panel but for the wide energy band from 2 to 12 keV. (Bottom) History of `ANG_DIST` with 8 s resolution. []{data-label="fig:sxs_lightcurve"}](rate_ANG_DST.pdf){height="0.85\linewidth"}
Figure \[fig:sxs\_lightcurve\] shows light curves around Fe K$\alpha$ line, wide energy band as well as the history of the `ANG_DIST`. The events concentrate around the time of the wobbling in both energy bands. There is no bright celestial target around the direction where the satellite pointed at this time. No background flare events can be seen for other instruments around this time. Figure \[fig:sxs\_expimg\] bottom panel shows the spatial distribution of the events in the energy band from 6.38 to 6.42 keV. The 19 events spatially concentrate toward the target position. This provides strong indication that these events originate from the target.
### SXI and HXI {#sssec:sxihxireduction}
With regard to the SXI data, false events originating from the cross talk issue are eliminated with the parameters in `sxipipeline` set as follows: $N_{min}$ of 6, $PHA_{sp}$ of 15, and $R$ of 0.7 [@Nakajima17b]. The SXI also suffers from a light leak due to optical/infrared light primarily when the minus Z axis of the spacecraft points to the day earth (MZDYE). Although our observation was free from the MZDYE periods, there was another moderate light leak during the sun illumination of the spacecraft. We also see possible charges left inside the CCDs after the passage of the South Atlantic Anomaly (SAA) as described in @Nakajima17b. The pulse heights of the events detected around the physical edge of the CCDs are weakly affected by these issues. The target was always near the physical edge of the CCD1 during the exposure. To minimize the effect of these problems, we choose only the data during the eclipse of the spacecraft and when the time after the passage of the SAA is larger than 1800 s [@Nakajima17b]. The pile-up fraction is estimated using `pileest` and the results is below 0.7% with a grade migration parameter of 0.1. No additional filtering is applied to the HXI-1 cleaned event files.
Analyses {#sec:ana}
========
All the spectral analyses described below are performed using `XSPEC` v12.9.0u [@1996ASPC..101...17A]. We adopt the spectral model `tbvarabs` for the photoelectric absorption using the interstellar medium abundances described in @2000ApJ...542..914W.
SXS Spectral Analysis {#ssec:sxs_ana}
---------------------
The spectrum obtained with the SXS in the 2–12 keV band is shown in the top panel of figure \[fig:sxs\_spec\]. The events are summed over all the 35 pixels and their total number is 752. The concentrations of events near 5.9, 9.7 and 11.5 keV originate from the instrumental background lines of Mn K$\alpha$, Au L$\alpha$ and L$\beta$, respectively. Due to the limited statistics of the events, we focus on the spectral analysis around a peak at 6.4 keV that is magnified in the bottom panel of figure \[fig:sxs\_spec\]. Most of the events fall within 6.39–6.41 keV and the primary peak is slightly above 6.40 keV. This distribution corresponds to the Fe K$\alpha_1$ and K$\alpha_2$ lines.
We estimate the number of non-X-ray background (NXB) events [@kilbourne17] included in the 6.4 keV line utilizing `sxsnxbgen`. This tool considers the magnetic cut-off rigidity (COR) weighting of the observation and extract events with identical filtering as the source data from the SXS archive NXB event file. Because the events in the energy band of 6.38–6.42 keV are detected in the specific pixels as shown in the bottom panel of figure \[fig:sxs\_expimg\], we only consider those pixels to calculate the NXB. The estimated NXB spectrum is overlaid on the source spectrum in the bottom panel of figure \[fig:sxs\_spec\]. The expected number of NXB counts in 6.38–6.42 keV range is $\leq$ 2 when we assume the same exposure time as the target.
![(Top) SXS spectrum summed over all the 35 pixels. Peaks around 5.9, 9.7, and 11.5 keV are the instrumental background of Mn K$\alpha$, Au L$\alpha$ and L$\beta$, respectively. Poisson error bars [@1986ApJ...303..336G] are presented. Note that the spectrum is binned to 4 eV. (Bottom) Same as the top panel but for the energy range near 6.4 keV. The sum of the fitted models of seven Lorentzian functions for the Fe K$\alpha$ lines and a power-law is shown in a solid red line, with each component shown in dashed lines and different colors. Although the fitting is performed using the original 0.5 eV per bin spectrum, we show the spectrum with a binning of 2 eV for display purposes. Blue data with filled triangles are the calculated NXB spectrum that is not subtracted from the source spectrum. []{data-label="fig:sxs_spec"}](ah100042040sxs_cl2_nocut_8chbin_counts.pdf "fig:"){height="0.85\linewidth"}\
![(Top) SXS spectrum summed over all the 35 pixels. Peaks around 5.9, 9.7, and 11.5 keV are the instrumental background of Mn K$\alpha$, Au L$\alpha$ and L$\beta$, respectively. Poisson error bars [@1986ApJ...303..336G] are presented. Note that the spectrum is binned to 4 eV. (Bottom) Same as the top panel but for the energy range near 6.4 keV. The sum of the fitted models of seven Lorentzian functions for the Fe K$\alpha$ lines and a power-law is shown in a solid red line, with each component shown in dashed lines and different colors. Although the fitting is performed using the original 0.5 eV per bin spectrum, we show the spectrum with a binning of 2 eV for display purposes. Blue data with filled triangles are the calculated NXB spectrum that is not subtracted from the source spectrum. []{data-label="fig:sxs_spec"}](ah100042040sxs_cl2_4chbin_6p37-6p43_7Lorentz_ThawExWid_NXB_counts_4paper.pdf "fig:"){height="0.85\linewidth"}
The K$\alpha$ line centroid near 6.4 keV implies neutral or near-neutral ionization state of Fe. If so, the line should be modeled with Lorentzian functions [@nla.cat-vn1050673] that analytically represent the natural shape of an emission line. It is well known that the K$\alpha$ lines of the $3d$ transition metals are highly asymmetric. @1997PhRvA..56.4554H applied seven Lorentzians to accurately represent the asymmetric K$\alpha$ line from neutral Fe. We assume the near-neutral state and then adopt the best-fit parameters in @1997PhRvA..56.4554H, which will be justified in section \[sec:discussion\]. The NXB spectrum is represented using a power-law model with its index fixed to zero. The power-law component is also included to the source spectrum with its parameters linked between the source and background. We set the following four parameters to be free: the energy at the maximum of the primary Lorentzian ($\alpha_{11}$ in Table II in @1997PhRvA..56.4554H), its width, the normalization factor commonly multiplied to all the seven Lorentzians, and the flux of the power-law component. The relative energy at the maximum of each Lorentzian is fixed as well as the relative width and amplitude. The continuum emissions from the target and the cosmic X-ray background are ignored from the statistical point of view. We adopt c-statistics [@1979ApJ...228..939C] for the spectral fitting. The original 0.5 eV per bin source and background spectra are fitted while the binned spectra are shown in figure \[fig:sxs\_spec\] for display purposes. The best-fit energy at the maximum of the primary Lorentzian is 6405.4 eV and its width is 3.5 eV (FWHM). This yields the Fe K$\alpha_1$ line centroid of 6404.3 eV, a value which is remarkably similar with that of neutral Fe (6403.1 eV) measured by @1997PhRvA..56.4554H.
\[tab:sxs\_specfit\]
$\ast$ Energy at the maximum of the primary Lorentzian ($\alpha_{11}$ in Table II in @1997PhRvA..56.4554H).\
$\dagger$ See text for a discussion of the probability distributions for $E_{\alpha 11}$ and $\sigma_{\alpha 11}$.
To investigate the probability distribution function in the parameter space, we performed Markov Chain Monte Carlo simulations within `XSPEC`. We adopt a proposal distribution of a Gaussian for the chain with a length of 10$^5$. Considering the distribution, the energy at the maximum of the primary component and its width are estimated to be 6405.4$^{+2.4}_{-2.5}$ eV and 3.5$^{+6.4}_{-1.6}$ eV, respectively. The best-fit parameters for the spectral fit are summarized in table \[tab:sxs\_specfit\]. This is the first observational result resolving Fe K$\alpha_1$ and K$\alpha_2$ lines for X-ray binary systems, which demonstrates the superb energy resolution of the microcalorimeter. The accuracy of the energy scale of the SXS is affected by the instrumental gain uncertainty. There had been no on-orbit full-array gain calibration before the observation of IGR J16318. A later calibration using the filter-wheel $^{55}$Fe sources was carried out after changing several cooler power settings (Eckart et al. in preparation). Because the MXS was not yet available, a dedicated calibration pixel that was located outside of the aperture and continuously illuminated by a collimated $^{55}$Fe source served as the only contemporaneous energy-scale reference. The time-dependent scaling required to correct the gain was applied to each pixel in the array. It was known prior to launch that the time-dependent gain-correction function for the calibration pixel generally did not adequately correct the energy scale of the array pixels. The relationship between changes of the calibration pixel and of the array was not fixed, but rather depended on the temperatures of the various shields and interfaces in the dewar. Therefore, although the relative drift rates across the array were characterized during the later calibration with the filter-wheel $^{55}$Fe source, the changes in cooler power settings between the IGR J16318 observation and that calibration limit the usefulness of that characterization. In fact, the measured relative gain drift predict a much larger energy-scale offset between the final two pointings of the Perseus cluster of galaxies than was actually observed.
To overcome our limited ability to extrapolate from the calibration pixel, we examined the whole-array Mn K$\alpha$ instrumental line [@kilbourne17] in source-free data taken from 7th March to 15th March, when the SXS was being operated with the same cooler settings [@tsujimoto17a] as those in the IGR J16318 observation. The SXS energy scale is found to be shifted by at most +1 $\pm$ 0.5 eV at 5.9 keV. Further insight into the gain uncertainty comes from examining the errors in the Mn K$\beta$ position in the filter-wheel $^{55}$Fe data after adjusting all the pixels gain scales based on the Mn K$\alpha$ line. The errors ranged within $-0.6$–+0.2 eV, which indicate the minimum scale of the gain uncertainty at 6.5 keV. We conclude that the gain shift with uncertainty of the line centroid of Fe K$\alpha$, which is between the Mn K$\alpha$ and K$\beta$ lines, is +1 $\pm$ 0.5 eV at the time of the observation of IGR J16318.
SXI and HXI Analysis {#ssec:sxihxi_ana}
--------------------
The SXI image in the energy band from 4.0 to 12 keV is shown in figure \[fig:sxiregion\]. This shows the only additional X-ray source in the FoV, based on the 2XMMi-DR3 catalog [@2012ApJ...756...27L]. Note that the additional filtering of the sun illumination of the spacecraft and the time after the passage of the SAA is not applied to the image because the filtering has only a small effect on the pulse height of each event. Another note is that the PSF shape of the target is not smooth because some pixels are affected by the cross-talk issue [@Nakajima17b] and have been filtered. In spite of the unintended off-axis pointing, the target was securely in the CCD1. Photon extracting regions are drawn with a magenta circle.
![SXI image in the energy band from 4.0 to 12 keV smoothed by a Gaussian of 6 pixels. Each CCD is designated as well as a cataloged X-ray source. The source spectrum extraction region is shown with a magenta circle. Regions shown by green rectangles with red lines are excluded in the extraction. []{data-label="fig:sxiregion"}](2nd_part_combined_woCRE_4to12keV_wo_calsrc.pdf){width="\linewidth"}
The hard X-ray image obtained by the HXI-1 in the energy band from 5.5 to 80 keV is shown in figure \[fig:hxi\_image\]. The circular region in magenta designates the same region as that in figure \[fig:sxiregion\]. Thanks to the moderate PSF of the hard X-ray telescope [@2016SPIE.9905E..12A], a number of events were detected even though the target is just on the edge of the FoV. The source and background spectra are extracted from the regions colored in yellow with solid and dashed lines, respectively.
![HXI-1 image after the standard screening in the energy band of 5.5 to 80 keV smoothed by a Gaussian of 8 pixels. Source and background regions are shown with a solid ellipse and a dashed polygon, respectively. The same sky region as in figure \[fig:sxiregion\] is designated with magenta circle as a reference. A region shown by yellow rectangle with red line is excluded in the source extraction. []{data-label="fig:hxi_image"}](ah100042040hx1_img_5p5to80keV_rev.pdf){width="\linewidth"}
Figure \[fig:sxihxi\_lightcurves\] shows the light curves of the SXI and HXI-1 extracted from the source regions designated in figure \[fig:sxiregion\] and figure \[fig:hxi\_image\], respectively. Background is not subtracted and aspect correction is not applied. Barycenter and dead time correction are applied for the HXI-1 data prior to the extraction. Note that the additional filtering of the sun illumination of the spacecraft and the time after the passage of the SAA is not applied for the SXI light curve because the filtering has only a small effect on the pulse height of each event. The event rate in the energy band dominated by fluorescence lines and continuum both exhibit time variability on a time scale of thousands of seconds, which is also seen in the previous observations . The root mean square fractional variation of the continuum band is 0.34 $\pm$ 0.03 (HXI-1) and $<$ 0.17 (3$\sigma$) (SXI), while that of the fluorescence line band is $<$ 0.25 (HXI-1) and $<$ 0.15 (SXI).
![Light curves of the SXI (top) and HXI-1 (bottom) with 400 s resolution. The energy bands dominated by the fluorescence lines (red) and continuum emission (green) are shown with the ratio between the two bands (blue). []{data-label="fig:sxihxi_lightcurves"}](2nd_part_combined_woCRE_lc_hr_4paper.pdf "fig:"){height="0.85\linewidth"}\
![Light curves of the SXI (top) and HXI-1 (bottom) with 400 s resolution. The energy bands dominated by the fluorescence lines (red) and continuum emission (green) are shown with the ratio between the two bands (blue). []{data-label="fig:sxihxi_lightcurves"}](ah100042040hx1_lc_hr_4paper.pdf "fig:"){height="0.85\linewidth"}
Pulsation search was performed both for the SXI and HXI-1 light curves in each band shown in figure \[fig:sxihxi\_lightcurves\] and also in the entire band. After the search from 1 s to one tenth of the exposure time of each instrument, we found no significant periodic pulsation. This prevents a conclusive determination that the compact object is a neutron star.
Because there is no apparent outburst during the exposure, we extract the spectra of the SXI and HXI-1 without any distinction of time. The NXB for the SXI is calculated using `sxinxbgen` that considers both the magnetic COR weighting of the observation and the position of the source extracting region in the CCD. To maximize the statistics, we subtract only the NXB component rather than extracting background spectrum from the surrounding region for the SXI. We extract all the events during the good time interval of each instrument and hence the extracted durations are not precisely coincident between the SXI and HXI-1. In figure \[fig:sxihxi\_spec\] top panel, we apply a model of `tbvarabs*{cutoffpl+gau+gau+gau}` (hereafter model A). We set the Fe abundance of the absorbing material to be free to reproduce both of the low-energy extinction and the Fe absorption edge, while the abundances of other elements are fixed to solar values. The difference from the model adopted in is that we represent the fluorescence lines from the excitation states with different total angular momenta (K$\alpha_1$ and K$\alpha_2$, K$\beta_1$ and K$\beta_3$) with a single Gaussian function, while introduce a Gaussian function for each fluorescence line. Considering that the Fe K$\alpha$ line width measured with the SXS is negligible for the SXI and HXI-1, the widths of the Gaussian functions are fixed to be zero. Furthermore, the line centroid of Ni K$\alpha$ is fixed so that the ratio of the centroids of Fe K$\alpha$ and Ni K$\alpha$ becomes the value in @1997PhRvA..56.4554H. We also introduce a constant factor that is multiplied to the HXI-1 data to account for possible inter-instrument calibration uncertainty of the effective area. An edge-like structure seen slightly below 30 keV is due to an edge in quantum efficiency of the CdTe double-sided strip detectors and hence is not seen in the unfolded spectrum shown in the bottom panel of figure \[fig:sxihxi\_spec\].
![(Top) Spectra obtained with the SXI (black) and HXI-1 (red). The best-fit spectral model is drawn with solid lines. Each model component is designated with dashed lines. (Bottom) Unfolded spectra using the best-fit model A summarized in table \[tab:sxihxi\_specfit\]. Color coding is the same as that in the top panel. []{data-label="fig:sxihxi_spec"}](SXI-HXI_tbvarabs_copl_compgau_3gaus_sunaltgt0_tsaagt1800.pdf "fig:"){height="0.85\linewidth"} ![(Top) Spectra obtained with the SXI (black) and HXI-1 (red). The best-fit spectral model is drawn with solid lines. Each model component is designated with dashed lines. (Bottom) Unfolded spectra using the best-fit model A summarized in table \[tab:sxihxi\_specfit\]. Color coding is the same as that in the top panel. []{data-label="fig:sxihxi_spec"}](SXI-HXI_tbvarabs_copl_compgau_3gaus_sunaltgt0_tsaagt1800_eeu.pdf "fig:"){height="0.85\linewidth"}
The best-fit parameters are summarized in table \[tab:sxihxi\_specfit\]. Comparison of the spectral parameters with those obtained from the Suzaku observation in 2006 shows that the flux of continuum and line components significantly decreased in the ten year interval while the equivalent widths increased. The unabsorbed luminosity in the 2–10 keV band is 1.0 $\times$ 10$^{34}$ and 5.0 $\times$ $10^{35}$ ergs s$^{-1}$ assuming the distance to the target of 0.9 and 6.2 kpc, respectively. This is much less than the Eddington limit of 1.8 $\times$ $10^{38}$ ergs s$^{-1}$ for a neutron star of 1.4 M$_{\solar}$ and is consistent with values derived for the vast majority of HMXBs, even if including correction for the partial blockage of the continuum source as discussed in section \[sec:discussion\]. The Fe K-shell absorption edge energy is another key parameter that strongly depends on the ionization state of the reprocessing materials. In order to explore this we add the `edge` model that gives
$$f'(E) = \left\{ \begin{array}{ll}
f(E) & (E < E_{\rm edge}) \\
f(E) \cdot {\rm exp}[-\tau_{\rm MAX}(E/E_{\rm edge})^{-3}] & (E \geq E_{\rm edge}),
\end{array} \right.$$ where $E_{\rm edge}$ and $\tau_{\rm MAX}$ are the edge position and the absorption depth at the edge, respectively. Because the `edge` model accounts for absorption at the edge position, we set the Fe abundance of the `tbvarabs` to zero in our spectral fitting. The results are given in table \[tab:sxihxi\_specfit\] in the column labelled model B.
Evaluating the flux of the possible Compton shoulder is performed by adding another Gaussian function to model A with its centroid and width (1$\sigma$) fixed to 6.3 keV and 50 eV, respectively [@2002MNRAS.337..147M]. There is no significant flux of the additional line with its 90% upper limit of 5.4 $\times$ $10^{-4}$ cm$^{-2}$ s$^{-1}$ that corresponds to the 90% upper limit of the equivalent width of 103 eV.
\[tab:sxihxi\_specfit\]
$\ast$ Exponential cutoff energy in the power-law model.\
Discussion {#sec:discussion}
==========
The Fe line in IGR J16318 contains information about the ionization state and kinematics of the emitting gas via the profile shape. It also contains information about the quantity and geometrical distribution of the emitting gas via the line strength, i.e., the flux or equivalent width. This does not necessarily yield unique determinations of interesting physical quantities, but can strongly constrain them under various scenarios. General discussions of the dependence of flux or equivalent width have been provided by many authors, e.g., @1985gecx.conf..153K, @1986LNP...266..249M, @2010ApJ...715..947T, and .
In particular, in the simplest case of a point source of continuum producing the Fe K line via fluorescence at the center of a spherical uniform cloud, simple analytic calculations show that the line equivalent width is approximately proportional to the equivalent hydrogen column density ($N_{\rm H}$) of the cloud for $N_{\rm H}$ $\leq$ 1.5 $\times$ $10^{24}$ cm$^{-2}$. At greater $N_{\rm H}$ the gas becomes Thomson thick and the equivalent width no longer increases. The maximum equivalent width is 1–2 keV and depends on the Fe elemental abundance and on the shape of the SED of the continuum source in the energy band above $\sim$ 6 keV. For solar Fe abundance and an SED consisting of a power-law with photon index of 2, the maximum attainable equivalent width is less than 2 keV. Numerical calculations for toroidal reprocessors show that the Thomson thin approximation breaks down at $N_{\rm H}$ much less than 1.5 $\times$ $10^{24}$ cm$^{-2}$ [@2010MNRAS.401..411Y].
Equivalent widths greater than 2 keV can be obtained if the reprocessor is not spherically symmetric around the continuum source, i.e., if there is an opaque screen along the direct line of sight to the continuum source. This is the most likely explanation for large equivalent widths observed from X-ray binaries during eclipse (e.g., [@2006ApJ...651..421W]), or Seyfert 2 galaxies [@1987ApJ...320L...5K; @2016ApJ...825...85K]. This provides a likely explanation for the large equivalent width observed from IGR J16318; it is crudely consistent with the column density we measure $N_{\rm H}$ $\simeq$ 2.1 $\times$ $10^{24}$ cm$^{-2}$ together with at least a partial blockage of the continuum source by a structure that has Thomson depth much greater than unity. Then we predict that the true luminosity of the source is greater than we infer from simple dilution at a distance of 0.9–6.2 kpc, by a factor $\>$ 2.
We derived the line centroid of Fe K$\alpha$ in spite of low photon statistics. The weighted average of the energies at the maxima of the seven Lorentzian functions is 6399.1$^{+2.5}_{-2.6}$ eV if we consider the gain shift and uncertainty of the SXS. Our result is consistent with those obtained with CCD detectors aboard XMM-Newton and Suzaku . However, the uncertainty of the measurement significantly improved with the SXS. We have to consider the systematic velocity and the orbital velocity of the reprocessor. According to the NIR spectroscopy, there is no significant systemic velocity of the companion star with $c\Delta\lambda/\lambda$ = $-110$ $\pm$ 130 km s$^{-1}$ [@2004ApJ...616..469F]. If we assume the masses of the companion star and the compact object of 30 ${\rm M}_{\odot}$ and 1.4 ${\rm M}_{\odot}$ respectively, the line-of-sight velocity of the compact object with respect to the companion star is within $\pm155$ km s$^{-1}$. Then the total Doppler velocity is expected to be $-110$ $\pm$ 200 km s$^{-1}$, corresponding to the shift of $2.3$ $\pm4.3$ eV.
The top panel of figure \[fig:centroid\_KshellE\] shows the theoretical value of the Fe K$\alpha$ line centroid (${\rm E}_{\rm line}$) versus ionization state . Comparing those with the measured values, the ionization state of FeI–X is preferred. This is in agreement with the other HMXBs reported by @2010ApJ...715..947T. On the other hand, the line centroid measured with the SXI and HXI-1 conflicts formally, at the 90% level with that measured with the SXS. Monitoring the pulse heights of the onboard calibration $^{55}$Fe source by the SXI [@Nakajima17b] reveals that the pulse heights disperse in the range of $\sim$ 2–3 ch that corresponds to $\sim$ 12–18 eV. This can be interpreted as a systematic uncertainty on the SXI energy scale and this brings the SXI+HXI-1 into marginal agreement with the SXS. The middle panel shows the absorption edge of the Fe (${\rm E}_{\rm edge}$) as a function of ionization state [@2004ApJS..155..675K; @1967RvMP...39..125B]. The edge energy measured with the SXI and HXI-1 strongly constrain the ionization state to be no higher than FeIII, which is consistent with that obtained with the Fe K$\alpha$ line centroid. Even if we consider the gain uncertainty of the SXI as noted above, the ionization state is no higher than FeIV. We also plot the difference between ${\rm E}_{\rm edge}$ and ${\rm E}_{\rm line}$ in the bottom panel because such difference is rather robust against the inaccurate energy scale. Although the result suggests the very cold reprocessor, FeI–IV is possible if we introduce a Doppler shift of $\sim$ 1000 km s$^{-1}$ (see later for a justification of this assumed value). also discuss the ionization state of Fe with the statistically best spectrum. Although their line centroid value itself does not reject the slightly ionized state, they claim that the reprocessing materials are neutral considering the systematic uncertainty of the gain of Suzaku/XIS [@2007PASJ...59S..23K]. Here we develop the discussion with the updated and upgraded data obtained with Hitomi.
![(Top) Fe K$\alpha$ line centroid (${\rm E}_{\rm line}$) as a function of the ionization state calculated by @2014ApJ...780..136Y from the expectation by (charge number $\le$ 8) and (charge number $\ge$ 9). Values measured with the SXS and SXI+HXI-1 are shown by the red and blue solid lines, respectively. The gain shift of +1 eV and the most probable systematic velocity of the reprocessor are corrected for the SXS. The dashed lines designate 90 % confidence level. (Middle) Fe K-shell ionization energy (${\rm E}_{\rm edge}$) as a function of the ionization state expected by @2004ApJS..155..675K (charge number $\ge$ 1) and @1967RvMP...39..125B (charge number = 0). Values measured with the combined spectra of the SXI and HXI-1 is shown by the blue solid line as well as the statistical error range (dashed line). (Bottom) Difference of ${\rm E}_{\rm edge}$ and ${\rm E}_{\rm line}$ is plotted as well as the measured value with the SXI and HXI-1. []{data-label="fig:centroid_KshellE"}](centroid_KshellE_4paper_rev4.pdf){height="\linewidth"}
@2004ApJS..155..675K calculated the abundance distribution of the Fe ions in a photoionized plasma as a function of the ionization parameter $\xi$ = $L/nR^2$ [@1969ApJ...156..943T], where $n$ is the gas density, $R$ is the distance between the X-ray source of ionizing radiation and the gas, and $L$ is the luminosity of the continuum emission. The range of ionization states FeI–IV is consistent with an ionization parameter value log($\xi$) $\lesssim$ $-2$. The distance between X-ray source and gas responsible for the Fe emission, $R$, can be estimated based on the X-ray time variability. estimated the distance to be $R$ $\simeq$ $10^{13}$ cm with XMM-Newton by the maximum delay observed between the Fe K$\alpha$ line and the continuum variations. Light curves obtained from other observations also exhibited that Fe K$\alpha$ line followed almost immediately the continuum. Applying the $R$ $\simeq$ $10^{13}$ cm, we estimate $n$ and the thickness of the reprocessing materials along the line of sight ($l$) to be $n$ $\gtrsim$ 3 $\times$ $10^{10}$ cm$^{-3}$ and $l$ = $N_{\rm H}/n$ $\lesssim$ 7 $\times$ $10^{13}$ cm, respectively. If we consider the $\sim$ 80 d orbit and the masses of the companion star and the compact object as above, the distance between them is 2 $\times$ $10^{13}$ cm. The maximum size of the reprocessor $l$ and $R$ may be comparable with the system size.
One of the most probable candidates for the reprocessor is the cold stellar wind from the massive companion star. The wind velocity ($v_w$) at the distance $r$ can be estimated assuming the typical $\beta$-law of $$v_w~=~v_{\infty}(1 - R_{*}/r)^{\beta},$$ where $v_{\infty}$ is the terminal velocity and $R_{*}$ is the stellar radius. Assuming the commonly used $\beta$ = 0.5 and $r$ = $2R_{*}$, we obtain $v_w/v_{\infty}$ $\sim$ 0.7. When we assign a typical $v_{\infty}$ of the early type stars of $\sim$ 1500–2000 km s$^{-1}$ [@1978ApJ...225..893A], $v_w$ $\sim$ 1050–1400 km s$^{-1}$ is obtained. The measured Fe K$\alpha$ line width is equivalent to $v$ = 160$^{+300}_{-70}$ km s$^{-1}$. This is much less than the Doppler broadening expected from speeds that are characteristic of similar systems. This indicates that the line emitting region does not cover the whole region of the stellar wind including the companion star. It suggests that the line may be produced in a relatively small region centered on the compact object. In this case, the line centroid will be Doppler-shifted depending on the orbital phase of the compact object. When we shift both of the line centroid and the K-shell edge energy by 25 eV that corresponds to $v_w$ of 1250 km s$^{-1}$, the two estimates of ionization state contradict each other. This implies that the preferred orbital phase is $\sim$ 0.25 or $\sim$ 0.75. However, $v_{\infty}$ distributes in a wide range even among members of the supergiant HMXBs . Furthermore, argue that highly absorbed HMXB systems have lower wind velocities than classical supergiant HMXBs. An atmosphere model for the donor of Vela X-1 by @sander17 also expects that the wind velocity at the neutron star location is significantly lower than that predicted by the $\beta$-law. More accurate determination of $v_{\infty}$ of the companion star is needed for further discussion. Another interesting possibility is discussed by @2015ApJ...810..102T for supergiant HMXB. These authors argue that Fe K$\alpha$ must be produced close to the photosphere of the donor star, where the wind is still in the acceleration zone, in the region facing the compact object. This case agrees with the fact that the reprocessor does not cover the X-ray source completely. The SXS established an empirical upper limit to the Fe K$\alpha$ width which would imply stellar wind velocities at distances of 1.06 $R_{*}$–1.10 $R_{*}$. This is in agreement with theoretical predictions on the onset of wind clumps given by @2013MNRAS.428.1837S.
To investigate the time variability of the line and continuum emissions obtained in this observation, we plot the ratio of the continuum flux to the fluorescence line flux as a function of the latter for the SXI in figure \[fig:color\_ratio\] top panel. The clear positive correlation indicates that the continuum component exhibits variability with a larger dynamical range than the line component, as measured with the fractional variation of the light curves in section \[ssec:sxihxi\_ana\]. In other words, at least part of the line emission does not follow the continuum variability on time scales less than 400 s. This is consistent with the results obtained by with XMM-Newton. One possible explanation for the positive correlation is that the continuum is produced in a compact region while the line emission takes place in a significantly extended region. Another possibility is the time variation of the column density on the line of sight. Because the X-ray flux around the Fe K band can be affected by the absorption column, time variation of the absorption column on the line of sight can cause time variation only in the low energy band. To clarify this, we check the correlation between the count light curves in the 8–13 keV and 13–50 keV band with the HXI-1 as shown in the bottom panel of figure \[fig:color\_ratio\]. The clear positive correlation is a hint of the intrinsic variation of the continuum rather than due to the changes in the intervening column density.
![(Top) Intensity ratio between the continuum and fluorescence line band versus the intensity in the former band for the SXI. The bin size is 400 s. (Bottom) Count light curves of 13–50 keV band obtained with the HXI-1 versus that in the 8–13 keV band. []{data-label="fig:color_ratio"}](2nd_part_combined_woCRE_hard_vs_hr_4paper.pdf "fig:"){height="\linewidth"}\
![(Top) Intensity ratio between the continuum and fluorescence line band versus the intensity in the former band for the SXI. The bin size is 400 s. (Bottom) Count light curves of 13–50 keV band obtained with the HXI-1 versus that in the 8–13 keV band. []{data-label="fig:color_ratio"}](ah100042040hx1_cont1_vs_cont2.pdf "fig:"){height="\linewidth"}
The absence of the Compton shoulder is confirmed as it was in the spectrum obtained by Suzaku , making a clear contrast with another strongly absorbed HMXB GX 301–2 . and point out that the absence of a Compton shoulder can be due to an inhomogeneous distribution of reprocessing material. Another possibility is the smearing of the Compton shoulder due to the free electrons with an temperature of several eV [@2003ApJ...597L..37W] and/or the scattering with neutral hydrogen [@1996AstL...22..648S; @1999AstL...25..199S]. In fact, mid-infrared observations of IGR J16318 by @2012ApJ...751..150C revealed a spectral component with a temperature $\sim$ 37,000–40,000 K. Since this temperature is higher than that of typical B1 supergiant stars, they suggest that the component corresponds to dense and hot material surrounding the stellar photosphere and irradiated by X-rays from the compact object. Deeper exposure with high spectral resolution like the SXS is required for the further understanding of the circumstellar environment of this system.
Summary {#sec:summary}
=======
In spite of observing challenges such as the large offset angle and the issues such as cross-talk for the SXI, we analyze photons from the target for all of the instruments that had been started up at the time of the Hitomi observation of IGR J16318. The microcalorimeter spectrum resolved the Fe K$\alpha_1$ and K$\alpha_2$ lines for the first time in an X-ray binary system and revealed that the line width is narrower than that compatible with the full range of speeds expected from a stellar wind. Combining the line centroid measured by the SXS and the energy of the Fe K-shell absorption edge by SXI+HXI-1, we put a constraint on the ionization state of the reprocessing materials to be in the range of FeI–IV. Judging from the ionization parameter, the density and thickness of the materials are estimated. As reported in the past observations, the absorption is extraordinarily strong ($N_{\rm H}$ $>$ $10^{24}$ cm$^{-2}$) and the Compton shoulder component is not apparent. These characteristics can be attributed to reprocessing materials which are distributed in a narrow solid angle or scattering primarily with warm free electrons or neutral hydrogen.
The Hitomi observation of IGR J16318 measured the width and energy of the Fe K fluorescence line with precision which are unprecedented for an X-ray binary. They reveal a line width and shift which are much less than the Doppler broadening and shift expected from speeds which are characteristic of similar systems. This was achieved using the SXS detection of 19 photons. If the aspect stability and accuracy of Hitomi pointing system had been accurate at the few arc minutes level, we would have obtained far more detailed diagnostics for the Fe K line and absorption edge diagnostics. However this was not achieved in the initial operations of the Hitomi mission. We now know that the physics of the Fe K line is considerably different for this object, and perhaps for other X-ray binaries, from that previously assumed despite over 40 years of detailed study. Thus, microcalorimeter observations of X-ray binaries in the future with the X-ray recovery mission will open up a new and exciting field of study.
H. Nakajima led this research in data analysis and writing manuscript. He also contributed to the SXI hardware design, fabrication, integration and tests, in-orbit operation, and calibration. K. Hayashida provided key comments on the whole discussion. He also made hardware and software contributions to the SXI as one of the instrument principal investigators. T. Kallman contributed for discussion primarily on the Fe line diagnostics and to elaborate the manuscript. T. Miyazawa worked for the fabrication and calibration of the Hard X-ray Telescope. H. Takahashi contributed to the timing analyses of the HXI. He also made software and hardware contribution to the HXI. M. Guainazzi led observation planning and gave critical comments mainly on the reprocessing materials. H. Awaki, T. Dotani, C. Ferrigno, L. C. Gallo, P. Gandhi, C. A. Kilbourne, P. Laurent, K. Mori, K. Pottschmidt, C. S. Reynolds, and M. Tsujimoto improved the manuscript.
We thank the support from the JSPS Core-to-Core Program. We acknowledge all the JAXA members who have contributed to the ASTRO-H (Hitomi) project. All U.S. members gratefully acknowledge support through the NASA Science Mission Directorate. Stanford and SLAC members acknowledge support via DoE contract to SLAC National Accelerator Laboratory DE-AC3-76SF00515. Part of this work was performed under the auspices of the U.S. DoE by LLNL under Contract DE-AC52-07NA27344. Support from the European Space Agency is gratefully acknowledged. French members acknowledge support from CNES, the Centre National d’Études Spatiales. SRON is supported by NWO, the Netherlands Organization for Scientific Research. Swiss team acknowledges support of the Swiss Secretariat for Education, Research and Innovation (SERI). The Canadian Space Agency is acknowledged for the support of Canadian members. We acknowledge support from JSPS/MEXT KAKENHI grant numbers JP15J02737, JP15H00773, JP15H00785, JP15H02070, JP15H02090, JP15H03639, JP15H03641, JP15H03642, JP15H05438, JP15H06896, JP15K05107, JP15K17610, JP15K17657, JP16J00548, JP16J02333, JP16H00949, JP16H03983, JP16H06342, JP16K05295, JP16K05296, JP16K05300, JP16K05296, JP16K05309, JP16K13787, JP16K17667, JP16K17672, JP16K17673, JP21659292, JP23340055, JP23340071, JP23540280, JP24105007, JP24244014, JP24540232, JP24684010, JP25105516, JP25109004, JP25247028, JP25287042, JP25400236, JP25800119, JP26109506, JP26220703, JP26400228, JP26610047, JP26670560, and JP26800102. The following NASA grants are acknowledged: NNX15AC76G, NNX15AE16G, NNX15AK71G, NNX15AU54G, NNX15AW94G, and NNG15PP48P to Eureka Scientific. H. Akamatsu acknowledges support of NWO via Veni grant. C. Done acknowledges STFC funding under grant ST/L00075X/1. A. Fabian and C. Pinto acknowledge ERC Advanced Grant 340442. P. Gandhi acknowledges JAXA International Top Young Fellowship and UK Science and Technology Funding Council (STFC) grant ST/J003697/2. Y. Ichinohe, K. Nobukawa, and H. Seta are supported by the Research Fellow of JSPS for Young Scientists. N. Kawai is supported by the Grant-in-Aid for Scientific Research on Innovative Areas “New Developments in Astrophysics Through Multi-Messenger Observations of Gravitational Wave Sources”. S. Kitamoto is partially supported by the MEXT Supported Program for the Strategic Research Foundation at Private Universities, 2014-2018. B. McNamara and S. Safi-Harb acknowledge support from NSERC. T. Dotani, T. Takahashi, T. Tamagawa, M. Tsujimoto and Y. Uchiyama acknowledge support from the Grant-in-Aid for Scientific Research on Innovative Areas “Nuclear Matter in Neutron Stars Investigated by Experiments and Astronomical Observations”. N. Werner is supported by the Lendület LP2016-11 grant from the Hungarian Academy of Sciences. D. Wilkins is supported by NASA through Einstein Fellowship grant number PF6-170160, awarded by the Chandra X-ray Center, operated by the Smithsonian Astrophysical Observatory for NASA under contract NAS8-03060.
We thank contributions by many companies, including in particular, NEC, Mitsubishi Heavy Industries, Sumitomo Heavy Industries, and Japan Aviation Electronics Industry. Finally, we acknowledge strong support from the following engineers. JAXA/ISAS: Chris Baluta, Nobutaka Bando, Atsushi Harayama, Kazuyuki Hirose, Kosei Ishimura, Naoko Iwata, Taro Kawano, Shigeo Kawasaki, Kenji Minesugi, Chikara Natsukari, Hiroyuki Ogawa, Mina Ogawa, Masayuki Ohta, Tsuyoshi Okazaki, Shin-ichiro Sakai, Yasuko Shibano, Maki Shida, Takanobu Shimada, Atsushi Wada, Takahiro Yamada; JAXA/TKSC: Atsushi Okamoto, Yoichi Sato, Keisuke Shinozaki, Hiroyuki Sugita; Chubu U: Yoshiharu Namba; Ehime U: Keiji Ogi; Kochi U of Technology: Tatsuro Kosaka; Miyazaki U: Yusuke Nishioka; Nagoya U: Housei Nagano; NASA/GSFC: Thomas Bialas, Kevin Boyce, Edgar Canavan, Michael DiPirro, Mark Kimball, Candace Masters, Daniel Mcguinness, Joseph Miko, Theodore Muench, James Pontius, Peter Shirron, Cynthia Simmons, Gary Sneiderman, Tomomi Watanabe; ADNET Systems: Michael Witthoeft, Kristin Rutkowski, Robert S. Hill, Joseph Eggen; Wyle Information Systems: Andrew Sargent, Michael Dutka; Noqsi Aerospace Ltd: John Doty; Stanford U/KIPAC: Makoto Asai, Kirk Gilmore; ESA (Netherlands): Chris Jewell; SRON: Daniel Haas, Martin Frericks, Philippe Laubert, Paul Lowes; U of Geneva: Philipp Azzarello; CSA: Alex Koujelev, Franco Moroso.
[^1]: The corresponding authors are Hiroshi <span style="font-variant:small-caps;">Nakajima</span>, Kiyoshi <span style="font-variant:small-caps;">Hayashida</span>, Tim <span style="font-variant:small-caps;">Kallman</span>, Takuya <span style="font-variant:small-caps;">Miyazawa</span>, Hiromitsu <span style="font-variant:small-caps;">Takahashi</span>, and Matteo <span style="font-variant:small-caps;">Guainazzi</span>
[^2]: https://heasarc.gsfc.nasa.gov/docs/hitomi/calib/
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We analyze the effects of the electronic coupling to bosonic modes in a d-wave superconductor. The role of the scattering due to boson on the momentum transfer between electronic states in the Brilloine zone is addressed. We consider specific examples of $B_{1g}$ phonon, breathing mode phonon and spin resonance at $(\pi,\pi)$. The Fourier spectrum of the energy derivative local density of states (LDOS) is calculated. To properly calibrate the effects of different modes we fix the quasipartilce renormalization at specific momentum points. It is found that the $B_{1g}$ mode with highly anisotropic momentum-dependent coupling matrix element gives rise to well definded features in the Fourier spectrum, at the energy of mode plus gap, with a momentum transfer along the Cu-O bond direction of cuprates. This result is in a striking contrast to the cases of the coupling to other modes and also to the case of no mode coupling. The origin of this difference is explored in detail. A comparison with the recent STM experiments is briefly discussed.'
author:
- 'Jian-Xin Zhu and A. V. Balatsky'
- 'T. P. Devereaux'
- Qimiao Si
- 'J. Lee, K. McElroy, and J. C. Davis'
date: 'May 28, 2005'
title: 'Fourier-Transformed Local Density of States and Tunneling into a $D$-Wave Superconductor with Bosonic Modes'
---
Introduction
============
Determining the nature of single particle excitations is of fundmental importance in our understanding of the superconductivity in high-$T_c$ cuprates. To address this issue, a number of spectroscopies have been extensively used, including the angle resolved photoemission spectroscopy (ARPES) and tunneling. The salient features observed in ARPES include: (a) Near the $(\pi,0)$ $(M)$ point in the Brillouin zone, the spectral function in the superconducting state shows an anomalous line shape, the so-called “peak-dip-hump” structure; [@Dessau91; @Ding96; @Campuzano99] (b) Near the $d$-wave node of the superconducting gap, the dispersion shows a characteristic “kink” near 50-70 meV. [@Bogdanov00; @Kaminski01; @Lanzara01; @Johnson01; @Zhou03] Recent ARPES experiments with improved resolution [@Kim03; @Gromko03; @Sato03; @Cuk04] have revealed another “kink” in dispersion of the antinodal electronic states, near the $M$ point. An unusual spectral dip-hump features similar to the ARPES spectrum have also been observed in the tunneling data. [@Huang89; @Renner95; @Renner96; @DeWilde98; @Mandrus91; @Yurgens99; @Zasadzinski00; @Zasadzinski01] All these features were suggested to indicate that the electron self-energy renormalization could be due to the electronic coupling to a bosonic mode. Two main scenarios have been presented to explain the experimental data. On one hand, the antinodal renormalization is found to be strongly enhanced below $T_c$. [@Kim03; @Gromko03; @Sato03] Such a strong temperature dependence and the dominance of the coupling strength near the $M$ point can be thought of as evidence for the coupling of electrons to the 41 meV spin resonance mode (of electronic origin). [@Dahm96; @Shen97; @Norman97; @Norman98; @Abanov99; @Eschrig00; @Norman01; @Manske01; @Abanov03] As seen by inelastic neutron scattering experiments in most of the cuprates, [@Rossat-Mignod91; @Mook93; @Fong95; @Bourges95; @Bourges96; @Fong99; @Dai96; @Fong97; @Bourges97; @Dai99; @He01; @He02; @Bourges98; @Fong00; @Bourges00; @Dai01] the spin mode intensity substantially turns on below $T_c$ (even though some intensity might be present in a normal state) and has a well-defined momentum of $(\pi,\pi)$. This scenario of the electronic coupling to spin resonance mode has also been used to explain the tunneling spectra in planar tunnel junctions. On the other hand, it has been suggested that a significant electronic coupling to the half-breathing in-plane Cu-O bond stretching phonon or to the out-of-plane out-of-phase O buckling $B_{1g}$ phonon, with an energy of approximately 70 and 35 meV, respectively, might be responsible for the dispersion anomalies at the nodal [@Lanzara01] and antinodal directions, [@Cuk04] respectively. These two phonon modes have shown strong lineshape renormalizations with doping and temperature in Raman and neutron measurements. [@Friedl90; @Pyka93; @Reznik95; @Chaplot95; @McQueeney99; @Chung03; @Sugai03; @Opel99] The advantage of this scenario is that it could naturally explain the band renormalization effect in materials where no spin resonance mode has been detected, in the normal state, and in the deeply overdoped region where the spin mode is neither expected nor observed. To be consistent with ARPES data, this scenario requires the electron-phonon interaction to be highly anisotropic [@Sandvik04; @Devereaux04] and its impact on the electrons to be strongly enhanced in the superconducting state.
The nature of the involved bosonic modes, being phononic or electronic and their role in the mechanism of superconductivity remains controversial. Detailed energy and momentum spectoscopy of the relevant bosonic modes might be very helpful in understanding the mechanism of superconductivity in high-$T_c$ cuprates. One of the direct spectroscopies that allows energy determination is Inelastic Electron Tunneling Spectroscopy (IETS). It is a well-established and powerful tool that allows the measurement of the characteristic energies of extended modes. Examples of applications of this technique, among many, include measurements of molecular stretching and vibrational modes in metal-insulator-metal tunnel junctions; [@Jakievic66; @Hansma82] observation of the collective magnetic resonance in tunneling in the superconducting state of high-$T_c$ materials [@Zasadzinski01; @Norman97] and observation of the tunneling features at energies that correspond to the phonon peaks, as seen in planar tunneling into superconductors. [@Scalapino69]
The IETS directly measures excitation energies. When electrons scatter off a collective mode, a contribution to the electron self energy occurs above a corresponding threshold value of the frequency determined by the mode frequency. Thus, in a tunneling experiment, for bias voltages exceeding the threshold, electrons can excite the mode. This additional scattering channel leads to a step in the density of states (DOS) and in the tunneling conductance. Low temperatures are required to avoid thermal smearing of the step in the conductance. The crucial quantity that reveals inelastic peaks is a second derivative of tunneling conductance with respect to bias voltage: $\frac{d^{2}I}{dV^{2}}(eV)$. Peaks in this quantity are shown to be connected to the energies of the modes, e.g peaks in phonon Density of States. [@Scalapino69]
In high-$T_c$ materials the energies of a number of collective modes, like phonon and spin modes are close. For example the $B_{1g}$ phonon mode has typical energy of 36-40 meV and spin resonance mode has an energy that ranges between 35-40 meV depending on doping. Hence discrimination between different modes based on only the energy of the observed mode is a challenge. Aside from energy resolved features, it would be useful to come up with the measurement that would allow one to measure typical momenta involved in the electron scattering in cuprates. The ARPES is one such spectroscopy.
Scanning Tunneling Spectroscopy (STM) is another technique that would allow one to resolve the momentum transfer between different electronic states. Recently, the Fourier-transformed scanning tunneling microscopy (FT-STM) [@Hoffman02b; @McElroy03] was introduced to map out the Fermi surface and the momentum dependence of the $d$-wave energy gap. The results are in good agreement with the ARPES data. While the ARPES provides the information about the energy dispersion of single-particle electronic states, the FT-STM probes the scattering processes between states with different in plane momenta.
In this paper, we show how another technique, the Fourier Transform Inelastic Electron Tunneling Spectroscopy (FT-IETS), would allow simultaneous momentum and energy resolutions of the tunneling electrons. As such, the technique might be useful to address the role of different modes in cuprates. Motivated by the progress of elastic FT-STM technique, earlier few of us (JXZ, QS, AVB) have suggested an FT-IETS technique, Ref. , in a specific model that describes the Fourier transformed local density of states (LDOS) in a $d$-wave superconductor with the electronic coupling to the $(\pi,\pi)$ spin resonance mode. Here we go beyond the previous analysis and address the question of what consequences the electron self-energy renormalization will have on the tunneling characteristics for a number of collective modes broadly considered in the literature: the $B_{1g}$ and breathing phonon modes, as well as the $(\pi,\pi)$ spin resonance mode. The central quantity we will focus on in our analysis is energy derivative of the FT LDOS, $\rho^{\prime}(\mathbf{q},E)$. This quantity corresponds to $\frac{d^{2}I}{dV^{2}}(\mathbf{q},eV)$ measured by the FT-STM, where $I$ is the local tunneling current and $V$ the voltage bias. Here we are not concerned with the mechanism of the superconductivity in the cuprates. Instead we assume from the outset a $d$-wave channel effective pairing interaction and study the additional effects due to the electronic coupling to various bosonic modes, including the $B_{1g}$ and breathing phonon modes, and the $(\pi,\pi)$ spin resonance mode. A comparison of the calculated momentum transfer structure with the FT-STM measurement may shed new light on which type of bosonic mode the electron excitations are coupled to strongly.
The rest of the paper is outlined as follows: In Sec. \[SEC:Model\], we develop a theoretical model in which the electrons are coupled to bosonic modes. We consider three types of modes: the $B_{1g}$ and breathing phonon modes with possible coupling matrix elements, and the $(\pi,\pi)$ spin resonance mode. Additional weak disorder is used as a marker so that the momentum transfer can be investigated. In Sec. \[SEC:Results\], the numerical results for the FT spectrum of the energy derivative LDOS are presented. Sec. \[SEC:Conclusion\] contains some concluding remarks.
Theoretical model {#SEC:Model}
=================
We start with a BCS-type model to describe two-dimensional electrons with a $d$-wave pairing symmetry, which is relevant to high-temperature cuprates: $$\mathcal{H}_{BCS}=\sum_{\mathbf{k},\sigma} \xi_{\mathbf{k}}
c_{\mathbf{k}\sigma}^{\dagger} c_{\mathbf{k}\sigma}
+\sum_{\mathbf{k}}(\Delta_{\mathbf{k}}
c_{\mathbf{k}\uparrow}^{\dagger}c_{-\mathbf{k}\downarrow}^{\dagger}
+\Delta_{\mathbf{k}}^{*}
c_{-\mathbf{k}\downarrow}c_{\mathbf{k}\uparrow})\;,$$ where $c_{\mathbf{k}\sigma}^{\dagger}$ ($c_{\mathbf{k}\sigma}$) creates (annihilates) a conduction electron of spin $\sigma$ and wavevector $\mathbf{k}$. The quantity $\xi_{\mathbf{k}}$ is the normal state energy dispersion. We adopt a six-parameter fit to the band structure used previously for optimally doped Bi-2212 systems, [@Norman95] having the form $$\begin{aligned}
\xi_{\mathbf{k}}&=&-2t_{1} (\cos k_x + \cos k_y) -4t_{2} \cos k_{x}
\cos k_y \nonumber \\
&& -2t_{3} (\cos 2k_x + \cos 2k_y) \nonumber \\
&&-4t_{4} (\cos 2k_x \cos k_y + \cos k_x \cos 2k_y) \nonumber \\
&& -4 t_{5} \cos 2k_x \cos 2k_y - \mu \;,\end{aligned}$$ where $t_1=1$, $t_{2}=-0.2749$, $t_{3}=0.0872$, $t_4=0.0938$, $t_5=-0.0857$, and $\mu=-0.8772$. Unless specified explicitly, the energy is measured in units of $t_1$ hereafter.
As in previous works, [@Eschrig00; @Zhu04; @Sandvik04; @Devereaux04] we assume that an effective $d$-wave pairing interaction has pre-existed from certain many-body effects. Therefore, even before the electronic coupling to the bosonic excitations, the $d$-wave superconducting order has already been established and the corresponding order parameter is given by: $$\Delta_{\mathbf{k}}=\frac{\Delta_{0}}{2}(\cos k_x -\cos k_y)\;.$$ The additional renormalization comes from the additional electron-bosonic mode interaction.
By introducing a two-component spinor operator, one defines a matrix Green’s function in the Nambu space. The bare Green’s function in the assumption of a real $d$-wave pair potential reads $$\hat{\mathcal{G}}_{0}^{-1}(\mathbf{k};i\omega_{n})=\left(
\begin{array}{cc}
i\omega_{n}-\xi_{\mathbf{k}} & -\Delta_{\mathbf{k}} \\
-\Delta_{\mathbf{k}} & i\omega_{n}+\xi_{\mathbf{k}}
\end{array}
\right)\;.$$ Here $\omega_{n}=(2n+1)\pi T$ is the fermionic Matsubra frequency.
Coupling to the collective modes {#SUBSEC:Collective}
--------------------------------
As mentioned before, there are many bosonic collective modes existing in cuprates. Here we focus on the out-of-plane out-of-phase buckling ($B_{1g}$) and the in-plane half-breathing phonon modes associated with the motion of the O ions, and the $(\pi,\pi)$ spin resonance mode.
We model the electronic coupling to the phonon modes by $$\mathcal{H}_{el-ph}=\frac{1}{\sqrt{N_{L}}}\sum_{\mathbf{k},\mathbf{q}
\atop \sigma} g_{\nu}(\mathbf{k},\mathbf{q})
c_{\mathbf{k}+\mathbf{q},\sigma}^{\dagger}
c_{\mathbf{k}\sigma}A_{\nu,\mathbf{q}}\;,$$ where $N_{L}$ is the number of lattice sites, $A_{\nu,\mathbf{q}}=b_{\nu\mathbf{q}} +
b_{\nu,-\mathbf{q}}^{\dagger}$ with $b_{\nu\mathbf{q}}^{\dagger}$ ($b_{\nu\mathbf{q}}$) creating (annihilating) one phonon of type $\nu$ (representing $B_{1g}$ or breathing mode) and wavevector $\mathbf{q}$. We consider two types of the coupling matrix elements. The first type are both $\mathbf{q}$ and $\mathbf{k}$ dependent: $$\begin{aligned}
g_{B_{1g}}(\mathbf{k},\mathbf{q})&=&\frac{g_{0}}{\sqrt{M(\mathbf{q})}}
\{\phi_{x}(\mathbf{k})\phi_{x}(\mathbf{k}+\mathbf{q}) \cos(q_y/2) -
\phi_{y}(\mathbf{k})
\phi_{y}(\mathbf{k}+\mathbf{q})\cos(q_x/2)\}\;,\\
g_{br}(\mathbf{k},\mathbf{q})
&=& g_{0} \sum_{\alpha=x,y}
\{\phi_{b}(\mathbf{k}+\mathbf{q})\phi_{\alpha}(\mathbf{k})\cos[(k_{\alpha}+q_{\alpha})/2]
-\phi_{b}(\mathbf{k})\phi_{\alpha}(\mathbf{k}+\mathbf{q})\cos(k_{\alpha}/2)\}\;,\end{aligned}$$ where $M(\mathbf{q})=[\cos^{2}(q_x/2) +
\cos^{2}(q_y/2)]/2$, and $$\begin{aligned}
\phi_{x}&=&\frac{i}{\mathcal{N}_{\mathbf{k}}} \biggl{[}
\xi_{\mathbf{k}} t_{x,\mathbf{k}} -
t_{xy,\mathbf{k}}t_{y,\mathbf{k}}\biggr{]}\;,\\
\phi_{y}&=&\frac{i}{\mathcal{N}_{\mathbf{k}}} \biggl{[}
\xi_{\mathbf{k}} t_{y,\mathbf{k}} -
t_{xy,\mathbf{k}}t_{x,\mathbf{k}}\biggr{]}\;,\\
\phi_{b}&=&\frac{1}{\mathcal{N}_{\mathbf{k}}} \biggl{[}
\xi_{\mathbf{k}}^{2} - t_{xy,\mathbf{k}}^{2}\biggr{]}\;,\end{aligned}$$ with $\mathcal{N}_{\mathbf{k}}=\{ [\xi_{\mathbf{k}}^{2} -
t_{xy,\mathbf{k}}^{2}]^{2} + [\xi_{\mathbf{k}} t_{x,\mathbf{k}} -
t_{xy,\mathbf{k}}t_{y,\mathbf{k}}]^{2} + [\xi_{\mathbf{k}}
t_{y,\mathbf{k}} - t_{xy,\mathbf{k}}t_{x,\mathbf{k}}]^{2}\}^{1/2}$, $t_{\alpha,\mathbf{k}}=-2t_{1}\sin(k_{\alpha}/2)$ and $t_{xy,\mathbf{k}}=-4t_{2} \sin(k_x/2)\sin(k_y/2)$. The $\mathbf{k}$ dependence has been argued to be crucial in the interpretation of ARPES data. [@Cuk04; @Devereaux04] The second type has only $\mathbf{q}$ dependence: [@Song95; @Scalapino95; @Nazarenko96; @Dahm96; @Bulut96; @Sandvik04] $$\vert g_{B_{1g}}(\mathbf{k},\mathbf{q})\vert^{2}=\vert g_{0}
\vert^{2} \biggl{[} \cos^{2}\biggl{(}\frac{q_x}{2} \biggr{)} +
\cos^{2} \biggl{(} \frac{q_y}{2}\biggr{)} \biggr{]}\;,$$ $$\vert g_{br}(\mathbf{k},\mathbf{q})\vert^{2}=\vert g_{0} \vert^{2}
\biggl{[} \sin^{2}\biggl{(}\frac{q_x}{2} \biggr{)} + \sin^{2}
\biggl{(} \frac{q_y}{2}\biggr{)} \biggr{]}\;.$$ This type of $\mathbf{q}$ has been the focus before in the context of a $d_{x^2-y^2}$-wave pairing mechanism. For convenience of notation, we refer to the phonon modes with the first type of coupling as $B_{1g}$-I and $br$-I ones, while those with the second type of coupling as $B_{1g}$-II and $br$-II ones.
The electronic coupling to the $(\pi,\pi)$ spin resonance mode is modeled as: $$\mathcal{H}_{el-sp}=\frac{g_0}{2N_{L}}\sum_{\mathbf{k},\mathbf{q}
\atop \sigma,\sigma^{\prime} }
c_{\mathbf{k}+\mathbf{q},\sigma}^{\dagger}(\mathbf{S}_{\mathbf{q}}\cdot
\bm{\sigma}_{\sigma\sigma^{\prime}})
c_{\mathbf{k},\sigma^{\prime}}\;,$$ where $\mathbf{S}$ is the spin operator for the $(\pi,\pi)$ mode, $\bm{\sigma}$ is the Pauli matrix vector.
We calculate the electronic self-energy due to the electron-bosonic excitation coupling up to the second order in the coupling matrix elements. For the electron-phonon coupling, the self-energy is given as: $$\begin{aligned}
\hat{\Sigma}(\mathbf{k};i\omega_{n})&=&-\frac{T}{N_{L}}\sum_{\mathbf{q},\nu}
\sum_{\Omega_{m}} g_{\nu}(\mathbf{k}-\mathbf{q},\mathbf{q})
g_{\nu}(\mathbf{k},-\mathbf{q}) \nonumber \\
&& \times \mathcal{D}_{\nu}(\mathbf{q};i\Omega_{m}) \hat{\tau}_{3}
\hat{\mathcal{G}}_{0}(\mathbf{k}-\mathbf{q};i\omega_{n}-i\Omega_{m})\hat{\tau}_{3}\;,\nonumber
\\\end{aligned}$$ where $\Omega_{m}=2m\pi T$ is the bosonic Matsubra frequency, $\hat{\tau}_{3}$ is the third component of the Pauli matrix in the Nambu space, the quantity $\mathcal{D}_{\nu}(\mathbf{q};i\Omega_{m})$ is the Fourier transform of the phonon Green’s function $\mathcal{D}_{\nu}(\mathbf{q};\tau)=-\langle T_{\tau}
[A_{\nu,\mathbf{q}}(\tau)A_{\nu,-\mathbf{q}}(0)]\rangle$ and is taken as $$\mathcal{D}_{\nu}(\mathbf{q};i\Omega_{m})=\frac{1}{2}
\left[\frac{1}{i\Omega_{m}-\Omega_{\nu}}-\frac{1}
{i\Omega_{m}+\Omega_{\nu}}\right]\;,$$ with $\Omega_{\nu}$ being the resonance frequency of the phonon modes.
For the electron-$(\pi,\pi)$-resonance spin-fluctuation coupling, the self-energy is given as: [@Zhu04] $$\hat{\Sigma}(\mathbf{k};i\omega_{n})=-\frac{3g_0^{2}T}{4N_{L}}
\sum_{\mathbf{q}} \sum_{\Omega_{m}} \chi(\mathbf{q};i\Omega_{m})
\hat{\mathcal{G}}_{0}(\mathbf{k}-\mathbf{q};i\omega_{n}-i\Omega_{m})
\;,$$ where $\chi(\mathbf{q};i\Omega_{m})$ is the Fourier transform of the spin-spin correlation function, $\chi(\mathbf{q};\tau)=-\langle
T_{\tau} [S_{\mathbf{q}}^{z}(\tau)S_{-\mathbf{q}}^{z}(0)]\rangle$ dynamical spin susceptibility. We treat the susceptibility in a phenomenological form: [@Eschrig00] $$\chi(\mathbf{q};i\Omega_{m})=-\frac{f(\mathbf{q})}{2}
\left[\frac{1}{i\Omega_{m}-\Omega_{0}}-\frac{1}
{i\Omega_{m}+\Omega_{0}}\right]\;. \label{EQ:Suscep}$$ Here the spin resonance mode energy is also denoted by $\Omega_0$. The quantity $f(\mathbf{q})$ describes the momentum dependence of the mode and is assumed to be enhanced at the $\mathbf{Q}=(\pi,\pi)$ point. Using the correlation length $\xi_{sf}$ (chosen to be 2 here), it can be written as $$f(\mathbf{q})=\frac{1}{1+4\xi_{sf}^{2}[\cos^{2}\frac{q_{x}}{2} +
\cos^{2}\frac{q_{y}}{2}]}\;.$$ The form of the dynamic susceptibility as given by Eq. (\[EQ:Suscep\]) is especially suitable for the optimally doped YBa$_2$Cu$_3$O$_{6+y}$ (YBCO) compounds in the superconducting phase, where the observed neutron resonance peak is almost resolution-limited in energy and fairly sharp in wavevector. The resonance peak in BSCCO is broadened in both energy and wavevector. In addition, given that the peak in BSCCO is still quite sharp in energy, we expect that the energy broadening of the resonance mode is not important for the present study. We have also neglected the incommensurate peaks seen in the inelastic neutron scattering experiments in YBCO (the part that disperses “downward” away from the resonance peak), [@Dai98; @Arai99; @Fong00; @Brinckmann99; @Kao00] since their spectral weight is significantly smaller than that of the resonance mode.
The dressed electron Green’s function $\mathcal{G}$, due to the renormalization effect of bosonic excitations, is given by: $$\hat{\mathcal{G}}^{-1}(\mathbf{k};i\omega_{n})
=\hat{\mathcal{G}}_{0}^{-1}(\mathbf{k};i\omega_{n})
-\hat{\Sigma}(\mathbf{k};i\omega_{n})\;.$$ To study the momentum transfer between the bosonic excitation renormalized electronic states, additional impurities or defects are required to scatter the electrons. The scattering from impurities is described by, $$H_{imp}=\sum_{l\sigma} U_{l} c_{l\sigma}^{\dagger} c_{l\sigma} \;,$$ where $U_{l}$ is the strength of the zero-ranged impurity scattering potential at the $l$-th site. For simplicity, we consider only the case of nonmagnetic scattering and assume the scattering potential from all these impurities are identical, i.e., $U_{l}=U_0$. The full Green’s function satisfies the following equation of motion: $$\hat{G}(i,j;i\omega_{n}) = \hat{\mathcal{G}}(i,j;i\omega_{n}) +
\sum_{l} U_{l} \hat{\mathcal{G}}(i,l;i\omega_{n})\hat{\tau}_{3}
\hat{G}(l,j;i\omega_{n})\;.$$ Since we are most interested in effects of the electron collective mode coupling, it is desirable to keep as small as possible the disturbance (i.e., quantum interference, formation of virtual resonance etc.) arising from the quasiparticle scattering off the impurities themselves. This suggests to consider a dilute concentration of weak impurities. In this limit, the Born approximation is applicable, and one arrives at: $$\hat{G}(i,j;i\omega_{n}) = \hat{\mathcal{G}}(i,j;i\omega_{n}) +
U_{0}\sum_{l} \hat{\mathcal{G}}(i,l;i\omega_{n})\hat{\tau}_{3}
\hat{\mathcal{G}}(l,j;i\omega_{n})\;.$$
Due to the impurity scattering, the correction to the local density of states (LDOS) at the $i$-th site, summed over two spin components, is $$\delta\rho(\mathbf{r}_{i},E)=-\frac{2U_0}{\pi}\sum_{l}
\mbox{Im}[\hat{\mathcal{G}}(i,l;E+i\gamma)\hat{\tau}_{3}
\hat{\mathcal{G}}(l,i;E+i\gamma)]_{11} \;. \label{ldos_correction}$$
Fourier Transform
-----------------
The local density of states is proportional to the local differential tunneling conductance (i.e., $dI/dV$). To look into the renormalization effect of collective bosonic excitations in the STM, the energy derivative of the LDOS, corresponding to the derivative of the local differential tunneling conductance (i.e., $d^{2}I/dV^{2}$), is more favorable to enhance the signal. For a fixed value of energy, one first gets a set of $\delta
\rho^{\prime}(i,E)$ (the prime $\prime$ means the energy derivative) in real space, and then performs the Fourier transform: $$\delta \rho^{\prime}(\mathbf{q},E) = \sum_{i} \delta
\rho^{\prime}(\mathbf{r}_i,E) e^{-i\mathbf{q}\cdot
\mathbf{r}_{i}}\;,$$ to obtain a map of the Fourier spectrum in $\mathbf{q}$ space, $$P(\mathbf{q},E)=\vert \delta \rho^{\prime}(\mathbf{q},E) \vert\;.$$
One can also prove that the relation between $\delta\rho(\mathbf{q},E)$ and that due to a single impurity $\delta\rho_{single}(\mathbf{q},E)$: $$\delta\rho^{\prime}(\mathbf{q},E)=F(\mathbf{q})\delta\rho_{single}^{\prime}(\mathbf{q},E)\;,$$ where $F(\mathbf{q})$ is the form factor for the spatial distribution of weak impurities \[$F(\mathbf{q})=1$ for the case of a single impurity\].
Results and discussions {#SEC:Results}
=======================
For the numerical calculation, we take the superconducting energy gap $\Delta_0=0.1$, the frequency of all collective modes $\Omega_{0}=0.15$. A quasiparticle lifetime broadening of $\gamma=0.005$ is used. A weak impurity scattering strength $U_0=0.1$. We take a large system size $N_{L}=1024\times 1024$ to achieve the high-momentum and energy resolution. The Fourier spectrum $P(\mathbf{q},E)$ is then constructed from $\delta
\rho^{\prime}(\mathbf{r}_i;E)$ within a given window of size $61\times 61$ centering the single impurity. We choose the coupling strength for all three types of collective modes in such a way that at the Fermi energy $E=0$, the frequency renormalization factor $Z$ appearing in the self energy $$\hat{\Sigma}(\mathbf{k};i\omega_{n})
=i\omega_{n}[1-Z(\mathbf{k};i\omega_{n})]\hat{\tau}_0 +
\chi(\mathbf{k};i\omega_{n})\hat{\tau}_{3} +
\Phi(\mathbf{k};i\omega_{n})\hat{\tau}_{1}\;,$$ has the same real-part value for the $B_{1g}$-I, $B_{1g}$-II, $br$-II phonon modes, and $(\pi,\pi)$-resonance spin fluctuation modes at the $M$ point, while for the $br$-I phonon mode at the wave-vector $(\pi/2,\pi/2)$. The calibrated value of the coupling strength $g_0$ for all these collective modes is summarized in Table \[TAB:g0\].
$B_{1g}$-I mode $br$-I mode $B_{1g}$-II mode $br$-II mode $(\pi,\pi)$ mode
----------------- ------------- ------------------ -------------- ------------------
1.05 1.48 0.75 0.75 2.30
: The calibrated value of the coupling strength $g_0$ for different type of collective modes.[]{data-label="TAB:g0"}
We present in Fig. \[FIG:DOS\] the results of the DOS and its energy derivative as a function of energy for a clean (i.e., $U_{0}=0$) $d$-wave superconductor with the electronic coupling to the $B_{1g}$-I, br-I, $(\pi,\pi)$ spin resonance modes. For comparison, the DOS for the case of no mode coupling is also shown. When there is no electron-mode coupling, there is a van Hove singularity peak appearing outside the superconducting gap edge. When the electrons are coupled to the $B_{1g}$ and $(\pi,\pi)$-spin resonance modes, the van Hove singularity peak is strongly suppressed. Instead, one sees a dip structure following the coherent peak at the gap edge. The distance between this dip and the coherent peak defines the resonance energy $\Omega_0$. However, there is very little suppression when the electrons are coupled to the $br$-I mode. The planar tunneling experiments indeed observed the peak-dip structure rather than the peak-peak (van Hove singularity) structure. For the band-structure parametrization we have adopted, this implies that the electronic coupling to the collective modes must exist, and the $B_{1g}$ and $(\pi,\pi)$ spin resonance mode are the most promising candidates. Unfortunately, as shown in Fig. \[FIG:DOS\], the dip structure due to the coupling to the $B_{1g}$ and $(\pi,\pi)$ spin resonance mode is almost identical. It would be very challenging to distinguish between these two modes in the planar tunneling experiments, which is measuring the momentum averaged DOS. Therefore, we propose to look at the momentum transfer structure between the band renormalized states, which can be measured by the FT-STM. To achieve this goal, we need to have a signal strong enough to be detectable in STM experiments. The derivative of the DOS, $\rho^{\prime}(E)$, would serve the purpose. As shown in the right column of Fig. \[FIG:DOS\], when the electrons are coupled to the $B_{1g}$ and spin resonance modes, there is a strong peak structure at $E=-(\Delta_0 + \Omega_0)$ exhibiting in the $\rho^{\prime}(E)$ spectrum, which has a one-to-one correspondence to the dip structure in $\rho(E)$ itself.
In Fig. \[FIG:FOURIER-1\], we present the results of the Fourier spectrum of the derivative of the LDOS, $\rho^{\prime}(\mathbf{q},E)$, at the energy $E=-(\Delta_0 +
\Omega_0)$ for a $d$-wave superconductor with the electronic coupling to the collective modes. For comparison, the same quantity is also shown (last panel) for the case of no mode coupling. Note that the case without the mode coupling, the energy $\Omega_0$ has no special meaning in the context of the electronic properties, and the energy $E=-(\Delta_0 + \Omega_0)$ is chosen merely for comparison to the case of mode coupling. The main results are: For all cases, there are strong intensity at the large momentum transfer near $(\pi,\pi)$. For the cases of the $br$-I coupling and no mode coupling, [@footnote-no-mode-coupling] similar feature at very small momentum transfer is obtained, relating to the fact that the DOS spectrum in two cases (see Fig. \[FIG:DOS\]) is similar to each other. For the cases of the coupling to the $B_{1g}$-II and $br$-II modes, there are intensity peaks at a finite momentum transfer along the diagonals. For the case of the coupling to the spin resonance mode, no strong feature is obtained at the intermediate value of momentum transfer, which is consistent with our previous calculation for this specific case. [@Zhu04] For the case of the coupling to the $B_{1g}$-I phonon mode, there exists not only the intensity peaks with a momentum transfer along the diagonal but also the ones with a momentum transfer along the bond directions of a square lattice.
The different Fourier spectrum patterns come from the detailed renormalization of electronic structure by the coupling to these modes. To better understand these patterns, we turn to a detailed analytical form of the Fourier transform. By putting aside the external form factor associated with a specific configuration of weak disorder, the Fourier spectrum is determined uniquely by the electronic single particle Green’s function and is found to be: $$\begin{aligned}
&\delta\rho(\mathbf{q};E)=\frac{u_0}{N_{L}} \sum_{\mathbf{k}} &
\nonumber \\
& \{ [A(\mathbf{k};E)B(\mathbf{k}+\mathbf{q};E) +
B(\mathbf{k};E)A(\mathbf{k}+\mathbf{q};E)] & \nonumber \\
& -[J(\mathbf{k};E)K(\mathbf{k}+\mathbf{q};E) +
K(\mathbf{k};E)J(\mathbf{k}+\mathbf{q};E)]\}\;,& \label{EQ:FT}\end{aligned}$$ where $$\begin{aligned}
A(\mathbf{k};E)&=&-\frac{2}{\pi}
\mbox{Im}[\mathcal{G}_{11}(\mathbf{k};E+i\gamma)]\;,
\label{EQ:A}\\
B(\mathbf{k};E)&=&\mbox{Re}[\mathcal{G}_{11}(\mathbf{k};E+i\gamma)]\;,
\label{EQ:B} \\
J(\mathbf{k};E)&=&-\frac{2}{\pi}
\mbox{Im}[\mathcal{G}_{12}(\mathbf{k};E+i\gamma)]\;,
\label{EQ:J} \\
K(\mathbf{k};E)&=&
\mbox{Re}[\mathcal{G}_{12}(\mathbf{k};E+i\gamma)]\;. \label{EQ:K}\end{aligned}$$ Here as shown in Eqs. (\[EQ:FT\]) through (\[EQ:K\]), the Fourier spectrum is determined by the convolution of the imaginary [@footnote-arpes] and the real parts of the single-particle $(\mathcal{G}_{11})$ and anomalous $(\mathcal{G}_{12})$ Green’s function in the superconducting state. The stronger intensity in $\delta\rho(\mathbf{q};E)$ (also $\delta\rho^{\prime}(\mathbf{q};E)$) will be contributed from the wave vector $\mathbf{q}$, which connects the simultaneously largest intensity in $A(\mathbf{k};E)$ and $B(\mathbf{k};E)$ maps, and in $J(\mathbf{k};E)$ and $K(\mathbf{k};E)$ maps. To be illustrative, we present in Fig. \[FIG:ARPES\] those maps for the electronic coupling to the $B_{1g}$-I and spin resonance modes, and also for the case without the mode coupling at $E=-(\Delta_0 + \Omega_{0})$. Notice that the joint intensity of $J(\mathbf{k};E)K(\mathbf{k}^{\prime};E)$ are smaller by an overall factor of 10 than that of $A(\mathbf{k};E)B(\mathbf{k}^{\prime};E)$ and the Fourier spectrum is mostly determined by the latter product. For the case of no mode coupling, the strongest weight in these quantities are located at the $M$ points of the first Brillouin zone, which leads to the strongest intensity of the Fourier spectrum with the momentum transfer $\mathbf{q}=(\pi,\pi)$ and the zero momentum transfer $\mathbf{q}=0$. If the electrons are coupled to the $B_{1g}$-I phonon mode, besides the strongest intensity on the closed ridges in $A(\mathbf{k};E)$ and $B(\mathbf{k};E)$ \[red area\], there are also moderately strong intensity on the two split beams around the $M$ points in these two maps \[bright green in the former and dark blue in the latter\]. The intensity $A(\mathbf{k};E)B(\mathbf{k}^{\prime};E)$ connected by $\mathbf{q}=\mathbf{k}^{\prime}-\mathbf{k}$ with $\mathbf{k}$ and $\mathbf{k}^{\prime}$ located at the ends of these beams becomes stronger. These wave vectors are just those in the $B_{1g}$-I panel of Fig. \[FIG:FOURIER-1\] at which the Fourier spectrum exhibits peaks \[green spots\]. However, if the electrons are coupled to the spin resonance mode, no such beams exist, which explains the lack of peaks in Fourier spectrum \[see the $(\pi,\pi)$ panel in Fig. \[FIG:FOURIER-1\]\].
Experimentally, only the peaks in the Fourier spectrum with the momentum transfer along the bond direction has been observed so far. [@JLee05] The scenario that the electrons are coupled to the $B_{1g}$ phonon mode and the coupling matrix is highly anisotropic (i.e., $B_{1g}$-I mode) bears the closest resemblance to the experimental observation, though not in a full agreement. In Fig. \[FIG:FOURIER-2\], we present the energy evolution of the spectrum pattern for the electronic coupling to the $B_{1g}$-I mode. It shows that the characteristic momentum transfer wave vector decreases slightly with the increased energy. However, the intensity at these wave vectors decreases rapidly when the energy moves away from the action point $-(\Delta_0 +\Omega_0)$. This is also not inconsistent with the experiment.
Concluding remarks {#SEC:Conclusion}
==================
There is considerable evidence that numerous ARPES and tunneling data can be interpreted in terms of the electronic coupling to bosonic modes with energy about 40 meV. Possible candidates for this mode are the $(\pi,\pi)$ spin resonance mode and various phonon modes. The planar junction tunneling has provided an accurate measurement of the energy scale of this mode. However, since the structure in the tunneling spectra due to the electronic coupling to $B_{1g}$ phonon mode or to the $(\pi,\pi)$ spin resonance mode does not have much difference, it is very difficult to determine the nature of the mode.
In this paper, we have analyzed the Fourier spectra of the energy derivative LDOS to investigate the momentum transfer structure arising from the electronic coupling to these different modes. This quantity nicely complements ARPES for the understanding of the electronic responses to these modes. It can now be obtained from the FT-STM experiments with elevated spatial resolution. In general, we found that the detailed momentum dependence of the coupling matrix element strongly influences the electronic properties. In particular, we have shown that if the $B_{1g}$ or breathing modes are coupled to the electrons with only a $\mathbf{q}$ dependence, the spectrum displays the peak structure along the diagonals with a small momentum transfer while the $(\pi,\pi)$ mode coupling does not produce much weight at the small momentum transfer region. On the other hand, if the electrons are coupled to the $B_{1g}$ phonon mode with a matrix element that depends not only on $\mathbf{q}$ but also on $\mathbf{k}$, the peak structure with a small momentum transfer can also appear along the Cu-O bond directions of the CuO$_{\rm 2}$ plane. Recent FT-STM experiments have indeed found a peak at a bond-directed momentum transfer.
Our calculations also show the structures at large momentum transfer for all cases of the electron-collective mode coupling. In other words, both types (electronic and phononic) of mode couplings produce structures near $(\pi,\pi)$, but only a coupling to the phonon modes yields additional structures at the small momentum transfers. No peak structures near $(\pi,\pi)$ have been observed in the experiments. The situation is somewhat similar to the elastic scattering case (i.e., in the absence of the collective mode coupling), where similar structures near $(\pi,\pi)$ also appear in the theoretical spectra [@Wang02; @Zhang03] but are not observed experimentally. It is likely that the lack of structures at large momentum-transfer in the elastic and inelastic experiments has a common origin. One possibility has to do with strong inhomogeneities, which may give rise to a dominant forward scattering and make only the structures at small momentum transfers observable. This amounts to the following form factor for the weak disorder configuration: $$F({\mathbf{q}})=\frac{1}{1 + r_c [\sin^{2}(q_x/2) +
\sin^{2}(q_y/2)]}\;,$$ where the parameter $r_c$ controls the range of the forward scattering in the $\mathbf{q}$ space. The overall modulation in the Fourier spectrum of the derivative LDOS $\delta \rho^{\prime}({\bf q}, \omega) \propto F(\mathbf{q})$ will be confined to small momenta if $F(\mathbf{q})$ is.
Finally, several remarks are in order: (i) The electron-collective mode coupling we have considered preserves the translational symmetry. Fourier spectra were studied after introducing elastic impurity scattering with weak scattering potential. Alternatively, local vibrational mode scattering will not only provide an inelastic scattering channel but also will break the translational symmetry at the beginning. [@Balatsky03] To fully understand the possible FT-STM experiments, it would be instructive to consider the electronic coupling to the distributed local vibrational modes. We leave this problem for separate analysis. (ii) There has also be increased interest in the quasiparticle scattering from the $\tau_1$ impurities. [@Nunner05] It would be very helpful to study the the FT spectrum through the $\tau_1$ scatterers. We delay this investigation to a future publication.
We thank D.-H. Lee, N. Nagaosa, M. R. Norman, D. J. Scalapino, and Z. X. Shen for very useful discussions. This work was supported by the US DOE (J.X.Z. and A.V.B.), the NSERC, the Office of Naval Research under Grant No. N00014-05-1-0127, and the A. von Humboldt Foundation (T.P.D.), the NSF under Grant No. DMR-0424125 and the Robert A. Welch Foundation (Q.S.), the Office of Naval Research under grant N00014-03-1-0674, the NSF under Grant No. DMR-9971502, the NSF-ITR FDP-0205641, and the Army Research Office under Grant No. DAAD19-02-1-0043 (K.M., J.L., and J.C.D).
[99]{}
D. S. Dessau, B. O. Wells, Z.X. Shen, W. E. Spicer, A. J. Arko, R. S. List, D. B. Mitzi, A. Kapitulnik, Phys. Rev. Lett. [**66**]{}, 2160 (1991).
H. Ding, A. F. Bellman, J. C. Campuzano, M. Randeria, M. R. Norman, T. Yokoya, T. Takahashi, H. Katayama-Yoshida, T. Mochiku, K. Kadowaki, G. Jennings, and G. P. Brivio, Phys. Rev. Lett. [**76**]{}, 1533 (1996).
J. C. Campuzano, H. Ding, M. R. Norman, H. M. Fretwell, M. Randeria, A. Kaminski, J. Mesot, T. Takeuchi, T. Sato, T. Yokoya, T. Takahashi, T. Mochiku, K. Kadowaki, P. Guptasarma, D. G. Hinks, Z. Konstantinovic, Z. Z. Li, and H. Raffy Phys. Rev. Lett. [**83**]{}, 3709 (1999).
P. V. Bogdanov, A. Lanzara, S. A. Kellar, X. J. Zhou, E. D. Lu, W. J. Zheng, G. Gu, J.-I. Shimoyama, K. Kishio, H. Ikeda, R. Yoshizaki, Z. Hussain, and Z. X. Shen, Phys. Rev. Lett. [**85**]{}, 2581 (2000).
A. Kaminski, M. Randeria, J. C. Campuzano, M. R. Norman, H. Fretwell, J. Mesot, T. Sato, T. Takahashi, and K. Kadowaki, Phys. Rev. Lett. [**86**]{}, 1070 (2001).
A. Lanzara, P. V. Bogdanov, X. J. Zhou, S. A. Kellar, D. L. Feng, E. D. Lu, T. Yoshida, H. Eisaki, A. Fujimori, K. Kishio, J.-I. Shimoyama, T. Noda, S. Uchida, Z. Hussain, Z.-X. Shen, Nature [**412**]{}, 510 (2001).
P. D. Johnson, T. Valla, A. V. Fedorov, Z. Yusof, B. O. Wells, Q. Li, A. R. Moodenbaugh, G. D. Gu, N. Koshizuka, C. Kendziora, Sha Jian, and D. G. Hinks, Phys. Rev. Lett. [**87**]{}, 177007 (2001).
X. J. Zhou, T. Yoshida, A. Lanzara, P. V. Bogdanov, S. A. Kellar, K. M. Shen, W. L. Yang, F. Ronning, T. Sasagawa, T. Kakeshita, T. Noda, H. Eisaki, S. Uchida, C. T. Lin, F. Zhou, J. W. Xiong, W. X. Ti, Z. X. Zhao, A. Fujimori, Z. Hussain, Z.-X. Shen, Nature [**423**]{}, 398 (2003).
T. K. Kim, A. A. Kordyuk, S. V. Borisenko, A. Koitzsch, M. Knupfer, H. Berger, and J. Fink, Phys. Rev. Lett. [**91**]{}, 167002 (2003).
A. D. Gromko, A. V. Fedorov, Y.-D. Chuang, J. D. Koralek, Y. Aiura, Y. Yamaguchi, K. Oka, Yoichi Ando, and D. S. Dessau, Phys. Rev. B [**68**]{}, 174520 (2003).
T. Sato, H. Matsui, T. Takahashi, H. Ding, H.-B. Yang, S.-C. Wang, T. Fujii, T. Watanabe, A. Matsuda, T. Terashima, and K. Kadowaki, Phys. Rev. Lett. [**91**]{}, 157003 (2003).
T. Cuk, F. Baumberger, D. H. Lu, N. Ingle, X. J. Zhou, H. Eisaki, N. Kaneko, Z. Hussain, T. P. Devereaux, N. Nagaosa, and Z.-X. Shen, Phys. Rev. Lett. [**93**]{}, 117003 (2004).
Q. Huang, J. E. Zasadzinski, K. E. Gray, J. Z. Liu, and H. Claus, Phys. Rev. B [**40**]{}, 9366 (1989).
Ch. Renner and Ø. Fischer, Phys. Rev. B [**51**]{}, 9208 (1995).
Ch. Renner, B. Revaz, J.-Y. Genoud, and Ø. Fischer, J. Low Temp. Phys. [**105**]{}, 1083 (1996).
Y. DeWilde, N. Miyakawa, P. Guptasarma, M. Iavarone, L. Ozyuzer, J. F. Zasadzinski, P. Romano, D. G. Hinks, C. Kendziora, G. W. Crabtree, and K. E. Gray, Phys. Rev. Lett. [**80**]{}, 153 (1998).
D. Mandrus, L. Forro, D. Koller, and L. Mihaly, Nature [**351**]{}, 460 (1991).
A. Yurgens, D. Winkler, T. Claeson, S.-J. Hwang, and J.-H. Choy, Int. J. Mod. Phys. B [**29-31**]{}, 3758 (1999).
J.F. Zasadzinski, L. Ozyuzer, N. Miyakawa, D. G. Hinks, K. E. Gray, Physica C [**341-348**]{}, 867 (2000).
J. F. Zasadzinski, L. Ozyuzer, N. Miyakawa, K. E. Gray, D. G. Hinks, and C. Kendziora, Phys. Rev. Lett. [**87**]{}, 067005 (2001).
T. Dahm, D. Manske, D. Fay, and L. Tewordt, Phys. Rev. B [**54**]{}, 12006 (1996); T. Dahm, D. Manske, and L. Tewordt, Phys. Rev. B [**58**]{}, 12454 (1998).
Z. X. Shen and J. R. Schrieffer, Phys. Rev. Lett. [**78**]{}, 1771 (1997).
M. R. Norman, H. Ding, J. C. Campuzano, T. Takeuchi, M. Randeria, T. Yokoya, T. Takahashi, T. Mochiku, and K. Kadowaki, Phys. Rev. Lett. [**79**]{}, 3506 (1997).
M. R. Norman and H. Ding, Phys. Rev. B [**57**]{}, R11089 (1998).
Ar. Abanov and A. V. Chubukov, Phys. Rev. Lett. [**83**]{}, 1652 (1999); Phys. Rev. B [**61**]{}, R9241 (2000).
M. Eschrig and M. R. Norman, Phys. Rev. Lett. [**85**]{}, 3261 (2000); Phys. Rev. B [**67**]{}, 144503 (2003).
M. R. Norman, M. Eschrig, A. Kaminski, and J. C. Campuzano, Phys. Rev. B [**64**]{}, 184508 (2001).
D. Manske, I. Eremin, and K. H. Bennemann, Phys. Rev. lett. [**87**]{}, 177005 (2001).
Ar. Abanov, A. V. Chubukov, and J. Schmalian, Adv. Phys. [**52**]{}, 119 (2003).
J. Rossat-Mignod, L. P. Regnault, C. Vettier, P. Bourges, P. Burlet, J. Bossy, J. Y. Henry, and G. Lapertot, Physica C [**185-189**]{}, 86 (1991).
H. A. Mook, M. Yethiraj, G. Aeppli, T. E. Mason, and T. Armstrong, Phys. Rev. Lett. [**70**]{}, 3490 (1993).
H. F. Fong, B. Keimer, P. W. Anderson, D. Reznik, F. Dogan, and I. A. Aksay, Phys. Rev. Lett. [**75**]{}, 316 (1995); H. F. Fong, B. Keimer, D. Reznik, D. L. Milius, and I. A. Aksay, Phys. Rev. B [**54**]{}, 6708 (1996).
P. Bourges, L. P. Regnault, J. Y. Henry, C. Vettier, Y. Sidis, and P. Burlet, Physica B [**215**]{}, 30 (1995).
P. Bourges, L. P. Regnault, Y. Sidis, and C. Vettier, Phys. Rev. B [**53**]{}, 876 (1996).
H. F. Fong, P. Bourges, Y. Sidis, L. P. Regnault, A. Ivanov, G. D. Gu, N. Koshizuka, and B. Keimer, Nature [**398**]{}, 588 (1999);
P. Dai, M. Yethiraj, H. A. Mook, T. B. Linemer, and F. Dogan, Phys. Rev. Lett. [**77**]{}, 5425 (1996).
H. F. Fong, B. Keimer, D. L. Milius, and I. A. Aksay, Phys. Rev. Lett. [**78**]{}, 713 (1997).
P. Bourges, H. F. Fong, L. P. Regnault, J. Bossy, C. Vettier, D. L. Milius, I. A. Aksay, and B. Keimer, Phys. Rev. Lett. [**78**]{}, 713 (1997).
P. Dai, H. A. Mook, S. M. Hayden, G. Aeppli, T. G. Perring, R. D. Hunt, and F. Dogan, Science [**284**]{}, 1344 (1999).
H. F. He, Y. Sidis, P. Bourges, G. D. Gu, A. Ivanov, N. Koshizuka, B. Liang, C. T. Lin, L. P. Regnault, E. Schoenherr, and B. Keimer, Phys. Rev. Lett. [**86**]{}, 1610 (2001).
H. F. He, P. Bourges, Y. Sidis, C. Ulrich, L. P. Regnault, S. Pailhes, N. S. Berzigiarova, N. N. Kolesnikov, and B. Keimer, Science [**295**]{}, 1045 (2002).
P. Bourges, in [*The Gap Symmetry and Fluctuations in High Temperature Superconductors*]{}, edited by J. Bok, G. Deutscher, D. Pavuna, and S. A. Wolf (Plenum, New York, 1998), p.349.
H. F. Fong, P. Bourges, Y. Sidis, L. P. Regnault, J. Bossy, A. Ivanov, D. L. Milius, I. A. Aksay, and B. Keimer, Phys. Rev. B [**61**]{}, 14773 (2000).
P. Bourges, B. Keimer, L. P. Regnault, and Y. Sidis, J. Supercond. [**13**]{}, 735 (2000).
P. Dai, H. A. Mook, R. D. Hunt, and F. Dogan, Phys. Rev. B [**63**]{}, 054525 (2001).
B. Friedl, C. Thomsen, and M. Cardona, Phys. Rev. Lett. [**65**]{}, 915 (1990).
N. Pyka, W. Reichardt, L. Pintschovius, G. Engel, J. RossatMignod, and J. Y. Henry, Phys. Rev. Lett. [**70**]{}, 1457 (1993).
D. Reznik, B. Keimer, F. Dogan, and I. A. Aksay, Phys. Rev. Lett. [**75**]{}, 2396 (1995).
S. L. Chaplot, W. Reichardt, L. Pintschovius, and N. Pyka, Phys. Rev. B [**52**]{}, 7230 (1995).
R. J. McQueeney, Y. Petrov, T. Egami, M. Yethiraj, G. Shirane, and Y. Endoh, Phys. Rev. Lett. [**82**]{}, 628 (1999); R. J. McQueeney, J. L. Sarrao, P. G. Pagliuso, P. W. Stephens, and R. Osborn, Phys. Rev. Lett. [**87**]{}, 077001 (2001).
J.-H. Chung, T. Egami, R. J. McQueeney, M. Yethiraj, M. Arai, T. Yokoo, Y. Petrov, H. A. Mook, Y. Endoh, S. Tajima, C. Frost, and F. Dogan, Phys. Rev. B [**67**]{}, 014517 (2003).
S. Sugai, H. Suzuki, Y. Takayanagi, T. Hosokawa, and N. Hayamizu, Phys. Rev. B [**68**]{}, 184504 (2003).
M. Opel, R. Hackl, T. P. Devereaux, A. Virosztek, A. Zawadowski, A. Erb, E. Walker, H. Berger, and L. Forró, Phys. Rev. B [**60**]{}, 9836 (1999).
A. W. Sandvik, D. J. Scalapino, and N. E. Bickers, Phys. Rev. B [**69**]{}, 094523 (2004).
T. P. Devereaux, T. Cuk, Z.-X. Shen, and N. Nagaosa, Phys. Rev. Lett. [**93**]{}, 117004 (2004).
R. C. Jaklevic and J. Lambe, Phys. Rev. Lett. [**17**]{}, 1139 (1966); D. J. Scalapino and S. M. Markus, Phys. Rev. Lett. [**18**]{}, 459 (1967).
P. Hansma, Ed., [*Tunneling Spectroscopy: Capabilities, Application and New Techniques*]{} (Plenum Press, New York, 1982).
D. J. Scalapino, in “The Electron Phonon Interaction and Strong-Coupling Superconductors”, in [*Superconductivity*]{}, Vol. 1, Chapter 10, p. 449, ed. R.D. Parks (Marcel Dekker, New York, 1969). See also W. L. McMillan and J. M. Rowell, “Tunneling and Strong Coupling Superconductivity”, the same volume, Chapter 10, p. 561 and references therein.
J. E. Hoffman, K. McElroy, D.-H. Lee, K. M. Lang, H. Eisaki, S. Uchida, and J. C. Davis, Science [**297**]{}, 1148 (2002).
K. McElroy, R. W. Simmonds, J. E. Hoffman, D. H. Lee, J. Orenstein, H. Eisaki, S. Uchida, J. C. Davis, Nature [**422**]{}, 592 (2003).
J.-X. Zhu, J. Sun, Q. Si, and A. V. Balatsky, Phys. Rev. Lett. [**92**]{}, 017002 (2004).
M. R. Norman, M. Randeria, H. Ding, and J. C. Campuzano, Phys. Rev. B [**52**]{}, 615 (1995).
J. Song and J. F. Annett, Phys. Rev. B [**51**]{}, 3840 (1995); [**52**]{}, 6930(E) (1995).
D. J. Scalapino, J. Phys. Chem. Solids [**56**]{}, 1669 (1995).
A. Nazarenko and E. Dagotto, Phys. Rev. B [**53**]{}, R2987 (1996).
N. Bulut and D. J. Scalapino, Phys. Rev. B [**54**]{}, 14971 (1996).
P. Dai, H. A. Mook, and F. Dogan, Phys. Rev. Lett. [**80**]{}, 1738 (1998); H. A. Mook, P. Dai, R. D. Hunt, and F. Dogan, J. Phys. Chem. Solids [**59**]{}, 2140 (1998); H. A. Mook, P. Dai, S. M. Hayden, G. Aeppli, T. G. Perring, and F. Dogan, Nature [**395**]{}, 580 (1998).
M. Arai, T. Nishijima, Y. Endoh, T. Egami, S. Tajima, K. Tomimoto, Y. Shiohara, M. Takahashi, A. Garrett, and S. M. Bennington, Phys. Rev. Lett. [**83**]{}, 608 (1999).
J. Brinckmann and P. A. Lee, Phys. Rev. Lett. [**82**]{}, 2915 (1999);
Y.-J. Kao, Q. Si, and K. Levin, Phys. Rev. B [**61**]{}, R11898 (2000); Y. Zha, K. Levin, and Q. Si, Phys. Rev. B [**47**]{}, 9124 (1993).
Here, with the experiment of Ref. in mind, we have done a Fourier transform of the LDOS in the entire field of view. If impurities are explicitly identified, as in the case of Zn-doped cuprates, a filtered Fourier transform can also be adopted. As discussed in Ref. , this procedure would consider the LDOS at the sites other than those of two diagonal strips; in other words, the Fourier transform would only incorporate sites from four disconnected trangles. Within this filtered Fourier transform procedure, the features near $(\pi,\pi)$ disappear in the absence of any mode coupling, but persist for all the cases involving a mode coupling.
Notice that the spectral weight $A(\mathbf{k},\omega)$ as given by Eq. (\[EQ:A\]) is the quantity that the ARPES experiments can measure. The role of the $(\pi,\pi)$-resonance spin fluctuation and the phonon modes on the superconducting state spectral function has been studied previously. This function gives the information about the single particle energy dispersion in the first Brillouin zone in a homogeneous system.
J. Lee [*et al.*]{} (unpublished).
Q.-H. Wang and D.-H. Lee, Phys. Rev. B [**67**]{}, 020511(R) (2003).
D. Zhang and C. S. Ting, Phys. Rev. B [**67**]{}, 100506 (2003).
A. V. Balatsky, Ar. Abanov, and J.-X. Zhu, Phys. Rev. B [**68**]{}, 214506 (2003), and references therein.
T. S. Nunner, B. M. Anderson, A. Melikyan, and P. J. Hirschfeld, cond-mat/0504693 (unpublished).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The study of the uniqueness and nondegeneracy of ground state solutions to semilinear elliptic equations is of great importance because of the resulting energy landscape and its implications for the various dynamics. In [@AIKN3], semilinear elliptic equations with combined power-type nonlinearities involving the Sobolev critical exponent are studied. There, it is shown that if the dimension is four or higher, and the frequency is sufficiently small, then the positive radial ground state is unique and nondegenerate. In this paper, we extend these results to the case of high frequencies when the dimension is five and higher. After suitably rescaling the equation, we demonstrate that the main behavior of the solutions is given by the Sobolev critical part for which the ground states are explicit, and their degeneracy is well characterized. Our result is a key step towards the study of the different dynamics of solutions of the corresponding nonlinear Schrödinger and Klein-Gordon equations with energies above the energy of the ground state. Our restriction on the dimension is mainly due to the existence of resonances in dimension three and four.'
author:
- 'Takafumi Akahori, Slim Ibrahim, Norihisa Ikoma, Hiroaki Kikuchi, Hayato Nawa'
title: Uniqueness and nondegeneracy of ground states to nonlinear scalar field equations involving the Sobolev critical exponent in their nonlinearities for high frequencies
---
Introduction {#section:1}
============
In this paper, we investigate the uniqueness and nondegeneracy of ground state to the semilinear elliptic equation of the form $$\label{eq:1.1}
- \Delta u + \omega u
=
|u|^{p-1}u+|u|^{\frac{4}{d-2}}u
\quad
\mbox{in $\mathbb{R}^{d}$}, \quad u \in H^1({\mathbb{R}^d}) :=H^1({\mathbb{R}^d}, \mathbb{C})$$ where $d\ge 3$, $\omega >0$ and $1< p < \frac{d+2}{d-2} = : 2^\ast - 1$. Here we call $u$ a *ground state* to provided $u$ is a nontrivial solution to satisfying $$\mathcal{S}_\omega (u)
= \inf \left\{ \mathcal{S}_\omega(v) :
\text{$v$ is a nontrivial solution to \eqref{eq:1.1}} \right\}$$ where the action $S_\omega$ is defined by $$\label{eq:1.2}
\mathcal{S}_{\omega}(u)
:=\frac{1}{2}\|\nabla u \|_{L^{2}}^{2}
+
\frac{\omega}{2}\|u\|_{L^{2}}^{2}
-
\frac{1}{p+1}\|u\|_{L^{p+1}}^{p+1}
-
\frac{1}{2^{*}}\|u\|_{L^{2^{*}}}^{2^{*}}.$$ Observe that critical points of $\mathcal{S}_\omega$ solve . In addition, a solution $u$ to is said to be *nondegenerate in ${H^1_{\rm rad}}({\mathbb{R}^d})$* when the linearized equation of at $u$ $$-\Delta \varphi + \omega \varphi =
p |u|^{p-1} \varphi + \frac{d+2}{d-2} |u|^{\frac{4}{d-2}} \varphi
\quad {\rm in} \ {\mathbb{R}^d}, \quad \varphi \in
{H^1_{\rm rad}}({\mathbb{R}^d}) := \{ u \in H^1({\mathbb{R}^d}) : \text{$u$ is radial} \}$$ has the trivial function $\varphi\equiv0$ as its unique solution. That is $${\rm Ker}\, L_u|_{{H^1_{\rm rad}}({\mathbb{R}^d})} = \left\{ 0 \right\}$$ where $L_u : H^1({\mathbb{R}^d}) \to (H^1({\mathbb{R}^d}))^\ast $ is defined by $$\label{eq:1.3}
L_u := -\Delta + \omega
- p |u|^{p-1} - \frac{d+2}{d-2} |u|^{\frac{4}{d-2}}.$$ Equation appears in studying standing wave solutions for the following nonlinear Schrödinger equation and Klein-Gordon equation: $$\begin{aligned}
\label{eq:1.4}
i \frac{\partial \psi}{\partial t} - \Delta \psi
=
|\psi|^{p-1}\psi+|\psi|^{\frac{4}{d-2}}\psi ,
\\[6pt]
\label{eq:1.5}
\frac{\partial^{2} \psi}{\partial t^{2}} - \Delta \psi + m \psi
=
|\psi|^{p-1}\psi+|\psi|^{\frac{4}{d-2}}\psi
. \end{aligned}$$ More precisely, when we look for solutions of the form $\psi(t,x) = e^{-i \lambda t} u(x)$ ($\lambda \in {\mathbb{R}}$), we observe that or is equivalent to solving with the choices $\lambda=\omega$ and $\lambda = \pm \sqrt{m-\omega}$, $\omega<m$.
Due to the presence of multiple powers in , or , these equations loose their scaling invariances and thus a main interest in studying them is to explore the implications of such perturbations, in particular the emergence of ground state solitary waves, the resulting energy landscape, and its role for the global dynamics.
Recently, the dynamics for and were intensively studied. When the energy of initial data is less than the ground state energy, only two scenarios can happen: finite time blow-up or scattering. For example, we refer to [@Akahori-Ibrahim-Kikuchi-Nawa1; @AIKN3; @Kenig-Merle; @Killip-Oh-Poco-Visan; @MXZ]. However, when the energy of initial data is slightly greater than the ground state energy, the dynamic is much more complicated, and the combination of finite time blow-up, scattering and non-dispersion behaviors are shown in forward or backward in time. We refer to [@AIKN3; @Nakanishi-Schlag1; @Nakanishi-Schlag2] for more details. In studying the dynamics around the ground state, basic properties of ground state such as the uniqueness and nondegeneracy play a crucial role. This is our main motivation for the present paper.
On the other hand, the uniqueness and nondegeneracy of positive solutions have applications to the study of nonlinear elliptic equations. For instance, let us consider the following singular perturbation problem $$\label{eq:1.6}
-{\varepsilon}^2 \Delta v + V(x) v = f(v) \quad {\rm in} \ {\mathbb{R}^d}$$ where $V(x): {\mathbb{R}^d}\to {\mathbb{R}}$ and $f: {\mathbb{R}}\to {\mathbb{R}}$ are given functions and $0 < {\varepsilon}\ll 1$ a parameter. When we try to find spike solutions to concentrating at some point $x_0 \in {\mathbb{R}^d}$, the uniqueness and nondegeneracy of positive solutions to $$-\Delta u + V(x_0) u = f(u) \quad {\rm in} \ {\mathbb{R}^d}$$ are keys in order to apply the Lyapunov–Schmidt reduction method. For instance, see [@FW-86; @O-88; @AM]. Since it suffices to consider positive ground states due to Proposition \[proposition:1.2\] for , the uniqueness and nondegeneracy of ground states to is regarded as a step toward those of positive solutions to . Therefore, to study those properties is fundamental and important.
In the case of a single power nonlinearity, and in his celebrated paper [@Kwong], Kwong proved the uniqueness and nondegeneracy of positive radial solutions to $$\label{eq:1.7}
-\Delta u + u = u^p \quad {\rm in} \ {\mathbb{R}^d}, \quad u(x) \to 0 \quad {\rm as} \ |x| \to \infty$$ where $d \geq 1$ and $ 1 < p < 2^\ast - 1$. For generalizations and related results, we refer to [@CL-91; @C; @KZ91; @MS; @PS; @ST] and references therein. Here it is important to mention that it is not simple to apply those results for except for some particular cases. In fact, to the best of our knowledge, Pucci and Serrin [@PS] studied the uniqueness of radial positive solutions to $$\Delta u + f(u) = 0 \quad \mbox{in $\mathbb{R}^{d}$},
\quad u(x) \to 0 \quad {\rm as} \ |x| \to \infty$$ and treated a general nonlinearity $f(u)$. We will see in Appendix \[section:C\] that the result of [@PS] can be applied to when $3 \leq d \leq 6$ and $2^\ast - 2 \leq p < 2^\ast - 1$ with $1<p$, and unfortunately, not in the case $d \geq 7$ and $\omega \gg 1$. See Proposition \[proposition:C.1\] and Remark \[remark:C.1\], and for other cases, we do not know whether or not the result of [@PS] can be applied. Furthermore, the nondegeneracy of radial solutions is not treated in [@PS].
In addition, the uniqueness of radial positive solutions to is delicate according to [@DdPG]. In [@DdPG], Dávila, del Pino and Guerra gave a numerical result which shows that the uniqueness of positive solutions to fails for $d=3$, $1<p<2^\ast - 2$ and $\omega \ll 1$.
About the uniqueness and nondegeneracy of ground states to , these properties were proved in [@AIKN3] under the assumptions $d\ge 4$, $1+\frac{4}{d}<p<2^\ast - 1$ and $\omega \ll 1$. We also mention that the papers [@Grossi; @GLP] studied the uniqueness and nondegeneracy of ground states to equations in bounded domains with single power type nonlinearity whose exponent is the critical one or close to it.
Recently, Coles and Gustafson [@ColesGustafson] showed the uniqueness of the ground state to when $d=3$, $3<p<5$ and $\omega \gg 1$. For more precise statement, see Remark \[remark:1.2\]. Here we also note that they also study the dynamics of the perturbed critical nonlinear Schrödinger equation.
From the above observations, our aim in this paper is to address the uniqueness and nondegeneracy of ground state to for $\omega \gg 1$ in the higher dimensional case. To state our result more precisely, we first recall the existence of ground state to :
\[proposition:1.1\] Assume either $d = 3$ and $3<p<5$ or else $d\ge4$ and $1<p<\frac{d+2}{d-2}$. Then, for any $\omega >0$ there exists a ground state to .
For the sake of clarity and self-content, a sketch of the proof of Proposition \[proposition:1.1\] will be given in Appendix \[section:A\], using simpler arguments than those in [@ASM; @Zhang-Zou].
\[remark:1.1\] In [@Akahori-Ibrahim-Kikuchi-Nawa1 Theorem 1.2], the nonexistence of ground state to was mistakenly claimed for $d=3$, $1+\frac{4}{d}< p < 2^\ast - 1$ and sufficiently large $\omega$. Indeed, in the proof of that Theorem, *(3.18)* was overlooked and now the above Proposition \[eq:1.1\] fixes that.
Using a standard argument for semilinear elliptic equation (see [@AIKN3; @GNN; @Lieb-Loss]), we can derive the following properties of the ground states to :
\[proposition:1.2\] Assume either $d = 3$ and $3<p<5$ or else $d\ge4$ and $1<p<\frac{d+2}{d-2}$. Then, for any $\omega >0$ and any ground state $Q_{\omega}$ to , the following properties hold:
1. $Q_\omega \in C^2({\mathbb{R}^d},\mathbb{C})$.
2. There exist $y \in {\mathbb{R}^d}$, $\theta \in {\mathbb{R}}$ and a radial positive function $\Phi_\omega$ such that $Q_\omega (x) = e^{i \theta} \Phi_\omega (x-y)$. In particular, $\Phi_\omega$ is a radial positive ground state to .
3. Each radial positive ground state to is strictly decreasing in the radial direction.
From Proposition \[proposition:1.2\], it suffices to study radial positive ground states to . Now, we state our main result:
\[theorem:1.1\] Assume $d\ge 5$ and $1 <p < \frac{d+2}{d-2}$. Then, there exists an $\omega_{*}>0$ such that for any $\omega>\omega_{*}$, the (radial) positive ground state to is unique and nondegenerate in ${H^1_{\rm rad}}({\mathbb{R}^d})$.
\[remark:1.2\] [(i)]{} When $3 \leq d \leq 6$ and $\frac{4}{d-2} \leq p < \frac{d+2}{d-2}$ with $1<p$, by Propositions \[proposition:1.1\], \[proposition:1.2\] and \[proposition:C.1\], equation admits a unique radial positive solution for any $\omega > 0$. Furthermore, combining this fact with Theorem \[theorem:1.1\], when $d=5,6$, $\frac{4}{d-2} \leq p < \frac{d+2}{d-2}$ with $1<p$ and $\omega > \omega_\ast$, we find that the unique positive solution to is nondegenerate in ${H^1_{\rm rad}}({\mathbb{R}^d})$.
[(ii)]{} It follows from [@Ni-Takagi Section 5] and Theorem \[theorem:1.1\] that for any $d\ge 5$, any $\omega > \omega_\ast$ and the radial positive ground state $\Phi_\omega$ to , we can prove $${\rm Ker}\, L_{{\Phi_\omega}}
= \left\{ \sum_{j=1}^d a_j \frac{\partial \Phi_\omega}{\partial x_j} :
a_j \in \mathbb{C}
\right\}.$$ See for the definition of $L_{{\Phi_\omega}}$.
[(iii)]{} In [@ColesGustafson], Coles and Gustafson established the uniqueness of the ground state to $$-\Delta u + \hat{\omega}({\varepsilon}) u =\varepsilon |u|^{p-1}u+|u|^{\frac{4}{d-2}}u \quad {\rm in} \ {\mathbb{R}}^3$$ where $3<p<5$, $0<{\varepsilon}\ll 1$, $\hat{\omega}({\varepsilon}) = \omega_1 {\varepsilon}^2 + o({\varepsilon}^2)$ and $\omega_1>0$. By scaling, the equation can be rewritten as $$-\Delta u + \omega_{{\varepsilon}} u = |u|^{p-1} u + |u|^{\frac{4}{d-2}} u \quad {\rm in} \ {\mathbb{R}}^3, \quad
\omega_{{\varepsilon}} \to \infty \quad {\rm as} \ {\varepsilon}\to 0.$$ Therefore, Theorem \[theorem:1.1\] is a higher-dimensional counterpart of the results in [@ColesGustafson], however, our argument is different from the one in [@ColesGustafson].
Next, we describe the difficulties and ideas of the proof of Theorem \[theorem:1.1\] as well as its comparison with the case $\omega \ll 1$ and the result of [@Grossi; @GLP]. Set $${\mathcal{G}_\omega}:= \left\{ {\Phi_\omega}:
\text{${\Phi_\omega}$ is a radial positive ground state to \eqref{eq:1.1}}. \right\}$$
Our aim is to show that ${\mathcal{G}_\omega}$ is a singleton and ${\rm Ker}\, L_{{\Phi_\omega}}|_{{H^1_{\rm rad}}({\mathbb{R}^d})} = \{0\}$ for ${\Phi_\omega}\in {\mathcal{G}_\omega}$.
In [@AIKN3], for $\omega \ll 1$ and ${\Phi_\omega}\in {\mathcal{G}_\omega}$, we use the following rescaling corresponding to the subcritical power $p$: $$T_{\omega}[\Phi_{\omega}]
:=\omega^{-\frac{1}{p-1}}\Phi_{\omega} \left( \omega^{-\frac{1}{2}} x \right),$$ so that $T_{\omega} [ {\Phi_\omega}]$ solves $$-\Delta u + u
-
|u|^{p-1}u
-
\omega^{\frac{2^{*}-(p+1)}{p-1}}|u|^{\frac{4}{d-2}}u
=0.$$ Next, we showed that $T_\omega [{\Phi_\omega}] \to U$ strongly in $H^1({\mathbb{R}^d})$ as $\omega \to 0$ where $U$ is a unique radial positive solution to . By [@Kwong], we know that $U$ is nondegenerate in ${H^1_{\rm rad}}({\mathbb{R}^d}) $, that is, $${\rm Ker}\, L^{\dagger}_{U} |_{{H^1_{\rm rad}}({\mathbb{R}^d})} = \left\{ 0 \right\}, \quad
L^{\dagger}_{U} := -\Delta + 1 - p U^{p-1}.$$ Hence, from the implicit function theorem, we observe that for $\omega \ll 1$, admits a unique radial positive ground state to which is nondegenerate in ${H^1_{\rm rad}}({\mathbb{R}^d})$.
On the other hand, when $\omega \gg 1$, the Sobolev critical term becomes dominant and we use the following rescaling (cf. [@Han; @Moroz-Muratov; @Pan-Wang; @PS; @ST]): $$\label{eq:1.8}
\widetilde{\Phi}_{\omega}(x)=M_{\omega}^{-1}\Phi_{\omega} \left(M_{\omega}^{-\frac{2}{d-2}}x\right),
\quad M_\omega := {\Phi_\omega}(0) = \| {\Phi_\omega}\|_{L^\infty({\mathbb{R}^d})}.$$ Then, we can verify that $$\label{eq:1.9}
-
\Delta \widetilde{\Phi}_{\omega}
+
\alpha_{\omega} \widetilde{\Phi}_{\omega}
-
\beta_{\omega} \widetilde{\Phi}_{\omega}^{p}
-
\widetilde{\Phi}_{\omega}^{\frac{d+2}{d-2}}
=
0, \quad
\alpha_{\omega}:=\omega M_{\omega}^{-\frac{4}{d-2}}, \quad
\beta_{\omega}:= M_{\omega}^{p-1-\frac{4}{d-2}}.$$ Next, we prove $\alpha_\omega, \beta_\omega \to 0$ and ${\widetilde{\Phi}_\omega}\to W$ strongly in $\dot{H}^1({\mathbb{R}^d})$ as $\omega \to \infty$ where $$\dot{H}^1({\mathbb{R}^d}) := \left\{ u \in L^{2^\ast} ({\mathbb{R}^d}) :
\| u \|_{\dot{H}^1} := \| \nabla u \|_{L^2({\mathbb{R}^d})} < \infty \right\}$$ and $W$ is *the Talenti function* with $W(0)=1$, that is, $$\label{eq:1.10}
W(x)
:=\left(1+\frac{|x|^{2}}{d(d-2)} \right)^{-\frac{d-2}{2}}, \quad
-\Delta W = W^{\frac{d+2}{d-2}} \quad {\rm in} \ {\mathbb{R}^d}.$$ Remark that the convergence is proved in ${\dot{H}^1}({\mathbb{R}^d})$, which is different from $H^1({\mathbb{R}^d})$. Moreover, in contrast to $U$, $W$ is degenerate in ${\dot{H}^1}_{\rm rad}({\mathbb{R}^d})$, and $$\label{eq:1.11}
{\rm Ker}\, L^{\ddagger}_W|_{{\dot{H}^1}_{\rm rad}({\mathbb{R}^d})} = \left\{ \Lambda W \right\}$$ where $$\label{11111}
L^{\ddagger}_W := -\Delta + \frac{d+2}{d-2} W^{\frac{4}{d-2}}, \quad
\Lambda W := \frac{d-2}{2}W+x\cdot \nabla W
= \frac{d}{d \lambda} \Big|_{\lambda =1}
\lambda^{\frac{d-2}{2}} W( \lambda x ).$$ From these facts, we need more delicate analysis to show the uniqueness and nondegeneracy for $\omega \gg 1$.
To overcome the above difficulties, we use arguments inspired by [@Grossi; @GLP]. We argue indirectly and suppose that Theorem \[theorem:1.1\] fails to hold. To derive a contradiction, key ingredients consist of a uniform decay estimate of elements of ${\widetilde{\mathcal{G}}_\omega}$ and Pohozaev’s identity corresponding to where $${\widetilde{\mathcal{G}}_\omega}:=
\{ {\widetilde{\Phi}_\omega}\colon
\Phi_\omega \in {\mathcal{G}_\omega}\}.$$
For the uniform spatial decay, we use the Kelvin transform to reduce the problem to a ball and apply Moser’s iteration scheme. One of differences from [@Grossi; @GLP] is the presence of the subcritical term $\beta_\omega {\widetilde{\Phi}_\omega}^p$ and we have to show that this term is harmless to employ the Moser iteration.
After showing the uniform decay, we derive a contradiction. In [@Grossi; @GLP], contradictions are derived through the information on boundary data. In our case, we investigate the behaviors of $\alpha_\omega {\widetilde{\Phi}_\omega}$ and $\beta_\omega {\widetilde{\Phi}_\omega}^p$ in with Pohozaev’s identity. To this end, we need not only the convergence of $\widetilde{\Phi}_{\omega}$ in $\dot{H}^{1}(\mathbb{R}^{d})$ but also in $L^{2}(\mathbb{R}^{d})$ which can be obtained from the uniform decay. This requires us to assume $d \geq 5$.
Now, we introduce the notation used in this paper. By $B_R$ we denote the open ball in $\mathbb{R}^{d}$ of center $0$ and radius $R$, namely, $B_{R}:=\{x\in \mathbb{R}^{d} : |x| < R \}$. For given positive quantities $a$ and $b$, the notation $a \lesssim b$ means the inequality $a\le C b$ for some positive constant $C$ depending only on $d$ and $p$.
This paper is organized as follows. In Section \[section:2\], we prove the convergence results of elements of ${\widetilde{\mathcal{G}}_\omega}$ as $\omega \to \infty$. Section \[section:3\] is devoted to deriving the uniform decay estimate for elements of ${\widetilde{\mathcal{G}}_\omega}$. Finally, in Section \[section:4\], we give a proof of Theorem \[theorem:1.1\]. For readers’ convenience, we include an Appendix where we prove Proposition \[proposition:1.1\] in Section \[section:A\], state a result of the Moser iteration technique in Section \[section:B\], and discuss when the result of [@PS] is applicable to in Section \[section:C\].
Convergence as $\boldsymbol{\omega\to \infty}$ {#section:2}
==============================================
Our aim in this section is to prove the following convergence result of rescaled ground states:
\[proposition:2.1\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, we have $$\lim_{\omega \to \infty}
\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\big\| {\widetilde{\Phi}_\omega}- W \big\|_{{\dot{H}^1}} = 0$$ where ${\widetilde{\Phi}_\omega}$ and $W$ are defined in and .
In order to prove Proposition \[proposition:2.1\], we introduce Nehari’s and Pohozaev’s functionals (associated to equation ) defined by: $$\begin{aligned}
\mathcal{N}_{\omega}(u)
&:=
\|\nabla u \|_{L^{2}}^{2}
+
\omega
\|u \|_{L^{2}}^{2}
-
\|u\|_{L^{p+1}}^{p+1}
-
\|u\|_{L^{2^{*}}}^{2^{*}},
\label{eq:2.1}\end{aligned}$$ and $$\begin{aligned}
\mathcal{P}_{\omega}(u)
&:=
\frac{1}{2^{*}} \|\nabla u \|_{L^{2}}^{2}
+
\frac{\omega}{2}
\|u \|_{L^{2}}^{2}
-
\frac{1}{p+1}\|u\|_{L^{p+1}}^{p+1}
-
\frac{1}{2^{*}}\|u\|_{L^{2^{*}}}^{2^{*}}
\label{eq:2.2}\end{aligned}$$ respectively. Recalling , the following linear combinations are useful in the study of ground states to : $$\begin{aligned}
\mathcal{S}_{\omega}(u)- \mathcal{P}_{\omega}(u)
&=
\frac{1}{d}\|\nabla u \|_{L^{2}}^{2},
\label{eq:2.3}
\\[6pt]
\label{eq:2.4}
\mathcal{P}_{\omega}(u)-\frac{1}{2^{*}}\mathcal{N}(u)
&=
\frac{\omega}{d} \|u \|_{L^{2}}^{2}
-
\frac{2^{*}-(p+1)}{2^{*}(p+1)}\|u\|_{L^{p+1}}^{p+1},
\\[6pt]
\label{eq:2.5}
\mathcal{P}_{\omega}(u)-\frac{1}{p+1}\mathcal{N}(u)
&=
\frac{p-1}{2(p+1)} \omega \|u \|_{L^{2}}^{2}
-
\frac{2^{*}-(p+1)}{2^{*}(p+1)}
\Big\{
\|\nabla u\|_{L^{2}}^{2}
-
\|u\|_{L^{2^{*}}}^{2^{*}}
\Big\}
.\end{aligned}$$
We record the following basic properties of solutions to :
\[proposition:2.2\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and $\omega>0$. Then, the following hold:
1. If $u$ is an $H^{1}$-solution to , then $$\label{eq:2.6}
\mathcal{N}_{\omega}(u)=\mathcal{P}_{\omega}(u)
=0,
\qquad S_\omega (u) = \frac{1}{d} \| \nabla u \|_{L^{2}}^{2}.$$
2. If ${\Phi_\omega}\in {\mathcal{G}_\omega}$, then $$\begin{aligned}
\label{eq:2.7}
\|\nabla {\Phi_\omega}\|_{L^{2}}
&\le
\|\nabla W \|_{L^{2}},
\\[6pt]
\label{eq:2.8}
\|{\Phi_\omega}\|_{L^{2}}
&\lesssim
\omega^{-\frac{1}{2}},
\\[6pt]
\label{eq:2.9}
|{\Phi_\omega}(x)|
&\lesssim
\omega^{-\frac{1}{4}} |x|^{-\frac{d-1}{2}} \quad
{\rm for\ all} \ x \in {\mathbb{R}^d}\setminus \{0\}
.\end{aligned}$$
See [@Berestycki-Lions] for the proof of the identities in . The inequality follows from and . The inequality follows from , and . Finally, we prove . Using the fundamental theorem of calculus, Hölder’s inequality and Hardy’s inequality, we see that $$\begin{aligned}
|x|^{d-1} |\Phi_{\omega}(x)|^{2}
&\le
\int_{0}^{|x|}
\left\{
(d-1)r^{d-1} \frac{|\Phi_{\omega}(r)|}{r} |\Phi_{\omega}(r)|
+
2r^{d-1} |\Phi_{\omega}'(r)| |\Phi_{\omega}(r)|
\right\}
\,dr
\\[6pt]
&\lesssim
\left(
\int_{0}^{|x|}
\left\{
\frac{|\Phi_{\omega}(r)|^{2}}{r^{2}}
+
|\Phi_{\omega}'(r)|^{2} \right\} r^{d-1} \,dr
\right)^{\frac{1}{2}}
\left(
\int_{0}^{|x|}
|\Phi_{\omega}(r)|^{2} r^{d-1} \,dr
\right)^{\frac{1}{2}}
\\[6pt]
&\lesssim
\| \nabla \Phi_{\omega}\|_{L^{2}} \|\Phi_{\omega}\|_{L^{2}}
.
\end{aligned}$$ Furthermore, applying the inequalities and to the right-hand side above, we obtain the desired result .
Next, we consider the rescaled ground states. Let $\widetilde{\Phi}_{\omega} \in {\widetilde{\mathcal{G}}_\omega}$. Then, $\widetilde{\Phi}_{\omega}$ satisfies equation and $$\label{eq:2.10}
\|\widetilde{\Phi}_{\omega}\|_{L^{\infty}}=\widetilde{\Phi}_{\omega}(0)=1=\|W\|_{L^{\infty}}=W(0).$$ Moreover, we see from , and that $$\begin{aligned}
\alpha_{\omega}\|\widetilde{\Phi}_{\omega}\|_{L^{2}}^{2}
+
\|\nabla \widetilde{\Phi}_{\omega}\|_{L^{2}}^{2}
-
\beta_{\omega}\|\widetilde{\Phi}_{\omega}\|_{L^{p+1}}^{p+1}
-
\|\widetilde{\Phi}_{\omega}\|_{L^{2^{*}}}^{2^{*}}
&=0,
\nonumber\end{aligned}$$and $$\begin{aligned}
\label{eq:2.11}
\frac{\alpha_{\omega}}{d}
\|\widetilde{\Phi}_{\omega} \|_{L^{2}}^{2}
-
\frac{2^{*}-(p+1)}{2^{*} (p+1)}
\beta_{\omega}
\|\widetilde{\Phi}_{\omega} \|_{L^{p+1}}^{p+1}
&
=0
.\end{aligned}$$
The following lemma tells us the asymptotic behavior of $M_{\omega}$ and $\alpha_\omega$ as $\omega \to \infty$:
\[proposition:2.3\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then $$\begin{aligned}
\label{eq:2.12}
\lim_{\omega \to \infty} \inf_{{\Phi_\omega}\in {\mathcal{G}_\omega}} M_{\omega} &= \infty,
\\
\label{eq:2.13}
\lim_{\omega \to \infty} \sup_{{\Phi_\omega}\in {\mathcal{G}_\omega}} \alpha_\omega &=0.\end{aligned}$$
First, we prove . Let $\omega>0$ and ${\Phi_\omega}\in {\mathcal{G}_\omega}$. Since $0$ is a maximum point of ${\Phi_\omega}$ by Proposition \[proposition:1.2\], we see $\Delta {\widetilde{\Phi}_\omega}(0) \leq 0$. Recalling ${\widetilde{\Phi}_\omega}(0)=1$, we see from that $$\label{eq:2.14}
\omega - M_{\omega}^{p-1} - M_{\omega}^{\frac{4}{d-2}} \le 0,$$ which implies .
Next, we prove . From and , we may assume $M_\omega \geq 1$ and $\omega \leq 2 M_\omega^{\frac{4}{d-2}}$. Furthermore, we see from the definition of $\alpha_\omega$ that $$0 \leq \alpha := \limsup_{\omega \to \infty} \sup_{{\Phi_\omega}\in {\mathcal{G}_\omega}} \alpha_\omega
\leq 2.$$
We prove by contradiction and suppose $0<\alpha\leq 2$. Then, we can take sequences $\{\omega_{n}\}$ and $\{{\Phi_{\omega_n}}\}$ such that $\lim_{n\to \infty}\omega_{n}=\infty$, ${\Phi_{\omega_n}}\in {\mathcal{G}_{\omega_n}}$ for each $n\ge 1$ and $\lim_{n\to \infty}\alpha_{\omega_{n}}=\alpha$. Since $\alpha>0$, we may also assume that $$\inf_{n\ge 1} \alpha_{\omega_{n}} \ge \frac{\alpha}{2} .$$ Combining the above inequality with the definition of ${\widetilde{\Phi}_{\omega_n}}$, and , we see that $$\begin{split}
\alpha \| {\widetilde{\Phi}_{\omega_n}}\|_{L^{2}}^{2}
+
\|\nabla {\widetilde{\Phi}_{\omega_n}}\|_{L^{2}}^{2}
&
\le
2\alpha_{n} \| {\widetilde{\Phi}_{\omega_n}}\|_{L^{2}}^{2}
+
\|\nabla {\widetilde{\Phi}_{\omega_n}}\|_{L^{2}}^{2}
\\[6pt]
&\lesssim
\omega_{n}\| \Phi_{\omega_{n}} \|_{L^{2}}^{2}
+
\|\nabla \Phi_{\omega_{n}} \|_{L^{2}}^{2}
\lesssim
1+
\|\nabla W\|_{L^{2}}^{2}.
\end{split}$$ Hence, $\{{\widetilde{\Phi}_{\omega_n}}\}$ is bounded in $H^{1}(\mathbb{R}^{d})$. We also have $\|{\widetilde{\Phi}_{\omega_n}}\|_{L^{\infty}}={\widetilde{\Phi}_{\omega_n}}(0)=1$ (see ).
Since ${\widetilde{\Phi}_{\omega_n}}$ satisfies equation with $\omega = \omega_n$ and $\beta_{\omega_n} \to 0$ as $n \to \infty$ due to , we find from the $W^{2,q}$ estimate and Schauder’s estimate (see [@GT]) that there exists a subsequence of $\{{\widetilde{\Phi}_{\omega_n}}\}$ (still denoted by the same symbol) and a radial function $\widetilde{\Phi} \in H^{1}(\mathbb{R}^{d})$ such that $$\lim_{n\to \infty}{\widetilde{\Phi}_{\omega_n}}= \widetilde{\Phi} \quad
\text{weakly in $H^1({\mathbb{R}^d})$ and strongly in
$C^2_{\rm loc}({\mathbb{R}^d})$}$$ and $$\label{eq:2.15}
\left\{ \begin{array}{l}
-\Delta \widetilde{\Phi} + \alpha \widetilde{\Phi}
= \widetilde{\Phi}^{\frac{d+2}{d-2}} \quad {\rm in} \ {\mathbb{R}^d},
\\[6pt]
\widetilde{\Phi}(0)=1.
\end{array} \right.$$ On the other hand, Pohozaev’s identity associated with the equation in implies that if $\alpha>0$, then $\widetilde{\Phi} \equiv 0$ (see, e.g., [@Berestycki-Lions Section 2.2]). This is a contradiction. Thus, $\alpha=0$ and we have completed the proof of the lemma.
Now, we are ready to prove Proposition \[proposition:2.1\].
By contradiction, assume that there exist a constant $\varepsilon_{0}>0$, a sequence $\{ \omega_n \}$ in $(0,\infty)$ and a sequence $\{{\widetilde{\Phi}_{\omega_n}}\}$ such that $\lim_{n\to \infty}\omega_{n}=\infty$, ${\widetilde{\Phi}_{\omega_n}}\in {\widetilde{\mathcal{G}}_{\omega_n}}$ and $$\label{eq:2.16}
\lim_{n\to \infty}\| {\widetilde{\Phi}_{\omega_n}}- W \|_{{\dot{H}^1}}
\ge \varepsilon_{0}.$$ We remark that $\{{\widetilde{\Phi}_{\omega_n}}\}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$, $\|{\widetilde{\Phi}_{\omega_n}}\|_{L^{\infty}}={\widetilde{\Phi}_{\omega_n}}(0)=1$, ${\widetilde{\Phi}_{\omega_n}}$ is a positive solution to with $\omega=\omega_{n}$ and $\lim_{n\to \infty}\alpha_{\omega_{n}}=\lim_{n\to \infty}\beta_{\omega_{n}}= 0$. Hence, as in the proof of Lemma \[proposition:2.3\], we can verify that there exist a subsequence of $\{{\widetilde{\Phi}_{\omega_n}}\}$ (still denoted by the same symbol) and a radial function $\widetilde{\Phi} \in \dot{H}^{1}(\mathbb{R}^{d})$ such that $$\lim_{n\to \infty}{\widetilde{\Phi}_{\omega_n}}= \widetilde{\Phi}
\quad
\mbox{weakly in ${\dot{H}^1({\mathbb{R}^d})}$ and strongly in $C^2_{\rm loc}({\mathbb{R}^d})$}$$ and $$\label{eq:2.17}
\left\{ \begin{array}{l}
\Delta \widetilde{\Phi}+\widetilde{\Phi}^{\frac{d+2}{d-2}}=0 \quad {\rm in} \ {\mathbb{R}^d},
\\[6pt]
\widetilde{\Phi} (0) = 1
.
\end{array} \right.$$ From the uniqueness of radial solutions to the problem (see [@CGS]), it follows that $\widetilde{\Phi} = W$.
On the other hand, we see from the weak lower semicontinuity of the $\dot{H}^{1}$-norm and Lemma \[proposition:2.2\] that $$\| W \|_{{\dot{H}^1}} = \| \widetilde{\Phi} \|_{{\dot{H}^1}}
\le
\liminf_{n\to \infty}\| {\widetilde{\Phi}_{\omega_n}}\|_{{\dot{H}^1}}
\le \| W \|_{{\dot{H}^1}}$$ and therefore $$\label{eq:2.18}
\lim_{n\to \infty}\| {\widetilde{\Phi}_{\omega_n}}\|_{{\dot{H}^1}}= \| \widetilde{\Phi} \|_{{\dot{H}^1}}=\|W \|_{{\dot{H}^1}}
.$$ Combining the weak convergence in $\dot{H}^{1}(\mathbb{R}^{d})$, $\widetilde{\Phi}=W$ and , we find that $$\lim_{n\to \infty} {\widetilde{\Phi}_{\omega_n}}= W
\quad
\mbox{strongly in ${\dot{H}^1({\mathbb{R}^d})}$}.$$ However, this contradicts and we have completed the proof.
Uniform decay estimate {#section:3}
======================
In this section, we discuss uniform decay properties of the rescaled ground states. In particular, we aim to derive the following crucial uniform decay estimate:
\[proposition:3.1\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, there exist two constants $\omega_{dec}>0$ and $C_{dec}>0$ such that for any $\omega >\omega_{dec}$ and any $x\in \mathbb{R}^{d}$, $$\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \widetilde{\Phi}_{\omega}(x) \le C_{dec} \left(1 +|x| \right)^{-(d-2)}
.$$
A proof of Proposition \[proposition:3.1\] will be given in Section \[subsection:3.3\]. First, we derive the following convergence result from Propositions \[proposition:2.1\] and \[proposition:3.1\]:
\[theorem:3.1\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, for any $q> \frac{d}{d-2}$, we have $$\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| \widetilde{\Phi}_{\omega} -W \|_{L^{q}}=0
.$$
Note first that Proposition \[proposition:2.1\] together with Sobolev’s inequality gives us $$\label{eq:3.1}
\lim_{\omega \to \infty}\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| {\widetilde{\Phi}_\omega}- W \|_{L^{2^\ast}} = 0
.$$ Moreover, it follows from that for each $\omega>0$ and $q > 2^\ast$, $$\| {\widetilde{\Phi}_\omega}- W \|_{L^q}^q \leq \int_{{\mathbb{R}^d}} 2^{q-2^\ast}
| {\widetilde{\Phi}_\omega}- W |^{2^\ast} d x = 2^{q-2^\ast} \| {\widetilde{\Phi}_\omega}- W \|_{L^{2^\ast}}^{2^\ast}.$$ Hence, we find that for any $q > 2^\ast $, $$\lim_{\omega \to \infty}
\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| {\widetilde{\Phi}_\omega}-W \|_{L^{q}}
=0.$$
Next, assume $\frac{d}{d-2}<q<2^{*}$ and fix $q_{0} \in (\frac{d}{d-2}, q)$. From Proposition \[proposition:3.1\] we see that for any sufficiently large $\omega$, $$\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| {\widetilde{\Phi}_\omega}\|_{L^{q_0}} \lesssim 1,$$ where the implicit constant depends on $q_{0}$. Furthermore, by Hölder’s inequality and , we get $$\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| {\widetilde{\Phi}_\omega}- W \|_{L^{q}}
\lesssim
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\| {\widetilde{\Phi}_\omega}- W \|_{L^{2^{*}}}^{\frac{2^{*}(q-q_{0})}{q(2^{*}-q)}}
= 0.$$ Thus, we have completed the proof of Corollary \[theorem:3.1\].
Exponential decay estimate {#subsection:3.1}
--------------------------
In this subsection, we derive an exponential decay estimate which we need in the proof of Proposition \[proposition:3.1\]. Let us begin with rephrasing the estimate in terms of the rescaled ground state: for every $\omega>0$, ${\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}$ and $x \in \mathbb{R}^{d}\setminus \{0\}$, $$\label{eq:3.2}
\begin{split}
\widetilde{\Phi}_{\omega}(x)
&=
M_{\omega}^{-1}\Phi_{\omega}(M_{\omega}^{-\frac{2}{d-2}}x)
\\[6pt]
&\lesssim
M_{\omega}^{-1} \omega^{-\frac{1}{4}}( M_{\omega}^{-\frac{2}{d-2}}|x|)^{-\frac{d-1}{2}}
= \alpha_\omega^{-\frac{1}{4}}|x|^{-\frac{d-1}{2}}.
\end{split}$$
Next, we state the main result in this subsection:
\[proposition:3.2\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, there exist constants $L_{0}>0$ and $C_0>0$ such that for any $ \omega >0$ and $|x| \geq L_{0} \alpha_\omega^{-\frac{1}{2}}$, $$\label{eq:3.3}
\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \widetilde{\Phi}_{\omega}(x)
\leq
C_{0} \alpha_\omega^{\frac{d-2}{4}} e^{-\frac{\sqrt{\alpha_\omega}}{2}|x|}.$$
Let $\omega>0$, ${\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}$ and $L_{0}>0$ be a large number to be specified later. Since ${\widetilde{\Phi}_\omega}$ is strictly decreasing in the radial direction by Proposition \[proposition:1.2\], we see from that if $|x| \geq L_{0} \alpha_{\omega}^{-\frac{1}{2}}$, then $$\label{eq:3.4}
{\widetilde{\Phi}_\omega}(x) \le {\widetilde{\Phi}_\omega}( L_{0} \alpha_\omega^{-\frac{1}{2}} )
\lesssim L_{0}^{-\frac{d-1}{2}}
\alpha_\omega^{\frac{d-2}{4}}.$$
Next, we rewrite as $$\label{eq:3.5}
-\Delta {\widetilde{\Phi}_\omega}+ \Big( \alpha_\omega - \beta_\omega {\widetilde{\Phi}_\omega}^{p-1}
- {\widetilde{\Phi}_\omega}^{\frac{4}{d-2}} \Big) {\widetilde{\Phi}_\omega}= 0 \quad
{\rm in} \ {\mathbb{R}^d}.$$ We see from and the definitions of $\alpha_{\omega}$ and $\beta_{\omega}$ that if $|x| \geq L_{0} \alpha_{\omega}^{-\frac{1}{2}}$, then $$\begin{aligned}
\label{eq:3.6}
\beta_{\omega} {\widetilde{\Phi}_\omega}^{p-1} (x)
&\lesssim
L_{0}^{-\frac{(d-1)(p-1)}{2}}
\alpha_{\omega}^{\frac{(d-2)(p-1)}{4}}
\beta_{\omega}
=
L_{0}^{-\frac{(d-1)(p-1)}{2}} \omega^{-\frac{4-(d-2)(p-1)}{4}}
\alpha_{\omega}
,
\\[6pt]
\label{eq:3.7}
{\widetilde{\Phi}_\omega}^{\frac{4}{d-2}}(x)
&\lesssim L_{0}^{-\frac{2(d-1)}{d-2}} \alpha_{\omega}. \end{aligned}$$ Furthermore, it follows from , , and the assumption $p<\frac{d+2}{d-2}=1+\frac{4}{d-2}$ that if we choose a sufficiently large $L_0$ depending only on $d$ and $p$, then $$\label{eq:3.8}
-\Delta {\widetilde{\Phi}_\omega}(x) + \frac{\alpha_\omega}{2} {\widetilde{\Phi}_\omega}(x) \leq 0$$ for all $|x| \geq L_{0} \alpha_\omega^{-\frac{1}{2}}$.
Now, we shall derive by using the comparison principle. To this end, let $R> L_{0} \alpha_\omega^{-\frac{1}{2}}$ and introduce a positive function $\psi_{R}$ on $\mathbb{R}$ as $$\psi_R(r) := \exp \Big( - \frac{\alpha_{\omega}^{\frac{1}{2}}}{2} \big( r - L_{0} \alpha_\omega^{-\frac{1}{2}} \big) \Big)
+
\exp \Big( \frac{\alpha_{\omega}^{\frac{1}{2}}}{2} ( r - R ) \Big).$$ It is easy to verify that $$\label{eq:3.9}
\left|\psi_{R}'(r)\right| \leq \frac{\alpha_{\omega}^{\frac{1}{2}}}{2} \psi_R(r),
\quad \psi_{R}''(r) = \frac{\alpha_\omega}{4} \psi_R(r).$$ We use the same symbol $\psi_{R}$ to denote the radial function $\psi_{R}(|x|)$ on $\mathbb{R}^{d}$. Then, we see from that if $L_{0} \ge 2(d-1)$ and $L_{0}\alpha_\omega^{-\frac{1}{2}} < r < R$, then $$\label{eq:3.10}
-\Delta \psi_{R}+\frac{\alpha_{\omega}}{2}\psi_{R}
=
-\psi_{R}''-\frac{d-1}{r}\psi_{R}'+\frac{\alpha_{\omega}}{2}\psi_{R}
\geq 0 .$$ Furthermore, it follows from , $\psi_R( L_{0} \alpha_\omega^{-\frac{1}{2}} ) \geq 1$ and $\psi_R(R)\ge 1$ that $$\label{eq:3.11}
{\widetilde{\Phi}_\omega}(R) \le {\widetilde{\Phi}_\omega}(L_{0}\alpha_{\omega}^{-\frac{1}{2}})
\lesssim
L_{0}^{-\frac{d-1}{2}} \alpha_\omega^{\frac{d-2}{4}}
\min
\big\{
\psi_R( L_{0}\alpha^{-\frac{1}{2}} ), \ \psi_R(R)
\big\}.$$ Hence, the comparison principle together with , and implies that if $L_{0}\alpha_\omega^{-\frac{1}{2}} \leq |x| \leq R$, then $${\widetilde{\Phi}_\omega}(x) \lesssim
L_{0}^{-\frac{d-1}{2}} \alpha_{\omega}^{\frac{d-2}{4}} \psi_R(|x|).$$ Since $R > L_{0}\alpha_{\omega}^{-\frac{1}{2}}$ is arbitrary, taking $R \to \infty$, we find that $${\widetilde{\Phi}_\omega}(x)
\lesssim L_{0}^{-\frac{d-1}{2}} \alpha_\omega^{\frac{d-2}{4}}
\exp \Big( - \frac{\alpha_{\omega}^{\frac{1}{2}}}{2}
\big( |x| - L_{0} \alpha_\omega^{-\frac{1}{2}}
\big)
\Big)
= L_{0}^{-\frac{d-1}{2}} e^{\frac{L_{0}}{2}}
\alpha_\omega^{\frac{d-2}{4}}
e^{- \frac{ \sqrt{\alpha_{\omega}}}{2} |x|}$$ for all $|x| \geq L_{0} \alpha_\omega^{-\frac{1}{2}}$, which is the desired estimate .
Kelvin transforms of rescaled ground states {#subsection:3.2}
-------------------------------------------
In this subsection, we consider the Kelvin transform of elements in ${\widetilde{\mathcal{G}}_\omega}$. We use $K[u]$ to denote the Kelvin transform of a function $u$, that is, $$K[u](x):= |x|^{-(d-2)} u \left(\frac{x}{|x|^{2}}\right)
.$$ Remark that $\|K[{\widetilde{\Phi}_\omega}]\|_{L^{\infty}(B_{1})} \lesssim 1$ implies $$\sup_{|x|\ge 1}{\widetilde{\Phi}_\omega}(x) \lesssim |x|^{-(d-2)}
.$$ Thus, to prove Proposition \[proposition:3.1\], it suffices to show that there exists $\omega_{dec}>0$ such that $$\label{eq:3.12}
\sup_{\omega>\omega_{dec}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| K[{\widetilde{\Phi}_\omega}]\|_{L^{\infty}(B_{1})} <\infty.$$
It is easy to verify that $K[{\widetilde{\Phi}_\omega}]$ satisfies $$\label{eq:3.13}
-\Delta K[{\widetilde{\Phi}_\omega}]
+
\alpha_{\omega}\frac{1}{|x|^{4}}
K[{\widetilde{\Phi}_\omega}]
=
\beta_{\omega}\frac{1}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p}
+
K[{\widetilde{\Phi}_\omega}]^{\frac{d+2}{d-2}}
,$$ where $$\label{eq:3.14}
\gamma:= 4 - (d-2)(p-1) > 0.$$ We also see from Lemma \[proposition:3.2\] that if $|x|\le \alpha_{\omega}^{\frac{1}{2}} /L_{0}$, then $$\label{eq:3.15}
K[{\widetilde{\Phi}_\omega}](x) \lesssim \alpha_{\omega}^{\frac{d-2}{4}}
|x|^{-(d-2)} e^{-\frac{\sqrt{\alpha_{\omega}}}{2} |x|^{-1}}
.$$ Furthermore, since the Kelvin transform is linear and preserves the ${\dot{H}^1({\mathbb{R}^d})}$ norm, we have $$\label{eq:3.16}
\| K[u] - K[v] \|_{{\dot{H}^1}} = \| u - v \|_{{\dot{H}^1}}$$ for any $u,v\in \dot{H}^{1}(\mathbb{R}^{d})$. Hence, Proposition \[proposition:2.1\] leads us to the following result:
\[proposition:3.3\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, it holds that $$\lim_{\omega\to \infty}\sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\| K [{\widetilde{\Phi}_\omega}] - K[W] \|_{\dot{H}^{1}}=0.$$
In order to use Moser’s iteration (see Proposition \[proposition:B.1\]) for , we need the following lemma:
\[proposition:3.4\] Assume $d\ge 3$, $1<p<\frac{d+2}{d-2}$ and ${\mathcal{G}_\omega}\neq \emptyset$. Then, it holds that $$\label{eq:3.17}
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \int_{|x|\le 4}
\Big| \frac{\beta_{\omega}}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p-1}(x)
\Big|^{\frac{d}{2}}dx
=
0.$$
It follows from Lemma \[proposition:2.3\] that there exists $\omega_{1}>0$ such that $ \alpha_\omega \le 1$ for all $\omega \ge \omega_{1}$ and ${\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}$. In what follows, we always assume that $\omega>\omega_{1}$. Also let $L_0 \geq 1$ be the constant appeared in Lemma \[proposition:3.2\]. We divide the integral into two parts: $$\begin{aligned}
I_{\omega,in}
&:=
\int_{|x|\le \sqrt{\alpha_\omega}/L_0}
\left| \frac{\beta_{\omega}}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p-1}(x)
\right|^{\frac{d}{2}}
dx,
\\[6pt]
I_{\omega,out}
&:=
\int_{ \sqrt{\alpha_\omega}/L_0 \le |x|\le 4}
\left| \frac{\beta_{\omega}}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p-1}(x)
\right|^{\frac{d}{2}}
dx.\end{aligned}$$
We first show $$\label{eq:3.18}
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} I_{\omega,out}=0.$$ To this end, set $s_{0}:=\frac{4}{(d-2)(p-1)}$ and note that thanks to $0 < p-1 < \frac{4}{d-2}$, we have $1 < s_0 < \infty$. Moreover, since $\frac{d(p-1)s_{0}}{2}=2^{*}$ and $1-\frac{1}{s_{0}}=\frac{\gamma}{4}$, we see from Hölder’s inequality and Lemma \[proposition:3.3\] that $$\begin{aligned}
I_{\omega,out}
&= \beta_\omega^{\frac{d}{2}}
\int_{ \sqrt{\alpha_\omega} / L_0 \le |x|\le 4}
|x|^{- \frac{d}{2} \gamma } K[{\widetilde{\Phi}_\omega}]^{ \frac{d(p-1)}{2} }(x)\,dx
\\[6pt]
&\le
\beta_\omega^{\frac{d}{2}}
\left\{
\int_{ \sqrt{\alpha_\omega} / L_0 \le |x|\le 4}
K[{\widetilde{\Phi}_\omega}]^{2^{*}}(x)\,dx
\right\}^{1/s_0}
\left\{
\int_{ \sqrt{\alpha_\omega} / L_0 \le |x|\le 4}
|x|^{-2d} \,dx
\right\}^{\frac{\gamma}{4}}
\\[6pt]
&\lesssim
\beta_\omega^{\frac{d}{2}}
\left\{
\int_{\sqrt{\alpha_\omega} / L_0}^{4}
r^{-d-1} \,dr
\right\}^{\frac{\gamma}{4}}
\lesssim
\beta_\omega^{\frac{d}{2}}
\alpha_\omega^{ -\frac{d\gamma}{8} }
=
\omega^{-\frac{d\gamma}{8}} .
\end{aligned}$$ Thus, holds.
Next, we consider $I_{\omega,in}$. We see from and $\gamma + (d-2)(p-1)= 4$ (see ) that $$\label{eq:3.19}
\begin{aligned}
I_{\omega,in}
&= \beta_\omega^{\frac{d}{2}}
\int_{|x| \le \sqrt{\alpha_\omega}/ L_0 }
|x|^{-\frac{d\gamma}{2}} K[{\widetilde{\Phi}_\omega}]^{\frac{d}{2}(p-1)}(x)\,dx
\\[6pt]
&\lesssim
\beta_\omega^{\frac{d}{2}}
\alpha_\omega^{ \frac{d(d-2)(p-1)}{8} }
\int_{|x| \le \sqrt{\alpha_\omega}/ L_0}
|x|^{- \frac{d}{2} \{ \gamma +(d-2)(p-1)\} }
\exp\left( - \frac{d(p-1) \sqrt{\alpha_\omega}}{ 4|x|} \right) dx
\\[6pt]
&\lesssim
\beta_\omega^{\frac{d}{2}}
\alpha_\omega^{ \frac{d(d-2)(p-1)}{8}}
\int_0^{\sqrt{\alpha_\omega}/L_0}
r^{-d-1}
\exp\left( \frac{-d(p-1)\sqrt{\alpha_\omega}}{4r}\right) \,dr
.
\end{aligned}$$ Furthermore, by the change of variables $s = \sqrt{\alpha_\omega} r^{-1}$, we find from that $$I_{\omega,in} \lesssim
\beta_\omega^{\frac{d}{2}}
\alpha_\omega^{ \frac{d(d-2)(p-1)}{8} - \frac{d}{2} }
\int_{L_0}^\infty s^{d-1} \exp\left( -\frac{d(p-1) s}{4}\right) ds
.$$ Since $\frac{d}{8} \gamma = \frac{d}{2} - \frac{d(d-2)(p-1)}{8}$, we see that $$\beta_\omega^{\frac{d}{2}}
\alpha_{\omega}^{ \frac{d(d-2)(p-1)}{8} - \frac{d}{2} }
= \beta_\omega^{\frac{d}{2}} \alpha_\omega^{-\frac{d}{8}\gamma}
= \omega^{-\frac{d\gamma}{8}}.$$ Hence, we conclude that $$\label{eq:3.20}
\lim_{\omega\to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} I_{\omega, in}=0
.$$ By and , we obtain the desired result .
Proof of Proposition \[proposition:3.1\] {#subsection:3.3}
----------------------------------------
Now we prove Proposition \[proposition:3.1\]:
As mentioned in , it suffices to show $$\sup_{\omega>\omega_{dec}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} \| K[{\widetilde{\Phi}_\omega}]\|_{L^{\infty}(B_{1})} <\infty.$$ Recall that $K[{\widetilde{\Phi}_\omega}]$ is a solution to . We shall prove by applying Proposition \[proposition:B.1\] to $K[{\widetilde{\Phi}_\omega}] $ with $$\label{eq:3.21}
a(x) = \frac{\alpha_\omega}{|x|^4}, \quad
b(x)
=
\frac{\beta_\omega}{|x|^\gamma} K[{\widetilde{\Phi}_\omega}]^{p-1}(x)
+ K[{\widetilde{\Phi}_\omega}]^{\frac{4}{d-2}} (x) .$$ First, we note that it follows from that for any $v \in H_{0}^{1}(B_{4})$, $$\label{eq:3.22}
\int_{B_{4}} \frac{\alpha_\omega}{|x|^4} K[{\widetilde{\Phi}_\omega}](x) |v(x)|\,dx
<\infty.$$ By , Proposition \[proposition:2.1\] and Sobolev’s inequality, one has $$\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left\| K[{\widetilde{\Phi}_\omega}] - K[W] \right\|_{L^{2^\ast}({\mathbb{R}^d})} = 0,$$ implying that the family $\{ K [ {\widetilde{\Phi}_\omega}]^{2^\ast} \}$ is uniformly integrable. Hence, it is not difficult to check that $$\label{eq:3.23}
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left\| K[{\widetilde{\Phi}_\omega}]^{\frac{4}{d-2}} - K[W]^\frac{4}{d-2} \right\|_{L^{\frac{d}{2}}({\mathbb{R}^d})} = 0.$$ Combining and with Lemma \[proposition:3.4\], we may apply Proposition \[proposition:B.1\] (i) to show that for every $q>1$ there exists $\omega_{dec,q} > 0$ such that $$\label{eq:3.24}
\sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left\| K[{\widetilde{\Phi}_\omega}]^{q} \right\|_{H^1(B_1)} \leq C_{q}.$$
Next, let $\eta$ be a non-increasing smooth function on $[0,\infty)$ such that $\eta(r)=1$ for $r \leq 1/2$ and $\eta(r)=0$ for $r\ge 1$. It is easily seen that for any $x \in \mathbb{R}^{d}\setminus \{0\}$ and any $q>1$, $$\label{eq:3.25}
K \big[\eta K[{\widetilde{\Phi}_\omega}]^{q} \big] (x)
=
|x|^{-(d-2)} \eta \left( \frac{1}{|x|} \right)
K[{\widetilde{\Phi}_\omega}]^{q} \left(\frac{x}{|x|^{2}} \right)
=
|x|^{(d-2)(q-1)}
\eta \left( \frac{1}{|x|} \right){\widetilde{\Phi}_\omega}^q (x).$$ It follows from , and that for any $q>1$, there exists $C_{q}>0$ such that $$\label{eq:3.26}
\sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left\| |x|^{(d-2)(q-1)}\eta \left(\frac{1}{|x|}\right) {\widetilde{\Phi}_\omega}^{q} \right\|_{{\dot{H}^1}}
= \sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}\left\| \eta K[{\widetilde{\Phi}_\omega}]^{q} \right\|_{{\dot{H}^1}}
\le C_{q}
.$$ Furthermore, by [@Berestycki-Lions Lemma A.III] and , it holds that $$\sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left( |x|^{(d-2)(q-1)} {\widetilde{\Phi}_\omega}^{q}(x) \right)
\lesssim C_{q} |x|^{-\frac{d-2}{2}}$$ for all $|x|\ge 2$. This implies that for any $q>1$ and any $|x|\le 1/2$, $$\label{eq:3.27}
\sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} K[{\widetilde{\Phi}_\omega}](x) \lesssim
C_{q}|x|^{-\frac{d-2}{2q}}.$$
To prove , we shall apply Proposition \[proposition:B.1\] (ii). Since we have , and , what remains to prove is that there exist $q_0>d/2$ and $\omega_{dec} > 0$ such that $$\label{eq:3.28}
\sup_{\omega > \omega_{dec}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\int_{|x| \leq 4}
\left| \frac{\beta_\omega}{|x|^\gamma} K[{\widetilde{\Phi}_\omega}]^{p-1} \right|^{q_0} dx
\le C_{q_0}.$$
To this end, we divide the proof of into two cases.
[**Case 1:**]{} $\frac{d}{d-2}<p<\frac{d+2}{d-2}$.
We first remark that the condition $\frac{d}{d-2} < p < \frac{d+2}{d-2}$ implies $ 0 < \gamma < 2 $. Therefore, we may choose $q_0>d/2$ and sufficiently large $q > 1$ such that $$\left[ \gamma + \frac{(d-2)(p-1)}{2q} \right] q_0 < d.$$ It follows from that $$\sup_{\omega > \omega_{dec,q}} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}}
\left(\frac{1}{|x|^\gamma} K[ {\widetilde{\Phi}_\omega}]^{p-1} (x)\right)^{q_0}
\lesssim C_q |x|^{ - \gamma q_0 - \frac{(d-2)(p-1)}{2q} q_0 } \in L^1(B_{1/2}).$$ From this, holds.
[**Case 2:**]{} $1<p\le \frac{d}{d-2}$.
We remark that $1 < p \leq \frac{d}{d-2}$ gives $ 2 \leq \gamma < 4$ and $\frac{d}{2} < \frac{2d}{\gamma}$. Let $q_0 \in ( \frac{d}{2} , \frac{2d}{\gamma} )$ and we claim that holds for this $q_0$. For this purpose, we remark that by $1 - \frac{d(p-1)}{2} < 1$, there exist $\theta > \frac{1}{2}$ and $s_1 > 1$ so that $$\label{eq:3.29}
\frac{4d \theta }{4\theta -1}>\gamma q_{0}, \quad
\frac{4d \theta }{(4\theta -1)\gamma q_{0}} \ge s_{1}>1, \quad
\frac{d(p-1)}{2} s_1 > s_1 - 1.$$ Next, we divide the integral into two parts: $$\begin{aligned}
J_{\omega,out}
&:=
\int_{ \alpha_{\omega}^{\theta} \le |x|\le 4}
\Big| \frac{\beta_{\omega}}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p-1}
\Big|^{q_{0}}
dx ,
\\[6pt]
J_{\omega,in}
&:=
\int_{|x|< \alpha_{\omega}^{\theta}}
\Big| \frac{\beta_{\omega}}{|x|^{\gamma}}
K[{\widetilde{\Phi}_\omega}]^{p-1}
\Big|^{q_{0}}
dx .\end{aligned}$$
We first consider $J_{\omega,out}$. Since $\frac{(p-1)q_{0}s_{1}}{s_{1}-1}>1$ holds due to $q_0 > \frac{d}{2}$ and , by Hölder’s inequality and , we see that $$\begin{split}
J_{\omega,out}
&= \beta_{\omega}^{q_{0}}
\int_{ \alpha_{\omega}^{\theta} \le |x|\le 4}
|x|^{- \gamma q_{0} }
K[{\widetilde{\Phi}_\omega}]^{(p-1)q_{0}}\,dx
\\[6pt]
&\le
\beta_{\omega}^{q_{0}}
\left\{
\int_{|x|\le 4}
K[{\widetilde{\Phi}_\omega}]^{\frac{(p-1) q_{0} s_{1}}{s_{1}-1}}
\,dx
\right\}^{1-\frac{1}{s_{1}}}
\left\{
\int_{ \alpha_{\omega}^{\theta} \le |x|\le 4}
|x|^{- \gamma q_{0} s_{1}} \,dx
\right\}^{\frac{1}{s_{1}}}
\\[6pt]
&\le C(q_{0},s_{1})
\beta_\omega^{q_{0}}
\left\{
\int_{\alpha_{\omega}^{\theta}}^{4}
r^{-\gamma q_{0} s_{1}+d-1} \,dr
\right\}^{\frac{1}{s_{1}}}
\end{split}$$ where $C(q_{0}, s_{1})>0$ depends only on $d$, $p$, $q_{0}$ and $s_{1}$. Recalling the definitions of $\alpha_{\omega}$, $\beta_{\omega}$ and $\gamma$, we obtain $$\label{eq:3.30}
\begin{aligned}
J_{\omega,out}
&\le
C(q_{0}, s_{1})\beta_{\omega}^{q_{0}}
\alpha_{\omega}^{\frac{(-\gamma q_{0} s_{1}+d)\theta}{ s_{1}}}
=
C(q_{0}, s_{1})
M_{\omega}^{-\frac{\gamma q_{0}}{d-2}}
\omega^{\frac{(-\gamma q_{0}s_{1} + d)\theta}{s_{1}}}
M_{\omega}^{\frac{4\theta \gamma q_{0} s_{1} - 4 d\theta}{(d-2)s_{1}}}
\\[6pt]
&=
C(q_{0}, s_{1})
\omega^{\frac{(-\gamma q_{0}s_{1} + d)\theta}{s_{1}}}
M_{\omega}^{-\frac{(4\theta -1) \gamma q_{0}}{(d-2) s_{1}}
\big\{ \frac{4d\theta}{(4\theta -1)\gamma q_{0}} - s_{1} \big\}}
.
\end{aligned}$$ Since $2 \leq \gamma < 4$ and $s_1 > 1$ imply $-\gamma q_0 s_1 + d < 0$, and yield $$\label{eq:3.31}
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} J_{\omega,out}=0.$$
Next, we consider $J_{\omega,in}$. Since $\theta > \frac{1}{2}$ and $\lim_{\omega \to \infty}\alpha_\omega = 0$, we may assume $\alpha^{-\theta}_\omega \leq L_0\alpha_\omega^{-\frac{1}{2}}$. Thus, from Lemma \[proposition:3.2\] (or ) and the definition of $\gamma$, it follows that $$\label{eq:3.32}
\begin{aligned}
J_{\omega,in}
&=
\beta_{\omega}^{q_{0}}
\int_{|x| \le \alpha_{\omega}^{\theta} }
|x|^{-\gamma q_{0}}
K[{\widetilde{\Phi}_\omega}]^{(p-1)q_{0}} \,dx
\\[6pt]
&\lesssim
\beta_{\omega}^{q_{0}}
\alpha_\omega^{ \frac{(d-2)(p-1)q_{0}}{4} }
\int_{|x| \le \alpha_{\omega}^{\theta}}
|x|^{-4 q_{0}}
\exp\left( - \frac{(p-1)q_{0} \sqrt{\alpha_\omega} }{ 2 |x|} \right) dx
\\[6pt]
&\lesssim
\beta_{\omega}^{q_{0}}
\alpha_\omega^{ \frac{(d-2)(p-1)q_{0}}{4} }
\int_{0}^{\alpha_{\omega}^{\theta}}
s^{- 4q_{0}+d-1}
\exp\left( - \frac{(p-1)q_{0} \sqrt{\alpha_\omega}}{2 s} \right) \,ds
.
\end{aligned}$$ Using the change of variables $t = \alpha_{\omega}^{\theta} s^{-1}$, we find from $\theta>1/2$ that $$\label{eq:3.33}
\begin{split}
J_{\omega,in}
&\lesssim
\beta_{\omega}^{q_{0}}
\alpha_{\omega}^{\frac{(d-2)(p-1)q_{0}}{4}-(4q_{0}-d)\theta}
\int_{1}^{\infty}
t^{4q_{0}-d-1}
\exp\left( -\frac{(p-1)q_{0} \alpha_{\omega}^{-(\theta-1/2)}}{2} t\right)\, dt
\\[6pt]
&\le
C(q_{0}, \theta)
\beta_{\omega}^{q_{0}}
\alpha_{\omega}^{\frac{(d-2)(p-1)q_{0}}{4}-(4q_{0}-d)\theta}
\int_{1}^{\infty}
\exp\left( -\frac{(p-1)q_{0} \alpha_{\omega}^{-(\theta-1/2)}}{4} t\right)\, dt
\\[6pt]
&\le
C(q_{0}, \theta)
\beta_{\omega}^{q_{0}}
\alpha_{\omega}^{\frac{(d-2)(p-1)q_{0}}{4}-(4q_{0}-d)\theta +\theta -\frac{1}{2}}
\exp\left( -\frac{(p-1)q_{0} \alpha_{\omega}^{-(\theta-1/2)}}{4}\right)
\\[6pt]
&\le C(q_{0}, \theta) \beta_{\omega}^{q_{0}}
\end{split}$$ where $C(q_{0}, \theta)$ denotes a positive constant depending only on $d$, $p$, $q_{0}$ and $\theta$. Thus, it follows from and $\lim_{\omega\to \infty}\beta_{\omega}=0$ that $$\label{eq:3.34}
\lim_{\omega \to \infty} \sup_{{\widetilde{\Phi}_\omega}\in {\widetilde{\mathcal{G}}_\omega}} J_{\omega,in}=0.$$ By and , we have and complete the proof.
Proof of Theorem \[theorem:1.1\] {#section:4}
================================
In this section, we shall give a proof of Theorem \[theorem:1.1\]. In the sequel, we assume $d \geq 5$, hence, by Proposition \[proposition:1.1\], we have ${\mathcal{G}_\omega}\neq \emptyset$ and the results of Sections \[section:2\] and \[section:3\] hold.
Our proof is based on the ideas in [@Grossi; @GLP] and we first prove the uniqueness by contradiction. Therefore, we suppose to the contrary that there exists a sequence $\{\omega_{n}\}$ in $(0,\infty)$ such that $\lim_{n\to \infty}\omega_{n}=\infty$ and for each $n\ge 1$, $\Phi_{n,1}, \Phi_{n,2}\in \mathcal{G}_{\omega_{n}}$ and $\Phi_{n,1}\neq \Phi_{n,2}$. For $j=1,2$, we set
$$\label{eq:4.1}
\begin{aligned}
M_{n,j}&:=\max_{\mathbb{R}^{d}}\Phi_{n,j}=\Phi_{n,j}(0),
\\[6pt]
\widetilde{\Phi}_{n,j}(x)&:=M_{n,j}^{-1}\Phi_{n,j}(M_{n,j}^{-\frac{2}{d-2}}x),
\\[6pt]
\alpha_{n,j} &:= \omega_n M_{n,j}^{-\frac{4}{d-2}},
\\[6pt]
\beta_{n,j} &:= M_{n,j}^{p-1-\frac{4}{d-2}},
\end{aligned}$$
and $$ \mu_{n}:=\frac{M_{n,2}}{M_{n,1}}.$$ We shall derive a contradiction. Let us begin with the following lemma:
\[proposition:4.1\] Assume $d\ge 5$ and $1<p<\frac{d+2}{d-2}$. Then, for $j=1,2$, $$\label{eq:4.2}
\lim_{n \to \infty} \frac{\beta_{n,j}}{\alpha_{n,j}}
=
\frac{2(p+1)}{4-(d-2)(p-1)}
\frac{\| W\|^2_{L^{2}}}{\|W\|_{L^{p+1}}^{p+1}}$$ where $\alpha_{n,j}, \beta_{n,j}$ are given in , and $W$ in . Furthermore, it holds that $$\label{eq:4.3}
\lim_{n\to \infty} \mu_{n} = 1 .$$
The claim follows from identity and Corollary \[theorem:3.1\]. Furthermore, an elementary computation together with shows .
In what follows, owing to Lemma \[proposition:4.1\], we may assume the that for all $n$ and $j=1,2$: $$\begin{aligned}
\label{eq:4.4}
&\beta_{n,j} \lesssim \alpha_{n,j},
\\[6pt]
\label{eq:4.5}
&\frac{1}{2}\le \mu_{n} \le \frac{3}{2}. \end{aligned}$$ Moreover, Proposition \[proposition:2.1\], Corollary \[theorem:3.1\] and Lemma \[proposition:4.1\] imply that $$\label{eq:4.6}
\lim_{n\to \infty}
\| {\widetilde{\Phi}_{n,1}}-W \|_{H^{1}}
=
\lim_{n\to \infty}
\| \mu_{n} {\widetilde{\Phi}_{n,2}}( \mu_{n}^{\frac{2}{d-2}}\cdot )
-W
\|_{H^{1}}
=0
.$$ Next, we define $$\begin{aligned}
\label{eq:4.7}
\Psi_{n}(x)
&:=
\Phi_{n,1}(M_{n,1}^{-\frac{2}{d-2}}x)-\Phi_{n,2}(M_{n,1}^{-\frac{2}{d-2}}x)
= M_{n,1} \Big\{ {\widetilde{\Phi}_{n,1}}(x) - \mu_n {\widetilde{\Phi}_{n,2}}(\mu_{n}^{\frac{2}{d-2}} x) \Big\},
\\[6pt]
\label{eq:4.8}
\wt{z}_{n}(x) &:=\frac{\Psi_{n}(x)}{\| \Psi_{n} \|_{L^{\infty}}}. \end{aligned}$$ Since $\widetilde{\Phi}_{n,1}$ and $\mu_n {\widetilde{\Phi}_{n,2}}(\mu_{n}^{\frac{2}{d-2}} x)$ are solutions to the same equation $$-\Delta u + \alpha_{n,1} u = \beta_{n,1} u^p + u^{(d+2)/(d-2)},$$ we can verify that $$\label{eq:4.9}
-\Delta \wt{z}_n = - \alpha_{n,1} \wt{z}_n + p \beta_{n,1} \int_0^1 V_n^{p-1} (x,\theta) d \theta \wt{z}_n
+ \frac{d+2}{d-2} \int_0^1 V_n^{\frac{4}{d-2}} (x,\theta) d \theta \wt{z}_n$$ where $$\label{eq:4.10}
V_{n}(x,\theta):=\theta \widetilde{\Phi}_{n,1}(x)+(1-\theta)\mu_{n} \widetilde{\Phi}_{n,2}(\mu_{n}^{\frac{2}{d-2}}x).$$ We first show that $\{\wt{z}_n\}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$:
\[proposition:4.2\] Assume $d\ge 5$ and $1 < p < \frac{d+2}{d-2}$. Then, $\{\wt{z}_n\}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$.
Since $\wt{z}_n \in H^1({\mathbb{R}^d})$, using $\wt{z}_n$ as test function to , we have $$\label{eq:4.11}
\begin{split}
\alpha_{n,1} \| \wt{z}_{n} \|_{L^{2}}^{2}
+
\| \nabla \wt{z}_{n} \|_{L^{2}}^{2}
&\lesssim
\beta_{n,1} \int_{{\mathbb{R}^d}} \int_{0}^1 V_n^{p-1}(x,\theta) d \theta \, |\wt{z}_{n}|^{2} \, dx
\\[6pt]
& \quad + \int_{{\mathbb{R}^d}} \int_{0}^{1} V_{n}^{\frac{4}{d-2}} (x, \theta)
\, d \theta \, |\wt{z}_{n} |^{2} \, dx.
\end{split}$$ We see from Young’s inequality, and that for any $\delta >0$, $$\label{eq:4.12}
\begin{split}
&\beta_{n,1}
\int_{{\mathbb{R}^d}} \int_{0}^1 V_n^{p-1}(x,\theta) d \theta \, |\wt{z}_{n}|^{2}
\,dx
\\[6pt]
\lesssim \, &
\beta_{n,1}
\delta^{\frac{4(p-1)}{\gamma}}
\int_{{\mathbb{R}^d}} |\wt{z}_{n}|^{2}
\, dx
+
\delta^{-\frac{4}{d-2}} \beta_{n,1}
\int_{{\mathbb{R}^d}} \int_{0}^{1} V_{n}^{\frac{4}{d-2}} (x,\theta) \, d \theta
\, |\wt{z}_{n}|^{2}
\, dx
\\[6pt]
\lesssim \, &
\alpha_{n,1}
\delta^{\frac{4(p-1)}{\gamma}} \| \wt{z}_{n} \|_{L^{2}}^{2}
+
\delta^{-\frac{4}{d-2}} \beta_{n,1}
\int_{{\mathbb{R}^d}} \int_{0}^{1} V_{n}^{\frac{4}{d-2}} (x,\theta) \, d \theta
\, |\wt{z}_{n}|^{2}
\, dx,
\end{split}$$ where the implicit constants are independent of $\delta$.
Next, set ${\varepsilon}_{0} :=\frac{4}{3(d-2)}$ and $p_{0} := \frac{2^\ast}{2-{\varepsilon}_0}$. Note that $$1-\frac{1}{p_{0}} = \frac{4+(d-2){\varepsilon}_0 }{2d} = \frac{8}{3d} .$$ Since $\|\wt{z}_n \|_{L^\infty} = 1$ holds by definition, it follows from Hölder’s inequality, Proposition \[proposition:3.1\] and Sobolev’s inequality that for all $n \geq 1$, $$\label{eq:4.13}
\begin{aligned}
\int_{{\mathbb{R}^d}} \int_0^{1} V_{n}^{\frac{4}{d-2}} (x, \theta) d \theta |\wt{z}_n|^{2} dx
&\le
\int_{{\mathbb{R}^d}} \int_0^{1} V_{n}^{\frac{4}{d-2}} (x, \theta) d \theta |\wt{z}_n |^{2-{\varepsilon}_0} d x
\\[6pt]
&\leq
\left\| \int_0^{1} V_{n}^{\frac{4}{d-2}} (x, \theta) d \theta
\right\|_{L^{\frac{2d}{4+(d-2)\varepsilon_{0}}}}
\| |\wt{z}_{n}|^{2-{\varepsilon}_0} \|_{L^{p_{0}}}
\\[6pt]
&\lesssim
\left\| (1+|x|)^{-4} \right\|_{L^{\frac{2d}{4+(d-2)\varepsilon_{0}}}}
\left\|\wt{z}_{n} \right\|_{L^{2^{*}}}^{2-\varepsilon_{0}}
\lesssim \| \nabla \wt{z}_n \|_{L^2}^{2-{\varepsilon}_0}.
\end{aligned}$$
Choosing $\delta$ sufficiently small depending on $d$ and $p$, we find from , , and $\lim_{n\to \infty}\beta_{n,1}=0$ that $$\| \nabla \wt{z}_{n} \|_{L^{2}}^{2}
\le
C \| \nabla \wt{z}_n \|_{L^2}^{2-{\varepsilon}_0}$$ where $C>0$ depends only on $d$ and $p$. Hence, $\{ \wt{z}_{n} \}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$.
Next, we derive a uniform decay estimate for $\{\wt{z}_n\}$.
\[proposition:4.3\] Assume $d\ge 5$ and $1<p<\frac{d+2}{d-2}$. Then, there exists $C_{1}>0$ such that for any $x \in {\mathbb{R}^d}$ and $n \geq 1$, $$\label{eq:4.14}
|\wt{z}_{n}(x)| \leq C_{1} |x|^{-(d-2)}
.$$
We see from and Lemma \[proposition:4.2\] that $\{K[\wt{z}_{n}] \}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$. Furthermore, it follows from that $$\label{eq:4.15}
\begin{split}
&-\Delta K[\wt{z}_{n}]
+ \frac{\alpha_{n,1}}{|x|^{4}} K[\wt{z}_{n}]
\\[6pt] = \, & \frac{1}{|x|^{4}}
\left[ p \beta_{n,1} \int_0^1 V_{n}^{p-1} \left(\frac{x}{|x|^{2}},\theta \right) d \theta +
\frac{d+2}{d-2} \int_0^1 V_{n}^{\frac{4}{d-2}} \left(\frac{x}{|x|^{2}},\theta\right) d \theta \right] K[\wt{z}_{n}].
\end{split}$$
To prove , we shall apply Proposition \[proposition:B.1\] (ii). We first remark that by Lemma \[proposition:3.2\], $\Psi_n$ decays exponentially and so does $\wt{z}_n$. Thus, for any $n\ge 1$ and any $v \in H_{0}^{1}(B_{4})$, we have $$\label{eq:4.16}
\int_{B_{4}} \frac{\alpha_{n,1}}{|x|^{4}}
|K[\wt{z}_{n}](x)||v(x)|\,dx <\infty
.$$
Next, it follows from that for each $r > 0$, $x \in {\mathbb{R}^d}\setminus \{0\}$ and $ \theta \in [0,1]$, $$\label{eq:4.17}
V_n^r \left( \frac{x}{|x|^2} , \theta \right) \leq
C_r \left\{\widetilde{\Phi}_{n,1}^r \left( \frac{x}{|x|^2} \right)
+ \mu_{n}^r \widetilde{\Phi}_{n,2}^r \left( \mu_{n}^{\frac{2}{d-2}} \frac{x}{|x|^2} \right)
\right\}$$ where $C_r$ depends only on $r$. When $r = p - 1$, by $\gamma = 4 - (d-2)(p-1)$, we get $$\begin{aligned}
\frac{\beta_{n,1}}{|x|^4}
\int_0^1 V_n^{p-1} \left( \frac{x}{|x|^2} , \theta \right) d \theta
& \leq C_p \frac{\beta_{n,1}}{|x|^4}
\left\{ \wt{\Phi}_{n,1}^{p-1} \left( \frac{x}{|x|^2} \right)
+ \mu^{p-1} \wt{\Phi}_{n,2}^{p-1}
\left( \mu_n^{\frac{2}{d-2}} \frac{x}{|x|^2} \right) \right\}
\\
& \leq C_p \frac{\beta_{n,1}}{|x|^{\gamma}}
\left\{ K[ \wt{\Phi}_{n,1} ]^{p-1} (x) +
\mu_n^{-(p-1)} K [ \wt{\Phi}_{n,2} ]^{p-1} \left( \mu_n^{-\frac{2}{d-2}} x \right)
\right\}.
\end{aligned}$$ Hence, recalling , in the proof of Proposition \[proposition:3.1\], we see that for some $q_0 > \frac{d}{2}$, $$\label{eq:4.18}
\sup_{n \geq 1}\left\| \frac{\beta_{n,1}}{|x|^4}
\int_0^1 V_n^{p-1} \left( \frac{x}{|x|^2} , \theta \right) d \theta \right\|_{L^{q_0}(B_4)}
< \infty.$$
On the other hand, when $r = \frac{4}{d-2}$, gives us that $$\frac{1}{|x|^4}V_n^{\frac{4}{d-2}} \left( \frac{x}{|x|^2} , \theta \right)
\leq
C_d \left\{ K[ \wt{\Phi}_{n,1} ]^{\frac{4}{d-2}} (x)
+ K[ \wt{\Phi}_{n,2} ]^{\frac{4}{d-2}} \left( \mu_n^{\frac{2}{d-2}} x \right) \right\}.$$ By Proposition \[proposition:3.1\], $\{ K[\wt{\Phi}_{n,j}] \}$ is bounded in $L^\infty({\mathbb{R}^d})$ for $j =1,2$. Hence, $$\label{eq:4.19}
\sup_{n \geq 1}
\left\| \frac{1}{|x|^4} \int_0^1V_n^{\frac{4}{d-2}} \left( \frac{x}{|x|^2} , \theta \right)
d \theta
\right\|_{L^\infty(B_4)} < \infty.$$
From , , and , we apply Proposition \[proposition:B.1\] (ii) to obtain $$\sup_{n\geq 1} \left\| K[\tilde{z}_n] \right\|_{L^\infty(B_1)} < \infty.$$ Therefore, Lemma \[proposition:4.3\] holds.
As a corollary of Lemma \[proposition:4.3\], we obtain
\[proposition:4.4\] Assume $d\ge 5$ and $1<p<\frac{d+2}{d-2}$. Then $$\sup_{n\ge 1}\|\wt{z}_{n}\|_{L^{2}}<\infty.$$
Before proving Theorem \[theorem:1.1\], we use the following identity which is easily obtained from elementary calculations:
\[proposition:4.5\] Assume $d\ge 5$ and $1\le q \le \frac{d+2}{d-2}$. Then, the following holds: $$\begin{aligned}
\int_{\mathbb{R}^{d}} W^{q}(x) \Lambda W(x) \,dx
&=
-\frac{4-(d-2)(q-1)}{2(q+1)}\|W\|_{L^{q+1}}^{q+1}
\end{aligned}$$ where $\Lambda$ is defined in .
Now, we derive a contradiction and prove the uniqueness part:
By Lemmas \[proposition:4.2\] and \[proposition:4.4\], $\{\wt{z}_n\}$ is bounded in $H^1({\mathbb{R}^d})$ and we may assume that $$\label{eq:4.20}
\lim_{n\to \infty}\wt{z}_n = \wt{z}_\infty \quad
{\rm weakly \ in\ } H^1({\mathbb{R}^d}).$$ Moreover, recalling that $\wt{z}_n$ satisfies , by elliptic regularity with Corollary \[theorem:3.1\], and , we can see that $$\label{eq:4.21}
\lim_{n\to \infty} \wt{z}_n = \wt{z}_\infty \quad {\rm in} \ C^2_{\rm loc}({\mathbb{R}^d}),$$ hence, $\wt{z}_\infty$ is a solution to $$\label{eq:4.22}
-\Delta \wt{z}_{\infty}
-
\frac{d+2}{d-2} W^{\frac{4}{d-2}}
\wt{z}_{\infty}
=
0
.$$ Thus, we find from , and the radial symmetry of $\wt{z}_{\infty}$ that either $\wt{z}_{\infty}\equiv 0$ or $\wt{z}_{\infty} = \kappa \Lambda W$ with $\kappa \neq 0$.
First, we suppose that $\wt{z}_{\infty}\equiv 0$. Then, it follows from and that $1= \| \wt{z}_n \|_{L^\infty} \to 0$, which is a contradiction.
Next, assume $\wt{z}_{\infty}=\kappa \Lambda W$ for some $\kappa \neq 0$. Using and , we see that $$\label{eq:4.23}
\begin{split}
&\frac{M_{n,1}}{\|\Psi_{n}\|_{L^{\infty}}}
\frac{\omega_{n}}{d}
\bigm(
\|\Phi_{n,1}\|_{L^{2}}^{2}-\|\Phi_{n,2}\|_{L^{2}}^{2}
\bigm)
\\[6pt]
= \, &
\frac{M_{n,1}}{\|\Psi_{n}\|_{L^{\infty}}}
\frac{4-(d-2)(p-1)}{2d(p+1)}
\bigm(
\|\Phi_{n,1}\|_{L^{p+1}}^{p+1}-\|\Phi_{n,2}\|_{L^{p+1}}^{p+1}
\bigm)
.
\end{split}$$ For the left-hand side of , using the change of variables, we observe that $$\label{eq:4.24}
\begin{aligned}
&\frac{M_{n,1}}{\|\Psi_{n}\|_{L^{\infty}}}
\frac{\omega_{n}}{d}
\bigm(
\|\Phi_{n,1}\|_{L^{2}}^{2}-\|\Phi_{n,2}\|_{L^{2}}^{2}
\bigm)
\\[6pt]
=&\,
M_{n,1}\frac{\omega_{n}}{d}
\int_{\mathbb{R}^{d}}
\frac{\Phi_{n,1}(x)-\Phi_{n,2}(x)}{\|\Psi_{n}\|_{L^{\infty}}}
(\Phi_{n,1}(x)+\Phi_{n,2}(x))\,dx
\\[6pt]
=&\,
\frac{\alpha_{n,1}}{d}
\int_{\mathbb{R}^{d}}
\widetilde{z}_{n}(x)
\Big[\wt{\Phi}_{n,1}(x)
+
\mu_{n} \wt{\Phi}_{n,2}(\mu_{n}^{\frac{2}{d-2}} x )\Big] \,dx
.
\end{aligned}$$ In a similar way, the right-hand side of becomes $$\label{eq:4.25}
\begin{aligned}
&
\frac{M_{n,1}}{\|\Psi_{n}\|_{L^{\infty}}}
\frac{4-(d-2)(p-1)}{2d(p+1)}
\bigm(
\|\Phi_{n,1}\|_{L^{p+1}}^{p+1}-\|\Phi_{n,2}\|_{L^{p+1}}^{p+1}
\bigm)
\\[6pt]
=&\,
M_{n,1}
\frac{4-(d-2)(p-1)}{2d}
\int_{\mathbb{R}^{d}} \int_{0}^{1}\Big[ \theta \Phi_{n,1}+(1-\theta)\Phi_{n,2}\Big]^{p}d\theta
\frac{\Phi_{n,1}-\Phi_{n,2}}{\|\Psi_{n}\|_{L^{\infty}}}
\,dx
\\[6pt]
=&\,
\beta_{n,1} \frac{4-(d-2)(p-1)}{2d}
\int_{\mathbb{R}^{d}} \int_{0}^{1}\Big[ \theta \wt{\Phi}_{n,1}(x)
+
(1-\theta) \mu_{n} \wt{\Phi}_{n,2}(\mu_{n}^{\frac{2}{d-2}} x ) \Big]^{p}d\theta \widetilde{z}_{n}(x)\,dx
.
\end{aligned}$$ Hence, we obtain the following identity from through with : $$\label{eq:4.26}
\begin{aligned}
&\frac{1}{2}
\int_{\mathbb{R}^{d}}
\wt{z}_{n}(x)
\Big[\wt{\Phi}_{n,1}(x)
+
\mu_{n} \wt{\Phi}_{n,2}(\mu_{n}^{\frac{2}{d-2}} x )\Big] \,dx
\\
= &
\frac{\beta_{n,1}}{\alpha_{n,1}}
\frac{4-(d-2)(p-1)}{4}
\int_{\mathbb{R}^{d}} \int_{0}^{1} V_{n}(x,\theta)^{p}d\theta \wt{z}_{n}(x)\,dx .
\end{aligned}$$
Since Corollary \[theorem:3.1\] and Lemma \[proposition:4.1\] yield $$\lim_{n\to \infty} {\widetilde{\Phi}_{n,1}}= W, \quad \lim_{n\to \infty} \mu_{n} \wt{\Phi}_{n,2}(\mu_{n}^{\frac{2}{d-2}} \cdot )
= W \quad {\rm strongly \ in} \ L^2({\mathbb{R}^d}),$$ recalling with $\wt{z}_\infty = \kappa \Lambda W$, we see from Lemma \[proposition:4.5\] that $$\label{eq:4.27}
\begin{split}
&
\lim_{n\to \infty}
\frac{1}{2}
\int_{\mathbb{R}^{d}}
\widetilde{z}_{n}(x)
\left[\wt{\Phi}_{n,1}(x)+\mu_{n} \wt{\Phi}_{n,2}( \mu_{n}^{\frac{2}{d-2}} x )\right] \,dx
=
\int_{\mathbb{R}^{d}}\kappa \Lambda W(x) W(x) \,dx
=-\kappa \|W\|_{L^{2}}^{2}.
\end{split}$$ In a similar way, we can check that
$$\lim_{n\to \infty} \int_{0}^{1} V_n^p (x,\theta) \,d\theta = W^{p}
\quad \mbox{strongly in $L^2({\mathbb{R}^d})$}$$ and $$\label{eq:4.28}
\begin{split}
\lim_{n\to \infty}
\int_{\mathbb{R}^{d}} \int_{0}^{1}
V_n(x,\theta)^p
d\theta \wt{z}_{n}(x)\,dx
&=
\int_{\mathbb{R}^{d}}
W^{p}(x) \kappa \Lambda W(x)\,dx
\\[6pt]
&=
-\kappa
\frac{4-(d-2)(p-1)}{2(p+1)}\|W\|_{L^{p+1}}^{p+1}
.
\end{split}$$ Putting , , and together, we find that $$\begin{split}
-\kappa \|W\|_{L^{2}}^{2}
&=
\frac{2(p+1)}{4-(d-2)(p-1)}
\frac{\| W\|_{L^{2}}^2}{\|W\|_{L^{p+1}}^{p+1}}
\\[6pt]
&\qquad \times
\frac{4-(d-2)(p-1)}{4}
\left\{ -\kappa
\frac{4-(d-2)(p-1)}{2(p+1)}\|W\|_{L^{p+1}}^{p+1} \right\}
\\[6pt]
&= -\kappa \frac{4 - (d-2)(p-1)}{4} \|W\|_{L^{2}}^{2}
.
\end{split}$$ This contradicts $\kappa \neq 0$, and the uniqueness in Theorem \[theorem:1.1\] holds.
Next, we shall prove the nondegeneracy in ${H^1_{\rm rad}}({\mathbb{R}^d})$.
From the uniqueness part, there exists $\wt{\omega}_\ast>0$ such that if $\omega > \wt{\omega}_\ast$, then admits a unique radial positive ground state and we denote it by ${\Phi_\omega}$. Our aim is to find $\omega_\ast \geq \wt{\omega}_\ast$ such that $$\label{eq:4.29}
{\rm Ker} \, L_{\Phi_\omega} |_{{H^1_{\rm rad}}({\mathbb{R}^d})} = \left\{ 0 \right\}
\quad {\rm for\ every} \ \omega > \omega_\ast.$$
In order to prove , we argue indirectly and suppose to the contrary that there exist $\{ \omega_n \}$ and $\{u_n\}$ such that $$\wt{\omega}_\ast < \omega_n \to \infty, \quad
u_n \in H^1_{\rm rad} ({\mathbb{R}^d}) \setminus \left\{ 0 \right\},
\quad L_{{\Phi_{\omega_n}}} u_n = 0.$$ Remark that we may assume $u_n: {\mathbb{R}^d}\to {\mathbb{R}}$ and $\| u_n \|_{L^\infty({\mathbb{R}^d})} = 1$ due to the linearity of $L_{{\Phi_{\omega_n}}}$ and elliptic regularity. Set $$v_n(x) := u_n \left( M_n^{-\frac{2}{d-2}} x \right), \quad
M_n := {\Phi_{\omega_n}}(0) = \| {\Phi_{\omega_n}}\|_{L^\infty}.$$ Then we observe that $$\label{eq:4.30}
-\Delta v_n + \alpha_n v_n =
\left[
p \beta_n {\widetilde{\Phi}_{\omega_n}}^{p-1} + \frac{d+2}{d-2}
{\widetilde{\Phi}_{\omega_n}}^{\frac{4}{d-2}}
\right] v_n \quad {\rm in} \ {\mathbb{R}^d}, \quad
\| v_n \|_{L^\infty} = 1$$ where $\alpha_n := \omega_n M_{n}^{-\frac{4}{d-2}}$, $\beta_n := M_n^{p-1-\frac{4}{d-2}}$ and ${\widetilde{\Phi}_{\omega_n}}(x) := {\Phi_{\omega_n}}( M_n^{-\frac{2}{d-2}} x )$. By Proposition \[proposition:3.1\], we can argue as in Lemma 4.2 to show that $$\label{eq:4.31}
\text{$\{v_n\}$ is bounded in ${\dot{H}^1({\mathbb{R}^d})}$}.$$
Next, let us consider the Kelvin transform of $v_n$ and write $K[v_n]$ for it. Then $K[v_n]$ satisfies $$-\Delta K[v_n] + \alpha_n \frac{1}{|x|^4} K[v_n]
= \left\{ p \frac{\beta_n}{|x|^{\gamma}} K[{\widetilde{\Phi}_{\omega_n}}]^{p-1}
+ \frac{d+2}{d-2} K[ {\widetilde{\Phi}_{\omega_n}}]^{\frac{4}{d-2}} \right\} K [v_n],$$ where $\gamma=4-(d-2)(p-1)$. We remark that each $v_n$ has an exponential decay and this fact can be proved reasoning as for Lemma \[proposition:3.2\]. Therefore, applying the argument in Lemma 4.3, we get the uniform decay estimate for $\{v_n\}$: $$\label{eq:4.32}
\sup_{n \geq 1} \left| v_n(x) \right| \leq C_0 \left( 1 + |x| \right)^{-(d-2)}
\quad {\rm for\ all} \ x \in {\mathbb{R}^d}.$$
By and , $\{v_n\}$ is bounded in $H^1({\mathbb{R}^d})$ and we may assume that there exists a $v_\infty \in {H^1_{\rm rad}}({\mathbb{R}^d})$ so that $$v_n \rightharpoonup v_\infty \quad \text{weakly in $H^1({\mathbb{R}^d})$}.$$ Using Proposition \[proposition:2.1\], , and elliptic regularity, we have $$\lim_{n\to \infty}v_n = v_\infty \quad \mbox{strongly in $C^2_{\rm loc}({\mathbb{R}^d})$}, \quad 1 = \| v_\infty \|_{L^\infty}, \quad
-\Delta v_\infty = \frac{d+2}{d-2} W^{\frac{4}{d-2}} v_\infty \quad {\rm in} \ {\mathbb{R}^d}.$$ Since $\| v_\infty \|_{L^\infty} = 1$ and $v_\infty \in {H^1_{\rm rad}}({\mathbb{R}^d})$, from , there exists a $\kappa \neq 0$ such that $$\label{eq:4.33}
v_\infty = \kappa \Lambda W.$$
Next, we consider $w_n(x) := x \cdot \nabla {\widetilde{\Phi}_{\omega_n}}(x)$. It is not difficult to check that $w_n$ satisfies $$\label{eq:4.34}
-\Delta w_n + \alpha_n w_n =
\left[ p \beta_n {\widetilde{\Phi}_{\omega_n}}^{p-1}
+ \frac{d+2}{d-2} {\widetilde{\Phi}_{\omega_n}}^{\frac{4}{d-2}} \right] w_n
+ 2 \left[ - \alpha_n {\widetilde{\Phi}_{\omega_n}}+ \beta_n {\widetilde{\Phi}_{\omega_n}}^p
+ {\widetilde{\Phi}_{\omega_n}}^{\frac{d+2}{d-2}} \right].$$ Thus, multiplying by $w_n$ and by $v_n$, it follows from the integration by parts that $$\label{eq:4.35}
\int_{{\mathbb{R}^d}}\left[ - \alpha_n {\widetilde{\Phi}_{\omega_n}}+ \beta_n {\widetilde{\Phi}_{\omega_n}}^p
+ {\widetilde{\Phi}_{\omega_n}}^{\frac{d+2}{d-2}} \right] v_n dx = 0.$$ Recall that $$\label{17/10/04/15:56}
-\Delta {\widetilde{\Phi}_{\omega_n}}+ \alpha_n {\widetilde{\Phi}_{\omega_n}}=
\beta_n {\widetilde{\Phi}_{\omega_n}}^p + {\widetilde{\Phi}_{\omega_n}}^{\frac{d+2}{d-2}} \quad {\rm in} \ {\mathbb{R}^d}.$$ Multiply by $v_{n}$ and by ${\widetilde{\Phi}_{\omega_n}}$, and then integrate them: putting these together, we obtain $$\begin{aligned}
\int_{{\mathbb{R}^d}}
\left[ \beta_n {\widetilde{\Phi}_{\omega_n}}^p + {\widetilde{\Phi}_{\omega_n}}^{\frac{d+2}{d-2}} \right] v_n dx
&= \int_{{\mathbb{R}^d}} \nabla {\widetilde{\Phi}_{\omega_n}}\cdot \nabla v_n + \alpha_n {\widetilde{\Phi}_{\omega_n}}v_n dx
\\
&= \int_{{\mathbb{R}^d}}
\left[
p \beta_n {\widetilde{\Phi}_{\omega_n}}^{p-1} + \frac{d+2}{d-2}
{\widetilde{\Phi}_{\omega_n}}^{\frac{4}{d-2}}
\right] v_n {\widetilde{\Phi}_{\omega_n}}dx,
\end{aligned}$$ which implies $$\int_{{\mathbb{R}^d}} {\widetilde{\Phi}_{\omega_n}}^{\frac{d+2}{d-2}} v_n d x
= - \frac{(d-2)(p-1)}{4} \int_{{\mathbb{R}^d}} \beta_n {\widetilde{\Phi}_{\omega_n}}^p v_n dx.$$ Combining this with , we find $$\int_{{\mathbb{R}^d}} {\widetilde{\Phi}_{\omega_n}}v_n dx
= \left[ 1 - \frac{(d-2)(p-1)}{4} \right] \frac{\beta_n}{\alpha_n} \int_{{\mathbb{R}^d}}
{\widetilde{\Phi}_{\omega_n}}^p v_n dx.$$ As $n \to \infty$, Corollary \[theorem:3.1\], Lemma \[proposition:4.1\] and yield $$\int_{{\mathbb{R}^d}} W \kappa \Lambda W d x
= \left[ 1 - \frac{(d-2)(p-1)}{4} \right]
\frac{2(p+1)}{4-(d-2)(p-1)} \frac{\|W\|_{L^2}^2}{\| W \|_{L^{p+1}}^{p+1}}
\int_{{\mathbb{R}^d}} W^p \kappa \Lambda W d x.$$ Since $\kappa \neq 0$, Lemma 4.5 gives a contradiction: $$- \| W \|_{L^2}^2 = - \left[ 1 - \frac{(d-2)(p-1)}{4} \right] \| W \|_{L^2}^2.$$ Thus, holds and we complete the proof of Theorem \[theorem:1.1\].
Existence of ground state {#section:A}
=========================
In this section, we sketch the proof of Proposition \[proposition:1.1\]. Since we restrict nonlinearities to combined power-type ones, the proof is much simpler than the general case dealt with in [@Zhang-Zou]. In particular, we can use a positive functional $\mathcal{I}_{\omega}$ given by $$\label{eq:A.1}
\begin{split}
\mathcal{I}_{\omega}(u)
&:=
\mathcal{S}_{\omega}(u)-\frac{1}{p+1}\mathcal{N}_{\omega}(u)
\\[6pt]
&=
\frac{p-1}{2(p+1)}
\Big\{ \|\nabla u \|_{L^{2}}^{2} + \omega \|u \|_{L^{2}}^{2} \Big\}
+
\frac{4-(d-2)(p-1)}{2d(p+1)}
\|u \|_{L^{2^{*}}}^{2^{*}}.
\end{split}$$ Moreover, we easily verify the following structures of $\mathcal{S}_{\omega}$ and $\mathcal{N}_{\omega}$ (cf. [@Willem Chapter 4]):
For any $u \in H^{1}(\mathbb{R}^{d})\setminus \{0\}$, there exists a unique $\lambda(u)>0$ such that $$\label{eq:A.2}
\mathcal{N}_{\omega}(\lambda u)
\left\{
\begin{array}{rcl}
>0 & \mbox{if}& 0< \lambda <\lambda(u),
\\
=0 & \mbox{if}& \lambda=\lambda(u),
\\
<0 & \mbox{if}& \lambda >\lambda(u).
\end{array}
\right.$$
For any $u \in H^{1}(\mathbb{R}^{d})\setminus \{0\}$, $$\label{eq:A.3}
\text{the function $\lambda \mapsto \mathcal{I}_\omega (\lambda u)$
is non-decreasing in $[0,\infty)$}.$$
Next, we introduce several variational values: $$\begin{aligned}
\label{eq:A.4}
\sigma
&:=
\inf\{ \|\nabla u \|_{L^{2}}^{2} : \mbox{$u \in \dot{H}^{1}(\mathbb{R}^{d})$
with $\|u\|_{L^{2^{*}}}=1$}\},
\\[6pt]
\label{17/07/08/11:28}
m_{\omega}
&:=\inf\{ \mathcal{S}_{\omega}(u) : \mbox{$u\in H^{1}(\mathbb{R}^{d})\setminus \{0\}$ with $\mathcal{N}_{\omega}(u)=0$} \},
\\[6pt]
\nonumber
\widetilde{m}_{\omega}
&:=\inf \big\{\mathcal{I}_{\omega}(u) :
\mbox{$u\in H^{1}(\mathbb{R}^{d})\setminus \{0\}$ with $\mathcal{N}_{\omega}(u)\le 0$}
\big\}
.\end{aligned}$$
By a standard argument (cf. [@Willem Chapter 4]), it is known that a minimizer for $m_\omega$ becomes a ground state to . Hence, in order to prove Proposition \[proposition:1.1\], it suffices to show the existence of minimizer for $m_\omega$.
We first state the relationship between $m_\omega$ and $\wt{m}_\omega$ (cf. [@Akahori-Ibrahim-Kikuchi-Nawa1 Proposition 1.2]):
\[proposition:A.1\] Assume $d\ge 3$ and $1<p<\frac{d+2}{d-2}$. Then, for any $\omega>0$, we have the following:
1. $m_{\omega}=\widetilde{m}_{\omega}>0$
2. Any minimizer for $\widetilde{m}_{\omega}$ is also a minimizer for $m_{\omega}$, and vice versa.
We shall prove claim (i). Since $\mathcal{I}_\omega (u) = \mathcal{S}_\omega(u)$ for every $u \in H^1({\mathbb{R}^d})$ with $\mathcal{N}_\omega (u) = 0$, it is clear that $\wt{m}_\omega \leq m_\omega$. For the opposite inequality $m_\omega \leq \wt{m}_\omega$, fix any $u \in H^1({\mathbb{R}^d}) \setminus \{0\}$ with $\mathcal{N}_\omega (u) \leq 0$. By , there exists a $\lambda \in (0,1]$ such that $\mathcal{N}_\omega (\lambda u) = 0$. By , $$m_\omega \leq \mathcal{S}_\omega (\lambda u) = \mathcal{I}_\omega(\lambda u)
\leq \mathcal{I}_\omega (u),$$ which yields $m_\omega \leq \wt{m}_\omega$. Thus, $m_\omega = \wt{m}_\omega$. It remains to prove that $\wt{m}_\omega >0$. Let $u \in H^{1}(\mathbb{R}^{d})\setminus \{0\}$ with $\mathcal{N}_{\omega}(u)\le 0$. Then, it follows from $\mathcal{N}_{\omega}(u)\le 0$ and Sobolev’s inequality that $$\min\{1, \omega \}
\|u\|_{H^{1}}^{2}
\lesssim
\|u\|_{H^{1}}^{p+1}+\|u\|_{H^{1}}^{2^{*}}
.$$ This implies that there exists a constant $c(\omega)>0$ such that $\|u\|_{H^{1}}^{2}\ge c(\omega)$ and therefore $\mathcal{I}_\omega (u) \gtrsim \min\{1,\omega\}c(\omega)$. Since $u$ is arbitrary, we find that $\wt{m}_\omega >0$.
Next, we shall prove claim [(ii)]{}. Since $\mathcal{I}_{\omega}=\mathcal{S}_{\omega}-\frac{1}{p+1}\mathcal{N}_{\omega}$ and $m_{\omega}=\widetilde{m}_{\omega}$, it suffices to prove that $\mathcal{N}_{\omega}(\widetilde{Q}_{\omega})=0$ for all minimizer $\widetilde{Q}_{\omega}$ for $\widetilde{m}_{\omega}$. Suppose the contrary that there exists a minimizer $\widetilde{Q}_{\omega}$ for $\widetilde{m}_{\omega}$ such taht $\mathcal{N}_{\omega}(\widetilde{Q}_{\omega})<0$. Then, it follows from that there exists a unique $\lambda_{0}\in (0,1)$ such that $\mathcal{N}_{\omega}(\lambda_{0}\widetilde{Q}_{\omega})=0$. Furthermore, we have $$\widetilde{m}_{\omega}\le \mathcal{I}_{\omega}(\lambda_{0}\widetilde{Q}_{\omega})
< \mathcal{I}_{\omega}(\widetilde{Q}_{\omega})=\widetilde{m}_{\omega},$$ which is a contradiction. Thus, $\mathcal{N}_{\omega}(\widetilde{Q}_{\omega})=0$.
Next, we state a key inequality to show the existence of minimizer for $m_\omega$ (cf. [@Zhang-Zou Lemma 2.2]):
\[proposition:A.2\] Assume that $d\ge 3$ and $3< p<5$, or $d\ge 4$ and $1<p<\frac{d+2}{d-2}$. Then, the following estimate holds $$\label{eq:A.5}
m_{\omega}< \frac{1}{d}\sigma^{\frac{d}{2}}
= \frac{1}{2} \| \nabla W \|_{L^2}^2 - \frac{1}{2^\ast} \| W \|_{L^{2^\ast}}^{2^\ast}
= \frac{1}{d} \| \nabla W \|_{L^2}^2
.$$
Let $\chi$ be an even smooth function on $\mathbb{R}$ such that $\chi(r)=1$ for $0\le r\le 1$, $\chi(r)=0$ for $r\ge 2$, and $\chi$ is non-increasing on $[0,\infty)$. Then, we define $$W_{\varepsilon}(x) := \varepsilon^{-\frac{d-2}{4}}W \left( \frac{x}{\sqrt{\varepsilon}} \right)
=
\varepsilon^{\frac{d-2}{4}}\left(\varepsilon+\frac{|x|^{2}}{d(d-2)}\right)^{-\frac{d-2}{2}},
\quad V_{\varepsilon}(x)
:=
\chi(|x|)W_{\varepsilon}(x).$$ Then, we can verify that $\sigma^{\frac{d}{2}} = \| \nabla W_{\varepsilon}\|_{L^2}^2 = \| W_{\varepsilon}\|_{L^{2^\ast}}^{2^\ast} $ and $$\begin{aligned}
\label{17/07/08/13:27}
\|\nabla V_{\varepsilon}\|_{L^{2}}^{2}
&=
\sigma^{\frac{d}{2}}+O(\varepsilon^{\frac{d-2}{2}})
,
\\[6pt]
\label{17/07/08/14:29}
\| V_{\varepsilon} \|_{L^{2^{*}}}^{2^{*}}
&=
\sigma^{\frac{d}{2}}+O(\varepsilon^{\frac{d}{2}})
.\end{aligned}$$ Moreover, we find that $$\label{17/07/08/20:47}
\|V_{\varepsilon}\|_{L^{q+1}}^{q+1}
=
\left\{ \begin{array}{rcl}
O (\varepsilon^{\frac{2d-(d-2)(q+1)}{4}})
&\mbox{if}&
\frac{2}{d-2} < q,
\\[6pt]
O(\varepsilon^{\frac{d}{4}}|\log{\varepsilon}|)
&\mbox{if}&
q=\frac{2}{d-2},
\\[6pt]
O(\varepsilon^{\frac{(d-2)(q+1)}{4}})
&\mbox{if}&
0< q< \frac{2}{d-2}
.
\end{array} \right.$$
Next, for a given $\varepsilon\in (0,1)$, we introduce a function $y_{\varepsilon} \colon (0,\infty) \to \mathbb{R}$ as $$y_{\varepsilon}(t):=\frac{1}{2}t^{2} \big\{
\|\nabla V_{\varepsilon}\|_{L^{2}}^{2}
+
\omega \|V_{\varepsilon}\|_{L^{2}}^{2}
\big\}
-\frac{t^{2^{*}}}{2^{*}}
\|V_{\varepsilon}\|_{L^{2^{*}}}^{2^{*}}
.$$ It is easy to verify that the function $y_{\varepsilon}$ attains its maximum only at the point $$\tau_{\varepsilon, \max}:=\frac{\big\{
\|\nabla V_{\varepsilon}\|_{L^{2}}^{2}
+
\omega \|V_{\varepsilon}\|_{L^{2}}^{2}
\big\}^{\frac{d-2}{4}} }{ \|V_{\varepsilon}\|_{L^{2^{*}}}^{\frac{d}{2}}
}.$$ It follows from the definition of $\sigma$ (see ), and that $$\frac{\|\nabla V_{\varepsilon}\|_{L^{2}}^{2}}{\|V_{\varepsilon}\|_{L^{2^{*}}}^{2}}
=
\frac{\sigma^{\frac{d}{2}} + O( \varepsilon^{ \frac{d-2}{2}} )}{\sigma^{\frac{d-2}{2}}+O(\varepsilon^{\frac{d}{2}})}
=
\sigma + O(\varepsilon^{\frac{d-2}{2}}).$$ Moreover, we see from and that $$\frac{\|V_{\varepsilon}\|_{L^{2}}^{2}}{\|\nabla V_{\varepsilon}\|_{L^{2}}^{2}}
=
\left\{ \begin{array}{rcl}
O(\varepsilon^{\frac{1}{2}})
&\mbox{if}& d=3,
\\[6pt]
O( \varepsilon|\log{\varepsilon}| )
&\mbox{if}& d=4,
\\[6pt]
O( \varepsilon)
&\mbox{if}& d\ge 5.
\end{array}\right.$$ Hence, we find that $$\label{17/07/08/16:06}
\begin{split}
y(\tau_{\varepsilon,\max})
&=
\frac{1}{d}
\frac{\|\nabla V_{\varepsilon}\|_{L^{2}}^{d}
}{\|V_{\varepsilon}\|_{L^{2^{*}}}^{d}}
\bigg( 1+ \omega \frac{\|V_{\varepsilon}\|_{L^{2}}^{2}}{ \|\nabla V_{\varepsilon}\|_{L^{2}}^{2}}
\bigg)^{\frac{d}{2}}
\\[6pt]
&=
\left\{ \begin{array}{rcl}
\frac{1}{d}\sigma^{\frac{d}{2}} + O(\varepsilon^{\frac{1}{2}})
&\mbox{if}& d=3,
\\[6pt]
\frac{1}{d}\sigma^{\frac{d}{2}} + O(\varepsilon |\log{\varepsilon}|)
&\mbox{if}& d=4,
\\[6pt]
\frac{1}{d}\sigma^{\frac{d}{2}} + O(\varepsilon)
&\mbox{if}& d\ge 5,
\end{array} \right.
\end{split}$$
On the other hand, for each $\varepsilon\in (0,1)$, there exists $\tau_{\varepsilon,0}>0$ such that $\mathcal{N}_{\omega}(\tau_{\varepsilon,0}V_{\varepsilon})=0$.
Now, we assume that $d=3$ and $3< p< 5$. Then, it follows from , and that $$\begin{split}
0
&=\mathcal{N}_{\omega}(\tau_{\varepsilon,0}V_{\varepsilon})
\\[6pt]
&=
\omega
\tau_{\varepsilon,0}^{2}O(\varepsilon^{\frac{1}{2}})
+
\tau_{\varepsilon,0}^{2}
\{\sigma^{\frac{3}{2}}+O(\varepsilon^{\frac{1}{2}})\}
-
\tau_{\varepsilon,0}^{p+1}
O(\varepsilon^{\frac{5-p}{4}})
-
\tau_{\varepsilon,0}^{6}
\{\sigma^{\frac{3}{2}}+O(\varepsilon^{\frac{3}{2}})\}.
\end{split}$$ Divide both sides above by $\tau_{\varepsilon,0}^{2}\{\sigma^{\frac{3}{2}}+O(\varepsilon^{\frac{3}{2}})\}$. Then, we obtain $$\tau_{\varepsilon,0}^{4}
=
\omega O(\varepsilon^{\frac{1}{2}})
+
1+O(\varepsilon^{\frac{1}{2}})
-
\tau_{\varepsilon,0}^{p-1}O(\varepsilon^{\frac{5-p}{4}})
.$$ Since $p-1 < 4$, this implies that for any $\omega>0$, $$\label{17/07/08/21:59}
\lim_{\varepsilon \to 0}\tau_{\varepsilon,0} \ge \frac{1}{2}.$$ Furthermore, it follows from the definition of $m_{\omega}$ (see ), , , and $2< p <2^{*}-1$ that $$\begin{split}
m_{\omega}
&\le \mathcal{S}_{\omega}(\tau_{\varepsilon,0} V_{\varepsilon})
=
y_{\varepsilon}(\tau_{\varepsilon,0})
-
\frac{\tau_{\varepsilon,0}^{p+1}}{p+1}
\|V_{\varepsilon}\|_{L^{p+1}}^{p+1}
\\[6pt]
&\le
y_{\varepsilon}(\tau_{\varepsilon,\max})
-
\frac{\tau_{\varepsilon,0}^{p+1}}{p+1}
c_{1} \varepsilon^{\frac{5-p}{4}}
=
\frac{1}{3}\sigma^{\frac{3}{2}} + O(\varepsilon^{\frac{1}{2}})
-c_{2} \varepsilon^{\frac{5-p}{4}}
\end{split}$$ for some positive constants $c_{1}$ and $c_{2}$ depending only on $p$. Thus, we find that if $p>3$ and $\varepsilon$ is sufficiently small depending only on $p$ and $\omega$, then $$m_{\omega} < \frac{1}{3}\sigma^{\frac{3}{2}}.$$ Similarly, we can prove that if $d\ge 4$, then claim is true.
Now, we are ready to prove Proposition \[proposition:1.1\].
By Lemma \[proposition:A.1\], it suffices to prove the existence of minimizer for $\widetilde{m}_{\omega}$. To this end, we consider a minimizing sequence $\{u_{n}\}$ for $\widetilde{m}_{\omega}$. We denote the Schwarz symmetrization of $u_{n}$ by $u_{n}^{*}$. Note that $\| \nabla u_n^\ast \|_{L^2} \leq \| \nabla u_n \|_{L^2}$ and $\| u_n^\ast \|_{L^q} = \| u_n \|_{L^q}$ hold for each $q \in [2,2^\ast]$. For example, see [@Lieb-Loss]. From these properties, we have $$\begin{aligned}
\label{eq:A.6}
&\mathcal{N}_{\omega}(u_{n}^{*})\le 0
\quad
\mbox{for any $n \ge 1$},
\\[6pt]
\label{eq:A.7}
&\lim_{n\to \infty}\mathcal{I}_{\omega}(u_{n}^{*})=\widetilde{m}_{\omega},
\\[6pt]
\nonumber
&
\left\| u_{n}^{*} \right\|_{H^{1}} < \infty
\quad
\mbox{for any $n \ge 1$}.\end{aligned}$$ Since $\{u_{n}^{*}\}$ is radially symmetric and bounded in $H^{1}(\mathbb{R}^{d})$, there exists a radially symmetric function $Q \in H^{1}(\mathbb{R}^{d})$ such that, passing to some subsequence, $$\label{eq:A.8}
\lim_{n\to \infty}u_{n}^{*}=Q
\quad
\mbox{weakly in $H^{1}(\mathbb{R}^{d})$ and
strongly in $L^{p+1}({\mathbb{R}^d})$}.$$ We shall show that $Q$ becomes a minimizer for $m_{\omega}$.
We first show $Q \not\equiv 0$. Suppose the contrary that $Q \equiv 0$. Then, it follows from (\[eq:A.6\]) and (\[eq:A.8\]) that, passing to some subsequence, $$\label{eq:A.9}
0\ge \lim_{n\to \infty}\mathcal{N}_{\omega}(u_{n}^{*})
\ge \lim_{n\to \infty}\left\{
\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^{2}
-\left\| u_{n}^{*} \right\|_{L^{2^{*}}}^{2^{*}}
\right\}.$$ If $\| \nabla u_n^* \|_{L^2} \to 0$, then $\| u_n^* \|_{L^q} \to 0$ for all $2<q \leq 2^\ast$. By and , one has $\| u_n^* \|_{L^2} \to 0$ and $\widetilde{m}_{\omega} = 0$. However, this contradicts $\widetilde{m}_{\omega} > 0$ (see Lemma \[proposition:A.1\]). Therefore, we may assume $\lim_{n\to\infty} \| \nabla u_n^* \|_{L^2} > 0$.
Now, with the definition of $\sigma$ gives us $$\lim_{n\to \infty}\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^{2}
\ge
\sigma
\lim_{n\to \infty}\left\| u_{n}^{*} \right\|_{L^{2^{*}}}^{2}
\ge
\sigma \lim_{n\to \infty}
\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^{\frac{2(d-2)}{d}}.$$ From $\lim_{n\to\infty} \| \nabla u_n^* \|_{L^2} >0$, it follows that $$\label{eq:A.10}
\sigma^{\frac{d}{2}}
\le \lim_{n\to \infty}\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^2.$$ Hence, we see from , , and that $$\begin{split}
\widetilde{m}_{\omega}
&=
\lim_{n\to \infty}\mathcal{I}_{\omega}(u_{n}^{*})
\\[6pt]
&\ge
\lim_{n\to \infty}
\left\{
\frac{p-1}{2(p+1)}\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^{2}
+
\frac{4-(d-2)(p-1)}{2d(p+1)}\left\| u_{n}^{*} \right\|_{L^{2^{*}}}^{2^{*}}
\right\}
\\[6pt]
&\ge \frac{1}{d}\lim_{n\to \infty}\left\| \nabla u_{n}^{*} \right\|_{L^{2}}^{2}
\ge \frac{1}{d}\sigma^{\frac{d}{2}}.
\end{split}$$ However, this contradicts . Thus, $Q \not\equiv 0$.
Next, we shall show that $\mathcal{N}_{\omega}(Q)=0$. Using the Brezis-Lieb Lemma [@Brezis-Lieb], we have $$\begin{aligned}
\label{eq:A.11}
\mathcal{I}_{\omega}(u_{n}^{*})
-
\mathcal{I}_{\omega}(u_{n}^{*}-Q)
-
\mathcal{I}_{\omega}(Q)
&=o_{n}(1),
\\[6pt]
\label{eq:A.12}
\mathcal{N}_{\omega}(u_{n}^{*})-\mathcal{N}_{\omega}(u_{n}^{*}-Q)
-\mathcal{N}_{\omega}(Q)
&=o_{n}(1)
.\end{aligned}$$ Furthermore, together with and the positivity of $\mathcal{I}_{\omega}$ implies that $$\label{eq:A.13}
\mathcal{I}_{\omega}(Q)\le \widetilde{m}_{\omega}.$$
Let us suppose $\mathcal{N}_{\omega}(Q)<0$ and derive a contradiction. Note that implies that $\mathcal{I}_{\omega}(Q)=\widetilde{m}_{\omega}$. Moreover, it follows from that there exists a unique $\lambda_{0}\in (0,1)$ such that $\mathcal{N}_{\omega}(\lambda_{0}Q)=0$. Hence, we have $$\widetilde{m}_{\omega}\le \mathcal{I}_{\omega}(\lambda_{0}Q)<\mathcal{I}_{\omega}(Q)=\widetilde{m}_{\omega}.$$ This is a contradiction.
Next, suppose that $\mathcal{N}_{\omega}(Q)>0$. Then, it follows from and that $\mathcal{N}_{\omega}(u_{n}^{*}-Q)<0$ for any sufficiently large $n$. Hence, we can take $\lambda_{n}\in (0,1)$ such that $\mathcal{N}_{\omega}(\lambda_{n}(u_{n}^{*}-Q))=0$. Furthermore, we see from and that $$\begin{split}
\widetilde{m}_{\omega}
&\le
\mathcal{I}_{\omega}( \lambda_{n} (u_{n}^{*}-Q))
\le \mathcal{I}_{\omega}(u_{n}^{*}-Q)
=\mathcal{I}_{\omega}(u_{n}^{*})
-
\mathcal{I}_{\omega}(Q)+o_{n}(1)
\\[6pt]
&= \widetilde{m}_{\omega} -
\mathcal{I}_{\omega}(Q)+o_{n}(1).
\end{split}$$ Hence, we conclude that $\mathcal{I}_{\omega}(Q)=0$ and $Q \equiv 0$. However, this is a contradiction. Thus $\mathcal{N}_{\omega}(Q)=0$.
Since $Q \not\equiv 0$ and $\mathcal{N}_{\omega}(Q)=0$, we have $$\label{eq:A.15}
m_{\omega}
\le \mathcal{S}_{\omega}(Q)=\mathcal{I}_{\omega}(Q).$$ Moreover, it follows from and Proposition \[proposition:A.1\] that $$\label{eq:A.16}
\mathcal{I}_{\omega}(Q)
\le
\liminf_{n\to \infty}\mathcal{I}_{\omega}(u_{n}^{*})
\le \widetilde{m}_{\omega}=m_{\omega}.$$ Combining (\[eq:A.15\]) and (\[eq:A.16\]), we obtain $\mathcal{S}_{\omega}(Q)=\mathcal{I}_{\omega}(Q)=m_{\omega}$. Thus, we have proved that $Q$ is a minimizer for $m_{\omega}$.
The Moser iteration {#section:B}
===================
Here we state a result used in sections \[section:3\] and \[section:4\] to obtain the uniform decay estimates.
\[proposition:B.1\] Assume $d\ge 3$. Let $a(x)$ and $b(x)$ be functions on $B_{4}$, and let $ u \in H^1(B_4)$ be a weak solution to $$\label{eq:B.1}
-\Delta u + a(x) u = b (x) u \quad {\rm in} \ B_4.$$ Suppose that $a(x)$ and $u$ satisfy that $$\label{eq:B.2}
a(x) \geq 0 \quad {\rm for\ a.a.}\ x \in B_4, \quad
\int_{B_4} a(x) |u(x)v(x)| d x < \infty \quad
{\rm for\ each\ } v \in H^1_0(B_4).$$
1. Assume that for any ${\varepsilon}\in (0,1)$, there exists $t_{{\varepsilon}}>0$ such that $$\label{eq:B.3}
\left\| \chi_{[|b|>t_{{\varepsilon}}]}b \right\|_{L^{d/2}(B_4)} \leq {\varepsilon}$$ where $[|b|>t] := \left\{ x \in B_4 : |b(x)| > t \right\}$, and $\chi_A(x)$ denotes the characteristic function of $A \subset {\mathbb{R}^d}$. Then, for any $q \in (0,\infty)$, there exists a constant $C(d,q, t_{{\varepsilon}})$ such that $$\| |u|^{q+1} \|_{H^1(B_1)} \leq C(d,q,t_{{\varepsilon}}) \| u \|_{L^{2^\ast}(B_4)}.$$
2. Let $s>d/2$ and assume that $ b \in L^{s}(B_{4})$. Then, there exists a constant $C(d,s, \|b\|_{L^{s}(B_{4})})$ such that $$\| u \|_{L^\infty(B_1)} \le C \left(d,s, \|b\|_{L^{s}(B_{4})} \right)
\| u \|_{L^{2^\ast}(B_4)}.$$
Here, the constants $C(d,q,t_{{\varepsilon}})$ and $C(d,s, \|b\|_{L^{s}(B_{4})})$ in *(i)* and *(ii)* remain bounded as long as $q$, $t_{{\varepsilon}}$ and $\|b\|_{L^{s}(B_{4})}$ are bounded.
By the assumption of Proposition \[proposition:B.1\], notice that $$\int_{B_4} a (x) \varphi u^2 dx \geq 0 \quad \text{for all } u \in H^1(B_4)$$ where $\varphi \in C^\infty_0(B_4)$ with $\varphi \geq 0$. Using this fact and arguing as in [@Liu-Liu-Wang Proof of Proposition 2.2] and [@GT] (cf. [@BK-79]). we may prove Proposition \[proposition:B.1\]. Therefore, we omit the details of the proof.
The Pucci-Serrin condition {#section:C}
==========================
In this section, we give the range of space dimension $d$ and the subcritical power $p$ for which [@PS Theorem 1] is applicable to the case of equation .
\[proposition:C.1\] Let $3 \leq d \leq 6$ and assume $ \frac{4}{d-2} \leq p < \frac{d+2}{d-2}$ with $1<p$. Then, for any $\omega >0$, the equation admits at most one positive radial solution.
In order to apply [@PS Theorem 1], what we need to check is [@PS (2.5)]. In our case, this condition becomes $$\label{eq:C.1}
\frac{d}{du} \left[ \frac{F(u)}{f(u)} \right] \geq \frac{d-2}{2d} \quad
{\rm for\ } u > 0, \ u \neq a$$ where $$f(u) : = - \omega u + u^p + u^q, \quad
F(u) := - \frac{\omega}{2} u^2 + \frac{u^{p+1}}{p+1} + \frac{u^{q+1}}{q+1},
\quad q := \frac{d+2}{d-2}, \quad f(a) = 0.$$
We first rewrite . Since $$\frac{d}{du}\left[ \frac{F(u)}{f(u)} \right]
= 1 - \frac{F(u) f'(u) }{ (f(u))^2 },$$ is equivalent to $$\label{eq:C.2}
0 \leq q (f(u))^2 - (q+1) F(u) f'(u) =:g(u) \quad {\rm for\ all}\ u > 0.$$
Next, we expand $g(u)$ as follows: $$g(u)
=
A_{2} \omega^{2} u^{2}
+
A_{p+1} \omega u^{p+1}
+
A_{q+1} \omega u^{q+1}
+
A_{2p} u^{2p}
+
A_{p+q} u^{p+q}$$ where $$\label{eq:C.3}
\begin{aligned}
A_{2}&:=\frac{q-1}{2},
\\[6pt]
A_{p+1}&:=\frac{(p^{2}-3p-2)q+p^{2}+p+2}{2(p+1)},
\\[6pt]
A_{q+1}&:=\frac{(q-1)(q-2)}{2},
\\[6pt]
A_{2p}&:=\frac{q-p}{p+1},
\\[6pt]
A_{p+q}&:=\frac{(q-p)(p+1-q)}{p+1}.
\end{aligned}$$ We remark that our assumption yields $ q - 1 \leq p < q$ and $2 \leq q$. Hence, it is easily seen that $$\label{eq:C.4}
A_2 >0, \ A_{q+1} \geq 0, \ A_{2p} > 0, \ A_{p+q} \ge 0.$$
To show , we divide the arguments into two cases:
[**Case 1:**]{} $d=3$.
In this case, we have $q = 5$ and $ 4 \leq p < 5$, which implies $$2 (p+1) A_{p+1} = 6p^2 - 14 p - 8 > 0 \quad
\text{for all $ 4 \leq p < 5$.}$$ By , holds.
[**Case 2:**]{} $d=4, 5, 6$.
In this case, we rewrite $g(u)$ as follows: $$g(u) = \omega^2 u^2 \left\{
A_2 + A_{p+1} \frac{u^{p-1}}{\omega} + A_{2p} \left( \frac{u^{p-1}}{\omega} \right)^2 \right\}
+ A_{q+1} \omega u^{q+1} + A_{p+q} u^{p+q}.$$ By , it suffices to show $$Q(r) := A_2 + A_{p+1} r + A_{2p} r^2 \geq 0 \quad
\text{for each $r \geq 0$ and $q-1 \leq p < q$}.$$
When $d=4$, one has $q = 3$ and $$Q(r) = 1 + \frac{2(p^2-2p-1)}{p+1} r + \frac{3-p}{p+1} r^2.$$ If $p^{2}-2p-1 \ge 0$, then $Q(r)> 0$ for all $r\ge 0$. On the other hand, if $p^{2}-2p-1 < 0$, then we obtain $2 \leq p < 1 + \sqrt{2} = :p_0$ and simple computations give $$\label{eq:C.5}
\min_{r\ge 0}{Q(r)}=
1 - \frac{(p^2-2p-1)^2}{(p+1)(3-p)}.$$ Set $$h(p) := (p+1)(3-p) - (p^2 - 2p - 1)^2
= -p^4 + 4p^3 - 3p^2 -2 p +2.$$ Note that $$h'(p) = -4p^3 + 12p^2 - 6p -2, \quad
h''(p) = -12p^2 + 24 p - 6 < 0 \quad {\rm in} \ [2,p_0].$$ We also observe that $$h'(2) = 2 >0, \quad h'(p_0) = -2p_0 + 2 < 0, \quad
h(2) = 2 = h(p_0) > 0.$$ Hence, $h(p) > 0$ in $[2,p_0]$ and by , we have $\min_{r \geq 0} Q(r) \geq 0$ and .
When $d=5$, we see $$Q(r) = \frac{1}{3} \left\{
2 + \frac{5p^2 - 9p - 4}{p+1} r
+ \frac{7-3p}{p+1} r^2
\right\} .$$ Remark that $5p^2 -9p - 4 < 0$ is equivalent to $9 - \sqrt{161} < 10 p < 9 + \sqrt{161}$ and also that $ \frac{4}{3}< \frac{9+\sqrt{161}}{10} < \frac{7}{3}$. Hence, if $ \frac{9+\sqrt{161}}{10} \leq p < \frac{7}{3}$, then $Q(r) \geq 0$ for all $ r \geq 0$.
When $\frac{4}{3} \leq p < \frac{9+\sqrt{161}}{10}$, observe that $$\min_{r \geq 0} Q(r)
= \frac{1}{3} \left\{ 2 - \frac{(5p^2 - 9p - 4)^2}{4(7-3p)(p+1)} \right\}.$$ Since $$\begin{aligned}
8(7-3p)(p+1) - (5p^2-9p-4)^2
&= - 25p^4 + 90 p^3 - 65p^2 - 40p + 40
\\
&=(p-1)^2 ( -25p^2 +40p + 40 ) =: (p-1)^2 h(p),
\end{aligned}$$ by $\frac{9+\sqrt{161}}{10} < \frac{9+13}{10} = \frac{11}{5}$ and $h(\frac{11}{5}) = 7 > 0$, we obtain $h(p) \geq 0$ for every $p \in [ \frac{4}{3} , \frac{9+\sqrt{161}}{10} ]$. Thus, $\min_{r\geq 0} Q(r) \geq 0$ and holds.
When $d=6$, we observe $$Q(r) = \frac{1}{2} \left\{ 1 - \frac{(3p+1)(2-p)}{p+1} r + \frac{2(2-p)}{p+1} r^2 \right\},
\
\min_{r \geq 0} Q(r) =
\frac{1}{2} \left\{ 1 - \frac{(2-p)(3p+1)^2}{8(p+1)} \right\}
.$$ Setting $h(p) := 8(p+1) - (2-p) (3p+1)^2$ for $1 \leq p < 2$, we obtain $$h(1) = 0, \quad
h'(p) = 27p^2 - 24 p -3 \geq 0 \quad \text{for each $1 \leq p <2$}.$$ Hence, $h(p) \geq 0$ for all $1 \leq p < 2$ and holds.
\[remark:C.1\] When $d \geq 7$ and $ 1 < p < \frac{d+2}{d-2} = q$, condition is not satisfied for $\omega \gg 1$. In fact, we have $1 < q < 2$ and $A_{q+1} < 0$ in . Fix an $\alpha \in ( \frac{1}{q-1}, \frac{1}{p-1} )$ and observe that $$\alpha ( p + q) < 1 + \alpha ( q + 1), \quad
2 + 2\alpha < 1 + \alpha ( q + 1).$$ Noting also $1+(p+1) \alpha < 1+(q+1)\alpha$ and $2 p \alpha < (p+q) \alpha$, we see that $$\begin{aligned}
g( \omega^{\alpha} )
&= A_2 \omega^{2+2\alpha} + A_{p+1} \omega^{1+ (p+1)\alpha}
+ A_{q+1} \omega^{ 1 + (q+1) \alpha } + A_{2p} \omega^{2\alpha p}
+ A_{p+q} \omega^{(p+q)\alpha}
\\
&= \omega^{1+(q+1)\alpha} \left( A_{q+1} + o(1) \right).
\end{aligned}$$ Since $A_{q+1} < 0$, we obtain $g(\omega^\alpha) < 0$ for $\omega \gg 1$ and is not satisfied.
It is worth noting that for any $\Phi_\omega \in {\mathcal{G}_\omega}$ we have $\Phi_\omega (0) = \| \Phi_\omega \|_{L^\infty} \sim \omega^{\frac{1}{p-1}}$ by Lemma \[proposition:4.1\]. Hence, breaks down even in the interval $[0,\Phi_\omega(0)]$.
Acknowledgement {#acknowledgement .unnumbered}
---------------
S.I. is partially supported by NSERC Discovery grant \# 371637-2014, and also acknowledges the kind hospitality of Tsuda University, Japan. The work of N.I. was supported by JSPS KAKENHI Grant Number JP16K17623 and JP17H02851. The work of H.K. was supported by JSPS KAKENHI Grant Number JP17K14223. The work of H.N. was supported by JSPS KAKENHI Grant Number 17H02859 and 15K13450.
[99]{}
T. Akahori, S. Ibrahim, H. Kikuchi and H. Nawa, Existence of a ground state and blow-up problem for a nonlinear Schrödinger equation with critical growth. Differential Integral Equations [**25**]{} (2012), no. 3-4, 383–402.
T. Akahori, S. Ibrahim, H. Kikuchi and H. Nawa, Global dynamics above the ground state energy for the combined power-type nonlinear Schrödinger equation with energy-critical growth at low frequencies, preprint.
C.O. Alves, M.A.S. Souto and M. Montenegro, Existence of a ground state solution for a nonlinear scalar field equation with critical growth. Calc. Var. Partial Differential Equations [**43**]{} (2012), no. 3-4, 537–554.
A. Ambrosetti and A. Malchiodi, Perturbation methods and semilinear elliptic problems on $\mathbf{R}^n$. Progress in Mathematics, [**240**]{}. Birkhäuser Verlag, Basel, 2006.
H. Berestycki and P.-L. Lions, Nonlinear scalar field equations. I. Existence of a ground state. Arch. Rational Mech. Anal. [**82**]{} (1983), no. 4, 313–345.
H. Brézis and T. Kato, Remarks on the Schrödinger operator with singular complex potentials. J. Math. Pures Appl. (9) [**58**]{} (1979), no. 2, 137–151.
H. Brézis and E. Lieb, A relation between pointwise convergence of functions and convergence of functionals. Proc. Amer. Math. Soc. [**88**]{} (1983), no. 3, 486–490.
H. Brézis and L. Nirenberg, Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents. Comm. Pure Appl. Math. [**36**]{} (1983), no. 4, 437–477.
L.A. Caffarelli, B. Gidas and J. Spruck, Asymptotic symmetry and local behavior of semilinear elliptic equations with critical Sobolev growth. Comm. Pure Appl. Math. [**42**]{} (1989), no. 3, 271–297.
C.C. Chen and C.S. Lin, Uniqueness of the ground state solutions of $\Delta u+f(u)=0$ in ${\bf R}^n,\;n\geq 3$. Comm. Partial Differential Equations [**16**]{} (1991), no. 8-9, 1549–1572.
C.V. Coffman, Uniqueness of the ground state solution for $\Delta u -u + u^{3}=0$ and a variational characterization of other solutions. Arch. Rational Mech. Anal. [**46**]{} (1972), 81–95.
M. Coles and S. Gustafson, Solitary Waves and Dynamics for Subcritical Perturbations of Energy Critical NLS. https://arxiv.org/pdf/1707.07219.pdf
J. Dávila, M. del Pino and I. Guerra, Non-uniqueness of positive ground states of non-linear Schröödinger equations. Proc. Lond. Math. Soc. (3) [**106**]{} (2013), no. 2, 318–344.
A. Floer and A. Weinstein, Nonspreading wave packets for the cubic Schrödinger equation with a bounded potential. J. Funct. Anal. [**69**]{} (1986), no. 3, 397–408.
B. Gidas, W.M. Ni and L. Nirenberg, Symmetry of positive solutions of nonlinear elliptic equations in $\mathbf{R}^n$. Mathematical analysis and applications, Part A, pp. 369–402, Adv. in Math. Suppl. Stud., 7a, Academic Press, New York-London, 1981.
D. Gilbarg and N.S. Trudinger, Elliptic partial differential equations of second order. Second edition. Grundlehren der Mathematischen Wissenschaften, [**224**]{}. Springer-Verlag, Berlin, 1983.
M. Grossi, A uniqueness result for a semilinear elliptic equation in symmetric domains. Adv. Differential Equations [**5**]{} (2000), no. 1-3, 193–212.
M. Grossi, C.S. Lin and S. Prashanth, A uniqueness result for a Neumann problem involving the critical Sobolev exponent. Math. Ann. [**325**]{} (2003), no. 4, 643–664.
Z.C. Han, Asymptotic approach to singular solutions for nonlinear elliptic equations involving critical Sobolev exponent. Ann. Inst. H. Poincaré Anal. Non Linéaire [**8**]{} (1991), no. 2, 159–174.
C.E. Kenig and F. Merle, Global well-posedness, scattering and blow-up for the energy-critical, focusing, non-linear Schrödinger equation in the radial case. Invent. Math. [**166**]{} (2006), no. 3, 645–675.
R. Killip, T. Oh, O. Pocovnicu and M. Vişan, Solitons and scattering for the cubic-quintic nonlinear Schrödinger equation on ${\mathbb{R}}^3$. Arch. Ration. Mech. Anal. [**225**]{} (2017), no. 1, 469–548.
M.K. Kwong, Uniqueness of positive solutions of $\Delta u - u + u^p = 0$ in $\mathbf{R}^n$. Arch. Rational Mech. Anal. [**105**]{} (1989), no. 3, 243–266.
M.K. Kwong and L.Q. Zhang, Uniqueness of the positive solution of [$\Delta u+f(u)=0$]{} in an annulus. Differential Integral Equations [**4**]{} (1991), no. 3, 583–599.
E. Lieb and M. Loss, Analysis. Second edition. Graduate Studies in Mathematics, [**14**]{}. American Mathematical Society, Providence, RI, 2001.
X.Q. Liu, J.Q. Liu and Z.Q. Wang, Quasilinear elliptic equations with critical growth via perturbation method. J. Differential Equations [**254**]{} (2013), no. 1, 102–124.
K. McLeod and J. Serrin, Uniqueness of positive radial solutions of $\Delta u + f(u) = 0$ in $\mathbf{R}^n$. Arch. Rational Mech. Anal. [**99**]{} (1987), no. 2, 115–145.
V. Moroz and C.B. Muratov, Asymptotic properties of ground states of scalar field equations with a vanishing parameter. J. Eur. Math. Soc. (JEMS) [**16**]{} (2014), no. 5, 1081–1109.
C. Miao, G. Xu and L. Zhao, The dynamics of the 3D radial NLS with the combined terms. Comm. Math. Phys. [**318**]{} (2013), no. 3, 767–808.
K. Nakanishi and W. Schlag, Global dynamics above the ground state energy for the focusing nonlinear Klein-Gordon equation. J. Differential Equations [**250**]{} (2011), no. 5, 2299–2333.
K. Nakanishi and W. Schlag, Global dynamics above the ground state energy for the cubic NLS equation in 3D. Calc. Var. Partial Differential Equations [**44**]{} (2012), no. 1-2, 1–45.
W.M. Ni and I. Takagi, Locating the peaks of least-energy solutions to a semilinear Neumann problem. Duke Math. J. [**70**]{} (1993), no. 2, 247–281.
Y.-G. Oh, Existence of semiclassical bound states of nonlinear Schrödinger equations with potentials of the class $(V)_a$. Comm. Partial Differential Equations [**13**]{} (1988), no. 12, 1499–1519 and Comm. Partial Differential Equations [**14**]{} (1989), no. 6, 833–834.
X.B. Pan and X. Wang, Blow-up behavior of ground states of semilinear elliptic equations $\mathbf{R}^n$ involving critical Sobolev exponents. J. Differential Equations [**99**]{} (1992), no. 1, 78–107.
P. Pucci and J. Serrin, Uniqueness of ground states for quasilinear elliptic operators. Indiana Univ. Math. J. [**47**]{} (1998), no. 2, 501–528.
J. Serrin and M. Tang, Uniqueness of ground states for quasilinear elliptic equations. Indiana Univ. Math. J. [**49**]{} (2000), no. 3, 897–923.
T. Tao, M. Vişan and X. Zhang, The nonlinear Schrödinger equation with combined power-type nonlinearities. Comm. Partial Differential Equations [**32**]{} (2007), no. 7-9, 1281–1343.
M. Willem, Minmax Theorems, Progress in Nonlinear Differential Equations and their Applications, [**24**]{}. Birkhäuser Boston, Inc., Boston, MA, 1996.
J. Zhang and W. Zou, The critical case for a Berestycki-Lions theorem. Sci. China Math. [**57**]{} (2014), no. 3, 541–554.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Two simple (rotator and one-particle) mechanistic models are suggested to describe simultaneously at a minimal level of sophistication two basic functions of F$_1$-ATPase: a motor regime driven by ATP hydrolysis and its inverted function as ATP synthesis. This description is consistent with the so-called rotary binding-change mechanism, a milestone of functioning ATP synthase, and uses a stepping (driving) function associated with two sequences of time instants, at which hydrolysis and synthesis reactions occur. It is useful to analyse experimental data and numerical simulations indeed predict corresponding dynamic behavior.'
address: |
$^{1}$Bogolyubov Institute for Theoretical Physics, 03143 Kyiv, Ukraine\
$^{2}$Section of Mathematical Physics, IMM, Technical University of Denmark, DK-2800 Lyngby, Denmark\
$^{3}$Department of Physical Chemistry, Chalmers University of Technology, S-412 96 Gothenburg, Sweden
author:
- 'A.V. Zolotaryuk$^{1,2},$ V.N. Ermakov$^{1,2}$, P.L. Christiansen$^2$, B. Nordén$^{3}$, and Y. Zolotaryuk$^{1,2}$'
title: '**Simple Mechanical Equivalents of Stepping Rotary Dynamics in F$_1$-ATPase** '
---
PACS numbers: 05.60.-k, 05.40.-a, 87.10.+e
Recently, the modeling of molecular motors - enzymes which transduce chemical energy into directed mechanical motion - based on the idea of gaining useful work by rectification of zero-mean noise fluctuations has attracted considerable attention [@a-r] and several models exploiting the so-called [*ratchet*]{} mechanism have been elaborated (see, e.g., Refs. [@ab; @p-a; @dv; @ma; @cf; @puk; @lm; @bko], to mention a few). In this context, ATP (adenosinetriphosphate) synthase shown schematically in Fiq. 1, being a realistic molecular engine of great interest nowadays [@bba], should also be studied from the point of view of biological physics [@fwa]. This machinery is composed of two rotary motors: a membrane-embedded unit F$_0$ and water-surrounded F$_1$-ATPase (also called F$_1$) connected by a coiled-coil $\gamma$ subunit. ATP synthase works as a reversible motor-pump machine: the proton flow through F$_0$ is believed to create a clockwise (when F$_1$ is viewed from the F$_0$ side) torque that turns $\gamma$, whereupon ATP molecules are sequentially synthesized at three catalytic sites, one on each $\beta$ subunit. Vice versa, ATP hydrolysis, which also occurs sequentially on $\beta$’s, but in the opposite direction, has been demonstrated to make $\gamma$ rotate backwards, converting F$_0$ into a proton pump. In this case, $\gamma$ is driven by sequential hinge-bending motions of $\beta$’s, like a crankshaft by the pistons in a car engine (for more details see Ref. [@wo]). In this paper, we focus on F$_1$-ATPase and consider its operation in both hydrolysis (as a motor) and synthesis (as a synthesizer) directions, using in parallel two simple mechanistic analogs: a plane rotator and a particle subjected to a periodic potential. Our description is consistent with Boyer’s binding-change mechanism [@bo; @a-w], the recent findings of Cherepanov [*et al.*]{} and Junge [*et al.*]{} [@cmj] on elasticity properties of the $\gamma$ subunit, as well as with the experimental results of Yasuda [*et al.*]{} [@y-y].
The structure of F$_1$-ATPase and the mechanics of motions within it are too complex to allow a detailed description of all interactions and motions of its different parts, i.e., the three $\beta$ and one $\gamma$ subunits [@bba; @wo; @bo; @a-w; @sej; @n-k; @k-y]. Therefore, it would be useful to describe this very sophisticated three-dimensional system, using simple mechanical equivalents (springs, particles, etc.) and keeping the main features of rotary dynamics found in previous studies, such as the modeling of Oster and Wang [@wo] and others [@bba; @cmj; @y-y; @sej; @n-k; @k-y]. This approach, “from complexity to simplicity”, is often used in biological physics. The typical example of such a modeling is the propagation of a nerve impulse on the giant axon of the squid, where insight in the dynamics of the original Hodgkin-Huxley equations [@hh] has been obtained by reduction to the minimal FitzHugh-Nagumo system [@fn].
Here the three-dimensional and four-body interaction in the $\beta_3\gamma$ subcomplex is suggested to be effectively described as a coupling of a planar rotator of length $R_0$ centered in the middle of an equilateral triangle (see left panels of Figs. 2 and 3) with three equivalent catalytic sites (denote them by numbers 1, 2, and 3) at the vertices of the triangle. Obviously, this rotor-stator interaction, resulting in a crankshaft-like rotation of $\gamma$ (and cooperative catalysis on $\beta$’s), is a periodic function: $U(\theta +2\pi)=U(\theta)$, where $\theta$ denotes the angular position of the rotator (positive if counter-clockwise). The mechanical equivalent of a driving torque on $\gamma$ can be chosen as a stretched spring or an effective particle displaced from equilibrium in the periodic potential $U(\theta)$ as illustrated in the left and the right panels of Figs. 2 and 3, respectively. At each site $i=1,2,3$, a spring $K_i$ is attached connecting this site with the rotator. All the three springs are supposed to be identical, but only one spring is allowed to be switched on, while the other two are switched off, at a time. Then sequential switching the springs will result in a power stroke on $\gamma$. Without loss of generality, the rotor-stator potential (given in units of the hydrolysis energy $W \simeq 80$ pN$\cdot$nm) can be written as $$U(\theta)= [r(\theta) -a]^2/ (l-a)^2 ,
\label{1}$$ where $a=d/R_0$ is the length of each spring $K_i$ being undistorted, $r(\theta)= \sqrt{1+(1+a)^2-2(1+a)\cos
\theta}$ the instantaneous spring length, and $l=r(\pm 2\pi/3)$.
Let us now describe how our spring system operates in both the hydrolysis and synthesis directions and how it can be coordinated with ATP hydrolysis/synthesis reactions at the catalytic sites of the $\beta$ subunits. According to Boyer’s hypothesis [@bo] supported by the structural studies by Walker and coworkers [@a-w], and recently by direct observation of Noji [*et al*]{}. [@n-k], each site 1, 2, or 3 can be found at least in one of the three states: T (ATP binding), E (empty), and D (ADP binding), at a time. Structurally, they are arranged as T, E, and D counter-clockwise (see Figs. 1-3) and can be put on the $\theta$ axis at the lattice (“catalytic”) sites with spacing $2\pi/3$, as shown in the right panels of Figs. 2 and 3. The dimensionless periodic (with period $2\pi$) potential $U(\theta)$ \[not necessary of the form (\[1\])\] is supposed to be “rigidly tied” to this lattice, so that its minima are always found at sites with state T. We assume this potential to satisfy the normalization conditions: $U$(T)$=0$ and $U$(D)$=1$. The last constarint means that the potential energy of the effective particle in state D is equal to the free hydrolysis energy $W$.
Consider first the hydrolysis direction when F$_1$ operates as a motor (see Fig. 2). Let initially the $\gamma$ subunit (rotor) be found at equilibrium, performing there thermal fluctuations. This state is represented by spring $K_1$ being undistorted and switched on state T, while the springs at the other sites ($K_2$ and $K_3$) are switched off. In the particle equivalent, this situation is represented in the right panel of Fig. 2(a) by a particle fluctuating in the vicinity of one of the minima of the potential $U(\theta)$, i.e., in state T. When an ATP molecule settles into site 2 which is found at this time in state E, it appears bound there, resulting in ATP binding: E $\rightarrow$ T. According to the rotary binding-change mechanism [@bo; @a-w], this transition implies the two conformational changes at the next two sites: release of the inorganic phosphate (P) and the ADP (adenosinediphosphate) molecule from site 2 (D $\rightarrow$ E), and hydrolysis of the ATP molecule located at site 3 (T $\rightarrow$ D). After these state transitions have occured, spring $K_1$ is switched off, while spring $K_2$ is switched on, causing the power stroke on the rotor as demonstrated by Fig. 2(b). Therefore the rotor is driven forward before it reaches a new equilibrium state, stepping forward by $2\pi/3$. Correspondingly, as shown in the right panel, the potential $U(\theta)$ steps forward (to the right) by $2\pi/3$, so that the particle appears to be lifted uphill at the level $U=1$, thereafter sliding down and finally dwelling in the next potential minimum before the next sequence of conformational changes takes place. However, if occasionally two sequential conformational transitions occur very close in time resulting in two almost simultaneous potential steps forward, the particle appears on the positive slope of the potential $U(\theta)$, sliding thereafter down backwards. When the time between two sequential transitions is still short, but long enough for the particle to make a descent close to equilibrium, the double potential step forward will result in the double sliding down on the negative slope. Indeed, both such occasional steps of $\gamma$ loaded by an actin filament were observed in experiments [@y-y; @n-k; @k-y]. We denote the sequence of time instants when hydrolysis events occur by $\{ t_i^+\}_{i=1}^\infty~$ and call it a “hydrolysis sequence”.
Consider now the case, when an external torque (e.g., from the side of F$_0$) is applied to the $\gamma$ subunit clockwise, as shown in the left panel of Fig. 3(a). According to Cherepanov [*et al.*]{} and Junge [*et al.*]{} [@cmj], the external energy of torsion is stored as an elastic strain energy of $\gamma$. When the rotator turns clockwise totally by $2\pi/3$, the torsional energy is liberated for the synthesis of one ATP molecule at the site being in state D. Then spring $K_1$ is switched off, while spring $K_3$ is on, bringing the system to the zero-energy level ($U=0$). Equivalently, the particle “drops down” to equilibrium state T \[see Fig. 3(b)\]. These mechanical equivalents are again in accordance with Boyer [@bo] for ATP synthesis, when the conformational change D $\rightarrow$ T implies the transitions at the other two sites: T $\rightarrow $ E and E $\rightarrow $ D. We denote the sequence of time instants when synthesis events occur by $\{ t_j^-\}_{j=1}^\infty$ and call it a “synthesis sequence”.
Summarizing, one can present the diagram shown in Fig. 4, which describes the elastic strain energy of the torsional rotor-stator interaction as a function of $\theta$, combining both the hydrolysis and synthesis directions. Arrows show the rotational direction under hydrolysis ($\theta >0$, [*motor*]{} regime) and synthesis ($\theta <0$, [*synthesizer*]{} regime).
In experiments [@y-y; @n-k; @k-y], the $\gamma$ subunit was loaded by an actin filament rotating in a viscous solution. Therefore, we approach the overdamped limit and the equation of motion for the rotator driven by a periodic potential $U(\theta)$ fluctuating stepwise forward and backwards by $2\pi/3$ (due to hydrolysis power strokes and an external load torque ${\cal T}_l$) reads $$\Gamma \dot{\theta}=- W \partial_{\theta}
U[\theta - 2\pi S(t)/3]- {\cal T}_l(\theta,t) + \xi(t) ,
\label{2}$$ where $\Gamma$ is a viscous friction coefficient, $U(\theta)$ is normalized by $U(2\pi n)=0$ and $U[2\pi(n \pm 1/3)]=1$, $n = 0, \pm 1, \ldots$ , and $\xi(t)$ is the Brownian torque with the auto-correlation function $\langle \xi(t)\xi(t')\rangle =2\Gamma
k_B T \delta(t-t')$, where $k_B$ denotes Boltzmann’s constant and $T$ is the absolute temperature. The stepping (driving) function $$S(t)= \sum_{i=1}^\infty\Theta(t-t^+_i)
- \sum_{j=1}^\infty\Theta(t-t^-_j) ,
\label{3}$$ where $\Theta(t) = 0$ for $t<0$ and $\Theta(t) = 1$ for $t \ge 0$, is defined through the two time sequences $\{ t_i^{+}\}_{i=1}^\infty$ and $\{ t_j^{-}\}_{j=1}^\infty$ indicating the time instants when hydrolysis and synthesis reactions occur, respectively.
The hydrolysis sequence $\{ t_i^+\}$ is a random process determined by ATP concentration (\[ATP\]). We define it through two characteristic duration times $\Delta t_{\mbox{\tiny D}}$ and $\Delta t_{\mbox{\tiny T}}$ by the recurrence relation $$t^+_{i+1} =t^+_i +\Delta t_{\mbox{\tiny D}} +2\zeta_{i+1}
\Delta t_{\mbox{\tiny T}},~i=1, 2, \ldots ,
\label{4}$$ where each $\zeta_i \in [0,1]$ is a random value with uniform distribution (since ATP concentration is supposed to be constant in the solution). The interval $\Delta t_{\mbox{\tiny T}}$ is large for low \[ATP\], but tends to zero as \[ATP\] is sufficiently high. Therefore one can assume that $ \Delta t_{\mbox{\tiny T}} = A_{\mbox{\tiny T}}/[\mbox{ATP}]$ with some constant $A_{\mbox{\tiny T}} > 0$. The interval $\Delta t_{\mbox{\tiny D}}$ triggers the hydrolysis reactions by release of P and ADP on the next site of the stator. In the limit of low concentration of nucleotides (ATP and ADP), this interval is short, whereas for high concentration, the release of the hydrolysis products is impeded and therefore some saturation for $\Delta t_{\mbox{\tiny D}}$ takes place. As a result, one can assume that $ \Delta t_{\mbox{\tiny D}} = \Delta t_{\mbox{\footnotesize st}}
[\mbox{ATP}]/( C_{\mbox{\tiny D}} + [\mbox{ATP}])$ with some constant $C_{\mbox{\tiny D}} > 0$ and $\Delta t_{\mbox{\footnotesize st}}$ being the duration time of one step. At zero temperature, it follows from the overdamped dynamics governed by Eq. (\[2\]) that $\Delta t_{\mbox{\footnotesize st}} \rightarrow \infty$ (an overdamped particle approaches a potential minimum after an infinitely long time). However, at nonzero temperature, the time for one step becomes [*finite*]{}, because in the vicinity of equilibrium, the particle is “captured” by thermal fluctuations. Thus, in the limit $a \rightarrow 0$, the potential (\[1\]) is reduced to the simple form $U=2(1-\cos\theta)/3$, admitting an explicit solution of Eq. (\[2\]) for each step if $\xi(t) \equiv 0$. Since the average amplitude of thermal fluctuations is $\sqrt{3{\cal D}/2}$, one finds that $\Delta t_{\mbox{\footnotesize st}} \simeq
(3t_0/2)\ln\left(\sqrt{3}\mbox{cot}\sqrt{3{\cal D}/8}\right)$, where $t_0=\Gamma/W$ is the time unit and ${\cal D}=k_B T/W$ the dimensionless strength of white noise in Eq. (\[2\]). Since, at room temperature $ k_BT \simeq 4$ pN$\cdot$nm, we thus have ${\cal D} \simeq 0.05$, so that $\Delta t_{\mbox{\footnotesize st}}$ can be estimated for each viscous load $\Gamma$.
Averaging Eq. (\[4\]), one finds that $\langle t^+_{j+1}-t^+_j \rangle
=\Delta t_{\mbox{\tiny D}} +
\Delta t_{\mbox{\tiny T}} $ and therefore the rotational rate of $\gamma$ (number of revolutions per second) is $ V = \left[ 3\left(\Delta t_{\mbox{\tiny D}} +
\Delta t_{\mbox{\tiny T}}\right)
\right]^{-1}$. Inserting here the dependences of $\Delta t_{\mbox{\tiny D,T}}$ on \[ATP\] given above, one finds $$V = { V_{\mbox{\footnotesize max}} [\mbox{ATP}] \over
K_{\mbox{\tiny M}}
+[\mbox{ATP}]^2/(C_{\mbox{\tiny D}} +[\mbox{ATP}]) }~,
\label{5}$$ where $ V_{\mbox{\footnotesize max}} =
(3\Delta t_{\mbox{\footnotesize st}})^{-1}=
\left(\Delta t_{\mbox{\tiny D}}\right)^{-1}$ is a maximal average velocity (as $[\mbox{ATP}] \rightarrow \infty$) and the constant $K_{\mbox{\tiny M}} = A_{\mbox{\tiny T}}/
\Delta t_{\mbox{\footnotesize st}}$ can be identified as the Michaelis constant, because Eq. (\[5\]) is reduced to the Michaelis-Menten law in the limit $C_{\mbox{\tiny D}} \rightarrow 0$. According to Yasuda [*et al.*]{} [@y-y], $K_{\mbox{\tiny M}} =0.8~\mu$M and $V_{\mbox{\footnotesize max}}=
4$ s$^{-1}$, and for these values the dependence (\[5\]) is shown in Fig. 5 (including the experimental data), with monotonic behavior for $ C_{\mbox{\tiny D}}
\le K_{\mbox{\tiny M}}$. Similarly [@y-y], one can assume that the one-step duration depends on the length of the filament $L$ as $\Delta t_{\mbox{\footnotesize st}} \simeq g_0 + g_1 L^3$, and substituting this expression into Eq. (\[5\]), one finds for small $ C_{\mbox{\tiny D}}$ the dependence $ V(L) \simeq \frac{1}{3}\left(A_{\mbox{\tiny T}}/[\mbox{ATP}]
+g_0 + g_1 L^3 \right)^{-1}$ plotted in Fig. 6, where the constants $ A_{\mbox{\tiny T}}$ and $g_{0,1}$ are fitted to the experiments [@y-y].
As described above, for a given random hydrolysis sequence (\[4\]), each synthesis instant $t_j^-$ is defined as the time when the rotor, being at some time in state T, rotates backwards by $2\pi /3$. Direct simulations of the dimensionless ($\tau = t/t_0$) equation (\[2\]) with the potential (\[1\]) are presented in Fig. 7. Here the intervals $\Delta \tau_{\mbox{\tiny D,T}} =
\Delta t_{\mbox{\tiny D,T}}/t_0$ in the sequence (\[4\]) are given through the constants $ C_{\mbox{\tiny D}}$ and $ K_{\mbox{\tiny M}}$, as well as the parameter $\Delta \tau_{\mbox{\footnotesize st}} =
\Delta t_{\mbox{\footnotesize st}}/t_0
= W/3\Gamma V_{\mbox{\footnotesize max}}$ to be evaluated from experiments. Thus, using that $\Gamma(L=1~\mu\mbox{m})
\simeq 1$ pN$\cdot$nm$\cdot$s and $V_{\mbox{\footnotesize max}}=
4$ s$^{-1}$ [@y-y], one finds $\Delta \tau_{\mbox{\footnotesize st}} \simeq 6.75.$ In the case without load (curves 1 and 2), the average velocity $\langle \theta \rangle /2\pi \tau
= \Gamma V/W$ is in good agreement with the direct observations (see Fig. 3 of Ref. [@y-y]) and the law plotted in Fig. 5 at $ C_{\mbox{\tiny D}} = K_{\mbox{\tiny M}}$, where the two velocities shown with dotted lines 1 and 2 correspond to curves 1 and 2 in Fig. 7, respectively. Note that the constant $ C_{\mbox{\tiny D}}$ in Eq. (\[5\]) implicitly describes the main feature of trajectories at low \[ATP\] (see curve 2): on average the number of multi-steps exceeds that of steps backwards, as observed experimentally [@n-k; @y-y]. Next, when a load torque ${\cal T}_l > 0$ is applied, the rotational rate decreases with increase of this torque, as illustrated by curve 3 in Fig. 7. Moreover, when the load exceeds some threshold value, the motor operates in inverse, as shown by curves 4 and 5. Similarly to Lattanzi and Maritan [@lm], the law (\[5\]) can be modified by subtracting a positive constant that controls the direction of rotation if ${\cal T}_l >0$.
Thus, we have developed two (rotator and one-particle) physical models of archetypal simplicity, which are consistent with the rotary binding-change mechanism [@bo; @a-w] and the elasticity properties of the $\gamma$ subunit [@cmj]. The cooperative rotational catalysis at the three $\beta$ subunits is described through two time sequences, each for switching hydrolysis and synthesis reactions, by adjusting the statistics of switching to satisfy the recent experimental results [@n-k; @y-y]. The models described in this paper are generic and simple; they do not depend on details of the periodic potential $U(\theta)$. In the hydrolysis (motor) direction, the $\gamma$ subunit works in a “passive” regime; only all the $\beta$ subunits are coordinated in the cooperative rotational catalysis. In the reversible (synthesis) direction, $\gamma$ is “active”, causing the corresponding (again cooperative, but in the inverse sequence) conformational changes after its strain energy of torsion reaches the free ATP hydrolysis energy. These important features are consistent with both Boyer’s binding-change mechanism [@bo; @a-w] and the findings of Cherepanov [*et al.*]{} and Junge [*et al.*]{} [@cmj]. The dependences of the model parameters on ATP concentration are general and physically motivated. In the framework of our description, the load torque ${\cal T}_l(\theta, t)$ generated by the F$_0$ part can further be involved explicitly resulting in a general motor/pump model of ATP synthase.
We also conclude that the puzzle, how does the binding-change mechanism work, may be essential not only for understanding the chemistry (dissipative catalysis) of creation of ATP, one of the most important processes in life, but also constitute a key physical problem behind the function of molecular motors, such as design of man-made molecular devices.
We acknowledge partial financial support from the European Union under the INTAS Grant No. 97-0368 and the LOCNET Project No. HPRN-CT-1999-00163. We thank A.C. Scott for stimulating and helpful discussions.
[99]{}
R.D. Astumian, Science [**276**]{}, 917 (1997); F. Jülicher, A. Ajdari, and J. Prost, Rev. Mod. Phys. [**69**]{}, 1269 (1997); P. Reimann, Phys. Reports [**361**]{}, 57 (2002).
R.D. Astumian and M. Bier, Phys. Lett. Lett. [**72**]{}, 1766 (1994).
J. Prost [*et al.*]{}, Phys. Lett. Lett. [**72**]{}, 2652 (1994).
I. Derényi and T. Vicsek, Proc. Natl. Acad. Aci. USA [**93**]{}, 6775 (1996); G.N. Stratopoulos, T.E. Dialynas, and G.P. Tsironis, Phys. Lett. A [**252**]{}, 151 (1999).
F. Marchesoni, Phys. Lett. A [**237**]{}, 126 (1998); M. Borromeo and F. Marchesoni, Phys. Lett. A [**249**]{}, 199 (1998).
S. Cilla and L.M. Floría, Physica D [**113**]{}, 157 (1998); S. Cilla, F. Falo, and L.M. Floría, Phys. Rev. E [**63**]{}, 031110 (2001).
M. Porto [*et al.*]{}, Phys. Rev. Lett. [**85**]{}, 491 (2000).
G. Lattanzi and A. Maritan, Phys. Rev. Lett. [**86**]{}, 1134 (2001).
C. Bustamante, D. Keller, and G. Oster, Acc. Chem. Res. [**34**]{}, 412 (2001).
For recent reviews on ATP synthase see special issue 2-3, Biochim. Biophys. Acta [**1458**]{} (2000).
H. Frauenfelder [*et al.*]{}, Rev. Mod. Phys. [**71**]{}, S419 (1999).
G. Oster and H. Wang, Biochim. Biophys. Acta [**1458**]{}, 482 (2000); H. Wang and G. Oster, Nature (London) [**396**]{}, 279 (1998).
P.D. Boyer, Biochim. Biophys. Acta [**1140**]{}, 215 (1993).
J.P. Abrahams [*et al.*]{}, Nature (London) [**370**]{}, 621 (1994).
D.A. Cherepanov [*et al.*]{}, FEBS Lett. [**449**]{}, 1 (1999); W. Junge [*et al.*]{}, [*ibid.*]{} [**504**]{}, 152 (2001).
R. Yasuda [*et al.*]{} Cell [**93**]{}, 1117 (1998).
D.S. Sabbert, S. Engelbrecht, and W. Junge, Proc. Natl. Acad. Sci. USA [**94**]{}, 4401 (1997).
H. Noji [*et al.*]{}, Nature (London) [**386**]{}, 299 (1997).
K. Kinosita [*et al.*]{}, Cell [**93**]{}, 21 (1998).
A.L. Hodgkin and A.F. Huxley, J. Physiol. [**117**]{}, 500 (1952).
R. FitzHugh, Biophys. J. [**1**]{}, 445 (1961); J. Nagumo [*et al.*]{}, Proc. IRE [**50**]{}, 2061 (1962).
FIGURE CAPTIONS
FIG. 1. Schematics of ATP synthase adapted from Ref. [@k-y]. An asymmetric $\gamma$ shaft rotates relatively to the hexamer formed by $\alpha$ and three $\beta$ subunits arranged alternatively. The other subunits, which constitute F$_0$ including the “anchor” part are not shown. The positive direction of $\gamma$ rotation and the directions of proton flow and rotational catalysis (sequential synthesis/hydrolysis reactions in the hexamer) are shown by the arrows.
FIG. 2. Spring (left panels) and particle (right panels) equivalents of the rotor-stator system evolving in hydrolysis direction. (a) Rotor (left) and particle (right) are found in equilibrium (in state T). (b) Power stroke caused by stretched spring $K_2$ will rotate $\gamma$ counter-clockwise (left). Sliding the particle down on the negative slope of periodic potential $U(\theta)$, after it has moved forward by $2\pi/3$ (right).
FIG. 3. Spring (left panels) and particle (right panels) equivalents of the rotor-stator system evolving in synthesis direction. (a) External torque drives the rotator clockwise and elastic strain energy of the system is stored in spring $K_1$ (left). Lifting particle uphill in potential $U(\theta)$ (right). (b) Release of elastic strain energy after synthesis takes place (left). After the particle energy has reached the value $U=1$, potential $U(\theta)$ steps backwards by $2\pi/3$, allowing the particle to “drop downhill” to zero energy level (right).
FIG. 4. Strain energy of the rotor-stator system against angular position $\theta$ in hydrolysis ($\theta >0$) and synthesis ($\theta <0$).
FIG. 5. Rotational rate of $\gamma$ against ATP concentration calculated for three values of $C_{\mbox{\tiny D}}$, using Eq. (\[5\]). Experimental results [@y-y] shown by circles are given for comparison.
FIG. 6. Rotational rate of $\gamma$ against the length of actin filament: experimental results (circles, squares, and triangles) and dependences $V(L)$ calculated for three values of ATP concentration.
FIG. 7. Typical trajectories for different regimes of F$_1$. Curves 1 and 2 demonstrate pure motor regime (without load) for high (curve 1; $\Delta \tau_{\mbox{\tiny D}}
=6.75$ and $\Delta \tau_{\mbox{\tiny T}} =0$) and low (curve 2; $\Delta\tau_{\mbox{\tiny D}} = 2$, $\Delta\tau_{\mbox{\tiny T}} =10$, \[ATP\] = 0.4 $\mu$M, and $C_{\mbox{\tiny D}}= K_{\mbox{\tiny M}}=0.8$ $\mu$M) ATP concentration. The other trajectories illustrate mixed motor/synthesizer regimes for different constant loads: below threshold (curve 3; $\Delta \tau_{\mbox{\tiny D}} =1.7$, $\Delta \tau_{\mbox{\tiny T}} =20$, and ${\cal T}_l =20$ pN$\cdot$nm), nearby threshold (curve 4; $\Delta \tau_{\mbox{\tiny D}} =6.75$, $\Delta \tau_{\mbox{\tiny T}} =0$, and ${\cal T}_l = {\cal T}_{\mbox{\footnotesize th}}
\simeq 38.4$ pN$\cdot$nm), and above threshold (curve 5; $\Delta \tau_{\mbox{\tiny D}} =6.75$, $\Delta \tau_{\mbox{\tiny T}} =0$, and ${\cal T}_l =0.48$ pN$\cdot$nm).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have carried out a spatial-kinematic study of three proto-planetary nebulae, IRAS 16594$-$4656, Hen 3-401, and Rob 22. High-resolution [${\rm H_2~}$]{}images were obtained with NICMOS on the [*HST*]{} and high-resolution spectra were obtained with the Phoenix spectrograph on Gemini-South. IRAS 16594$-$4656 shows a “peanut-shaped” bipolar structure with [${\rm H_2~}$]{}emission from the walls and from two pairs of more distant, point-symmetric faint blobs. The velocity structure shows the polar axis to be in the plane of the sky, contrary to the impression given by the more complex visual image and the visibility of the central star, with an ellipsoidal velocity structure. Hen 3-401 shows the [${\rm H_2~}$]{}emission coming from the walls of the very elongated, open-ended lobes seen in visible light, along with a possible small disk around the star. The bipolar lobes appear to be tilted 10$-$15$\arcdeg$ with respect to the plane of the sky and their kinematics display a Hubble-like flow. In Rob 22, the [${\rm H_2~}$]{}appears in the form of an “S”-shape, approximately tracing out the similar pattern seen in the visible. [${\rm H_2~}$]{}is especially seen at the ends of the lobes and at two opposite regions close to the unseen central star. The axis of the lobes is nearly in the plane of the sky. Expansion ages of the lobes are calculated to be $\sim$1600 yr (IRAS 16594$-$4656), $\sim$1100 yr (Hen 3-401), and $\sim$640 yr (Rob 22), based upon approximate distances.'
author:
- 'Bruce J. Hrivnak, Nathan Smith, Kate Y. L. Su, and Raghvendra Sahai'
title: 'A STUDY OF H$_2$ EMISSION IN THREE BIPOLAR PROTO-PLANETARY NEBULAE: IRAS 16594$-$4656, HEN 3$-$401, AND ROB 22 '
---
INTRODUCTION {#intro}
============
Proto-planetary nebulae (PPNs) are objects in transition between the asymptotic giant branch (AGB) and planetary nebula (PN) stages of stellar evolution. Studies over the past decade, particularly imaging studies with the [*Hubble Space Telescope*]{} ([*HST*]{}), have shown that this transition stage is key to understanding the shaping of PNs [@balick02]. In the interacting stellar winds model [@kwok82], which was later generalized to include an equatorial density gradient [@balick87], it was assumed that a fast wind from the central star interacted with the detached, slowly-moving remnant AGB envelope to shape the nebula. [*HST*]{} images of PPNs show that many and perhaps most of them display a basic bipolar structure [@ueta00; @su01; @sahai07; @siod08]. These are shapes that appear to be further developed in the PN stage. In addition, a point symmetry is often seen; @sahai98a attribute this to collimated jets, which may be episodic and change their direction, and which they advocate as the main shaping agents of PNs.
Studying [${\rm H_2~}$]{}in PPNs is one of the best ways to probe the presence of a fast, perhaps collimated, post-AGB wind and its effect in shaping the nebula. [${\rm H_2~}$]{}is the main constituent of the detached circumstellar envelope around a PPN. While a fast wind (V$>$100 km s$^{-1}$) has the energy to dissociate the [${\rm H_2~}$]{}on direct impact, it can also produce shocks in the medium that move at a slower speed and collisionally excite the [${\rm H_2~}$]{}. Recent surveys have shown the presence of shock-excited [${\rm H_2~}$]{}in bipolar PPNs [@garher02; @kelly05]. Detailed high-resolution [${\rm H_2~}$]{}studies have been published of approximately a half dozen PPNs [@sahai98; @kastner01; @cox03; @davis05; @hri06; @vds08].
In this paper we present the results of a detailed [${\rm H_2~}$]{}emission study of three bipolar PPNs, IRAS 16594$-$4656 (“Water Lily Nebula”), Hen 3$-$401 (IRAS 10178$-$5958), and Rob 22 (IRAS 10197$-$5750). All three have relatively hot central stars for PPNs, B and A spectral types, and all possess large infrared excesses due to circumstellar dust [@partha89; @garlar99a]. All three are or would be classified as DUPLEX (“DUst Prominent, Longitudinally-EXtended”) nebulae in the classification system of @ueta00 [see also @siod08].
Rob 22 has a nebula that looks like a pair of butterfly wings. At high spatial resolution, the lobes show a filamentary structure in visible light. In addition to bipolar structure, some point symmetry is apparent in the lobes. A large halo extends out to 25$\arcsec$ [@sahai99a]. The nebula appears to be viewed essentially edge on, with a dark lane obscuring the star in visible light. In the more detailed morphological classification system of @sahai07, this object is classified as Bcw,ml,ps(s),h(e): bipolar nebula with closed lobes, central obscuring waist, minor lobes present, point-symmetric shape, and enlongated halo. The spectral type is A2 Ie, based on light reflected off the lobes of the nebula [@allen78]. The circumstellar envelope appears to have a dual chemistry, with an oxygen-rich region evidenced by OH maser emission [OH284.18-0.79; @allen80] and crystalline silicate emission [@mol02] and a carbon-rich region evidenced by PAH emission [@mol97].
The nebula of Hen 3-401 possesses long, narrow lobes. It seems to be viewed at a slightly larger angle with respect to the plane of the sky, with the central star appearing faintly between the lobes [@sahai99b]. The morphological classification of this object is Bow\*(0.6),sk: bipolar nebula with open lobes, central obscuring waist with star evident (at 0.6 $\mu$m), with a skirt-like structure [@sahai07]. The spectral type of the star is Be, and the spectrum shows strong Balmer emission lines and lots of permitted and forbidden low-excitation emission lines [@allen78; @garlar99b]. The circumstellar envelope appears to be carbon-rich, since it shows CO emission [@loup90; @buj91] and PAH emission [@partha01] but not OH [@silva93] emission or crystalline silicates [@partha01].
In IRAS 16594$-$4656, the star is clearly seen and relatively bright compared to the nebula, giving the impression that the nebula is at a larger orientation angle. The nebula has the appearance of a basic bipolar structure with a pronounced point symmetry, consisting of three pairs of oppositely-directed, slightly curved, thin lobes. Also seen is a smaller elliptical structure oriented from northwest to southeast [@hri99]. The morphological classification of this object is Mcw\*(0.6),an,ps(m,an),h(a): multipolar nebula with closed lobes, central obscuring waist with star evident (at 0.6 $\mu$m), ansae present, point symmetry with two or more pairs of diametrically opposed lobes and with ansae, and a halo with centrosymmetric arcs [@sahai07]. Its spectrum shows Balmer emission lines [@garlar99a] and has been classified as B7 based on the H$\alpha$ line [@vds00]. @rey02, on the basis of high-resolution spectra, fit the spectrum with T$_{\rm eff}$=14,000 K and log[*g*]{}=2.1, slightly hotter than $\beta$ Ori (B8 I); this indicates a spectral type of B5-7 Ie. He also observed many emission lines. The nebula appears to be carbon rich, showing PAH features and displaying the 21 $\mu$m and 30 $\mu$m features seen in carbon-rich evolved stars [@garlar99a; @volk02]. @vds08 recently published a detailed kinematic and morphologic study of this object using H$_2$ emission, and we will refer to their results in the present study.
We begin with an examination of the high-resolution [${\rm H_2~}$]{}images (see @sahai00 for preliminary images of Rob 22 and Hen 3$-$401 and @hri04 for preliminary images of IRAS 16594$-$4656), then present new high-resolution, long-slit [${\rm H_2~}$]{}spectra, and then follow with a discussion of the kinematics. From these, we are able to determine properties of the collimated winds in these three PPNs and how they have shaped the surrounding nebulae.
HIGH-RESOLUTION H$_2$ IMAGING OBSERVATIONS {#h2 obs}
==========================================
High-resolution near-infrared imaging observations of the three PPNs were obtained with the [*Hubble Space Telescope*]{} ([*HST*]{}) and the Near-Infrared Camera and Multiobject Spectrometer (NICMOS) camera 2 (NIC2). The camera has a resolution of $0\farcs076\times0\farcs075~{\rm pixel^{-1}}$ and a field of view of 195 $\times$ 193. Observations were made with a narrow-band ${\rm H_2}$ filter (F212N) and also with an adjacent continuum filter (F215N) for the purpose of distinguishing the [${\rm H_2~}$]{}1$-$0 S(1) emission from the scattered starlight.
Hen 3-401 and Rob 22 were observed on 1998 under General Observer program ID No. 7840 (PI: Kwok) using the original [*HST*]{} NICMOS camera. Observations were also made with the F160W (H) and F222M (K) filters for both of these objects. IRAS 16594$-$4656 was observed in 2002 under General Observer program ID No. 9366 (PI: Hrivnak) using the refurbished [*HST*]{} NICMOS camera. Observations were also made with the F110W (J) filter. The observing log is listed in Table \[obslog\].
The data were taken in the multiple non-destructive mode (MULTIACCUM) to obtain a high dynamic range without saturating the detector. Three sets of observations were made through each filter at each position on the array, and each target was imaged at three (Hen 3-401, Rob 22) or four (IRAS 16594-4656) different positions of the NIC2 array using a predefined dither pattern.
The data were processed using the standard NICMOS [*calnica*]{} pipeline, which applied bias subtraction, dark subtraction, flat-fielding, and cosmic ray removal. The more recent data for IRAS 16594$-$4656 were re-calibrated using the HST archival On-the-Fly Reprocessing system, which provides recalibration using the most up-to-date reference files. The individual files were then corrected for the “pedestal effect” using IRAF/STSDAS [*pedsub*]{} task to flattening the sky value in the four quadrants of the NIC2 array. We then combined the twelve images with each filter using the IRAF STSDAS/Dither package with the parameters suggested in the [*HST Dither Handbook*]{} [@koekemoer02]. The data for Hen3$-$401 and Rob 22, which were taken and reduced earlier, were reduced with the same philosophy but in a more manual fashion, as described by @su03.
The reduced F212N and F215N images were then flux calibrated using the appropriate calibration values for each filter before and after the NICMOS refurbishing. These calibration values are listed in Table \[obscal\]. The individual F212N and F215N images for the three PPNs, with the thermal backgrounds removed, are displayed in Figure \[h2\_raw\].
For [IRAS 16594$-$4656]{}, the relatively-bright central star is surrounded by a distinct peanut-shaped nebula in the F212N image, while only some smaller and fainter nebulosity surrounds the star in the F215N image. Thus one can immediately see that most of the nebular light in the F212N image is due to the [${\rm H_2~}$]{}emission rather than scattered starlight. One can also see faint emission from two of the pairs of longer, thin lobes, with brighter emission (“ansae”) at their ends. For [Rob 22]{}, the F212N and F215N images appear rather similar. They each show, rather than a central star, emission coming from two closely-spaced regions separated by a dark equatorial band. These bright regions may represent light from the central star scattered toward us from above and below an obscuring disk or torus. The two bright regions lie roughly along the line that connects the approximately point-symmetric outer ends of the two lobes. The dark space between the two bright regions might thus be due to the torus. The nebula appears to be a little brighter in the F212N image than in the F215N image, especially in the outer regions, indicating that in the F212N image the nebula is seen mostly in scattered light but with some contribution from [${\rm H_2~}$]{}emission. For [Hen 3$-$401]{}, the nebula is seen more clearly and is larger in the F212N image than in the F215 image, indicating that in the F212N image, it is seen in both scattered light and [${\rm H_2~}$]{}emission. The star is bright in each image.
H$_2$ IMAGES {#h2 images}
============
To produce images of these three PPNs showing only [${\rm H_2~}$]{}emission requires the removal of the scattered light from the F212N image, which can be accomplished through the subtraction of an appropriately scaled image of the adjacent continuum (F215N image). The scaling is necessary to correct for differences in the source continuum in the two bandpasses. Differences due to the filter profiles and detector response were corrected by the flux calibration. In a similar previous study [@hri06], we examined the question of how best to scale the F215N continuum image for subtraction from the F212N image to remove the scattered continuum light. The object of that study, IRAS 17150$-$3224, had no other emission lines in these two bandpasses, and thus a simple correction for the continuum levels was appropriate. In these cases, we also need to be concerned about other emission features which may be contributing to the flux in the two filters.
In the following sub-sections we discuss the details of the continuum removal and the resulting [${\rm H_2~}$]{}image for each object.
IRAS 16594$-$4656
-----------------
A spectrum for IRAS 16594$-$4656 in the 2 $\mu$m region (2.10$-$2.27 $\mu$m) has been published by @vds03. It was obtained with a long slit 1$\farcs$0 wide and oriented E-W. The spectrum shows a relatively strong [${\rm H_2~}$]{}feature at 2.12 $\mu$m with a much weaker Br$\gamma$ emission line. Since the Br$\gamma$ is at a wavelength that places it at the edge of a wing of the F215N filter profile, its contribution is insignificant in this image. This is illustrated in Figure \[filters\], which shows the spectrum of the object plotted together with the filter profiles. Thus, for the continuum removal, we convolved a linear fit to the observed spectral continuum (with the [${\rm H_2~}$]{}emission removed) with the filter profiles and the detector response using the STSDAS/SYNPHOT [*calcphot*]{} task. This resulted in a scale factor of 1.016 that we used to scale the F215N image prior to subtraction of the flux-calibrated images.
The resultant [${\rm H_2~}$]{}image of IRAS 16594$-$4656 is displayed in Figure \[h2\_images\]a. It shows a limb-brightened, double-lobed bipolar structure with the west lobe larger and brighter (ratio of 1.6) than the east lobe. Approximate sizes (10$\sigma$ level) are 2$\farcs$6 by 2$\farcs$8 for the west lobe and 2$\farcs$2 by 2$\farcs$5 for the east lobe. In addition, one can see two pairs of more distant [${\rm H_2}$]{}-emitting ansae, each pair collinear on a line passing through the star. The ansae to the SW are brighter and farther from the star than those to the NE, with distances of 6$\farcs$2 and 5$\farcs$4 for the ansae at PA=33$\arcdeg$ and of 5$\farcs$8 and 4$\farcs$7 for the those at PA=54$\arcdeg$. These ansae are found at the ends of the thin lobes seen in the visible light images [@hri99; @su01], and will be discussed together with the visible image in Section \[discussion\]. The emission at the position of the star appears to approximately cancel out, indicating that there is little, if any, [${\rm H_2~}$]{}emission from the star.
Hen 3-401
---------
Spectra of the 2.12 $\mu$m [${\rm H_2~}$]{}line and the Br$\gamma$ line of Hen 3-401 have been published by @garher02. They observed this object twice with a 4$\farcs$5-wide slit oriented at P.A. = 70$\arcdeg$ and 90$\arcdeg$, approximately along the bipolar axis, and they found the [${\rm H_2~}$]{}emission to be clearly extended. The [${\rm H_2~}$]{}line is stronger than the Br$\gamma$ line by about a factor of 1.5, and since the latter lies in the wing of the F215N filter, its contribution to the flux in the F215N image is very small. This is illustrated in Figure \[filters\]. Neglecting the contribution of the line results in a scale factor of 0.994 (the spectrum is essentially flat in this region).
The resulting [${\rm H_2~}$]{}image of Hen 3-401 is shown in Figure \[h2\_images\]b. [${\rm H_2~}$]{}emission extends along the two almost cylindrical, slightly barrel-shaped lobes, each with a size of 2$\farcs$5 $\times$ 10$\arcsec$. The east lobe is brighter, with a ratio of 1.2. The two lobes are limb-brightened and appear to be brightest 2$\farcs$5$-$5$\farcs$3 (E) and 1$\farcs$5$-$4$\farcs$3 (W) from the star. The ends of the lobes are open. Little, if any, of the [${\rm H_2~}$]{}emission appears to come from the central star.[^1]
Rob 22
------
Rob 22 was also observed spectroscopically in this region by @garher02. They observed with the slit oriented at P.A. = 90$\arcdeg$, in this case almost perpendicular to the bipolar lobes. With a slit width of 4$\farcs$5, they likely included most of the nebula but missed some of the [${\rm H_2~}$]{}emission from the very ends of the lobes. Their spectra show the Br$\gamma$ line to be stronger than the H$_2$ 2.12 $\mu$m [${\rm H_2~}$]{}line by about a factor of 2.5, but since it still falls on the extreme edge of the filter profile, its contribution will be small. Nevertheless, we examined the scaling of the F215N image with and without consideration of this emission line. Neglecting the contribution of the line results in a scale factor of 0.979, while including it results in a scale factor of 0.967. We will use the scale factor with the line included, but note that the difference is only at the 1% level. Note, however, that the difference can be larger in localized regions, since the Br$\gamma$ and H$_2$ emissions likely have different distributions.
The resulting [${\rm H_2~}$]{}image of Rob 22 is shown in Figure \[h2\_images\]c. The nebula appears to have somewhat of a point-symmetric “S” shape, with a bright clump in the inner part of each lobe near the position of the unseen central star (clump to the NE is brighter) and a ridge of [${\rm H_2~}$]{}along the the end of each lobe (especially the S lobe). The north lobe is brighter, with a ratio of 1.2. Much of the extended structure within a few arcsec of the star disappears after continuum subtraction, indicating that the features at the bases of the polar lobes nearest the star are due to scattered light from a dusty flared disk or torus, not H$_2$ emission.
HIGH-RESOLUTION [${\rm H_2~}$]{}SPECTRA {#h2 spectra}
=======================================
Observations and Reduction
--------------------------
High dispersion, long-slit spectral observations were made of the three sources in the [${\rm H_2~}$]{}$v$=1$-$0 S(1) line at 2.1218 $\mu$m using the near-infrared echelle spectrograph Phoenix [@hinkle98] on the 8-m Gemini-South telescope on Cerro Pachon, Chile. Phoenix has a 1024$\times$256 InSb detector with a pixel scale of 0$\farcs$085 $\times$ 1.26 km s$^{-1}$ at a wavelength of $\sim$2 $\micron$. The slit width was 0$\farcs$34, providing an effective velocity resolution of 6 km s$^{-1}$ (R$\sim$50,000). The spectrum covers a range of $\sim$1300 km s$^{-1}$ (0.0086 $\mu$m) centered on this [${\rm H_2~}$]{}line. The area of the sky projected by the long slit onto the detector is $\sim$13$\farcs$4 in length, so for the observations of IRAS 16594$-$4656 and Hen 3-401, more than one exposure was taken moving along the slit, yielding a longer effective slit to cover each target. The observations were made in 2003 under three queue observing programs (GS-2002B-Q-53, GS-2003A-Q-27, GS-2003B-Q-41). Image quality on the different nights varied from 0$\farcs$7 to 1$\farcs$0 as measured in the R band.
Several spectra were obtained for each of the PPNs at various positions. For each object, one spectrum was taken with the slit oriented through the star along the major axis of the lobes. Additional spectra were taken at various offsets and orientations. For IRAS 16594$-$4656, two of the slit positions were oriented to pass though the star and the pairs of [${\rm H_2~}$]{}clumps. An observing log listing the slit positions is given in Table \[ph\_obs\], and the orientations of the slit for these various observations are shown in Figure \[ph\_obs\].
The weather for these observations was mostly clear, although some patchy clouds were present during observations of Rob 22 and IRAS 16594$-$4656. Sky subtraction was accomplished by obtaining one exposure of blank sky near each source immediately before or after a pair of exposures of the target. Spectral flat fields were obtained using an internal quartz emission lamp. For each target, a bright standard star was observed on the same night with the same grating setting in order to correct for telluric absorption and for flux calibration. Numerous telluric lines were also used for wavelength calibration, adopting the telluric spectrum available from NOAO[^2]. We calculated the apparent velocities adopting a rest wavelength of 21218.356 Å for the H$_2$ $v$=1–0 S(1) line (Bragg, Brault, & Smith 1982), and these velocities were then corrected to a heliocentric reference frame. Uncertainty in the resulting velocities is roughly $\pm$1 km s$^{-1}$, dominated by scatter in the dispersion solution for the numerous telluric lines across the observed wavelength range.
Position-Velocity Diagrams
--------------------------
### IRAS 16594$-$4656
Spectra of IRAS 16594$-$4656 were obtained at five different positions on the nebula, as illustrated in Figure 3. The resulting position-velocity (PV) diagrams are shown in Figure \[spec\_16594\]. In all cases the brightest emission displays a closed ellipse, but with different sizes and strong variations in intensity around the ellipse. In slit position [*a*]{} (PA=72$\arcdeg$), which is along the major bipolar axis, and slit position [*b*]{} (PA=52$\arcdeg$), which is rotated 20$\arcdeg$ from the major axis, the inner ellipses seem to be tilted slightly, with the W-SW lobe slightly blue-shifted compared to the E-NE lobe. However, this is a small effect, and the polar axis appears to lie very close to the plane of the sky.
Slit positions [*c*]{} (PA=33$\arcdeg$) and [*d*]{} (PA=345$\arcdeg$) also pass through the star, but are oriented closer to the minor axis (slit position d is along the minor axis), and the size of the bright inner ellipse is correspondingly smaller. Slit position [*e*]{} (PA=345$\arcdeg$) is parallel to the minor axis but offset from the star through the larger W lobe. The PV diagram at this position is consistent with a slice through a shell expanding with a velocity of $\sim$8-9 km s$^{-1}$ and with the N side brighter.
Intensity contours are shown in Figure \[spec\_16594\] to display the faintest emission levels. Slit position [*a*]{} shows that the [${\rm H_2~}$]{}emission extends $\pm$5 along the major axis. Slit positions [*b*]{} and [*c*]{} pass through the clumps at the ends of extended arms in the nebula, and in both spectra these clumps are seen in [${\rm H_2~}$]{}emission. In slit position [*b*]{}, the centroid velocities of the two clumps differ by 10 km s$^{-1}$ ($-$36 to $-$26), with each about 5 km s$^{-1}$ from the mean velocity of $\sim$$-$30 km s$^{-1}$. In slit position [*c*]{}, the velocity of the two clumps is about the same. These velocities can be explained by a model in which the clumps are ejected by rotating jets. The clumps at slit position [*b*]{} appear closer to the star than those at slit position [*c*]{}; this can be a result of a projection effect and/or temporal effect, assuming that they are moving outward from the central star at the same speeds. If the clumps in each point-symmetric pair began moving from the location of the star at the same time, then they should be at the same distance from the star. However, in both cases the brighter SW clumps appear farther from the star than the NE clumps, so the assumption of constant velocity for the clumps may not be correct. Alternatively, the clumps may simply be non-uniform brightness enhancements at the ends of the filaments.
The bright ellipses are consistent with limb-brightened lobes. When allowance is made for the resolution of the spectra, it appears that the interiors of the lobes are devoid of [${\rm H_2~}$]{}emission. In the four slit positions that include the star (Fig. \[spec\_16594\], panels $a-d$), the blue-shifted side of the inner ellipse is brighter. This may be due to extinction within the lobes or intrinsically brighter [${\rm H_2~}$]{}emission from the side facing us. The elliptical morphologies of the PV diagrams for the four slit positions through the star, especially the one along the major axis, are consistent with an expanding ellipsoid. Thus they differ from the “peanut-shape” PV diagram that one might expect from two bipolar expanding bubbles. Thus the expansion along the minor axis does not appear to be significantly retarded by a constraining higher-density torus.
We investigated the velocity field in more detail using the morpho-kinematic modeling program [*SHAPE*]{} [@stef06]. In this program, the morphology of the nebula and the velocity field can be specified and a resulting PV diagram produced. The shape of the nebula was assumed to be a three-dimensional symmetrical dumbbell based on the H$_2$ (cross-sectional) image. It was then assumed that the H$_2$ emissivity uniformly followed the contours of the nebula. We began by investigating a velocity field in which the velocity increased linearly with distance (so-called “Hubble flow”), which is seen in some PNs and is seen in Hen 3$-$401 (Section 4.2). This produced a PV diagram that had a peanut shape, contrary to the elliptical shape seen (see Figure \[SHAPE\_16594\]. left and right panels). When we assumed on the other hand a constant velocity (9 km s$^{-1}$), the PV diagram agreed more closely with the observed ellipse but was slightly diamond-shaped. A very good fit was obtained by the addition of a slight Hubble flow, with [*V*]{} = (8 + 2.4[*r*]{}/1$\arcsec$) km s$^{-1}$, as shown in the middle panels of Figure \[SHAPE\_16594\]. We determine, based on adding an inclination to the model, that the tilt of the polar axis is no more than $\sim$10$\arcdeg$ from the plane of the sky.
Slices in velocity across the P-V diagrams at position along the slit = 0 (along the line of sight through the star) with width 0$\farcs$77 are displayed in Figure \[slice\_16594\]. These show a uniform expansion of the shell with an expansion velocity of V$_{\rm exp}$ = 9 km s$^{-1}$. The velocity center is at $-$30 km s$^{-1}$. The expansion velocity of the lobe at 0$\farcs$9 west of the star is similar; this is consistent with uniform expansion of an elliptical bubble.
These observations reveal a systemic heliocentric radial velocity of V$_{\rm hel}$([${\rm H_2}$]{}) = $-$30 km s$^{-1}$, which translates to V$_{\rm LSR}$ = $-$26 km s$^{-1}$, with V$_{\rm exp}$ = 9 km s$^{-1}$. These velocities can be compared with the millimeter-wave CO $\it J$=1$-$0 emission line of V$_{\rm
LSR}$(CO) = $-$26 km s$^{-1}$ with V$_{\rm exp}$ = 16 km s$^{-1}$ [@loup90]. The CO expansion represents the velocity of the outer remnant of the AGB mass loss expanding into the ambient ISM, while the [${\rm H_2~}$]{}expansion represents the velocity of the shock front at the edge of the inner bipolar cavity along the line-of-site direction, which is essentially perpendicular to the polar axis. The difference between these two expansion velocities appears to reflect the two different kinematic regions. Heliocentric radial velocity measurements of individual lines in the visible spectrum showed a range of $-$10 to $-$40 km s$^{-1}$, with average values of $\sim$$-$20 km s$^{-1}$ for the absorption lines and $\sim$$-$32 km s$^{-1}$ for the emission lines [@rey02].
### Hen 3-401 {#hen3 PV}
Spectra of Hen 3-401 were obtained at four different positions on the nebula. The resulting position-velocity diagrams for the three oriented along the bipolar axis are shown in Figure \[spec\_hen3\]. (The fourth one, passing through the star but oriented perpendicular to the axis, showed no extended structure, only [${\rm H_2~}$]{}emission at the position of the star with the systemic velocity). In each of these three slit positions, we see a pair of convex arcs from each lobe. They are not closed at the ends, indicating open lobes out to at least $\pm$7 from the star. At the two offset positions, some emission appears to partially fill the interior of the lobes; this may simply be contamination from the very edges of the lobes due to seeing.
The bipolar axis is obviously tilted from the plane of the sky, with the W lobe moving toward us, and with velocities along the lobes that are proportional to distance from the star (Hubble flow). The pairs of convex arcs represent emission from the thin front and back walls of slightly barrel-shaped lobes. The difference between the pair in each lobe is about 30 km s$^{-1}$ for the slit passing through the star (Fig. \[spec\_hen3\]b) but about 23 km s$^{-1}$ in the two offset positions; this is consistent with seeing the radial expansion along the bipolar axis but only a projected component of it in the offset positions. If the expansion really is a Hubble flow, then the relationship between observed expansion velocities of the front and back sides of the lobes in Figure \[spec\_hen3\]b compared to their apparent spatial size can constrain the dynamical age of the nebula for a given distance. If we assume axial symmetry — so that the lobes have roughly the same length-to-width ratio as in H$_2$ images (this seems justifiable based in the similar morphology in Figures \[spec\_hen3\]b and \[hen3\_multi\]) — then one can stretch the PV plot in Figure \[spec\_hen3\]b to have the correct aspect ratio. Figure \[spec\_hen3\]b thus represents a slice through the structure of the nebula in the plane along our line-of-sight [see @smith06 for a more detailed explanation in the case of the bipolar lobes of Eta Carinae]. This yields a tilt of the polar axis from the plane of the sky for Hen 3-401 of roughly 15$\arcdeg$($\pm$5$\arcdeg$). In that case, the lateral expansion speed of the lobes (the radial expansion speed perpendicular to the polar axis) is V$_{\exp}\approx$15.5($\pm$1) km s$^{-1}$. In the images, the polar lobes have a width of roughly 2$\farcs$5, which corresponds to 3.7$\times$10$^{11}$($D_{kpc}$) km, where $D_{kpc}$ is the heliocentric distance in kpc to Hen 3-401. Thus, the age of the nebula would be only about 380 $\times D_{kpc}$ yr if the expansion has been ballistic. Assuming a distance of 3 kpc [@buj91] thus yields a nebular age of 1100 yr.
There is some weak [${\rm H_2~}$]{}emission coming from the star as well. In fact, the very weak central peak appears to show some extended structure, with a possible tilt $-$ i.e. it is slightly blueshifted toward the east and redshifted toward the west, as shown in Figure \[spec\_hen3\_zoom\] in an expanded view of the region around the central star. This may be emission from an inner ring or disk surrounding the central star. This same type of tilted structure from the central [${\rm H_2~}$]{}peak is also seen in similar spectra of the young PN M 2-9. In fact, it is remarkable that all aspects of the overall kinematic structure of [${\rm H_2~}$]{}emission from Hen 3-401 in panel [*b*]{} of Figure \[spec\_hen3\] are almost identical to those of M 2-9 [@smith05b]. Even the lateral expansion velocity of the lobes is roughly the same as for M 2-9. Although the apparent size of Hen 3-401 is smaller by a factor of $\sim$4.5, it distance is larger by about the same amount, so the two objects have the same size, nebular age, and overall appearance. A detailed comparison of their central stars might therefore be quite illuminating.
At most positions, [${\rm H_2~}$]{}emission from the far sides of the lobes is fainter than on the near side, especially where the far side is behind an area of bright emission on the near side. A similar effect is seen in M2-9, where it is probably due in part to extinction by dust in the lobes [@smith05a]. However, an asymmetry does appear in the brightness of the east-west lobes of Hen 3-401, with the brighter peak from the near side closer to the star in the west lobe, position $\approx$ 2 while in the east lobe it is at position $\approx$ 3$\arcsec$; this is in accord with what is seen in the [${\rm H_2~}$]{}image (Section 3.2).
Slices in velocity across the positions $+3\arcsec$ and $-3\arcsec$ with width 0$\farcs$26 are displayed in Figure \[slice\_hen3\]. Velocity widths of the arcs are narrower along the axis (Figure \[spec\_hen3\]b) than in the offset positions. This is due to projection effects in a Hubble flow. For a hollow cylinder with walls of constant thickness, the path length through the walls reaches a minimum for a sight line passing through the polar axis, translating to the smallest spread in velocity.
These observations reveal a radial velocity for the system of V$_{\rm hel}$([${\rm H_2}$]{}) = $-$24 km s$^{-1}$ (V$_{\rm LSR}$=$-$36 km s$^{-1}$), with V$_{\rm exp}$ = 15 km s$^{-1}$. These are in good agreement with the millimeter-wave CO $\it J$=1$-$0 emission-line values of V$_{\rm LSR}$(CO) = $-$36 km s$^{-1}$ with V$_{\rm exp}$ = 16 km s$^{-1}$ by @loup90 (although @buj91 measured the somewhat different value of V$_{\rm LSR}$(CO) = $-$28 km s$^{-1}$ with V$_{\rm exp}$ = 15 km s$^{-1}$ using the CO $\it
J$=1$-$0 and CO $\it J$=2$-$1 lines).
### Rob 22
Figure \[spec\_rob22\] shows the long-slit Phoenix spectra of Rob 22, with one slit positioned along the middle of the lobes through the center of the nebula (panel [*b*]{}) and offset slit positions on either side. The very weak [${\rm H_2~}$]{}emission and the complex continuum structure in the long-slit spectra indicate that even a narrowband F212N image is actually dominated by dust structures seen in reflected continuum light. This is true over most of the nebula, with the exceptions being the N and S ends of the polar lobes. In the center slit position ([*b*]{}) there appear to be two continuum sources separated by $\sim$0$\farcs$6, with the northern one the brighter, while in the slit position to the west there is one continuum source. These are consistent with the [${\rm H_2~}$]{}image. These continuum sources likely represent light from the obscured central star scattered toward us by regions of enhanced density. The three panels show a slight positional offset in the emission pattern, in agreement with the “S” shape of the nebula.
The spectra of the nebula show very faint [${\rm H_2~}$]{}emission from two hollow expanding lobes with an ellipsoidal velocity structure. The southern lobe is brighter and larger in [${\rm H_2}$]{}, in agreement with the [${\rm H_2~}$]{}image. The nebula appears to have a polar axis very close to the plane of the sky, although the southern lobe is slightly blue-shifted and thus tilted slightly toward us. The lateral expansion speed of the lobes (difference in velocity between the front and back of the S lobe) is approximately 30 km s$^{-1}$. Thus V$_{\rm exp}$ $\approx$ 15 km s$^{-1}$ and V$_{\rm hel}$([${\rm H_2}$]{}) $\approx$ +5 km s$^{-1}$, which transforms to V$_{\rm LSR}$ $\approx$ $-$7 km s$^{-1}$. These are in general agreement with the OH maser measurement of V$_{\rm
LSR}$(OH) = $-$6$\pm$3 km s$^{-1}$, V$_{\rm exp}$=20 km s$^{-1}$ [@allen80; @silva93]. However, they differ from the molecular-line CO $\it J$=1$-$0 measurement of V$_{\rm
LSR}$(CO)=$-$0.1 km s$^{-1}$, V$_{\rm exp}$ $\sim$ 35 km s$^{-1}$ [@buj91], although the investigators commented on the severe galactic CO contamination in this direction.
DISCUSSION
==========
IRAS 16594$-$4656
-----------------
The PV diagrams for [IRAS 16594$-$4656 ]{}indicate an expanding ellipsoidal nebula with its major axis nearly in the plane of the sky. This is a different orientation and morphology than had been concluded based upon the visual images of the nebula. In the visual images (see Fig. \[16594\_multi\]a), the star appears bright and the SW lobe(s) appears brighter and perhaps larger than the NE lobe(s). This is in contrast to nebulae like AFGL 2688 (Egg nebula), IRAS 17150$-$3224, IRAS 17441$-$2411, and Rob 22, in which the star is obscured and the two lobes look more symmetrical in size, suggesting that the polar axis is very nearly in the plane of the sky. Inspection of the visual images had previously led us to conclude that [IRAS 16594$-$4656 ]{}was inclined at an intermediate orientation ($\sim$30$\arcdeg$) with respect to the plane of the sky with the SW lobe oriented in the direction of the observer [@su01]. Even the new [${\rm H_2~}$]{}image, with the SW lobe appearing larger than the NE lobe, suggests this, assuming symmetry in the lobes.
However, other more recent observations at other wavelengths support our conclusion based upon kinematics that the axis of the main bipolar lobes, as delineated in the H$_2$ image, is nearly in the plane of the sky. A polarization study of this source [@ueta05] indicates a hollow shell with reflection from the walls. The polarization structure is aligned with the major axis of the nebula, with low polarization perpendicular to this, suggesting the presence of a dusty torus. A high-resolution mid-infrared study clearly shows the limb-brightened ends of such a torus [@volk06]. The mid-infrared image does not show the elliptical structure expected of a torus seen at an intermediate angle, but rather shows a bar with bright ends, consistent with a torus seen nearly edge-on. The bright mid-IR emission is located at the two ends of the pinched waist seen in the H$_2$ image, showing the presence of enhanced dust density which shapes the outflow. The approximately similar amounts of polarizations seen at opposite sides of the nebula are consistent with a nebula in which the bipolar axis is close to the plane of the sky rather than at an intermediate angle, and the authors suggest an angle of 15$\arcdeg$ with respect to the plane of the sky [@ueta05]. Thus both the polarization and mid-infrared images are consistent with the result of this kinematic study, and the corresponding one of @vds08, that the bipolar lobes are aligned nearly in the plane of the sky.
The [${\rm H_2~}$]{}image shows limb-brightened lobes, consistent with the idea that they were carved out by collimated outflows. Based on the angular diameter across the lobes of 2$\farcs$0 and the measured V$_{exp}$= 9 km s$^{-1}$, one can calculate an approximate expansion age for the lobes of 530 $\times D_{kpc}$ yr. Using the polar velocity determined using [*SHAPE*]{}, along with the angular diameter between the two ends of the bipolar lobes of 4$\farcs$0, leads to a expansion age of 740 $\times D_{kpc}$ yr This leads to an estimated age for the lobes of 1400-1900 yr, using a distance of 2.6 kpc based on an assumed total luminosity of 10$^4$ L$_{\sun}$ [@hri00].
Comparing the morphology of the nebula in different wavelength regions shows several striking features. In visible light (Fig. \[16594\_multi\]a), the nebula shows a basic bipolar structure, but with a series of three pairs of point-symmetric thin lobes. Thus at this wavelength it has been classified as point-symmetric and multipolar. Portions of several circumstellar rings are also seen [@hri01]. In the J band (Fig. \[16594\_multi\]b) the thin lobes are present and a portion of at least one ring is seen faintly. Seen more prominently, however, is a somewhat elliptical morphology extending from the northwest to the southeast that is also present in the visible image. However, it is not actually an ellipse, since the axes are not separated by 180$\arcdeg$; the northwest lobe is at a position angle of 305$\arcdeg$ and the southeast one at 110$\arcdeg$. The published H-band image is similar [@su03], with the star relatively brighter compared to the fainter thin lobes. The H$_2$ image (Fig. \[16594\_multi\]c), in contrast, shows a clear bipolar, “peanut” morphology with shorter, more spherical lobes. The major axis is at a position angle of 75$\arcdeg$ with the west lobe larger. Some of the structures that define these peanut-shaped lobes can be seen upon close examination in the V and H images. Among these is the bright “cap” at the end of the eastern lobe. As mentioned earlier, the mid-IR images at 10 and 20 $\mu$m reveal two bright regions just outside the pinched waist of the “peanut.” The mid-IR emission also faintly delineates the edges of the two peanut-shaped lobes seen in H$_2$. @volk06, in comparing the mid-IR images to this H$_2$ image, concluded that the coincidence of the faint mid-IR emission with the H$_2$ emission from the walls of these lobes suggested that the mid-IR emission was likely partially shock heated. The walls of these peanut-shaped lobes are clearly outlined in polarized light [@ueta07], agreeing with the position of the dust seen in the mid-IR.
In a composite image (Fig. \[16594\_multi\]d), we see the enhanced H$_2$ emission from the walls of the peanut-shaped bipolar lobes and the H$_2$ at the ends of the long, thin lobes. The non-uniformity in the emission from the walls may indicate lower density regions, through which jets may have carved holes from which the long, thin lobes may have emerged. This composite image also shows clearly several different structures seen in this object at different wavelengths, reinforcing the idea that the morphology of this nebula is more complex than simply symmetric bipolar lobes. It is more complex than either of the other PPNs of this study.
In their excellent study of this object, @vds08 used their long-slit H$_2$ spectra together with our spectra (from the archive) to produce channel maps and from these to determine a more detailed three-dimensional image of the shell structure of IRAS 16594$-$4656. While it basically consists of two hollow bipolar lobes joined at the waist, it is more complex, with some holes and extended structures (see their Fig. 8). These correspond to the brightness variations and the “cap” seen in our H$_2$ image. They also analyzed \[OI\] and \[CI\] emission lines, and making the reasonable assumption that these lines were produced in the same shock as the H$_2$ lines, they used these as diagnostics of the physical conditions in the shock region. From these, they deduce that the emission arises in very high density regions, 3 $\times$ 10$^6$ $\le$ n$_e$ $\le$ 5 $\times$ 10$^7 c$m$^{-3}$. At these conditions of density, temperature [T$_{rot}$$\approx$1400 K; @vds03], and velocity (8 km s$^{-1}$), the excitation can be due to either C-type or J-type shocks [@vds08]. They also made long-slit observations of the \[FeII\] 1.645 $\mu$m emission line, which indicate that this emission arises, not in the lobes, but from a wind close to the star.
Hen 3-401 {#hen-3-401-1}
---------
This nebula is very highly collimated, with the lobes extending $\sim$14$\farcs$5 from the central star with a length to width ratio for each lobe of $\sim$7. The lobes in this object do not appear to be closed, but rather have “tattered” ends. The PV diagrams show a Hubble flow in the velocity in the lobes. This can arise due to an eruptive ejection (ballistic motions) or (self-similar) hydrodynamic growth patterns [@balick02]. The velocities show that the eastern lobe is tilted away from us and the western lobe toward us. From the complementary image and spectroscopy, an estimate can be made of the velocity at the ends of the lobes and the inclination angle of the bipolar axis. With the measured length of the open-ended lobes of 14$\farcs$5 and the width of 2$\farcs$2 [from the sharpened V image, @sahai99b], and the lateral lobe expansion velocity of 15.5 km s$^{-1}$, one can calculate a minimum expansion velocity at the ends of the lobes of $\sim$200 km s$^{-1}$. The PV diagram along the bipolar axis (Fig. \[spec\_hen3\]b) shows that the velocity at the ends of the two lobes differ by 20 km s$^{-1}$ from that of the star. This implies an inclination angle of $\leq$ 6$\arcdeg$ to the plane of the sky. This is somewhat smaller than we found from the appearance of the PV diagram in Section \[hen3 PV\], but both clearly indicate a small inclination angle.
In visible light, a “skirt” structure is seen surrounding the inner parts of the bipolar lobes and extending coaxially with the lobes for $\sim$3$\arcsec$ from the star [discussed by @sahai99b]. This skirt also appears in the H-band image but not in H$_2$. This is seen in Figure \[hen3\_multi\]. The western lobe is the brighter in scattered light, with brightness ratios in the different bands of 2.4 [V; @sahai99b], 1.4 (J), and 1.4 (2.15 [${\rm \mu m}$]{}). If one assumes an intrinsic symmetry in the two lobes, then this suggests some obscuration by an extended equatorial disk. In contrast, the eastern lobe is brighter in [${\rm H_2}$]{}, with a ratio of 1.2. This may be a result of the differing local shock conditions in the nebula. There is a suggestion of a torus from the optical image [@sahai99b] and from the polarization measurements [@scar95; @ueta07]. A recent high-resolution mid-infrared imaging study shows that the mid-IR emission arises mostly from a 1$\arcsec$ core, with weaker emission arising from the walls extending out to about 3$\arcsec$ from the star [@muthu06]. These authors suggest that the core contains a flared disk, based upon the temperature gradient detected in the core.
In the H band, the star is more prominent than in the visible and the inner region of the lobes is slightly less prominent. In visible light, there is a bright inner edge to the lobes, with enhanced light out to the ends of the skirt. This is also seen in the H band but not the H$_2$ image. A faint halo is seen in both the visible and H-band images. The H$_2$ emission is particularly bright 2.0$-$5.5$\arcsec$ from the star. These different regions of brightness are seen in the composite image (Fig. \[hen3\_multi\]d).
Based upon the extremely large intensity ratio of the [${\rm H_2~}$]{}1$-$0 S(1) to 2$-$1 S(1) line of 26 [@garher02], it is clear that the excitation mechanism is due to shocks. These presumably arise as the fast wind, that carved out the narrow lobes, produces shocks in the walls that excite the [${\rm H_2}$]{}. H$\alpha$ wings are observed to extend out to $\sim$1600 km s$^{-1}$ [@garlar99b].
The young PN M 2-9 is very similar to Hen 3-401 in visible appearance, as has been remarked upon by @garlar99b, and also in its mid-infrared appearance [although with not quite as pinched a waist; @smith05a] and its kinematics. It possesses smooth, almost cylindrical lobes, a bright central star, and a thin [${\rm H_2~}$]{}zone [@hora94; @smith05b]. A recent study of winds driven by magnetic pressure from a toroidal field was able to reproduce well the morphology of these two objects [@garseg05].
Rob 22
------
The PV diagrams for [Rob 22 ]{}show an expanding nebula with the [${\rm H_2~}$]{}emission arising from the ends of the lobes and the walls. The bipolar axis lies nearly in the plane of the sky with the south lobe expanding slightly toward us compared with the north lobe. Given the expansion velocity of $\sim$15 km s$^{-1}$ and the angular size of the lobes of $\sim$2.0$\arcsec$, this leads to a kinematic age of $\sim$320 $\times D_{kpc}$ yr. For a distance of $\sim$2 kpc [@allen80], this yields an age for the lobes of $\sim$640 yr.
In the visual image [Figure \[rob22\_multi\]a; @sahai99a], the south lobe is brighter than the north lobe with a ratio of 1.6. In the near-infrared, the north lobe is slightly brighter, with ratios of 1.05 in H (F160W; Figure \[rob22\_multi\]b), 1.1 in 2.15 $\mu$m, and 1.2 in the [${\rm H_2~}$]{}image (Figure \[rob22\_multi\]c). If the reduced relative brightness in the south lobe in the V band is due to extinction by the dense equatorial disk, then the disk would need to be extremely large if it were orthogonal to the axis of the outflow, which as we have seen is very close to the plane of the sky. It may be more likely that there is simply an intrinsic asymmetry in the lobes that contributes most of the difference.
If we compare the different images (Fig. \[rob22\_multi\]), we see a strong point symmetry in the “S” shape. Even at 7.7 and 8.7 $\mu$m this “S” shape is seen [@muthu04]. Such a pattern can be formed by collimation of an outflow by a precessing disk. The H$\alpha$ emission has wings to at least $\pm$450 km s$^{-1}$ [@allen80].
The V image shows a distinct dark band separating the lobes. In the H image, this dark lane is not so prominent, and a bright, compact clump appears. This bright clump is red but not point-like, and likely represents warm dust heated by the obscured central star. The nebula is brighter close to the star, with a particularly bright region NE of the center of the nebula. A similar morphology to the H-band image is seen in a recent adaptive optics very high resolution (R$\sim$0.06$\arcsec$) K-band image obtained by @lag08; this shows some finer detail and in particular a bright linear extension toward the SE from the very bright region near the center, defining the southern part of the “V”-shaped central obscured region. The halo is evident in both the V and H images. A faint region in the N lobe (referred to as a “spur” by @sahai99a) is seen in both the V and H images, although in the V image it is outlined by emission farther out in the lobe while in the H image it is outlined by emission closer to the center. In the [${\rm H_2~}$]{}image, the emission is particularly strong in the same region NE of the center, with a less bright region on the other side (SW) of the center. In addition, the [${\rm H_2~}$]{}is particularly bright along the ends of the lobes. This can arise due to shock excitation produced as the collimated wind, that may have produced the S-shaped nebula, interacts with the AGB wind of the halo. Polarization measurements show scattering from the ends of the lobes, in addition to the presence of a equatorial dust torus [@ueta07].
The faint region “spur” in the N lobe is evident in all three images. It is interesting that in the OH observations [@dyer01], the more negative velocities arise from this faint region; perhaps it represents an evacuated hole punched through the nebula.
The OH emission appears to be confined primarily to the equatorial regions [@sahai99a; @dyer01; @zij01], where it may represent the expansion or rotation of a disk. The mixed C and O circumstellar chemistry might also suggest a disk, as this explanation has been advocated as the cause of the mixed chemistry in several other post-AGB objects, such as the Red Rectangle [@win03]. The detection of hot dust, manifested in the spectral energy distribution of this object (but not in the other two), is consistent with a disk. Or it may be that this object is in transition between an O-rich chemistry to a C-rich one, with the PAH emission arising closer to the star, as may be the case for PN with \[WC\] central stars.
Comparison With Other Spatial-Kinematic Studies of PPNs
-------------------------------------------------------
A small number of other bipolar PPNs have also been studied spatially or kinematically in [${\rm H_2~}$]{}. We discuss these for comparison.
AFGL 2688 (“Egg” nebula) has been imaged using [*HST*]{}-NICMOS [@sahai98] and shows [${\rm H_2~}$]{}emission from clumps in its spindle-shaped closed lobes, especially near their outer ends. This is consistent with the emission arising along the walls of the lobes. It also shows strong emission from an extended equatorial region. A kinematic study by @kastner01 shows that the highest velocity [${\rm H_2~}$]{}along the polar axis resides close to the star and that the velocity of the [${\rm H_2~}$]{}[*decreases*]{} along the polar axis away from the star. They find a similar effect for AFGL 618. They interpret this as resulting from a faster collimated wind colliding with slower-moving AGB material. The equatorial [${\rm H_2~}$]{}shows one side to be blue- and the other red-shifted, which @kastner01 interpret indicating a combination of expansion and rotation. An alternate interpretation would be multiple bipolar jets along the equator [@cox03] rather than rotation, and this seems to be the more likely explanation. This decrease with distance along the polar axis seen in AFGL 2688 and AFGL 618 is in contrast with what is seen in Hen 3-401. Hen 3-401 displays a clear trend of the [${\rm H_2~}$]{}velocity [*increasing*]{} in velocity with distance away from the star, in accord with what one would expect from an unhindered ejection of material from the region of the central star.
@davis05 recently made spatially-resolved high-resolution (R$\sim$20,000 or 16 km s$^{-1}$) [${\rm H_2~}$]{}observations of four bipolar PPNs. They observed them with the spectrograph slit oriented along the bipolar axis and then offset slightly to each side, similar to our observing pattern for Hen 3$-$401 and Rob 22. In two of these objects, M 1$-$2 and IRAS 17441$-$2411, one sees a distinct velocity difference between the two lobes ($\Delta$V of $\sim$55 and $\sim$10 km s$^{-1}$, respectively), for IRAS 17243$-$1755 [${\rm H_2~}$]{}is detected only weakly from one lobe, and for IRAS 17150$-$3224 the two lobes have similar velocities (bipolar axis in the plane of the sky). For this last object, they also suggested that there may be rotation in the lobes based on systematic velocity differences seen in the offset positions.
[@hri06] also studied IRAS 17150$-$3224 with higher spectral resolution, and they obtained a high-resolution (0$\farcs$2) [*HST*]{}-NICMOS [${\rm H_2~}$]{}image. The image shows the [${\rm H_2~}$]{}emission to come primarily from clumps near the ends of the lobes, but these are not resolved in the spectra, which basically show the integrated emission and velocity of the clumps. The two lobes have a similar velocity, as in the above study, with slight velocity differences in the offset positions due to small effects of the different [${\rm H_2~}$]{}clumps rather than rotation.
As we have referred to previously, @vds08 also recently studied the kinematics of IRAS 16594$-$4656. They also observed with the Phoenix spectrograph on Gemini-South at even somewhat higher spectral and spatial resolution than we did. Their spectra show a similar picture of an expanding H$_2$ ellipsoid with V$_{exp}$ = 8 km s$^{-1}$. They combined their spectra with ours to produce a detailed three-dimensional image of the shell structure.
In none of these earlier studies does one see such clearly-resolved spatial-kinematic structures as are seen in the new Gemini-South Phoenix data of IRAS 16594$-$4656, Hen 3$-$401, and Rob 22. In all three the emission from the walls of the lobes is clearly traced in the PV diagrams. The small slit size and high spectral resolution, together with the high spatial resolution images, result in exceptional kinematic detail for a PPN.
SUMMARY AND CONCLUSIONS
=======================
The high-resolution [*HST*]{}-NICMOS [${\rm H_2~}$]{}images of these three PPNs have allowed us to determine the spatial location of their [${\rm H_2~}$]{}emission and the spatially-resolved [${\rm H_2~}$]{}spectroscopic observations have allowed us to determine the kinematics of the [${\rm H_2~}$]{}emitting regions. In all three cases, the systemic and expansion velocities are similar to those determined from the molecular-line CO or OH measurements. We find the following results for these three PPNs.
[*IRAS 16594$-$4656*]{}: While the V and H-band images show a complex multi-lobe structure, the [${\rm H_2~}$]{}images shows a clear bipolar, peanut shape, although with variations in density (“holes”) and structure [see @vds08]. The [${\rm H_2~}$]{}emission originates along the walls of the lobes (sides and ends), with fainter emission from more distant (ejected?) clumps. The PV diagram shows the [${\rm H_2~}$]{}to arise in an expanding ellipsoidal velocity structure, which is in contrast to the bilobes of the density structure. The kinematics indicates that the bipolar lobes are nearly in the plane of the sky (i$\approx$10$\arcdeg$); this differs from the earlier interpretations of the lobes as being at some intermediate orientation, but it is consistent with recent mid-IR imaging and near-IR polarization studies. The lobes are estimated to have an age of $\sim$1600 yr.
[*Hen 3-401*]{}: The [${\rm H_2~}$]{}emission originates from the sides of lobes, which have open ends. The lobes are tilted somewhat to the plane of the sky ($\sim$10$-$15$\arcdeg$), with the western lobe moving toward us, and they show an increasing velocity with radial distance (Hubble flow). An estimated age for the lobes is $\sim$1100 yr. The open ends and unhindered outflow may be the consequence of its higher degree of collimation (greater linear momentum). [*Rob 22*]{}: The [${\rm H_2~}$]{}emission originate primarily from ends of the S shaped nebula and from regions of the S shape near the obscured central star. The nebula is nearly in the plane of the sky, consistent with the absence of a visible star due to obscuration by a disk. An age of $\sim$630 yr is estimated for the lobes.
[${\rm H_2~}$]{}surveys of PPNs have shown that [${\rm H_2~}$]{}emission is commonly found in those with a bipolar morphology. As can be seen from this study, the combination of [${\rm H_2~}$]{}high-resolution images and spatially-resolved, high-resolution spectra provides valuable insight into the structure and shaping mechanisms for these bipolar nebulae.
We thank Steve Ridgway, Bernadette Rogers, Kevin Volk, and Claudia Winge for making the Phoenix queue observations, Ken Hinkle for assistance in planning the Phoenix observing programs, and Anibal García-Hernández and Griet Van de Steene for making available to us their medium-resolution 2 $\mu$m spectra in digital form. We thank Nico Koning for making the SHAPE images and Wenxian Lu for help with the image measurements. The comments of the referee were helpful in improving the presentation. We acknowledge grants from NASA that provided partial support for B.J.H. (GO-07840.02-A, GO-09366.01-A), R.S. (GO-07840.01-A ,GO-09463.01-A, GO-09801.01-A), N.S. (HF-01166.01A), and K.Y.L.S. (GO-09366.03-A) from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. B.J.H. also acknowledges the support of the National Science Foundation under Grant No. 0407087 and R.S. thanks NASA for partially funding this work by a NASA LTSA award (no. 399-20-40-06). Some of the research described in this paper was carried out by R.S. at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. This research made use of the SIMBAD database, operated at CDS, Strasbourg, France, and NASA’s Astrophysics Data System.
Allen, D.A. 1978, , 184, 601 Allen, D.A., Hyland, A.R., & Caswell, J.L. 1980, , 192, 505 Balick, B. 1987, , 94, 671 Balick, B., & Frank, A. 2002, , 40, 439 Bragg, S.L., Brault, J.W., & Smith, W.H. 1982, ApJ, 263, 999 Bujarrabal, V., & Bachiller, R. 1991, , 242, 247 Cox, P., Huggins, P. J., Maillard, J.-P., Muthu, C., Bachiller, R., & Forveille, T. 2003, , 586, L87 Davis, C.J., Smith, D.M., Gledhill, T.M., & Varricatt, W.P. 2005, , 360, 104 Dyer, K.K., Goss, W.M., & Kemball, A.J. 2001, , 121, 2743 García-Hernández, D.A., Manchado, A., García-Lario, P., Domíngueq-Tagle, C., Conway, G.M., & Prada, F. 2002, , 387, 955 García-Lario, P., Manchado, A., Ulla, A., & Manteiga, M. 1999a, , 513, 941 García-Lario, P., Riera, A., & Manchado, A. 1999b, , 526, 854 García-Segura, G., López, J.A., & Franco, J. 2005, , 618, 919 Hinkle, K. H., Cuberly, R. W., Gaughan, N. A., et al. 1998, Proc. SPIE, 3354, ed. A.M. Fowler, 810 Hora, J. L., & Latter, W. B. 1994, , 437, 281 Hrivnak, B. H., Kelly, D. M., & Su, K. Y. L. 2004, in ASP Conf. Ser. 313, Asymmetric Planetary Nebulae III, ed. M. Meixner, J.H. Kastner, B. Balick, N. Soker (San Franciso: ASP), 175 Hrivnak, B. H., Kelly, D. M., Su, K. Y. L., Kwok, S., & Sahai, R. 2006, , 650, 237 Hrivnak, B. H., Kwok, S. & Su, K. Y. L. 1999, , 524, 849 Hrivnak, B. H., Kwok, S. & Su, K. Y. L. 2001, , 121, 2775 Hrivnak, B. H., Volk, K. & Kwok, S. 2000, , 535, 275 Kastner, J.H., Weintraub, D.A. , Gatley, I., & Henn, L. 2001, , 546, 279 Kelly, D.M., & Hrivnak, B.J., 2005, , 629, 1040 Koekemoer, A. M., et al. 2002, “HST Dither Handbook”, Version 2.0 (Baltimore: STScI) Kwok, S. 1982, , 258, 280 Lagadec, E., Chesneau, O., Zijlstra, A., & Mekarnia, D. 2008, in Asymmetric Planetary Nebulae IV, ed. R.L.M. Corradi, A. Manchado, , N. Soker, in publication Loup, C., Forveille, T., Nyman, L. $\AA$., & Omont, A. 1990, , 227, L29 Muthu, C., Kwok, S., & Volk, K. 2004, Bul. AAS, 36, 1570 Muthumariappan, C., Kwok, S., & Volk, K. 2006, , 640, 353 Molster, F., Waters, L.B.F.M., de Jong, T., Prusti, T., Zijlstra, A., & Meixner, M. 1997, IAU Symp. 180: Planetary Nebulae, ed. H.J. Habing, H.J.G.L.M. Lamers (Kluwer: Dordrecht), 361 Molster, F., Waters, L.B.F.M., Tielens, A.G.G.M., & Barlow, M.J. 2002, , 382, 184 Parthasarathy, M., García-Lario, P., Gauba, G. et al. 2001, , 376, 941 Parthasarathy, M., & Pottasch, S.R. 1989, , 154, L16 Reyniers, M. 2002, Ph.D. Thesis, Catholic University of Leuven Sahai, R., Bujarrabal, V., & Zijlstra A. 1999b, , 518, L115 Sahai, R., Hines, D. C., Kastner, J. H., Weintraub, D. A., Trauger, J. T., Rieke, M. J., Thompson, R. I., & Schneider, G. 1998, , 492, L163 Sahai, R., Morris, M., Sánchez Contreras, C., & Claussen, M. 2007, , 134, 2200 Sahai, R., Su, K.Y.L., Kwok, S., Dayal, A. & Hrivnak, B.J. 2000, in ASP Conf. Ser. 199, Asymmetric Planetary Nebulae II - From Origins to Microstructures, ed. J.H. Kastner, N. Soker, S.A. Rappaport (San Franciso: ASP), 167 Sahai, R., & Trauger, J.T. 1998, , 116, 1357 Sahai, R., Zijlstra A., Bujarrabal, V., & te Lintel Hekkert, P. 1999a, , 117, 1408 Scarrott, S.M., & Scarrott, R.M.J.,1995 , 277, 277 Silva, A.M., Azcárate, I.N., Poppel, W.G.L., & Likkel, L. 1993, , 275, 510 Siódmiak, N., Meixner, M., Ueta, T., Sugerman, B.E.K., Van de Steene, G.C., & Szczerba, R. 2008, , 677, 382 Smith, N. 2006, , 644, 1151 Smith, N., & Gehrz, R.D. 2005, , 129, 969 Smith, N., Balick, B., & Gehrz, R.D. 2005, , 130, 853 Steffen, W., & López, J.A. 2006, , 42, 99 Su, K. Y. L., Hrivnak, B. J., & Kwok, S. 2001, , 122, 1525 Su, K. Y. L., Hrivnak, B. J., Kwok, S., & Sahai, R. 2003, , 126, 848 Ueta, T., Meixner, M., & Bobrowsky, M. 2000, , 528, 861 Ueta, T., Murakawa, K., & Meixner, M. 2005, , 129, 1625 Ueta, T., Murakawa, K., & Meixner, M. 2007, , 133, 1345 Van de Steene, G. C., & van Hoof, P. A. M. 2003, , 406, 773 Van de Steene, G. C., Ueta, T., van Hoof, P. A. M., Reyniers, M., & Ginsburg, A.G. 2008, , 480, 775 Van de Steene, G. C., Wood, P.R., & van Hoof, P. A. M. 2000, in ASP Conf. Ser. 199, Asymmetric Planetary Nebulae II - From Origins to Microstructures, ed. J.H. Kastner, N. Soker, S.A. Rappaport (San Franciso: ASP), 191 van Winckel, H. 2003, , 41, 391 Volk, K., Hrivnak, B.J., Su, K.Y.L. & Kwok, S. 2006, , 651, 294 Volk, K., Kwok, S., Hrivnak, B.J., & Szczerba, R. 2002, , 567, 412 Zijlstra, A.A., Chapman, J.M., te Lintel Hekkert, P., Likkel, L., Comeron, F., Norris, R.P., Molster, F.J., & Cohen, R.J. 2001, , 322, 280
[lclccccc]{} 10178$-$5958 & Hen 3-401 &1998 Mar 6&1728 &1728 &&576 &1152\
10197$-$5750 & Rob 22 &1998 May 18&1344 &1344 &&320 &768\
16594$-$4656 & &2002 Aug 13&1215 &1215 &192 &&\
[lcccccc]{} F212N-new& 2.1213 & 0.0087 & 1.9135E-14 & 2.8722E-5 & 664.7 & 0.19\
F215N-new& 2.1487 & 0.0079 & 2.1031E-14 & 3.2389E-5 & 645.1 & 0.20\
F212N-old& 2.1213 & 0.0088 & 2.4272E-14 & 3.6432E-5 & 664.7 & 0.20\
F215N-old& 2.1487 & 0.0079 & 2.5533E-14 & 3.9323E-5 & 645.1 & 0.20\
F110W-new& 1.1235 & 0.1630 & 2.9059E-15 & 1.2234E-6 &1784.4 & 0.12\
F160W-old& 1.6060 & 0.1177 & 2.3956E-15 & 2.0610E-6 &1040.7 & 0.14\
[lcrccc]{} Rob 22 & 2003 Jan 14 & 17 & through star & 500 & 2\
Rob 22 & 2003 Jan 14 & 17 & 0$\farcs$7 W of star & 500 & 2\
Rob 22 & 2003 Jan 14 & 17 & 0$\farcs$7 E of star & 500 & 2\
Hen 3-401 & 2003 Dec 17& 73 & through star & 500 & 2\
Hen 3-401 & 2003 Dec 17& 73 & 0$\farcs$7 S of star & 500 & 2\
Hen 3-401 & 2003 Dec 17& 73 & 0$\farcs$7 N of star & 500 & 2\
Hen 3-401 & 2003 Dec 17& 163 & through star & 500 & 2\
IRAS 16594$-$4656 & 2003 May 08& 52 (b) & through star & 600 & 2\
IRAS 16594$-$4656 & 2003 May 08& 33 (c) & through star & 600 & 2\
IRAS 16594$-$4656 & 2003 May 08& 345 (e) & 0$\farcs$9 W of star & 600 & 2\
IRAS 16594$-$4656 & 2003 May 15& 345 (d) & through star & 600 & 3\
IRAS 16594$-$4656 & 2003 May 15& 72 (a) & through star & 600 & 2\
[^1]: @garlar99b previously displayed the [*HST*]{} [${\rm H_2~}$]{}data, using a scale factor based upon the field stars, and obtained a similarly appearing [${\rm H_2~}$]{}image.
[^2]: ftp://ftp.noao.edu/catalogs/arcturusatlas/ir/
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'H. Gupta, P. Rimmer, J. C. Pearson, S. Yu, E. Herbst, N. Harada, E. A. Bergin, D. A. Neufeld, G. J. Melnick, R. Bachiller, W. Baechtold, T. A. Bell, G. A. Blake, E. Caux, C. Ceccarelli, J. Cernicharo, G. Chattopadhyay, C. Comito, S. Cabrit, N. R. Crockett, F. Daniel, E. Falgarone, M. C. Diez-Gonzalez, M.-L. Dubernet, N. Erickson, M. Emprechtinger, P. Encrenaz, M. Gerin, J. J. Gill, T. F. Giesen, J. R. Goicoechea, P. F. Goldsmith, C. Joblin, D. Johnstone, W. D. Langer, B. Larsson, W. B. Latter R. H. Lin, D. C. Lis, R. Liseau, S. D. Lord, F. W. Maiwald, S. Maret, P. G. Martin, J. Martin-Pintado, K. M. Menten, P. Morris, H. S. P. Müller, J. A. Murphy, L. H. Nordh, M. Olberg, V. Ossenkopf, L. Pagani, M. Pérault, T. G. Phillips, R. Plume, S.-L. Qin, M. Salez, L. A. Samoska, P. Schilke, E. Schlecht, S. Schlemmer, R. Szczerba, J. Stutzki, N. Trappe, F. F. S. van der Tak, C. Vastel, S. Wang, H. W. Yorke, J. Zmuidzinas, A. Boogert, R. Güsten, P. Hartogh, N. Honingh, A. Karpov, J. Kooi, J.-M. Krieg, R. Schieder'
- 'P. Zaal'
title: 'Detection of OH$^+$ and H$_2$O$^+$ towards Orion KL'
---
Introduction
============
The Heterodyne Instrument for Far Infrared (HIFI) on the [*Herschel Space Observatory*]{} [^1] provides a unique opportunity to fully assess the first steps of the oxygen chemistry in a wide variety of sources. Initial HIFI observations quickly detected widespread absorption by OH$^+$ and H$_{2}$O$^{+}$ toward the star forming regions DR21, W31C, and W49N (Ossenkopf et al. 2010; Gerin et al. 2010; Neufeld et al. 2010, this issue). Prior to the HIFI observations, OH$^+$ had only been detected in absorption toward Sgr B2(M) (Wyrowski et al. 2010). Similarly, previous observations of H$_{2}$O$^{+}$ were limited to its detection in comet tails (e.g., Herzberg & Lew 1974; Wehinger et al. 1974), demonstrating the importance of photoionization in producing this ion in the absence of H$_2$. And until recently, only upper limits had been reported on the column density of H$_2$O$^+$ in the diffuse interstellar gas (Smith, Schempp, & Federman 1984).
By contrast, the recent HIFI detections of OH$^+$ and H$_{2}$O$^{+}$ in warm diffuse gas with a fairly small fraction of molecular hydrogen, elucidated the role of O$^+$ in initiating the oxygen-hydrogen chemistry. This chemistry is thought to begin with the production of H$^+$ and H$_{3}^+$ [*via*]{} cosmic ray or X-ray ionization of hydrogen, followed by charge transfer to produce O$^+$. Rapid hydrogen abstraction reactions of O$^+$ with H$_2$ then yield OH$^+$ and H$_{2}$O$^{+}$, and terminate with the production of H$_{3}$O$^{+}$. In diffuse molecular clouds, which have high electron abundances, the H$_3$O$^+$ is destroyed [*via*]{} dissociative recombination to yield OH and H$_2$O. In dense molecular clouds, both the ionization fraction and the atomic hydrogen abundance are comparatively lower, and the sequence of reactions, expected to start at H$_{3}^+$ and OH$^+$, yields a larger abundance of H$_{3}$O$^{+}$. This picture is probably overly simplistic for molecular clouds such as Orion KL, which are composed of both diffuse and dense gas.
Orion KL is the brightest infrared region in the Orion-Monoceros molecular cloud complex located less than 500 pc from the sun (Menten et al. 2007). In the foreground of Orion KL is the Orion Nebula, an HII region known to contain a cluster of thousands of young stars which produce a substantial flux of X-ray photons (Getman et al. 2005). Molecular line studies reveal three main regions in Orion KL: i. a core of very dense and hot gas ($n \sim 10^{7}$ cm$^{-2}$, $T \sim 200$ K); ii. cool, quiescent gas between systemic velocities of 8 km s$^{-1}$ and 10 km s$^{-1}$, surrounded by high-velocity outflows ($\ge 100$ km s$^{-1}$); and iii. a highly inhomogeneous and turbulent outflow source containing both high-velocity ($\ge 30$ km s$^{-1}$) and low-velocity ($\sim 18$ km s$^{-1}$) gas (Blake et al. 1987; Genzel & Stutzki 1989; O’Dell et al. 2008).
In this [*Letter*]{} we report the detection of absorption lines of OH$^+$ and H$_{2}$O$^{+}$, and an upper limit on the column density of H$_{3}$O$^{+}$ toward Orion KL. In addition to molecular absorption at a systemic velocity of 9 km s$^{-1}$, these observations find broad blueshifted absorption by OH$^+$ and H$_{2}$O$^{+}$ extending to large negative velocities. This is consistent with previously observed lines of H$_2$O with ISO (Lerate et al. 2006), as well as those of HF and [*para*]{}-H$_{2}^{18}$O detected recently with HIFI, and attributed to the low-velocity molecular outflow (Phillips et al. 2010).
Observations and data reduction
===============================
The observations were done in March 2010 as part of the Key Program [*Herschel/HIFI Observations of Extraordinary sources: The Orion and Sagittarius Star-forming Regions*]{} (HEXOS). The dual beam switch (DBS) observing mode was used, with the DBS reference beams lying approximately $3^\prime$ east and west of the Orion KL position $\alpha_{J2000} = 5^h35^m14.3^s$ and $\delta_{J2000} =
-5^{\circ}22'33.7''$. Spectra were taken with the Wide Band Spectrometer (WBS) with a Nyquist-limited frequency resolution of approximately 1.1 MHz over a 4 GHz wide IF band; the HIFI beams in bands 4, 5, and 6 have half-power beam widths of $21''$, $19''$, and $13''$ and main beam efficiencies of $0.670$, $0.662$, and $0.645$ (HIFI Observers’ Manual, v 2.0). The spectra were reduced through the standard Herschel Pipeline to Level 2 using HIPE version 2.4 (Ott 2010). The double sideband (DSB) spectra so obtained were then deconvolved (Comito & Schilke 2002) to single sideband (SSB) spectra using the [*doDeconvolution*]{} task in HIPE. The SSB spectra were converted to the FITS format and analyzed with the CLASS90 package. Although two orthogonal polarizations were observed simultaneously, only spectra from the H polarization in bands 4a and 6b and the V polarization in band 5a are shown, because of the smaller standing waves in these polarizations.
Spectroscopy
============
The spectroscopy of OH$^+$, H$_2$O$^+$, and H$_3$O$^+$ has been discussed in detail in the recent detection papers (Ossenkopf et al. 2010; Gerin et al. 2010). Here we summarize the essential aspects of the rotational spectra of these ions. The OH$^+$ ion has a $^{3}\Sigma^{-}$ electronic ground state, the two unpaired electron spins ($S=1$) yielding three components of the $N=1-0$ transition. The nuclear spin of the hydrogen atom ($I_{H}=1/2$) further splits each component into hyperfine components. The H$_2$O$^+$ ion has $C_{2v}$ symmetry and a $^{2}B_{1}$ ground state which results in the lowest level having [*ortho*]{} symmetry. The spin of the unpaired electron ($S=1/2$) results in two fine-structure components, each exhibiting a complex hyperfine pattern due to the spins of the two equivalent hydrogen nuclei ($I_{H}=1/2$). Rotational spectroscopy of H$_2$O$^+$ is limited to two laser magnetic resonance (LMR) studies (Strahan et al. 1986; M[ü]{}rtz et al. 1998). Here, we adopt the values of M[ü]{}rtz et al., which we and others have checked independently to be accurate to about 2 km s$^{-1}$ in equivalent radial velocity (see Neufeld et al. 2010 and Schilke et al. 2010, this volume). H$_3$O$^+$ is a closed-shell symmetric top molecule with a large amplitude inversion near 1.65 THz, resulting in a spectrum similar to NH$_3$ with transitions between symmetric and antisymmetric inversion states (Yu et al. 2009). Table 1 lists the observed transitions of the three ions, along with their line strengths and spontaneous emission rates.
\[tab:spec\]
[rcrrrrr]{}
------------------------------------------------------------------------
$Transition$ & $Frequency$ & $E_l$ & $g_l$ & $g_u$ &$\mu^{2}S$ & $10^{2}A_{ij}$\
& ( MHz) & (cm$^{-1}$) & & & ($\rm{D}^2$) & (s$^{-1}$)\
\
\
$F=1/2-1/2$ & $909045.2\pm1.5$ & 0.004 & 4 & 6 & 1.20 & 1.05\
$ 1/2-3/2$ & $909158.8\pm1.5$ & 0 & 2 & 4 & 2.40 & 0.52\
\
$F=5/2-3/2$ & $971803.8\pm1.5$ & 0 & 4 & 6 & 10.24 & 1.82\
$ 3/2-1/2$ & $971805.3\pm1.5$ & 0.004 & 2 & 4 & 5.69 & 1.52\
$ 3/2-3/2$ & $971919.2\pm1.0$ & 0 & 4 & 4 & 1.14 & 0.30\
\
\
$F=3/2-1/2$ & $1115150.0\pm1.8$ & 0.004 & 2 & 4 & 4.14 & 1.67\
$1/2-1/2$ & $1115186.0\pm1.8$ & 0.004 & 2 & 2 & 3.32 & 2.68\
$5/2-3/2$ & $1115204.0\pm1.8$ & 0 & 4 & 6 & 11.23 & 3.02\
$3/2-3/2$ & $1115263.0\pm1.8$ & 0 & 4 & 4 & 3.35 & 1.35\
$1/2-3/2$ & $1115298.7\pm1.8$ & 0 & 2 & 2 & 0.42 & 0.34\
\
$F=3/2-1/2$ & $1139541.1\pm1.8$ & 0.004 & 2 & 2 & 0.42 & 3.61\
$1/2-1/2$ & $1139560.6\pm1.8$ & 0.004 & 2 & 4 & 3.35 & 1.44\
$5/2-3/2$ & $1139653.5\pm1.8$ & 0 & 4 & 2 & 3.32 & 2.86\
$3/2-3/2$ & $1139673.3\pm1.8$ & 0 & 4 & 4 & 4.14 & 1.78\
\
$0_{0}^{-} - 1_{0}^{+}$ & $984711.9\pm0.1$ & 5.1 & 4 & 12 & 8.30 & 2.30\
$1_{1}^{-} - 1_{1}^{+}$ & $1655834.8\pm0.3$ & 0 & 6 & 6 & 6.22 & 5.48\
$2_{2}^{-} - 2_{2}^{+}$ & $1657248.4\pm0.3$ & 29.6 & 10 & 10 & 4.67 & 7.32\
Results
=======
Figure \[fig1\] shows the absorption lines of OH$^+$ and H$_2$O$^+$ toward Orion KL, as well as lines of HF and [*para*]{}-H$_{2}^{18}$O for comparison. The strongest hyperfine components of OH$^+$ and H$_2$O$^+$ appear at the source velocity of 9 km s$^{-1}$, which matches well that of the HF line in Orion KL. Additionally, lines of both ions show broad blue absorption wings extending to about $-75$ km s$^{-1}$, more extended than the HF absorption, but comparable to that of [*para*]{}-H$_{2}^{18}$O ($\sim -80$ km s$^{-1}$). We attribute the extended absorption of the ions to originate mainly from the low velocity molecular outflow. We failed to detect any emission or absorption from H$_3$O$^+$, and discuss the non-detection in $\S$ \[sec:disc\].
The high density of molecular lines in Orion KL makes contamination by unrelated lines a common problem. The absorption lines detected here are blended with weak to moderately strong emission lines of abundant “weeds”, including CH$_3$OH and SO$_2$. Efforts are underway to model and remove the emission from the contaminants by a method similar to that of Phillips et al. (2010); in the interim, the following approach was taken.
To better gauge the absorption, the contaminants were masked and intensities interpolated across the masked channels (Fig. \[fig1\]). The velocity-integrated optical depths of the ionic lines were obtained by normalizing the SSB spectra with the continuum and integrating over the velocity ranges for the source and the outflow, the interpolation yielding errors of $20\%-30\%$. On the assumptions that the absorption covers the source completely, and the molecules are in the lower state, the total column density ($N$) was then derived using the expression:
$$\int \tau d{\rm v}~({\rm km~s^{-1}}) = \frac{A_{ul}g_u \lambda^3}{8\pi g_l} N,$$
where $A_{ul}$ is the spontaneous emission rate, $g_u$ and $g_l$ are the upper and lower state degeneracies, and $\lambda$ is the transition wavelength.
We estimate column densities of OH$^+$ and H$_2$O$^+$ at 9 km s$^{-1}$ of $9 \pm 3 \times 10^{12}$ cm$^{-2}$ and $7 \pm 2\times 10^{12}$ cm$^{-2}$, and those in the outflow of $1.9 \pm 0.7\times 10^{13}$ cm$^{-2}$ and $1.0 \pm 0.3\times 10^{13}$ cm$^{-2}$. The column densities of OH$^+$ are more than an order of magnitude lower, and those of H$_2$O$^+$ are $2-6$ times lower than toward W31C and W49N (Gerin et al. 2010; Neufeld et al. 2010, this issue). From the least congested spectra of H$_3$O$^+$ at 984.7 and 1657.2 GHz (see Table 1), and an assumed excitation temperature of 100 K, we derive $3\sigma$ upper limits of $2.4 \times 10^{12}$ cm$^{-2}$ and $8.7 \times 10^{12}$ cm$^{-2}$ for the column density of [*ortho*]{} and [*para*]{}-H$_3$O$^+$, nearly an order of magnitude lower than in W31C (Gerin et al. 2010).
The abundance ratios of the three ions in Orion KL can be compared to the same ratios observed in W31C and W49N. The OH$^+$/H$_2$O$^+$ ratio is found to be $1.3 \pm 0.6$ in the source and $1.8 \pm 0.8$ in the outflow. This ratio is $2-15$ times lower than that measured toward W31C and W49N. The lower limit of 1.4 for the H$_2$O$^+$/H$_3$O$^+$ ratio, however, is nearly 2 times larger than in W31C.
Discussion {#sec:disc}
==========
The column densities of OH$^+$, H$_2$O$^+$, and H$_3$O$^+$ in Orion KL differ markedly from those in the diffuse gas toward W31C and W49N. In contrast with W31C and W49N, OH$^+$ and H$_2$O$^+$ are significantly more abundant relative to H$_3$O$^+$, for which we are only able to obtain an upper limit. The absolute column densities of OH$^+$ and H$_2$O$^+$ are also lower compared with W31C and W49N. A likely explanation for the low column densities of the three ions is that they are present in fairly dense material, both in the quiescent gas and the outflow. Unlike the quiescent gas, the Orion KL outflow is exposed to a strong ionizing flux from the foreground HII region; the enhanced ionization flux enhances the formation of ions, but the resultant large fractional ionization leads to a fast and efficient removal of molecular ions by dissociative recombination with electrons.
The observed velocity profiles of OH$^+$ and H$_2$O$^+$ in Orion KL support the above conclusion. As Fig. \[fig1\] shows, the OH$^+$ and H$_2$O$^+$ absorption tracks the HF absorption to velocities of about $-45$ km s$^{-1}$. This absorption also seems to follow closely, to about $-80$ km s$^{-1}$, the [*para*]{}-H$_{2}^{18}$O absorption in the outflow, suggesting that like HF and [*para*]{}-H$_{2}^{18}$O, OH$^+$ and H$_2$O$^+$ probably exist mainly in the low velocity outflow (Phillips et al. 2010). In fact, the molecular outflow accounts for over half of the observed column density of OH$^+$ and H$_2$O$^+$.
The conditions required to explain our observations may be more extreme than one might suppose. First, the molecular ions probably reside in gas of lower density ($n \le 10^5$ cm$^{-3}$) than that necessary to thermally excite the observed transitions—these have high spontaneous emission rates ($ > 10^{-2}$ s$^{-1}$; Table 1), and hence large critical densities ($10^{7}-10^{9}$ cm$^{-3}$) . This is supported by the observation that OH$^+$ and H$_2$O$^+$ are seen [*only*]{} in absorption. Second, the temperatures in the outflow gas are probably high.
We consider two scenarios in which the ions may be formed in the low velocity outflow. In the first, a large radiation flux impinges directly on the Orion KL outflow, which contains large water abundances (Melnick et al. 2010). The far UV flux that illuminates this gas can have values approaching $4\times 10^4$ times the average interstellar radiation field (Walmsley et al. 2000; Young Owl et al. 2000). In addition, the central region of the Orion Nebula has numerous sources of energetic X-ray photons (Getman et al. 2005; Preibisch et al. 2005), which can contribute to the surface ionization of this photon-dominated region (PDR). We estimate that at A$_V=1$ into the PDR, the ionization rate $\zeta_{X} \sim 3 \times 10^{-15}$ s$^{-1}$. [^2] Under these conditions, water can undergo photoionization to form H$_2$O$^+$ directly, enhancing the abundance of this species.
In the second scenario, the outflow penetrates the extended foreground HII region. The abundant H$^+$ can now undergo charge exchange with H$_2$O to yield H$_2$O$^+$. In either scenario, the high electron density probably results in a net reduction in the abundances of molecular ions, consistent with the observations: low column densities of OH$^+$ and H$_2$O$^+$, and the upper limit for H$_3$O$^+$.
We have attempted to model the first scenario using the Meudon PDR code (Le Petit et al. 2006). However, the model suffers difficulties while reproducing the observed column densities of the three ions. First, it requires a relatively low gas density ($n \sim 10^3$ cm$^{-3}$) in regions where OH$^+$ and H$_2$O$^+$ are produced, as larger assumed densities yield too much H$_3$O$^+$. Second, it requires a very large ionization rate ($\zeta > 1- 2\times10^{-14}$ s$^{-1}$) to maintain a ratio of atomic to molecular hydrogen near unity; otherwise, too much H$_{3}$O$^{+}$ is once again produced. The two parameters are nearly an order of magnitude different from others inferred from previous observations: $n \ge 10^4$ cm$^{-3}$ and $\zeta < 10^{-14}$ s$^{-1}$ (Genzel & Stutzki 1989; Lerate et al. 2008; Muench et al. 2008 and references therein). Nevertheless, a recent study on molecular hydrogen rotational excitation in the Orion bar infers a cosmic ray ionization rate of $7 \times 10^{-14}$ s$^{-1}$ (Shaw et al. 2009). The same study also invokes warm gas temperatures of 400-700 K; the lower value is contained in our model for the edge of the PDR. A critical evaluation of our model awaits further work and a thorough exploration of the parameter space, and will be presented in a future paper.
We are unable to confirm previous tentative detections of H$_3$O$^+$ toward Orion KL (Hollis et al. 1986; Wootten et al. 1986; Wootten et al. 1991; Phillips et al. 1992; Timmermann et al. 1996; Lerate et al. 2006). Of these, Phillips et al. (1992) present the best evidence: 3 emission lines at 307, 364, and 396 GHz, lying 45, 85, and 105 cm$^{-1}$ above ground; but they do not rule out the possibility of blends with other lines. The lines we observed are at lower energies (see Table 1), and are expected to be as strong or stronger than those observed by Phillips et al. (1992). The upper limits derived here for [*ortho*]{} and [*para*]{}-H$_3$O$^+$ are more than an order of magnitude lower than the column densities reported by Phillips et al. (1992). Timmermann et al. (1996) reported detection of the $4_{3}^{-}-3_{3}^{+}$ line near $70~\mu\rm{m}$ with the Kuiper Airborne Observatory, but the velocity of the line differs by more than $-60$ km s$^{-1}$ from predicted values. Lerate et al. (2006) detected the $2_{1}^{-}-1_{1}^{+}$, $2_{0}^{-}-1_{0}^{+}$, and $1_{1}^{-}-1_{1}^{+}$ lines with ISO: the first, near 2.98 THz is $80$ km s$^{-1}$ higher than the systemic velocity of 9 km s$^{-1}$; the second, near 2.97 THz, is 1 km s$^{-1}$ higher than the frequencies predicted by Yu et al. (2009); and the third, at 1655835 MHz, covered by our observations, is obscured by a strong $2_{12}-1_{10}$ [*ortho*]{}-H$_2^{18}$O line at 1655831 MHz.
Conclusions
===========
Our observations toward Orion KL have found OH$^+$ and H$_2$O$^+$ aborption at the quiescent 9 km s$^{-1}$ component and extended absorption in the low velocity molecular outflow associated with this source. This is, to our knowledge, the first detection of these ions toward a source with a large fraction of molecular gas. Given the complex and inhomogeneous nature of Orion KL, however, there are probably regions where the densities are sufficiently low and the excitation conditions optimal for these reactive ions to exist at detectable levels. Another possibility is that depletion of some of the gas-phase species onto the grains can result in lower abundances of water, leading to small column densities of OH$^{+}$ and H$_{2}$O$^{+}$. A surprising observation—and one remarkably different from that toward W31C—is the non-detection of H$_3$O$^+$. In our model of the outflow, we attribute this mainly to a very high ionization rate, which produces an almost equal abundance of atomic and molecular hydrogen at the assumed density.
HIFI has been designed and built by a consortium of institutes and university departments from across Europe, Canada and the United States under the leadership of SRON Netherlands Institute for Space Research, Groningen, The Netherlands and with major contributions from Germany, France and the US. Consortium members are: Canada: CSA, U.Waterloo; France: CESR, LAB, LERMA, IRAM; Germany: KOSMA, MPIfR, MPS; Ireland, NUI Maynooth; Italy: ASI, IFSI-INAF, Osservatorio Astrofisico di Arcetri- INAF; Netherlands: SRON, TUD; Poland: CAMK, CBK; Spain: Observatorio Astron—mico Nacional (IGN), Centro de Astrobiolog’a (CSIC-INTA). Sweden: Chalmers University of Technology - MC2, RSS & GARD; Onsala Space Observatory; Swedish National Space Board, Stockholm University - Stockholm Observatory; Switzerland: ETH Zurich, FHNW; USA: Caltech, JPL, NHSC. Support for this work was provided by NASA through an award issued by JPL/Caltech. A part of the work described in this paper was done at the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Copyright 2010$\copyright$ California Institute of Technology. All rights reserved.
Blake, G. A., Sutton, E. C., Mason, C. R., & Phillips, T. G. 1987, , 315, 621
Botschwina, P, Rosmus, P., & Reinsch, E.-A. 1985, Chem. Phys. Lett., 102, 299
Comito, C., & Schilke, P. 2002, , 395, 357
Feigelson, E. D., Getman, K. V., Townsley, L., et al. 2005, , 160, 379
Genzel, R., & Stutzki, J. 1989, , 27, 41
Gerin, M., De Luca, M., Black, J. H., et al. 2010, A&A, arXiv:1005.5653
Getman, K. V., Feigelson, E. D., Grosso, M., et al. 2005, , 160, 353
Herzberg, G., & Lew, H. 1974, A&A, 31, 123
Lerate, M.R., Barlow, M.J., Swinyard, B.M., et al. 2006, MNRAS, 370, 597
Lerate, M.R., Yates, J., Viti, S., et al. 2008, MNRAS, 387, L1660
Le Petit, F., Nehm[é]{}, C., Le Bourtelot, J., & Roueff, E. 2006, , 164, 506
Maloney, P. R., Hollenbach, D. J., & Tielens, A. G. G. M. 1996, , 466, 561
Melnick, G. J., Tolls, V., Neufeld, D. A., et al. A&A, in press
Menten, K. M., Reid, M. J., Forbrich, J., & Brunthaler, A. 2007, A&A, 474, 515
Muench, A., Getman, K., Hillenbrand, L., & Preibisch, T. 2008 in ASP Monograph Publications 4, Handbook of Star Forming Regions, Vol. I: The Northern Sky ed. B Reipurth (San Francisco, CA: ASP), 483
M[ü]{}ller, H. S. P., Schl[ö]{}der, F., Stutzki, J., & Winnewisser, G. 2005, J. Mol. Struct., 742, 215
M[ü]{}rtz, P., Zink, L. R., Evenson, K. M., & Brown, J. M. 1998, , 109, 9744
Neufeld, D. A., Goicoechea, J. R., Sonnentrucker, P., et al. 2010, arXiv:1007.0987
O’Dell, C. R., Muench, A., Smith, N., & Zapata, L. 2008 in ASP Monograph Publications 4, Handbook of Star Forming Regions, Vol. I: The Northern Sky ed. B Reipurth (San Francisco, CA: ASP), 544
Ossenkopf, V., M[ü]{}ller, H. S. P., Lis, D. C., et al. 2010, A&A, 518, L111
Ott, S. 2010, in ASP Conference Series, Astronomical Data Analysis Software and Systems XIX, Y. Mizumoto, K.-I. Morita, and M. Ohishi, eds., in press
Phillips, T. G., van Dishoeck, E. F., Keene, J. 1992 , 399, 533
Phillips, T. G., Bergin, E. A., Lis, D. C., et al. 2010, A&A, 518, L109
Preibisch, T., McCaughrean, M. J., Grosso, M., et al. 2005, , 160, 582
Schilke, P., Comito, C., M[ü]{}ller, H. S. P., et al. 2010, arXiv:1007.0670
Shaw, G., Ferland, G. J., Henney, W. J., et al. 2009 , 701, 677
Smith, W. H., Schempp, V. W., & Federman, S. R. 1984, , 277, 196
Strahan, S. E., Mueller, R. P., & Saykally, R. J. 1986, , 85, 1252
Timmermann, R., Nikola, T., & Poglitsch, A. 1996, , 463, L109
Walmsley, C. M., Natta, A., Oliva, E. & Testi, L. 2000, A&A, 364, 301
Wehinger, P. A., Wyckoff, S., Herbig, G. H., Herzberg, G & Lew, H. 1974 , 190, L43
Werner, H.-J, Rosmus, P., & Reinsch, E.-A. , 1983, 79, 905
Wootten, A., Boulanger, F., Bogey, M. et al. 1986 A&A, 166, L15
Wootten, A., Mangum, J. G., Turner, B. E. et al. 1991, , 380, L79
Wu, S. Chen, Y. Yang, X., Guo, Y., Liu, Y., Li, Y., Buenker, R. J., & Jensen, P. 2004, J. Mol. Spec. 225, 96
Wyrowski, F., Menten, K. M., Guesten, R., & Belloche, A. 2010, arXiv:1004.2627
Young Owl, R. C., Meixner, M. M., Wolfire, M., Tielens, A. G. G. M. & Tauber, J. 2000, , 540, 886
Yu, S., Drouin, B. J., Pearson, J. C. & Pickett, H. M. 2009, , 180, 119
[^1]: [*Herschel*]{} is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA.
[^2]: The surface brightness of the central region (dominated by $\theta ^1C$) is estimated to be $3\times10^{34}$ ergs s$^{-1}$pc$^{-2}$ (Feigelson et al. 2005). On the assumption that the molecular cloud lies 0.1 pc from this cluster, the expressions of Maloney et al. (1996) yield an X-ray ionization rate of about 2.8 $\times$ 10$^{-16}$N$_{22}^{-1}$ (where N$_{22}$ is the hydrogen column density in units of 10$^{22}$ cm$^{-2}$). Thus at A$_V = 1$, $ \zeta_{X} \sim 3 \times 10^{-15}$ s$^{-1}$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The answer to the question in the title is: in search of new physics beyond the Standard Model, for which there are many motivations, including the likely instability of the electroweak vacuum, dark matter, the origin of matter, the masses of neutrinos, the naturalness of the hierarchy of mass scales, cosmological inflation and the search for quantum gravity. So far, however, there are no clear indications about the theoretical solutions to these problems, nor the experimental strategies to resolve them. It makes sense now to prepare various projects for possible future accelerators, so as to be ready for decisions when the physics outlook becomes clearer. Paraphrasing George Harrison, “ If you don’t [*yet*]{} know where you’re going, any road [*may*]{} take you there."\
\
[*Contribution to the 2017 Hong Kong UST IAS Programme and Conference on High-Energy Physics.*]{}\
\
KCL-PH-TH-2017-18, CERN-TH-2017-080
address: |
Theoretical Particle Physics and Cosmology Group, Physics Department,\
KingÕs College London, London WC2R 2LS, UK;\
Theoretical Physics Department, CERN, CH-1211 Geneva 23, Switzerland\
John.Ellis@cern.ch
author:
- JOHN ELLIS
title: '**WHERE IS PARTICLE PHYSICS GOING?**'
---
Introduction
============
The bedrock upon which our search for new physics beyond the Standard Model (SM) is founded is our ability to make precise predictions within the Standard Model, notably for the LHC experiments. The predictions of many hard higher-order perturbative QCD calculations have been confirmed, as seen in Fig. \[fig:heaven\], providing confidence in predictions for the production of the Higgs boson [@Mistlberger], and for the backgrounds to many searches for new physics.
![\[fig:heaven\] *Many SM processes have been measured at the LHC, and have cross sections that are generally in excellent agreement with QCD calculations [@ATLASSM].* ](heaven){height="6cm"}
The Flavour Sector
==================
Many measurements in the flavour sector are also consistent with the predictions of the Cabibbo-Kobayashi-Maskawa (CKM) model [@CKMfitter; @UTfit], e.g., there are many consistent measurements of the unitarity triangle, as seen in the left panel of Fig. \[fig:CKM\]. Historically, the angle $\gamma$ has been the least constrained experimentally, but the LHCb Collaboration has recently published a combined measurement [@LHCbgamma] that dominates the world average and is consistent with the other unitarity triangle measurements.
![\[fig:CKM\] *Left panel: Compilation of experimental constraints on the CKM unitarity triangle [@CKMfitter]. Compilation of constraints on possible new physics contributions to operator coefficients [@Wolfgang].*](CKMfitter "fig:"){height="5.5cm"} ![\[fig:CKM\] *Left panel: Compilation of experimental constraints on the CKM unitarity triangle [@CKMfitter]. Compilation of constraints on possible new physics contributions to operator coefficients [@Wolfgang].*](flavouranomaly "fig:"){height="5.1cm"}\
That said, there are several anomalies in the flavour sector of varying significance. For example, there are strengthening indications of violations of $e/\mu$ lepton universality in $B \to K e^+ e^-$ and $B \to K \mu^+ \mu^-$ decays, [@Kll] and of $\tau/(\ell = e$ or $\mu)$ universality in $B \to D^{(*)} \tau \nu$ decays [@Dtaunu] - to which my attitude is ‘wait and see’, as lepton non-universality has held up very well so far. Much attention has been attracted to the $P_5^\prime$ angular distribution in $B \to K^* \mu^+ \mu^-$ decay [@P5prime], which may be accompanied by an anomaly in the $q^2$ distribution in $B \to \phi \mu^+ \mu^-$ decay, leading to the constraints on possible new physics contributions to operator coefficients shown in the right panel of Fig. \[fig:CKM\] [@Wolfgang]. These both appear at $q^2 \lesssim 5$ GeV$^2$, and I do not know how seriously to take them, in view of my lack of understanding of the non-perturbative QCD corrections in this region. My ignorance also makes it difficult for me to judge the significance of the apparent discrepancy between theory [@CTS; @Buras] and experiment for $\epsilon^\prime/\epsilon$. Finally, a new kid on the flavour block has been the interesting search for $H \to \mu \tau$ decay [@Htaumu] discussed below, though this may be reverting towards the SM with the latest Run 2 results [@CMSHtaumu2].
Higgs Physics
=============
The Higgs Mass
--------------
The most fundamental Higgs measurement is that of its mass. The combined LHC Run 1 results of ATLAS and CMS based on $H$ decays into $\gamma \gamma$ and $Z Z^* \to 2 \ell^+ 2 \ell^-$ yielded [@ATLAS+CMS] $$m_H \; = \; 125.09 \pm 0.21 ({\rm stat.}) \pm 0.11 ({\rm syst.}) \, ,
\label{mHRun1}$$ and the preliminary CMS result from Run 2 is consistent with this, with slightly smaller errors [@CMSRun2]: $$m_H \; = \; 125.26 \pm 0.20 ({\rm stat.}) \pm 0.08 ({\rm syst.}) \, ,
\label{mHRun2}$$ It is noteworthy that statistical uncertainties dominate, and we can look forward to substantial reductions in the future, determining $m_H$ at the [*per mille*]{}. Accurate knowledge of the Higgs mass is important for precision tests of Standard Model (and other) predictions and, as discussed later, is crucial for understanding the (in/meta)stability of the electroweak vacuum.
Higgs Couplings
---------------
The couplings of the Higgs boson to Standard Model particles are completely specified and, consequently, there are definite predictions for its production processes and decay branching ratios [@LHCHXSWG]. Concretely, one expects gluon-gluon fusion to dominate over vector-boson fusion, production in association with a vector boson and in association with a $t {\bar t}$ pair. The dominant $H$ decay mode is predicted to be into $b {\bar b}$, with much smaller branching ratios for $\gamma \gamma$ and $Z Z^* \to 2 \ell^+ 2 \ell^-$.
Much progress was made in Run 1 probing these predictions [@ATLAS+CMS2], but much remains to be done. Higgs decays to $\gamma \gamma, ZZ^*, WW^*$ and $\tau^+ \tau^-$ have been measured in gluon-gluon fusion, and there is solid evidence for vector-boson fusion, but the associated production mechanisms have yet to be confirmed. Moreover, there is no confirmation yet of the expected dominant $H \to b {\bar b}$ decay mode: LHC evidence is at the level of 2.6 $\sigma$ [@Tevatronbbbar], and the Tevatron experiments have reported evidence at the 2.8-$\sigma$ level. There is indirect evidence for the expected $H t {\bar t}$ vertex via the measurements of gluon-gluon fusion and $H \to \gamma \gamma$ decay, but no significant evidence via associated $H t {\bar t}$ or single $H t ({\bar t})$ production. Also on the agenda is the search for $H \to \mu^+ \mu^-$, which is predicted in the SM to appear at a level close to the current experimental sensitivity.
Fig. \[fig:Mepsilon\] is one way of displaying the available information on Higgs couplings [@EY; @ATLAS+CMS2]. It is a characteristic prediction of the SM that the couplings to other particles should be related to their masses, $\propto m_f$ for fermions and $\propto m_V^2$ for massive vector bosons. The black solid line is a fit where $m \to m^{(1+ \epsilon)}$ in the couplings: we see that the combined ATLAS and CMS data are highly consistent with the SM expectation that $\epsilon = 0$, shown as the blue dashed line.
![*A fit by the ATLAS and CMS Collaborations to a parametrization of the mass-dependence of the Higgs couplings: $m \to m^{(1+ \epsilon)}$ [@ATLAS+CMS2]. The Standard Model predictions are connected by a dotted line, the red line is the best fit, and the green and yellow bands represent the 68 and 95% CL fit ranges.*[]{data-label="fig:Mepsilon"}](Mepsilon){width="7cm"}
The couplings in Fig. \[fig:Mepsilon\] are all flavour-diagonal. The SM predicts that flavour-violating Higgs couplings should be very small, but measurements of flavour-violating processes at low energies would allow [*either*]{} $H \to \mu \tau$ [*or*]{} $H \to e \mu$ with branching ratio $\lesssim 10$%, whereas the branching ratio for $H \to e \mu$ must be $\lesssim 10^{-5}$ [@BEI]. The was some excitement after Run 1 when the combined CMS and ATLAS data indicated a possible 2-$\sigma$ excess [@Htaumu]. This has not reappeared in early Run 2 data [@CMSHtaumu2], but remains an open question.
Elementary Higgs Boson, or Composite?
=====================================
There has been a long-running theoretical debate whether the Higgs boson could be as elementary as the other particles in the SM, or whether it might be composite. The elementary option encounters quadratically-divergent loop corrections to the mass of the Higgs boson, which are frequently (usually?) postulated to be cancelled by supersymmetric particles [@susy] appearing at the TeV scale [@hierarchy] - which have not yet been seen.
On the other hand, the composite option has been favoured by many with memories of the (composite) Cooper pairs underlying superconductivity, and the (composite) pions associated with quark-antiquark condensation in QCD [@techni]. A composite Higgs would require a novel set of strong interactions, and early models tended to have a scalar particle much heavier than the Higgs that has been discovered, and to be in tension with the precision electroweak data. These difficulties can be circumvented by postulating that the Higgs is a pion-like pseudo-Nambu-Goldstone boson of a partially-broken larger symmetry that is restored at some higher energy scale [@PNGBH].
A phenomenological framework that is convenient for characterizing the experimental constraints on such as possibility is provided by the following form of effective Lagrangian that preserves a custodial SU(2)$_V$ symmetry that guarantees $\rho \equiv m_W/m_Z \cos \theta_W = 1$ up to quantum corrections [@NLEL]: $$\begin{aligned}
{\cal L} & = & \frac{v^2}{4} {\rm Tr} D_\mu \Sigma D^\mu \Sigma \left(1 + 2 \kappa_V \frac{H}{v} + b \frac{H^2}{v^2} + \dots \right)
- m_i \bar{\psi}^i_L \Sigma \left( \kappa_F \frac{H}{v} + \dots \right) + {\rm h.c.} \nonumber \\
&& + \frac{1}{2} \partial_\mu H \partial^\mu H + \frac{1}{2} H^2 + d_3 \frac{1}{6} \left( \frac{3 m_H^2}{v} \right) H^3
+ d_4 \frac{1}{24} \left( \frac{3 m_H^2}{v} \right) H^4 + ... \, ,
\label{nonlinearH}\end{aligned}$$ where $H$ is the field of the physical Higgs boson and the massive vector bosons are parametrized by the $2 \times 2$ matrix $\Sigma = \exp( i \frac{\sigma_a \pi_a}{v} )$. The terms in (\[nonlinearH\]) are normalized so that the coefficients $\kappa_V, b, \kappa_F, d_i = 1$ in the SM. The question for experiment is whether any of these coefficients exhibit a deviation that might be a signature of some composite Higgs model.
As seen in the left panel of Fig. \[fig:compo\], measurements of Higgs properties (yellow and orange ellipses) and precision electroweak data (blue ellipses) play complementary roles in constraining the $H$ couplings to vector bosons $\kappa_V$ and fermions $\kappa_F$ in (\[nonlinearH\]) [@Gfitter]. These constraints can be translated into lower limits on the possible compositeness scale in various models, as seen in the right panel of Fig. \[fig:compo\] [@SS].
![*Left panel: A fit of the LHC $H$ couplings to vector bosons and fermions $(\kappa_V, \kappa_F)$ using $H$ measurements (orange and yellow ellipses), and in combination with precision electroweak data (blue ellipses) [@Gfitter]. Right panel: Constraints from LHC Run 1 and early Run 2 data on the compositeness scale in various models [@SS].*[]{data-label="fig:compo"}](Gfitterkappas "fig:"){width="5.5cm"} ![*Left panel: A fit of the LHC $H$ couplings to vector bosons and fermions $(\kappa_V, \kappa_F)$ using $H$ measurements (orange and yellow ellipses), and in combination with precision electroweak data (blue ellipses) [@Gfitter]. Right panel: Constraints from LHC Run 1 and early Run 2 data on the compositeness scale in various models [@SS].*[]{data-label="fig:compo"}](VS "fig:"){width="5.75cm"}
Stability of the Electroweak Vacuum
===================================
If the Higgs is indeed elementary, the measurements (\[mHRun1\], \[mHRun2\]) of $m_H$, combined with those of $m_t$, raise important questions about the stability and history of the electroweak vacuum, suggesting the necessity of new physics beyond the SM [@Medellin]. The issue is that the Higgs quartic self-coupling $\lambda$ is renormalized not only by itself, which tends to increase it as the energy/mass scale increases, but also by the Higgs coupling to the top quark, which tends to drive it to smaller (even negative) values at higher scales $Q$, as seen in the left panel of Fig. \[fig:down\]GeV [@DDEEGIS]. At leading order: $$\lambda (Q) \; \simeq \; \lambda (v) - \frac{3 m_t^4}{2 \pi v^4} \log \left( \frac{Q}{v} \right) \, ,
\label{down}$$ The right panel of Fig. \[fig:down\] displays the results of one calculation of the regions of the $(m_H, m_t)$ plane where the electroweak vacuum is stable, metastable or unstable, and yields the following estimate of the ‘tipping point’ $\Lambda_I$ where $\lambda$ goes negative [@BDGGSSS]: $$\begin{aligned}
\log_{10} \left( \frac{\Lambda_I}{\rm GeV}\right) & = & 9.4 + 0.7 \left(\frac{m_H}{\rm GeV} -125.15 \right) \nonumber \\
& - & 1.0 \left( \frac{m_t}{\rm GeV} - 173.34 \right) + 0.3 \left( \frac{\alpha_s(m_Z) - 0.1184}{0.0007} \right) \, .
\label{LambdaI}\end{aligned}$$ The dominant uncertainty in the calculation of $\Lambda_I$ is due to that in $m_t$, followed by that in $\alpha_s(m_Z)$ (which enters in higher order in the calculation), the uncertainty due to the measurement of $m_H$ being relatively small. The final result is an estimate $$\log_{10} \left( \frac{\Lambda_I}{\rm GeV} \right) \; = 9.4 \pm 1.1 \, ,
\label{LambdaIvalue}$$ indicating that we are (probably) doomed, unless some new physics intervenes.
![*Left panel: Top quark loops renormalize the Higgs self-coupling $\lambda$ negatively, suggesting that it takes negative values at field values $\gtrsim 10^{9}$ GeV[@DDEEGIS], leading to instability of the Higgs potential in the SM. Right panel: Measurements of $m_t$ and $m_H$ indicate that the SM vacuum is probably metastable, although there are important uncertainties in $m_t$ and $\alpha_s$ [@BDGGSSS].*[]{data-label="fig:down"}](GoingNegative.png "fig:"){width="6.1cm"} ![*Left panel: Top quark loops renormalize the Higgs self-coupling $\lambda$ negatively, suggesting that it takes negative values at field values $\gtrsim 10^{9}$ GeV[@DDEEGIS], leading to instability of the Higgs potential in the SM. Right panel: Measurements of $m_t$ and $m_H$ indicate that the SM vacuum is probably metastable, although there are important uncertainties in $m_t$ and $\alpha_s$ [@BDGGSSS].*[]{data-label="fig:down"}](Buttazzo.png "fig:"){width="6cm"}
Some people discount this ‘problem’ on the grounds that the prospective lifetime of the vacuum is much longer than its age. However, there is another issue, namely that fluctuations in the Higgs field in the very early Universe would have been much larger than now, and would probably have driven almost everywhere in the Universe into an anti-De Sitter phase from which there would have been no escape [@aDS]. One could postulate that our piece of the Universe happened to be extraordinarily lucky and avoid this fate, but it seems more plausible that some new physics intervenes before the instability scale $\Lambda_I$. Possible such remedies include higher-dimensional operators in the SM effective field theory (see the next Section), a non-minimal Higgs coupling to gravity, or a threshold for new physics such as supersymmetry [@ER] (see later).
The SM Effective Field Theory
=============================
An alternative way of analyzing the Higgs and other data is to assume that all the known particles (including the Higgs boson) are SM-like, and look for the effects of physics beyond the SM via an effective field theory (the SMEFT) containing higher-dimensional SU(2)$\times$U(1)-invariant operators constructed out of SM fields, e.g., of dimension 6 [@SMEFT]: $${\cal L}_{eff} \; = \; \sum_n \frac{c_n}{\Lambda^2} {\cal O}_n \, ,
\label{dim6}$$ where the characteristic scale of new physics is described by $\Lambda$, with the $c_n$ being unknown dimensionless coefficients. Data on Higgs properties, precision electroweak data, triple-gauge couplings (TGCs), etc., can all be combined to constrain the SMEFT operator coefficients in a unified and consistent way. Table \[tab:dim6\] shows which observables currently provide the greatest sensitivities to some of these operators [@ESY].
\[tab:dim6\]
The left panel of Fig. \[fig:SMEFT\] shows how the coefficients of the SMEFT operators in Table \[tab:dim6\] were constrained by Run 1 Higgs data including kinematical variables (blue bar) and by Run 1 measurements of TGCs (red bar) [@ESY]. The green bar gives the resulting ranges when each operator is switched on individually, and the black bar is for a global fit marginalizing over all the listed operators. The right panel of Fig. \[fig:SMEFT\] manifests the complementarity between the Higgs and LEP-2 TGC data for constraining the anomalous couplings $\delta g_{1Z}$ and $\delta g_\gamma$ [@Falkowski]. Because of its power to constrain new physics appearing in many observables in a consistent way, the SMEFT is the preferred framework for assessing the sensitivities of future analyses of precision LHC measurements to physics beyond the SM, whose motivations are discussed in the next Section.
![*Left panel: The 95% CL ranges for fits to individual SMEFT operator coefficients (green bars), and the marginalised 95% CL ranges for global fits combining data on the LHC $H$ signal strength data with the kinematic distributions for associated $H+V$ production (blue bars), or with the LHC TGC data (red bars), and combining all the data (black bars) [@ESY]. Right panel: The 68 and 95% CL ranges allowed by a fit to the anomalous TGCs $(\delta g_{1,z}, \delta \kappa_\gamma)$ using LEP-2 TGC data (orange and yellow), LHC Higgs data (green) and their combination (blue) [@Falkowski].*[]{data-label="fig:SMEFT"}](SMEFTL "fig:"){width="6.5cm"} ![*Left panel: The 95% CL ranges for fits to individual SMEFT operator coefficients (green bars), and the marginalised 95% CL ranges for global fits combining data on the LHC $H$ signal strength data with the kinematic distributions for associated $H+V$ production (blue bars), or with the LHC TGC data (red bars), and combining all the data (black bars) [@ESY]. Right panel: The 68 and 95% CL ranges allowed by a fit to the anomalous TGCs $(\delta g_{1,z}, \delta \kappa_\gamma)$ using LEP-2 TGC data (orange and yellow), LHC Higgs data (green) and their combination (blue) [@Falkowski].*[]{data-label="fig:SMEFT"}](Falkowskietal "fig:"){width="5.5cm"}
The Standard Model is not Enough [@Bond]
========================================
There are many reasons to anticipate the existence of physics beyond the SM, of which I list just 7 here. 1) The prospective instability of the electroweak vacuum discussed earlier. 2) The astrophysical and cosmological necessity for dark matter. 3) The origin of matter itself, i.e., the cosmological baryon asymmetry. 4) The masses of neutrinos. 5) The naturalness of the hierarchy of mass scales in physics. 6) A mechanism (or replacement) for cosmological inflation to explain the great size and age of the Universe. 7) A quantum theory of gravity.
The good news is that LHC experiments are tackling most of these issues during Run 2. The bad news is that there is no consensus among theorists how to resolve them. Until recently, supersymmetry found the most theoretical favour, but the negative results from early Run 2 supersymmetry searches have caused some to waver. Not me, however - I still think that it is the most comprehensive and promising framework for new physics beyond the SM. In the words of the famous World War 1 cartoon [@cartoon] “If you knows of a better ’ole, go to it.” I do not, so I will stay in the supersymmetric ’ole.
Supersymmetry
=============
Indeed, I would even argue that Run 2 of the LHC has provided us with 3 new motivations for supersymmetry. i) It stabilizes the electroweak vacuum [@ER]. ii) It made a successful prediction for the Higgs mass, namely that it should weigh $\lesssim 130$ GeV in simple models [@susymH]. iii) It predicted correctly that the Higgs couplings measured at the LHC should be within a few % of their SM values [@EHOW]. These new motivations are additional to the classic ones from the naturalness of the mass hierarchy [@hierarchy], the availability of a natural dark matter candidate [@EHNOS], the welcome help of supersymmetry in making grand unification possible [@susyGUTs], and its apparent necessity in string theory, which I regard as the only serious candidate for a quantum theory of gravity.
At this point, I must ’fess up to two pieces of bad news. One is that theorists have also not reached any consensus on the most promising supersymmetric model, largely because there is no favoured scenario for supersymmetry breaking. Alternatives range from models in which this is assumed to be universal at some GUT scale (such as the CMSSM) to models in which all the soft supersymmetry-breaking parameters are treated entirely phenomenologically as unknown parameters at the electroweak scale (the pMSSM). The other piece of bad news is that the LHC experiments have found not even a hint of supersymmetry, despite many searches making different assumptions about the supersymmetric spectrum [^1].
In the following, the negative results of the searches are combined with other measurements to constrain the parameter spaces of a couple of representative supersymmetric models.
Probing a Supersymmetric SU(5) GUT
----------------------------------
The first model we study here is a supersymmetric SU(5) GUT in which the soft supersymmetry-breaking gaugino masses are assumed to be universal at the GUT scale, whereas the soft supersymmetry-breaking scalar masses are generation-independent but allowed to be different for the spartners of fermions in the $\mathbf{\bar{5}}$ and $\mathbf{10}$ representations [@MCSU5].
Fig. \[fig:glsq\] displays the regions of the $(m_{\tilde g}, m_{\tilde \chi^0_1})$ plane (left panel) and the $(m_{\tilde u_R}, m_{\tilde \chi^0_1})$ plane (right panel) that are allowed in a global fit in this supersymmetric SU(5) GUT at the 95% CL (blue contours) and favoured at the 68% CL (red contours), as well as the best-fit point (green stars). The black lines are the nominal 95% CL limits set by LHC searches, assuming simplified decay patterns with 100% branching ratios, and the coloured shadings represent the actual dominant decays found in different regions of parameter space. We see in the left panel that gluino masses $\gtrsim 1900$ GeV are indicated, with a best-fit value of $\simeq 2400$ GeV, whereas the ${\tilde u_R}$ mass may be $\sim 400$ GeV lighter. One curiosity is a small strip in the right panel where $m_{\tilde u_R} - m_{\tilde \chi^0_1}$ is small and $m_{\tilde u_R} \lesssim 650$ GeV. In this strip the dark matter (DM) density is brought into the range allowed by astrophysics and cosmology by squark-neutralino coannihilation, and this compressed-spectrum region is on the verge of exclusion by LHC searches.
\
\
The best-fit spectrum in this SU(5) GUT model is shown in Fig. \[fig:bestfit\]. We see that all the squarks have masses below $\sim 2200$ GeV at the best-fit point, where they would be within the range of future LHC runs. This analysis included the results from the first $\sim 13$/fb of LHC data at 13 TeV, and Fig. \[fig:compare\] compares the profiled $\chi^2$ likelihood functions for $m_{\tilde g}$ (left panel) and $m_{\tilde u_R}$ (right panel) found in this analysis (solid blue lines) with those found in an analysis restricted to 8 TeV data (dashed blue lines) [^2]. We see that, whilst the 13 TeV have had a significant impact, they have not yet been a game-changer. There is still plenty of room for discovering supersymmetry in future LHC runs in this model, though there are no guarantees!
One of the interesting experimental possibilities in this and related models is that the next-to-lightest supersymmetric particle (NLSP) might be the lighter stau slepton, with a mass that could be so close to that of the ${\tilde \chi^0_1}$ that it might have a long enough lifetime to decay at a separated vertex, or even escape from the detector as a massive charged non-relativistic particle, as illustrated in Fig. \[fig:staudecay\] [@MCSU5].
Probing the Minimal Anomaly-Mediated Supersymmetry-Breaking Model
-----------------------------------------------------------------
Another model we have studied recently is the minimal anomaly-mediated supersymmetry-breaking (mAMSB) model [@MCmAMSB]. In this case, the supersymmetric spectrum is relatively heavy. If one assumes that the lightest supersymmetric particle (LSP) is a wino that provides all the cosmological DM, it must weigh about 3 TeV, leading to a relatively heavy spectrum as seen in the left panel of Fig. \[fig:mAMSB\], though the spectrum could be lighter if the LSP is a Higgsino, or if it provides only a fraction of the dark matter, as seen in the right panel of Fig. \[fig:mAMSB\]. We also see that the soft supersymmetry-breaking scalar mass $m_0$ in the mAMSB model must be quite large if the LSP provides all the dark matter, $m_0 \gtrsim 4$ TeV, though it could be smaller if there is some other contribution to the dark matter.
Fig. \[fig:reaches\] displays the reaches of the LHC and a 100-TeV $pp$ collider (FCC-hh) in the $(m_{\tilde g}, m_{\tilde \chi^0_1})$ plane (left panel) and the $(m_{\tilde q_R}, m_{\tilde \chi^0_1})$ plane (right panel) in the mAMSB [@MCmAMSB]. We see that most of the allowed region of the mAMSB parameter space lies beyond the reach of the LHC, though it may be within reach of FCC-hh [@FCC-hh-BSM].
Direct Dark Matter Searches
===========================
Besides missing-energy searches at the LHC, the best prospects for exploring supersymmetry may be in the direct search for dark matter via scattering on nuclei in deep-underground laboratories [@DDMAspen]. Possible ranges of the LSP mass and the spin-independent cross section for LSP scattering on a proton target, $\sigma_p^{\rm SI}$, in the supersymmetric SU(5) and mAMSB models discussed above are shown in the left [@MCSU5] and right [@MCmAMSB] panels of Fig. \[fig:directDM\], respectively. In both panels the range of $\sigma_p^{\rm SI}$ excluded by the latest results from the PandaX [@PandaX] and LUX [@LUX] experiments is shaded green. The estimated sensitivities of the planned LZ [@LZ] and XENON1/nT [@XENON] experiments are also shown, as is the neutrino ‘floor’ below which neutrino-induced backgrounds dominate. As in previous plots, the ranges allowed at the 95% CL (favoured at the 68% CL) are surrounded by blue and red contours, respectively, while the coloured shadings within them correspond to different mechanisms for bringing the LSP density into the cosmological range (discussed in [@MCSU5; @MCmAMSB], and the best-fit points are marked by green stars.
We see that values of $\sigma_p^{\rm SI}$ anywhere from the present experimental limit down to below the neutrino ‘floor’ are possible in both the SU(5) and mAMSB cases. There are decent prospects for discovering direct DM scattering in the LZ and XENON1/nT experiments, but again no guarantees.
It is interesting to compare the sensitivities of LHC searches for mono-jet and other searches with those of direct searches for DM scattering, which can be done in the frameworks of simplified models for DM [@SDMM]. The results of the comparison depend, in particular, on the form of the coupling of the intermediate particle mediating the interactions between the DM and SM particles. Fig. \[fig:DMcomparison\] compares the sensitivities of LHC mono-jet and $\sigma_p^{\rm SI}$ constraints in the case of a vector-like mediator (left panel) and LHC mono-jet searches and constraints on the spin-dependent scattering cross section, $\sigma_p^{\rm SD}$, in the case of an axial-vector mediator (right panel) [@CMSSDMM]. We see that in the vector-like case the direct DM searches currently have more sensitivity except for small DM masses, whereas in the axial-vector case the LHC has greater sensitivity over a wide range of DM masses. These examples illustrate the complementarity of the LHC and direct searches in the quest for dark matter.
A Plea for Patience
===================
The LHC will continue to operate for another 15 to 20 years, with the objective of gathering two orders of magnitude more data than those analyzed so far. Thus it has many opportunities to discover new physics beyond the Standard Model, e.g., in Higgs studies and in searches for new particles beyond the Standard Model such as supersymmetry and/or dark matter. Some lovers of superymmetry may be tempted to lose faith. However, it is worth remembering that the discovery of the Higgs boson came 48 years after it was postulated, whereas the first interesting supersymmetric models in four dimensions were written down at the end of 1973 [@susy], only just over 43 years ago! Moreover, the discovery of gravitational waves came just 100 years after they were predicted. Sometimes one must be patient.
In the mean time, what are the prospects for new accelerators to follow the LHC?
Electron-Positron Colliders
===========================
Fig. \[fig:electronpositron\] shows the estimated luminosities as functions of the centre-of-mass energy for various projected $e^+ e^-$ colliders. We see that linear colliders (ILC [@ILC], CLIC [@CLIC]) could reach higher energies, but circular colliders (CEPC [@CEPC], FCC-ee [@FCC-ee]) could provide higher luminosities at low energies. This means that CLIC, in particular, might be the accelerator of choice if future LHC runs reveal some new particles with masses $\lesssim 1$ TeV, or if the emphasis will be on probing decoupled new physics via SMEFT effects that grow with the centre-of-mass energy [@ERSY-CLIC], whereas FCC-ee would be advantageous [@EY-FCC-ee] if high-precision Higgs and $Z$ measurements are to be prioritized.
The left panel of Fig. \[fig:FCC-eeCLIC\] compares the estimated sensitivities of FCC-ee and ILC measurements of Higgs and electroweak precision measurements to the coefficients of some dimension-6 operators in the SMEFT [@EY-FCC-ee]. The green bars are for fits to individual operator coefficients, and the red bars are after marginalization in global fits. We see that both FCC-ee (darker bars) and ILC (lighter bars) could reach far into the multi-TeV region. The right panel of Fig. \[fig:FCC-eeCLIC\] shows the estimated sensitivities of CLIC measurements to other combinations of dimension-6 SMEFT operators [@ERSY-CLIC], highlighting the advantages conferred by high-energy running at CLIC.
Higher-Energy Proton-Proton Colliders
=====================================
Circular colliders with circumferences approaching 100 km are being considered in China (CEPC/SppC [@CEPC]) and as a possible future CERN project (FCC-ee/hh [@FCC]). One could imagine filling the tunnel with two successive accelerators, as was done with LEP and then the LHC in CERN’s present 27-km tunnel.
Fig. \[fig:FCC-pp\] provides two illustrations of the possible physics reach of the FCC-hh project for a $pp$ collider. In the left panel we see the ways in which various Higgs production cross sections grow by almost two orders of magnitude with the centre-of-mass energy [@FCC-hh-H], offering many possibilities for high-precision measurements of Higgs production mechanisms and decay modes in collisions at 100 TeV. In particular, these might offer the opportunity to make the first accurate direct measurements of the triple-Higgs coupling. In the right panel we see the discovery reaches for squark and gluino discovery at FCC-pp [@FCC-hh-BSM]. The reaches for both these sparticles extend beyond 10 TeV and offer, e.g., the prospects for detecting the heavy spectrum of the mAMSB model shown in Fig. \[fig:reaches\].
In my opinion, the combination of high precision and large kinematic reach offered by large circular colliders is unbeatable as a vision for the future of high-energy physics, offer the twin possibilities of exploring the 10 TeV scale directly in $pp$ collisions at centre-of-mass energies up to 100 TeV and indirectly via the high-precision $e^+ e^-$ measurements mentioned in the previous Section.
Summary
=======
Despite the impressive progress already made, many things are still to be learnt about the Higgs boson, including its expected dominant $b \bar{b}$ decay modes, rare decays into lighter particles and the triple-Higgs coupling. The best tool for interpreting Higgs and other electroweak measurements is the SMEFT, and possible future $e^+ e^-$ colliders offer good prospects for higher-precision measurements beyond the sensitivities of the LHC.
Like that of Mark Twain, rumours of the death of supersymmetry are exaggerated. I still think that it is the best framework for TeV-scale physics beyond the SM at the TeV scale. Simple supersymmetric models have been coming under increasing pressure from LHC searches, but other models with heavier spectra are still quite healthy. There are good prospects for discovering supersymmetry in future LHC runs and in direct dark matter detection experiments, but no guarantees. Maybe we will have to wait for a future higher-energy $pp$ collider before discovering or abandoning supersymmetry?
In the mean time, we look forward to whatever indications the full LHC Run 2 date may provide before choosing what collider we would like to build next, but the answer to the question in the title may well be “round in circles".
Acknowledgments {#acknowledgments .unnumbered}
===============
The author’s work was supported partly by the STFC Grant ST/L000326/1. He thanks his collaborators on topics discussed here, and thanks the Institute of Advanced Study of the Hong Kong University of Science and Technology for its kind hospitality.
[0]{} For recent high points in this effort, see C. Anastasiou, C. Duhr, F. Dulat, F. Herzog and B. Mistlberger, Phys. Rev. Lett. [**114**]{} (2015) 212001 doi:10.1103/ PhysRevLett.114.212001 \[arXiv:1503.06056 \[hep-ph\]\]; J. Currie, E. W. N. Glover and J. Pires, Phys. Rev. Lett. [**118**]{} (2017) no.7, 072002 doi:10.1103/PhysRevLett.118.072002 \[arXiv:1611.01460 \[hep-ph\]\]. ATLAS Collaboration, [https://twiki.cern.ch/twiki/bin/view/AtlasPublic/]{} [StandardModelPublicResults]{}; see also CMS Collaboration, [https://twiki.cern.ch/]{} [twiki/bin/view/CMSPublic/PhysicsResultsCombined]{}.
J. Charles [*et al.*]{} \[CKMfitter Collaboration\], Phys. Rev. D [**91**]{} (2015) no.7, 073007\
doi:10.1103/PhysRevD.91.073007 \[arXiv:1501.05013 \[hep-ph\]\],\
and [http://ckmfitter.in2p3.fr]{}.
UTfit Collaboration, [http://www.utfit.org/UTfit/]{}.
R. Aaij [*et al.*]{} \[LHCb Collaboration\], JHEP [**1612**]{} (2016) 087\
doi:10.1007/JHEP12(2016)087 \[arXiv:1611.03076 \[hep-ex\]\]. For a recent review, see S. Bifani, [https://indico.in2p3.fr/event/13763/]{} [session/9/contribution/104/material/slides/0.pdf]{}.
For a recent review, see G. Wormser, [https://indico.in2p3.fr/event/13763/]{} [session/9/contribution/105/material/slides/0.pdf]{}.
S. Descotes-Genon, L. Hofer, J. Matias and J. Virto, JHEP [**1606**]{} (2016) 092 doi:10.1007/JHEP06(2016)092 \[arXiv:1510.04239 \[hep-ph\]\]. W. Altmannshofer, C. Niehoff, P. Stangl and D. M. Straub, arXiv:1703.09189 \[hep-ph\]. Z. Bai [*et al.*]{} \[RBC and UKQCD Collaborations\], Phys. Rev. Lett. [**115**]{} (2015) no.21, 212001 doi:10.1103/PhysRevLett.115.212001 \[arXiv:1505.07863 \[hep-lat\]\]. A. J. Buras, M. Gorbahn, S. JŠger and M. Jamin, JHEP [**1511**]{} (2015) 202 doi:10.1007/JHEP11(2015)202 \[arXiv:1507.06345 \[hep-ph\]\]. V. Khachatryan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**749**]{} (2015) 337 doi:10.1016/j.physletb.2015.07.053 \[arXiv:1502.07400 \[hep-ex\]\]; G. Aad [*et al.*]{} \[ATLAS Collaboration\], JHEP [**1511**]{} (2015) 211 doi:10.1007/JHEP11(2015)211 \[arXiv:1508.03372 \[hep-ex\]\]. CMS Collaboration, [https://cds.cern.ch/record/2159682/files/]{}\
[HIG-16-005-pas.pdf]{}.
G. Aad [*et al.*]{} \[ATLAS and CMS Collaborations\], Phys. Rev. Lett. [**114**]{} (2015) 191803 doi:10.1103/PhysRevLett.114.191803 \[arXiv:1503.07589 \[hep-ex\]\]. S. Oda, [https://indico.in2p3.fr/event/13763/session/0/contribution/39/]{}\
[material/slides/0.pdf]{}.
D. de Florian [*et al.*]{} \[LHC Higgs Cross Section Working Group\], arXiv:1610.07922 \[hep-ph\]. G. Aad [*et al.*]{} \[ATLAS and CMS Collaborations\], JHEP [**1608**]{} (2016) 045 doi:10.1007/JHEP08(2016)045 \[arXiv:1606.02266 \[hep-ex\]\]. T. Aaltonen [*et al.*]{} \[CDF and D0 Collaborations\], Phys. Rev. D [**88**]{} (2013) no.5, 052014 doi:10.1103/PhysRevD.88.052014 \[arXiv:1303.6346 \[hep-ex\]\]. J. Ellis and T. You, JHEP [**1306**]{} (2013) 103 doi:10.1007/JHEP06(2013)103 \[arXiv:1303.3879 \[hep-ph\]\]. G. Blankenburg, J. Ellis and G. Isidori, Phys. Lett. B [**712**]{} (2012) 386 doi:10.1016/j.physletb.2012.05.007 \[arXiv:1202.5704 \[hep-ph\]\]. J. Wess and B. Zumino, Phys. Lett. [**49B**]{} (1974) 52 doi:10.1016/0370-2693(74)90578-4; Nucl. Phys. B [**70**]{} (1974) 39 doi:10.1016/0550-3213(74)90355-1; Nucl. Phys. B [**78**]{} (1974) 1 doi:10.1016/0550-3213(74)90112-6. L. Maiani, [*Proc. Summer School on Particle Physics*]{}, Gif-sur-Yvette, 1979 (IN2P3, Paris, 1980) p. 3; G. ’t Hooft, in: G. ’t Hooft et al., eds., [*Recent Developments in Field Theories*]{} (Plenum Press, New York, 1980); E. Witten, Nucl. Phys. [**B188**]{} (1981) 513; R.K. Kaul, Phys. Lett. [**109B**]{} (1982) 19.
See, for example: S. Weinberg, Phys. Rev. D [**13**]{} (1976) 974. See, for example: N. Arkani-Hamed, A. G. Cohen, E. Katz and A. E. Nelson, JHEP [**0207**]{} (2002) 034 \[arXiv:hep-ph/0206021\]. G. F. Giudice, C. Grojean, A. Pomarol and R. Rattazzi, JHEP [**0706**]{} (2007) 045 doi:10.1088/1126-6708/2007/06/045 \[hep-ph/0703164\]; R. Contino, C. Grojean, M. Moretti, F. Piccinini and R. Rattazzi, JHEP [**1005**]{} (2010) 089 doi:10.1007/JHEP05(2010)089 \[arXiv:1002.1011 \[hep-ph\]\]. M. Baak [*et al.*]{} \[Gfitter Group\], Eur. Phys. J. C [**74**]{} (2014) 3046\
doi:10.1140/epjc/s10052-014-3046-5 \[arXiv:1407.3792 \[hep-ph\]\]. V. Sanz and J. Setford, arXiv:1703.10190 \[hep-ph\]. J. Ellis, J. R. Espinosa, G. F. Giudice, A. Hoecker and A. Riotto, Phys. Lett. B [**679**]{} (2009) 369 doi:10.1016/j.physletb.2009.07.054 \[arXiv:0906.0954 \[hep-ph\]\]. G. Degrassi, S. Di Vita, J. Elias-Miro, J. R. Espinosa, G. F. Giudice, G. Isidori and A. Strumia, JHEP [**1208**]{} (2012) 098 \[arXiv:1205.6497 \[hep-ph\]\]. D. Buttazzo, G. Degrassi, P. P. Giardino, G. F. Giudice, F. Sala, A. Salvio and A. Strumia, JHEP [**1312**]{} (2013) 089 doi:10.1007/JHEP12(2013)089 \[arXiv:1307.3536 \[hep-ph\]\]. A. Hook, J. Kearney, B. Shakya and K. M. Zurek, JHEP [**1501**]{} (2015) 061 doi:10.1007/JHEP01(2015)061 \[arXiv:1404.5953 \[hep-ph\]\]. J. R. Ellis and D. Ross, Phys. Lett. B [**506**]{} (2001) 331 \[arXiv:hep-ph/0012067\]. W. Buchmuller and D. Wyler, Nucl. Phys. B [**268**]{} (1986) 621; A. Pomarol and F. Riva, JHEP [**1401**]{} (2014) 151 \[arXiv:1308.2803 \[hep-ph\]\]. J. Ellis, V. Sanz and T. You, JHEP [**1503**]{} (2015) 157 doi:10.1007/JHEP03(2015)157 \[arXiv:1410.7703 \[hep-ph\]\]. A. Falkowski, M. Gonzalez-Alonso, A. Greljo and D. Marzocca, Phys. Rev. Lett. [**116**]{} (2016), 011801 doi:10.1103/PhysRevLett.116.011801 \[arXiv:1508.00581 \[hep-ph\]\]. Paraphrasing J. Bond, [http://www.imdb.com/title/tt0143145/fullcredits/]{}.
B. Bairnsfather, [http://www.independent.co.uk/news/uk/home-news/the-]{}\
[captain-who-gave-britain-its-ultimate-weapon-during-world-war-one-]{}\
[laughter-9833596.html]{}.
J. R. Ellis, G. Ridolfi and F. Zwirner, Phys. Lett. B [**257**]{} (1991) 83; H. E. Haber and R. Hempfling, Phys. Rev. Lett. [**66**]{} (1991) 1815; Y. Okada, M. Yamaguchi and T. Yanagida, Prog. Theor. Phys. [**85**]{} (1991) 1. J. R. Ellis, S. Heinemeyer, K. A. Olive and G. Weiglein, JHEP [**0301**]{} (2003) 006 doi:10.1088/1126-6708/2003/01/006 \[hep-ph/0211206\]. J. R. Ellis, J. S. Hagelin, D. V. Nanopoulos, K. A. Olive and M. Srednicki, Nucl. Phys. B [**238**]{} (1984) 453. doi:10.1016/0550-3213(84)90461-9 J. Ellis, S. Kelley and D.V. Nanopoulos, Phys. Lett. [**B260**]{} (1991) 131; U. Amaldi, W. de Boer and H. Furstenau, Phys. Lett. [**B260**]{} (1991) 447; P. Langacker and M. Luo, Phys. Review [**D44**]{} (1991) 817; C. Giunti, C. W. Kim and U. W. Lee, Mod. Phys. Lett. A [**6**]{} (1991) 1745. E. Bagnaschi [*et al.*]{}, Eur. Phys. J. C [**77**]{} (2017) no.2, 104 doi:10.1140/epjc/s10052-017-4639-6 \[arXiv:1610.10084 \[hep-ph\]\]. E. Bagnaschi [*et al.*]{}, arXiv:1612.05210 \[hep-ph\]. T. Golling [*et al.*]{}, \[arXiv:1606.00947 \[hep-ph\]\]. P. Cushman [*et al.*]{} \[WIMP Dark Matter Direct Detection Working Group\], arXiv:1310.8327 \[hep-ex\]. A. Tan [*et al.*]{} \[PandaX-II Collaboration\], Phys. Rev. Lett. [**117**]{} (2016) no.12, 121303 doi:10.1103/PhysRevLett.117.121303 \[arXiv:1607.07400 \[hep-ex\]\]. D. S. Akerib [*et al.*]{} \[LUX Collaboration\], Phys. Rev. Lett. [**118**]{} (2017) no.2, 021303 doi:10.1103/PhysRevLett.118.021303 \[arXiv:1608.07648 \[astro-ph.CO\]\]. D. S. Akerib [*et al.*]{} \[LZ Collaboration\], arXiv:1509.02910 \[physics.ins-det\]. E. Aprile [*et al.*]{} \[XENON Collaboration\], JCAP [**1604**]{} (2016) no.04, 027 doi:10.1088/1475-7516/2016/04/027 \[arXiv:1512.07501 \[physics.ins-det\]\]. O. Buchmueller, M. J. Dolan and C. McCabe, JHEP [**1401**]{} (2014) 025 doi:10.1007/JHEP01(2014)025 \[arXiv:1308.6799 \[hep-ph\]\]; O. Buchmueller, S. A. Malik, C. McCabe and B. Penning, Phys. Rev. Lett. [**115**]{} (2015) no.18, 181802 doi:10.1103/PhysRevLett.115.181802 \[arXiv:1505.07826 \[hep-ph\]\]; S. A. Malik [*et al.*]{}, [*White Paper from 2014 Brainstorming Workshop held at Imperial College London*]{}, Phys. Dark Univ. [**9-10**]{} (2015) 51 doi:10.1016/j.dark.2015.03.003 \[arXiv:1409.4075 \[hep-ex\]\]; J. Abdallah [*et al.*]{}, Phys. Dark Univ. [**9-10**]{} (2015) 8 doi:10.1016/j.dark.2015.08.001 \[arXiv:1506.03116 \[hep-ph\]\]; M. Bauer [*et al.*]{}, [*White Paper from 2016 Brainstorming Workshop held at Imperial College London*]{}, arXiv:1607.06680 \[hep-ex\]; D. Abercrombie [*et al.*]{} \[ATLAS/CMS Dark Matter Forum\], arXiv:1507.00966 \[hep-ex\]. A. Boveia [*et al.*]{} \[LHC Dark Matter Working Group\], arXiv:1603.04156 \[hep-ex\]. CMS Collaboration, [https://cds.cern.ch/record/2208044/files/DP2016[\_]{}057.pdf]{}.
G. Aarons [*et al.*]{} \[ILC Collaboration\], arXiv:0709.1893 \[hep-ph\]; ILC TDR, H. Baer, T. Barklow, K. Fujii, Y. Gao, A. Hoang, S. Kanemura, J. List and H. E. Logan [*et al.*]{}, arXiv:1306.6352 \[hep-ph\]; D. M. Asner [*et al.*]{}, arXiv:1310.0763 \[hep-ph\]. CLIC CDR, eds. M. Aicheler, P. Burrows, M. Draper, T. Garvey, P. Lebrun, K. Peach, N. Phinney, H. Schmickler, D. Schulte and N. Toge, CERN-2012-007,\
[http://project-clic-cdr.web.cern.ch/project-CLIC-CDR/]{}; M. J. Boland [*et al.*]{} \[CLIC and CLICdp Collaborations\], doi:10.5170/CERN-2016-004 arXiv:1608.07537 \[physics.acc-ph\]; H. Abramowicz [*et al.*]{}, arXiv:1608.07538 \[hep-ex\]. CEPC-SPPC Study Group, IHEP-CEPC-DR-2015-01, IHEP-TH-2015-01, IHEP-EP-2015-01. M. Bicer [*et al.*]{} \[TLEP Design Study Working Group Collaboration\], JHEP [**1401**]{} (2014) 164\[arXiv:1308.6176 \[hep-ex\]\]; see also [http://tlep.web.cern.ch/content/]{}\
[machine-parameters]{}.
J. Ellis, P. Roloff, V. Sanz and T. You, arXiv:1701.04804 \[hep-ph\]. J. Ellis and T. You, JHEP [**1603**]{} (2016) 089 doi:10.1007/JHEP03(2016)089 \[arXiv:1510.04561 \[hep-ph\]\]. FCC Collaboration, [https://fcc.web.cern.ch/Pages/default.aspx]{}.
R. Contino [*et al.*]{}, arXiv:1606.09408 \[hep-ph\].
[^1]: On the other hand, they have found no hint of any other physics beyond the SM, despite a similar myriad of searches.
[^2]: The grey lines are for the NUHM2 model - see [@MCSU5] for details.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Building speech recognizers in multiple languages typically involves replicating a monolingual training recipe for each language, or utilizing a multi-task learning approach where models for different languages have separate output labels but share some internal parameters. In this work, we exploit recent progress in end-to-end speech recognition to create a single multilingual speech recognition system capable of recognizing any of the languages seen in training. To do so, we propose the use of a universal character set that is shared among all languages. We also create a language-specific gating mechanism within the network that can modulate the network’s internal representations in a language-specific way. We evaluate our proposed approach on the Microsoft Cortana task across three languages and show that our system outperforms both the individual monolingual systems and systems built with a multi-task learning approach. We also show that this model can be used to initialize a monolingual speech recognizer, and can be used to create a bilingual model for use in code-switching scenarios.'
bibliography:
- 'strings.bib'
- 'refs.bib'
title: 'Towards Language-Universal End-to-End Speech Recognition'
---
multilingual, language-universal
Introduction {#sec:intro}
============
As voice-driven interfaces to devices and information become mainstream, increasing the global reach of speech recognition systems becomes increasingly important. There are two primary challenges that arise in expanding the language coverage of a speech application. First, since conventional speech recognition systems require each model to be trained independently, as the number of supported languages grows, the effort required to train, deploy, and maintain so many models in a production environment will increase dramatically. In addition, for second- and third-tier languages with fewer resources available, issues with data scarcity arise. For each language, building a speech recognition system requires a large collection of transcribed speech recordings from many speakers to train an acoustic model, linguistic expertise to create a pronunciation dictionary, and vast amounts of text data to train a language model.
Over the years, prior work has attempted to address the issues faced by low-resource language via a transfer learning approach [@lin2009study; @tuske2013investigation; @ghoshal2013multilingual; @huang2013cross; @heigold2013multilingual; @miao2014improving; @besacier2014automatic; @gales2015unicode]. In these approaches, language-specific deep networks are trained in which the the parameters in the lower layers of the network are shared across languages. This approach can also be interpreted as an instance of multi-task learning, where information across tasks, i.e. languages, is shared to create a more informative internal representation, less prone to over-fitting. Another common approach for creating models in low resource languages is to adapt a neural acoustic model that has been well trained on a high-resource language. This is typically done by replacing output layer of the well-trained model and re-training the model to predict the targets of low-resource languages [@tuske2013investigation; @ghoshal2013multilingual; @huang2013cross; @vu2013multilingual; @heigold2013multilingual; @wang2015transfer]. All of these models have been based on the conventional acoustic modeling strategy based on senones, and therefore still require a pronunciation lexicon to map words to phonemes and then senones.
Recently, new approaches to speech recognition that work in a so-called *end-to-end* manner have been proposed. In these systems, a neural network is trained to convert a sequence of acoustic feature vectors into a sequence of graphemes rather than senones. Unlike sequences of senone predictions, which need to be decoded using a pronunciation lexicon and a language model, the grapheme sequences can be directly converted to word sequences without any additional models or machinery. The end-to-end models proposed in the literature operate using a Connectionist Temporal Classification framework [@graves2006connectionist; @graves2014towards; @hannun2014deep; @miao2015eesen; @zweig2017advances], an attention-based encoder-decoder framework [@bahdanau2014neural; @chorowski2014end; @chorowski2015attention; @chan2015listen], or both [@kim2017joint]. Thus far, the research in end-to-end systems has focused on monolingual scenarios and to the best of our knowledge, there have been no studies of grapheme-based models capable of recognizing multiple languages. There has been recent work that shows how a low-resource graphemic system can be initialized with a well-trained high-resource model, as was done for the senone-based models described previously [@kunze2017transfer].
In this paper, we make significant progress towards a language-universal speech recognizer by creating a single end-to-end system capable of recognizing any language it has been trained on. Our model exploits the recent progress in end-to-end approaches to output character sequences directly, without requiring pronunciation lexicons. As in other multilingual systems, we apply a transfer learning approach to share model parameters among multiple languages. The novel aspects of our proposed model are twofold. First, we use a single universal character set that can be shared among all languages rather than separate language-specific output layers. Second, we propose a language-specific gating mechanism in the network that can increase the network’s modeling power by using multiplicative interactions to modulate the network’s internal representations in a language-specific way. We evaluate our proposed model on the Microsoft Cortana personal assistant task and show that our system outperforms separate language-specific models as well as the conventional multi-task learning approach.
The remainder of this paper is organized as follows. In Section \[sec:ctc\], we briefly review the end-to-end speech recognition paradigm. In Section \[sec:multilingual\_ctc\], we describe approaches to multilingual CTC including multi-task learning and our proposed model that features end-to-end learning with a universal character set and language-dependent gating units. In Section \[sec:exp\], we evaluate our proposed model through a series of experiments on the Microsoft Cortana task across three languages. Finally, we summarize our findings and discuss future avenues of research in Section \[sec:conclusion\].
End-to-end modeling using CTC {#sec:ctc}
=============================
In this work, we perform end-to-end speech recognition using a CTC-based approach with graphemes (characters) as the output symbols. With CTC, the neural network is trained according to a maximum-likelihood training criterion computed over all possible segmentations of the utterance’s sequence of feature vectors to its sequence of labels [@graves2006connectionist]. Core to CTC training is the presence of the *blank* symbol which can be interpreted as a “don’t care" symbol in the output label sequence. In CTC, the label sequences can contain blanks and repeated characters without penalty.
Given a sequence of acoustic feature vectors, $\bm x$ and the corresponding graphemic label sequence, $\bm y$, CTC trains the model to maximize the probability distribution over all possible label sequences $\bm{\pi}$. $$\begin{aligned}
P(\bm{y}|\bm{x}) = & \sum_{\bm{\pi} \in \Phi(\bm{y})}P(\bm{\pi}|\bm{x}) \approx \sum_{\bm{\pi} \in \Phi(\bm{y})} \prod_{t=1}^T P(\pi _t|\bm{x}).
\end{aligned}$$ Likelihoods are computed using the well-known forward-backward algorithm and the gradient of the likelihood is used to updated the network parameters using back propagation. After model training, decoding is performed in a greedy manner, where the most likely symbol at each time frame is hypothesized. The final sequence is obtained via a post-processing step where any character repetitions and blank symbols are removed from the output. Further improvements can be obtained by incorporating an external language model, at the character or word level [@graves2014towards; @hannun2014deep; @miao2015eesen; @zweig2017advances].
In this work, we consider two neural architectures for a language-universal end-to-end model that both use the CTC objective function.
Multilingual end-to-end models {#sec:multilingual_ctc}
==============================
Multi-task CTC with language-specific character sets
----------------------------------------------------
Multi-task learning (MTL) has been proposed as a means for improving generalization performance in low-data scenarios. Training a model on multiple related tasks simultaneously serves as an inductive bias to improve the model’s performance. Many prior approaches to training acoustic models for low resource languages have used MTL as a means of sharing some of the acoustic model parameters across languages, e.g. all the parameters up to the final output layer [@lin2009study; @heigold2013multilingual; @miao2014improving]. This forces the model to learn commonalities across languages which provides effective regularization and prevents over-fitting. To the best of our knowledge, this is the first effort to evaluate MTL with an end-to-end model and a CTC objective function. Our MTL architecture is similar to previous senone-based models, where the lower layers are shared among the different languages while the output layer or layers are trained to be language specific. This architecture is shown in Figure \[fig:arch\]a. Note that while the figure shows a language-specific output layer and final hidden layer, the choice of where to branch the model is a design choice that should be determined by experimentation. Multi-task learning for CTC can be realized as a combination of the individual single-task objective functions. We define $l$ to be an index over the different languages in the training data, $\bm{x}_{l}$ as a sequence of input feature vectors, and $\bm{y}_l$ the corresponding grapheme sequence. We can then define the CTC objective function for the $l$th language as $$\mathcal{F}_{l}(\theta_s,\theta_l) \triangleq -\ln P(\bm{y}_{l}|\bm{x}_{l},\theta_s,\theta_l)$$ where $\theta_l$ are the language-specific model parameters, and $\theta_s$ are the model parameters shared across languages. The multi-task objective function over all languages can then be defined as the summation of the individual negative log likelihoods of all $L$ languages, $$\mathcal{F} \triangleq \sum_{l=1}^{L} \mathcal{F}_{l}(\theta_s,\theta_l)$$
Multilingual CTC with a universal character set {#sec:union}
-----------------------------------------------
In many instances, present-day languages evolved from a common ancestry. It is therefore natural that they share some common graphemes and phonemes. For example, the English character set is a subset of the Spanish character set and knowing one language helps to speak and write the other. With this as motivation, we propose a multilingual architecture that uses a “universal" output label set consisting of the union of all characters from the multiple languages, as illustrated in Figure \[fig:arch\]b. Unlike the MTL approach, this is a single model with a universal character set that is trained on all multilingual data. Characters that are common across multiple languages are trained based on all relevant data while language-specific characters are only trained with data from that language. Given the training data from multiple languages and the universal label set, $\bm{y}_{u}$, the CTC loss for the models to be minimized is defined as, $$\mathcal{L}_{U}(\theta) \triangleq -\ln P(\bm{y}_{u}|\bm{x}_1,\dots,\bm{x}_L, \theta)$$ In this model, we assume *a priori* that we know the language identity of the utterances in both training and decoding. We use this information to restrict the predictions to only those characters present in the corresponding language and mask the activations from the other irrelevant characters. Let $Y_u$ be the universal label set with $K$ distinct labels. The label set for any particular language $Y_l$ is a subset of $Y_u$ ($Y_l \subseteq Y_u$). Given a language indicator, $l$, we can mask out the activations from unwanted characters using a $K$-dimensional binary mask defined as $$\begin{aligned}
M[ l, k ] =&
\begin{cases}
0, \text{if} \; k \not\in Y_{l} \\
1, \text{if} \; k \in Y_{l}
\end{cases}
\end{aligned}$$ This mask is then applied to the network outputs to compute the log-likelihoods for the CTC objective function. Note that this masking operation is applied in both training and decoding.
Language-specific gating units {#sec:arch2}
------------------------------
One possible drawback to the proposed universal character approach is that the same grapheme may have different underlying phonetic realizations in different languages. For example, the letter $i$ in English typically corresponds to the sound /ih/ in English, but /iy/ in Italian. Thus, the model needs to adequately capture language-specific information in order to properly account for such differences. Adding a language identification feature as an auxiliary input to the model is the simplest way to do so. However, we found empirically that this only provides minimal improvement. Instead, we propose to use the language identity to modulate the network’s internal multilingual representations in a language-specific manner. To do so, the outputs of each hidden layer are processed by a series of language-dependent gates before being passed to the next layer in the model. Specifically, we first create a one-hot language indicator vector $d_l$ for each language $l$, $$\label{e1}
d_l = [0 \hspace{0.2cm} 0 \hspace{0.2cm} 1].$$ Then, we compute the gate value based on the language indicator vector $d_l$ and the current output values of $h_i$, the $i$th hidden layer $$\label{e2}
g(h_i, d_l) = \sigma(U h_i + V d_l + b)$$ where $U, V$, and $b$ are trainable parameters. The language-gated hidden activations are then calculated as $$\label{e3}
\hat {h_i} = g(h_i, d_l) \odot h_i$$ Finally, $\hat {h_i}$ and $d_l$ are concatenated and input to the next layer. $$\label{e4}
\tilde {h_i} = [\hat {h_i} : d_l]$$
EXPERIMENTS {#sec:exp}
===========
Experimental corpora
--------------------
We investigated the performance of the proposed language-universal model on English (EN), German (DE), and Spanish (ES) data from Cortana, Microsoft’s personal assistant. For each language, we used 150 hours of training data, 10 hours of validation data, and 10 hours of test data. We used 80-dimensional log-mel filterbank coefficients as acoustic features, derived from 25 ms frames with a 10 ms frame shift. We concatenate three consecutive feature vectors to input to the network and employ frame-skipping [@sak2015fast] which decimates the original frame rate by a factor of three. Thus, each feature vector is presented to the network exactly once. Following [@zweig2017advances], we used a label symbol inventory consisting of the individual characters and their double-letter units. An initial capitalized letter rather than a space symbol was used to indicate word boundaries. This resulted in 81 distinct labels for English, 93 labels for German, and 97 labels for Spanish. Our universal label set for these three languages had 108 distinct labels and 81 overlapping labels. No pronunciation lexicon or language model was used in any of the experiments.
Training and decoding
---------------------
Our language-universal encoder was a 4-layer Bidirectional Long Short-Term Memory (BLSTM) network [@hochreiter1997long; @graves2013hybrid] with 320 cells in each layer and direction. A linear projection layer followed each BLSTM layer. All the weights in the models were initialized with a uniform distribution in the range of \[-0.05, 0.05\], and were trained using stochastic gradient descent with momentum. We used a learning rate of 0.0004 and gradient-clipping threshold per sample of 0.0003. Early stopping on the validation set was used to select the best model. For the decoding, the most likely sequence of characters was generated by the model in a greedy manner. The final output sequence was then obtained by removing any blank symbols or repetitions of characters from the output and replacing any capital letter with a space and its lowercase counterpart.
Gated language-universal end-to-end models
------------------------------------------
\[tab:result1\]
-------------- ------- ------- ------ ---------- ----------
Training Total Model Test CER % Rel.
Languages Hrs Arch Lang % Impr.
DE 150 stl 23.3 -
DE + EN 300 mtl 22.3 4.0
DE + EN 300 univ 22.5 3.2
DE + EN + ES 450 univ 22.8 2.1
DE 300 stl **15.8** **32.2**
ES 150 st 13.7 -
ES + EN 300 mtl 13.1 4.4
ES + EN 300 univ 12.9 5.8
ES + EN + DE 450 univ 13.1 3.9
ES 300 stl **11.7** **14.4**
-------------- ------- ------- ------ ---------- ----------
: Initial CER results for conventional single-task (stl), multi-task (mtl), and universal character set (univ) networks using monolingual and multilingual training data.
We first evaluated both multi-task and language-universal architectures on German and Spanish data, along with conventional monolingual training. As seen in Table \[tab:result1\], we obtained a small performance gain over the monolingual model by using multilingual training data. In addition, we observed that models trained with multi-task learning with language-specific output labels and models trained with a universal label set performed comparably, demonstrating the potential of a universal label set. However, with a conventional BLSTM topology, the benefit gained from the multilingual data was far less than simply doubling the amount of training data from the target language to 300 hours. More concerning, the performance actually degraded when we increased the multilingual data from two to three language. We also extensively evaluated various configurations for multi-task learning, where all or only some of the layers were shared, and found no significant improvements over the results reported in Table \[tab:result1\]. These initial results indicate that simply training conventional models with multilingual data is not very helpful. However, we note that these results may differ in extremely low resource scenarios where the data available per language could be an order of magnitude smaller.
\[tab:result2\]
We next evaluated the proposed language-universal model with language-specific gating. The results in Table \[tab:result2\] show that the proposed gated model significantly outperforms the monolingual model as well as the conventional multilingual models. Our model showed 10.7%, 11.4%, and 14.1% relative improvement in CER compared to a monolingual for EN, DE, and ES respectively. The relative improvements in WER, which was calculated without using any language model, were 7.0%, 8.6%, and 11.1%, for EN, DE, and ES respectively. Furthermore, when language-specific gating is used, additional improvement is obtained when the number of training languages increased from two to three.
Figure \[fig:gate\_function\_analysis\] shows the relative improvement over the baseline WER obtained by adding language information to the model in various ways. The leftmost bar shows the result of simply augmenting the input to each layer with the one-hot language vector $d_l$. The next three bars show the performance of different gating functions driven by the current hidden state, the language identity, or both, respectively. Finally, the rightmost bar shows the approach shown in Equations \[e3\]-\[e4\]. As the figure indicates, all approaches provide gains over the baseline model, but the proposed approach results in the largest improvement. We also investigated the effect of the gating mechanism on different layers and observed that the best performance was obtained when gating is applied to every layer (not shown).
![The relative improvement in WER obtained by using language identification information as an auxiliary input and/or a gating mechanism. []{data-label="fig:gate_function_analysis"}](figure/gate_functions_label3.png){width="8.5cm"}
Initial model for subsequent language-specific models
-----------------------------------------------------
While the primary goal of this work was to create a language-universal model, we found that this model is also a good initial model for creating a language-specific monolingual model when the training data is limited. Table \[tab:result3\] shows the CER on the DE test set obtained from different pre-training strategies used to initialize the model prior to fine-tuning on the DE training data. Notably, initializing with the gated language-universal model outperformed all of the other approaches, including bootstrapping from an EN model trained with significantly more data. The second best performance was obtained by the gated language-universal model directly, even without further fine-tuning on the DE training data.
\[tab:result3\]
--------------------- ------- -----------------------
% CER
Training Data Hours DE
– – 23.3
EN 150 24.3
EN 1000 21.4
EN + DE 300 21.1
EN + ES + DE + gate 450 **19.4** / **20.6**\*
--------------------- ------- -----------------------
: CER obtained for a monolingual model with different pre-training strategies. The model was then fine-tuned on 150 hours of the target language (DE), except \* where no fine-tuning was performed.
Bilingual end-to-end models
---------------------------
One promising aspect of a universal output character set is the potential to create end-to-end systems that can dynamically code-switch between languages. This can be done by training a model on bilingual training data with the union of the output symbols of the two languages while omitting the language-specific aspects of the model. Following this approach, we created a bilingual model from 150 hours of English (EN) and 150 hours of Spanish (ES) using the union of the character sets from the two languages.
Because a test set of bilingual mixed-code speech was not available, we evaluated this model on the EN and ES test sets separately. Table \[tab:result4\] shows that bilingual models perform as well as the monolingual models even if no language identification information is provided to the network. Thus, this approach is very promising as a method for creating end-to-end systems capable of code-switching during decoding.
\[tab:result4\]
---------------------------- ----- ------ ------
ES EN
monolingual EN or ES 150 39.9 45.0
bilingual EN + ES w/o Gate 300 38.8 45.1
---------------------------- ----- ------ ------
: WER on ES and EN test sets obtained from monolingual and bilingual end-to-end systems.
CONCLUSION {#sec:conclusion}
==========
We proposed a language-universal end-to-end speech recognition system capable of recognizing speech from any language seen in training. Key aspects of this model include the use of a universal character set and language-specific gating units. Because this model can support multiple languages, it has the potential to simplify model deployment in production environments. In addition, no pronunciation lexicon is required, which is beneficial for second- and third-tier languages, where such resources may be unavailable or limited. Our model was shown to outperform models trained on monolingual training data, as well as with the multi-task learning approach employed in previous work. Moreover, we found that our language-universal speech recognizer is a strong initial model for subsequent monolingual training, and demonstrated the potential of this system for decoding speech in the presence of code-switching. Moving forward, we plan to explore how to add a new language to an already-trained multilingual system and evaluate the performance improvements that can be obtained by using an external language model.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The growth of cooperatively rearranging regions was invoked long ago by Adam and Gibbs to explain the slowing down of glass-forming liquids. The lack of knowledge about the nature of the growing order, though, complicates the definition of an appropriate correlation function. One option is the point-to-set correlation function, which measures the spatial span of the influence of amorphous boundary conditions on a confined system. By using a swap Montecarlo algorithm we measure the equilibration time of a liquid droplet bounded by amorphous boundary conditions in a model glass-former at low temperature, and we show that the cavity relaxation time increases with the size of the droplet, saturating to the bulk value when the droplet outgrows the point-to-set correlation length. This fact supports the idea that the point-to-set correlation length is the natural size of the cooperatively rearranging regions. On the other hand, the cavity relaxation time computed by a standard, nonswap dynamics, has the opposite behavior, showing a very steep increase when the cavity size is decreased. We try to reconcile this difference by discussing the possible hybridization between MCT and activated processes, and by introducing a new kind of amorphous boundary conditions, inspired by the concept of frozen external state as an alternative to the commonly used frozen external configuration.'
author:
- Andrea Cavagna
- 'Tomás S. Grigera'
- Paolo Verrocchio
bibliography:
- 'all.bib'
title: Dynamic relaxation of a liquid cavity under amorphous boundary conditions
---
Introduction
============
It is common wisdom that the spectacular slowing down of supercooled liquids at low temperature is caused by the growth of a correlation length of some sort. The underlying idea is that of cooperativity: at lower temperatures, larger regions (termed cooperatively rearranging regions) must move together in order to fully relax [@glassthermo:adam65]. Unfortunately, the standard tools used in critical phenomena to detect a growing correlation length fail in glass-forming liquids, as it is not at all clear [*a priori*]{} what the order parameter should be. No obvious domain or structure can be observed in a low temperature liquid to distinguish it from a high temperature one. If order is growing in glass-formers, it must be some sort of amorphous order, and the corresponding order parameter must be nonstandard.
Recently, some progress has been achieved by using [@mosaic:bouchaud04; @self:prl07; @self:nphys08]. The idea goes as follows [@mosaic:bouchaud04]. Consider a low-temperature equilibrium configuration of a liquid and freeze all particles outside a certain region. This region (or cavity) is then let free to evolve and thermalize, subject to the pinning field produced by all the frozen particles surrounding it. Clearly, the smaller the region the stronger the effect of the pinning field, hence keeping the region in a very restricted portion of its own phase space. The idea, then, is to check how large the region must be to emancipate from the boundary conditions, [*i.e.*]{} to regain ergodicity and thermalize into a state different from the surrounding one. The advantage of this method is that the system chooses its own definition of ‘order’ by means of the amorphous boundary conditions, and we do not need to have any [ *a priori*]{} knowledge of the nature of such order. Practically speaking, the procedure amounts to measure, as a function of the size $R$ of the region, the correlation between the original region’s configuration (that of the frozen surrounding) and that achieved after the region has equilibrated subject to the . This quantity is called point-to-set correlation function [@cs:mezard06; @dynamics:montanari06], $q(R)$, and it has shown an interesting feature [@self:prl07; @self:nphys08]: its decay length-scale, $\xi_s$, increases on lowering $T$. Regions smaller than $\xi_s$ *cannot* relax completely, even given infinite time, due to the presence of the pinning .
Here, in order to get some information about the dynamics of the cooperatively rearranging regions, we study the dynamical behavior of a cavity under ABC. Of course, we do expect that the equilibration time of the cavity must be equal to its bulk value for large enough values of $R$. What is not trivial is at what specific value of $R$ the saturation occurs and whether the saturation occurs from above or from below, [*i.e.*]{} whether the equilibration time decreases or increases when the cavity gets larger. As we shall see, we obtain different results according to the specific dynamics we use: by means of a swap Montecarlo dynamics, where particles of different species can be exchanged in order to accelerate the dynamics, we find a clear saturation from below of the equilibration time taking place at $R\sim\xi_s$ (Section \[sec:swap-dynam-conf\]). This result seems to support the idea that $\xi_s$ is indeed the cooperativity length scale of the system (Sections \[sec:rfot-interpr-swap\]– \[sec:an-unexp-ineq\]). If, on the other hand, we use a standard, nonswap algorithm, we find that the dynamics slows down very steeply when the cavity size is decreased (Section \[sec:nonsw-dynam-conf\]).
We will discuss the meaning of these results and how this difference may be related to the actual relaxation mechanism active for the cooperatively rearranging regions in the bulk. In particular, we will show that the hybridization between the Mode Coupling Theory (MCT) and the activated relaxation channels can give rise to a nonmonotonic behavior of the relaxation time as a function of the cavity radius (Section \[sec:contr-betw-swap\]). We will also put forward the hypothesis that the standard amorphous boundary conditions, where the external configuration is frozen, introduce an artificial slowing down of the cavity dynamics, which can be overcome by switching to the more physical *frozen state* boundary conditions (Sections \[sec:fc-vs-vs\]– \[sec:cavity-dynamics-with\]). After briefly commenting on some relevant experiments (Section \[sec:some-exper-evid\]) we summarize our conclusions in Section \[sec:conclusions\].
Model and observables
=====================
We perform simulations of a 3-$d$ soft-spheres binary mixture [@soft-spheres:bernu87] with parameters as in Ref. [@self:nphys08]. The mode-coupling temperature [@rev:goetze92] for this system is $\TMC=0.226$ [@soft-spheres:roux89]. Our largest system has $N=16384$ particles in a box of length $L=25.4$. We run simulations at $T=0.482,0.350,0.246,0.214,0.202$. We first equilibrate the whole system with periodic boundary conditions to generate a set of equilibrium configurations, and then run the amorphous boundary simulations by picking an equilibrium configuration and artificially freezing all particles but those occupying a spherical cavity of radius $R=$ 1.06, 1.68, 1.92, 2.12, 2.28, 2.61, 2.87, 3.29, 3.62, 4.15, 4.57, 5.75, 7.2, 9.14, and 10.95. We use at least 16 samples for each $T$ and $R$.
Our main physical observable is the overlap, which measures the correlation between the running configuration and the reference one at $t=0$. The cavity is partitioned in small cubic boxes and $n_i$ is the number of particles in box $i$. The side $\ell$ of the cells is such that $n_i=\{0,1\}$. We measure the overlap within a small cubic volume $v$ located at the center of the sphere [@self:nphys08], $$q(t;R)\equiv \frac{1}{\ell^3 \, N_i} \sum_{i \in
v} n_i(t) \, n_i(0) \ ,$$ where the sum runs over all boxes and $N_i$ is the number of boxes in the central volume. To minimize statistical uncertainty without losing the local nature we choose $N_i = v/\ell^3 = 125$. On average, the overlap of two identical configurations is $1$, while for totally uncorrelated configurations $q = q_0 = \ell^3 = 0.062876$. The asymptotic value of the overlap, $q(R)\equiv \langle
q(t\to\infty;R)\rangle$, averaged over many realizations of the boundary conditions, is the point-to-set correlation function [@mosaic:bouchaud04; @dynamics:montanari06; @self:prl07; @self:nphys08; @review:biroli09].
In order to define a time-scale we measure the connected auto-correlation function of the overlap fluctuations, $$C(t;R) =\frac{
\left\langle \left( q(t_0+t;R)- q(R) \right) \left( q(t_0;R)- q(R)
\right) \right\rangle }
{
\left\langle \left( q(t_0;R)- q(R) \right)^2 \right\rangle
}
\ .
\label{connessa}$$ To estimate the equilibration time $\tau(R)$ we use the method discussed by A. Sokal in [@review:sokal97], based on the integral of the correlation function. More specifically, the relaxation time $\tau$ is found by solving the equation, $$\tau = \int_0^{\alpha \tau} dt \ C(t;R) \ ,$$ where the optimal value of $\alpha$ has been found to be $20$. In this way one is sure to sample the phenomenon on a time window that is self consistently much larger than the relaxation time.
Swap dynamics in the confined cavity {#sec:swap-dynam-conf}
====================================
We first focus on the results obtained with swap dynamics [@self:pre01]. With a swap Monte Carlo dynamics we propose (with probability 0.1) a move that swaps the position of two different particles. Provided that the radii of the two different species are not too different, so that the swap move is not always rejected, this kind of move decreases significantly the time needed by a single particle to break its cage. On the other hand, the swap dynamics has less of an impact on collective rearrangements, and indeed the swap relaxation time increases dramatically close to the glass transition, as the nonswap time.
Fig. \[fig:corr-allT\] shows the swap auto-correlation function $C(t;R)$ at various values of $R$ for our lowest temperature $T=0.202$. We stress that for those values of $R$ such that the order parameter $q(R)\neq 0$, ergodicity is broken [@self:nphys08]. In this case the [*connected*]{} correlation function describes the equilibrium dynamics *within a restricted region* of the cavity’s phase space.
Before estimating the relaxation time it is very important to be sure that the autocorrelation function does not depend on the size of the time window $\Delta t$ used to measure it. To this aim in Fig. \[fig:fss\] we show the autocorrelation function at our lowest temperature and at different values of the time window $\Delta t$, at two values of $R$: there is no significant dependence of $C(t;R)$ on $\Delta t$.
In Fig. \[fig:tau149\] we report the swap equilibration time $\tau(R)$ for our lowest temperature, $T=0.202$. Three features of this curve stand out: i) the swap equilibration time saturates for large $R$ to a value independent of the cavity size; ii) the swap equilibration time grows with $R$, so that saturation occurs from below; iii) growth and saturation are separated by a rather sharp kink at a well-defined value of $R$. The first fact is obvious: the effect of the boundary conditions is expected to fade away for large $R$, so that $\tau(R)$ must eventually reach its bulk value, which is exactly what happens. The remarkable point is that $\tau(R)$ reaches its bulk value for $R \sim \xi_s$, where $\xi_s$ is the point-to-set correlation length measured in Ref. [@self:nphys08].
This result can immediately be interpreted in terms of cooperativity: For $R<\xi_s$ the whole region is correlated, because the effect of the amorphous border breaks the ergodicity. For $R>\xi_s$, the effect of the border fades away and the region is able to decorrelate by breaking up into smaller correlated sub-parts: in this regime relaxation factorizes. Hence, it seems that the point-to-set correlation length $\xi_s$ does indeed play a role in the cooperative dynamics of the system. In the next Sections we will address this point more precisely.
RFOT interpretation of the swap equilibration time {#sec:rfot-interpr-swap}
==================================================
According to the of the glass transition, whether or not a region of radius $R$ relaxes depends on the balance between the surface tension $Y$ that develops when that region actually rearranges and the configurational entropy $\Sigma$ unleashed by the rearrangement: if $Y>T\Sigma R^{d-\theta}$ ($d$ is the space dimension, $\theta$ is the surface tension —or stiffness— exponent) the surface cost is larger than the entropic gain and the region does not rearrange. On the other hand, if $Y<T\Sigma
R^{d-\theta}$ the entropic gain outweighs the surface energy cost and the region has a thermodynamic advantage to rearrange. The rearranging size where entropy and surface tension balance, $\xi_s=
\left(Y/T\Sigma\right)^{1/(d-\theta)}$, is the static correlation length of .
Therefore, within a cavity with amorphous boundary conditions of radius $R<\xi_s$ has broken ergodicity, and can only explore the state imposed by the boundary conditions [@mosaic:bouchaud04]. In this regime the equilibration time is the time needed to explore that one state, which is roughly equal to the $\beta$-relaxation time, $\tau(R)\sim \tau_\beta$ [^1]. For $R>\xi_s$, instead, rearrangement occurs and ergodicity of the cavity is restored. In this regime the region is larger than the minimal rearranging size, so that relaxation factorizes: different subregions of size $\xi_s$ will rearrange independently from each other, and the equilibration time will be equal to its bulk value, i.e. $\tau(R)\sim \tau_0 \exp(\xi_s^\psi/T)$, where $\tau_0$ is an Arrhenius prefactor and $\psi$ is the exponent regulating the barrier growth.
Hence, within the sharp description, where the surface tension has just one value, $Y$, it is predicted a step-like jump of $\tau(R)$ at $R=\xi_s$, from $\tau_\beta$ up to $\tau_0
\exp(\xi_s^\psi/T)$, $$\tau(R) \sim
\begin{cases}
\tau_\beta & R< \xi_s \\
\tau_0\, e^{\xi_s^\psi/T}
& R>\xi_s \ .
\end{cases}
\label{assuppa}$$ Such stepwise behavior is not what we observed in Fig. \[fig:tau149\]. In order to reconcile data and theory, we note that for the typical temperatures and sizes studied in simulations surface tension fluctuations are relevant [@self:nphys08]. If the surface tension fluctuates [^2] (i.e. different give different $Y$), local excitations can have different sizes and therefore different relaxation times. When we measure these quantities by averaging over many different sets of ABC we smooth out the sharp behavior of .
![Schematic view of the second integral in equation . The upper panel represents the two functions within the integral, the lower panel is the resulting relaxation time.[]{data-label="cartoon1"}](cartoon-ramp){width="\columnwidth"}
To see more precisely how this happens, let us write the surface tension distribution as $P(Y;Y_c)$, where now $Y$ is the fluctuating surface tension, whereas $Y_c$ is its typical scale, defined by the peak of the distribution. This means that a region of radius $R$ will rearrange or not rearrange, depending on the value of $Y$; accordingly, its relaxation time can be the either the in-state time $\tau_\beta$, or the time needed to activately rearrange the region, $$\tau(R,Y) \sim
\begin{cases}
\tau_\beta & Y> T\Sigma R^{d-\theta} \\
\tau_0\, \exp\left[\frac{1}{T}\left(Y/T\Sigma\right)^\frac{\psi}{d-\theta}\right]
& Y< T\Sigma R^{d-\theta} .
\end{cases}
\label{sukia}$$ The macrosopic equilibration time will be given by an average over $Y$ of the time in , $$\begin{split}
\tau(R) & = \tau_\beta
\int_{T\Sigma R^{d-\theta}}^\infty P(Y;Y_c)\,dY \ + \\
+\ &\quad \tau_0 \int_0^{T\Sigma R^{d-\theta}} P(Y;Y_c)\;
\exp\left[\frac{1}{T}\left(Y/T\Sigma\right)^\frac{\psi}{d-\theta}\right]
\, dY .
\end{split}
\label{naso}$$ The first term in corresponds to regions surrounded by large surface tension, which do not rearrange, and it equals at most $\tau_\beta$. The second term corresponds to the low surface tension regions that do rearrange, and at low temperatures this term is large. Clearly, if $P(Y;Y_c)=\delta(Y-Y_c)$ we recover the step-like behavior of $\tau(R)$ described in . If, on the other hand, $P(Y;Y_c)$ is broad, the result is nontrivial.
With a fluctuating surface tension we can still define a typical mosaic correlation length, $\xi_s=\left(Y_c/T\Sigma\right)^{1/(d-\theta)}$ [@self:nphys08]. This relation suggests an obvious change of variables useful to recast equation in a simpler form: $$\tau(R) = \tau_\beta
\int_{R}^\infty P(r;\xi_s)\,dr + \tau_0 \int_0^{R} P(r;\xi_s)\;
e^{r^\psi/T} \, dr ,
\label{nasone}$$ where we are now integrating over all possible sizes of the rearranging regions. $P(r;\xi_s)$ is the distribution of sizes, which is of course peaked on $\xi_s$. To understand the behavior of the function $\tau(R)$ let us assume that $P(r;\xi_s)$ has a compact support, being different from zero only in the interval $r\in [\xi_s-\delta:\xi_s+\delta]$, and zero elsewhere. We have three regimes of $R$ (see Fig. \[cartoon1\]):
i\) for $R<\xi_s-\delta$ the first integral in is $1$ and the second integral is $0$, so that $\tau(R) = \tau_\beta$;
ii\) for $\xi_s-\delta < R < \xi_s+ \delta$ the weight shifts from the first to the second integral; because of the exponential, which is large at low $T$, $\tau$ grows with growing $R$, thus giving rise to a ramp that brings the relaxation time to a value considerably larger than $\tau_\beta$;
iii\) for $R> \xi_s+\delta$, the first integral is $0$, whereas the second one has reached its saturation value; to know this value, at low $T$ we can use the saddle point approximation: the maximum of the integrand occurs approximately for $r\sim \xi_s$, so that $\tau(R)
\sim \tau_0 \ e^{\xi_s^\psi/T}$. This last quantity is nothing else than the bulk relaxation time, $\tau_\mathrm{bulk}$.
What we have just described is a smooth growth of $\tau(R)$ from $\tau_\beta$ up to the bulk relaxation time $\tau_\mathrm{bulk}$, taking place in a range of $R$ around $\xi_s$, $$\tau(R) \sim
\begin{cases}
\tau_\beta & \text{for $R < \xi_s-\delta $} \\
\text{growth} & \text{for $\xi_s-\delta < R <\xi_s+\delta$} \\
\tau_0\; e^{\xi_s^\psi/T} & \text{for $R > \xi_s+\delta $} \ .
\end{cases}
\label{topa}$$ It is difficult to better specify the behavior of $\tau(R)$ in its increasing regime with no knowledge of the distribution $P(r;\xi_s)$ (or, equivalently, of $P(Y;Y_c)$). Still, in the saddle point limit (low $T$) there is something we can say: the second integral in is dominated by the exponential, and for $R<\xi$ the saddle-point coincides with the right edge of the integration domain, $r_{\mathrm SP}=R$. In this case we have, $$\tau(R) \sim \tau_0\ e^{R^\psi/T}, \qquad \xi_s-\delta < R <\xi_s+\delta .$$ The behavior described by is in agreement with what we have found in our swap simulations (Fig. \[fig:tau149\]). The relaxation time grows with the radius of the cavity, and it saturates to its bulk value at $R\sim \xi_s$, so that we can use the saturation point as an estimate of the static correlation length $\xi_s$. In Fig. \[fig:tau-allT\] we report the cavity swap relaxation time normalized by its bulk value for several different temperatures. We can see that the saturation point moves to larger values of $R$ at lower temperatures, a phenomenon consistent with the expectation that the correlation length grows when cooling the system. This fact consolidates the idea that the point where the cavity relaxation time saturates is indeed the same static correlation length as extracted from the point-to-set correlation function.
We test this interpretation by plotting in Fig. \[fig:xi\_vs\_xi\] the length scale of saturation of the swap relaxation time vs. the value of the static correlation length extracted by the point-to-set correlation function computed in [@self:nphys08]. Considering that both length scales have a degree of arbitrariness in their measurement, we normalize them in order to be equal at one specific temperature (see the caption of Fig. \[fig:xi\_vs\_xi\]). Even though we definitely would need a wider temperature range to say something certain, we can conclude that the two length scales track each other quite reasonably. This supports the idea that the point-to-set correlation length (an eminently static concept) can actually be measured also by using the swap relaxation time of a cavity subject to amorphous boundary conditions.
When cooler is faster {#sec:when-cooler-faster}
=====================
Both the stepwise behavior of and the smooth growth of have an interesting consequence: at some values of $R$ a colder cavity may be [*faster*]{} than a hotter cavity. How this happens is pictorially explained in Fig. \[cartoon2\]. By lowering the temperature, $\xi_s$ increases, so we push to the right the support of $P(r;\xi_s)$, and therefore the range of $R$ over which the growth of $\tau(R)$ occurs; at the same time, the bulk relaxation time increases, so that the low $T$ curve must saturate at a higher level than the high $T$ curve. This mechanism gives rise to a crossing of the cold and hot relaxation times, so that in the region of $R$ between the cold and hot value of $\xi_s$, we have that the lower $T$ cavity has a smaller relaxation time than the higher $T$ cavity.
![Schematic view showing how an inversion of the cold and hot relaxation times can take place. By lowering the temperature two (related) phenomena occur: i) the correlation length increases, so that the distribution $P(r;\xi_s)$ moves overall to the right (it also becomes more peaked, see [@self:jstatmech09], but this is irrelevant here); ii) the asymptotic bulk relaxation time increases, so at saturation $\tau(R)$ reaches a higher level. These two phenomena give rise to a regime, between the two correlation lengths, where the relaxation time of the colder cavity is lower than that of the hotter cavity.[]{data-label="cartoon2"}](cartoon-inversion){width="\columnwidth"}
This odd phenomenon is confirmed by our swap simulations. In Fig. \[fig:inversion\] we show the cavity swap relaxation time at two different values of $T$. It can be seen quite clearly that for certain values of the radius the cold cavity is faster than the warm cavity. In the inset of Fig. \[fig:inversion\] we directly show the two autocorrelation functions for one specific value of $R$, just to make clear that the effect does not depend on the particular definition of $\tau$.
As we have seen, this interesting phenomenon is quite naturally explained in the context of RFOT. In the sharp scenario, the inversion of cold and warm relaxation times is a direct consequence of the presence of two qualitatively different times: the short in-state time, $\tau_\beta$, and the long out-state relaxation time, $\tau_\mathrm{bulk}$. The existence of these two times means that at a certain value of $R$ a cold cavity may still be trapped into its original state, therefore having a [*short*]{} in-state relaxation time, whereas a warm cavity may be unlocked, and therefore have a longer relaxation time. We remark, once again, that one is comparing qualitatively different times: the in-state time $\tau_\beta$ is the time needed to relax within a state, with no cooperative rearrangement, while the relaxation time of a large cavity, $\tau_\mathrm{bulk}$ is the time needed for a full rearrangement. Such distinction is sharp, and easy to detect, only in the stepwise scenario of equation . On the other hand, as we have seen, in the real case $\tau(R)$ (averaged over many samples) is a smooth function, with a ramp connecting the in-state time to the bulk time, so that it is harder to distinguish the two different processes from the full $\tau(R)$ curve. The inversion of cold and hot relaxation times is therefore an interesting remnant of the presence of these two different time scales.
An unexpected inequality {#sec:an-unexp-ineq}
========================
In order to have a finite bulk equilibration time, we need the second integral in equation to be finite for $R\to\infty$. Therefore $P(Y;Y_c)$ must decay sufficiently fast to suppress the Arrhenius factor. If we make the reasonable assumption, $$P(Y;Y_c) \sim e^{-(Y/Y_c)^\nu} , \qquad Y \gg 1 ,$$ we must have, $$\nu \geq \frac{\psi}{d-\theta} .
\label{peppe}$$ As we have seen, the distribution $P(Y;Y_c)$ implies an equivalent distribution of the rearranging regions’ size, $P(R;\xs)$, inequality (\[peppe\]) means that $P(R;\xi_s)$ must decay fast enough to suppress the growth of the equilibration times for large $R$. This is reasonable. In [@self:nphys08] it was shown that the exponent $\nu$ is related to the anomaly exponent $\zeta$ that rules the nonexponential decay of the point-to-set correlation function $q(R)$, $$q(R) \sim e^{-(R/\xi_s)^\zeta} ,$$ with $$\zeta = \nu(d-\theta) ,
\qquad \zeta \geq 1 .$$ where $\theta$ is the surface tension (or stiffness) exponent. This leaves us with the inequality, $$\zeta \geq \psi .
\label{zanza}$$ On increasing the temperature the anomaly $\zeta$ must go to $1$, as the point-to-set correlation function $q(R)$ becomes a pure exponential [@self:nphys08]. If $\psi$ is temperature-independent, relation (\[zanza\]) then implies, $$\psi \leq 1 .
\label{luxury}$$ We note that the value $\psi\sim 1$ previously reported in [@self:jcp09a] satisfies (\[luxury\]). Of course, if we allow $\psi$ to depend on $T$ (as $\zeta$ does), then there would be no reason for to be valid in general, whereas would still hold.
Nonswap dynamics in the confined cavity {#sec:nonsw-dynam-conf}
=======================================
The dynamical behavior of the cavity when we switch off the swap moves is completely different from what we have seen until now: in contrast to the swap case, the relaxation is [*slower*]{} the smaller the cavity. In the bulk, the dynamics without swap is known to be significantly slower than with swap [@self:pre01] (this is why swap has been introduced in the first place). However, in the cavity, not only is nonswap dynamics slower, but the whole dynamical behavior as a function of $R$ is reversed.
We observe this phenomenon in Fig. \[fig:nonswap\], where we report the connected overlap as a function of time in the nonswap case for different values of $R$. The connected overlap is obtained by subtracting from $q(t)$ its equilibrium infinite time limit, $q(R)$, obtained with swap. The asymptotic value of the connected overlap must be equal to zero for all $R$ and this makes it easier to compare different sizes on the same plot. Smaller cavities are dramatically slower than larger ones. Under these conditions, it is clear that we cannot compute the overlap autocorrelation function in the nonswap case, as the system is robustly out of equilibrium. The only time correlation function that we can use is the overlap itself, $q(t)$, and to extract a relaxation time, $\tau(R)$, we cross $q(t)$ with an arbitrary value, $\bar q$. For those (few) values of $R$ for which this procedure is viable, we report $\tau(R)$ in Fig. \[fig:taunonswap\].
In smaller cavities, below $R\sim 4$, the nonswap overlap is completely stuck to a level which is [*above*]{} its equilibrium value. We can clearly see this by using a BIC (Beta Initial Condition) test. The idea is to initialize the cavity in a configuration $\beta$ which has overlap equal to zero with the $\alpha$ configuration used to thermalize the system, and which is frozen in the boundary condition. In this way, the BIC overlap $q_{\alpha\beta}(t)$ is zero at time zero, and it must increase to the same asymptotic value as the standard overlap $q_{\alpha\alpha}(t)$. When thermalization of the cavity is achieved the two overlaps must meet at the same equilibrium value, $q(R)$ [^3].
![BIC test comparison between swap and nonswap dynamics. In the BIC test the configuration is initialized both in the same configuration $\alpha$ as the frozen boundary (upper, full curve) and in a different configuration $\beta$ with respect to the frozen boundary (lower, dashed curve). The upper and lower curves must reach the same asymptotic value $q(R)$ for infinite times. The BIC test is positive for the swap dynamics; all of our swap data, fore every value of $R$ and $T$, have passed the BIC test. On the other hand, the BIC test is negative for the nonswap dynamics. Nevertheless, the nonswap time series is stationary, making it impossible to estimate a reasonable value of the relaxation time. $R=2.27$, $T=0.246$.[]{data-label="fig:stuck"}](new_testbic_portrait_BN){width="\columnwidth"}
A positive BIC test is shown in the upper panel of Fig. \[fig:stuck\] for the swap dynamics at small R: the two overlap branches meet at their asymptotic value, $q(R)$. We have run BIC tests for all our values of $R$ and $T$ in the swap case, always getting a positive result (the same holds for the data of [@self:nphys08]). In the lower panel of the same figure we see what happens in the nonswap case for the same value of $R$: despite the fact that the overlap is stationary for several decades, it is definitely not thermalized, as there is a clear and significant gap between the two branches, none of which reaches the equilibrium value $q(R)$ (dotted line). Hence, at this value of $R$ and of $T$ it is not even possible to roughly estimate $\tau$: no extrapolation of $q(T)$, however wild, makes sense with these data.
We notice that this slowing down happens also at relatively high temperatures: the effect of the confinement on the relaxation time is really drastic, and the difference between swap and nonswap dynamics stark. Incidentally, we note that without swap dynamics it would be impossible to measure the point-to-set correlation function (which is the equilibrium value of $q$), due to this hyper-slowing down. The slowing down of the dynamics in a confined cavity was noted before in ref. [@confinement:berthier11] for molecular dynamics.
In the next Sections we address the conflict between the swap and nonswap results.
The contradiction between swap and nonswap {#sec:contr-betw-swap}
==========================================
At this point we are left with a contradictory scenario. On one hand, with swap Monte Carlo the relaxation time grows up to its bulk value when increasing the cavity radius $R$, seemingly saturating when $R$ reaches the point-to-set correlation length $\xi_s$. This behavior suggests that $\xi_s$ is indeed the typical size of the cooperatively rearranging regions, which dominate activated dynamics at low temperatures. On the other hand, with standard nonswap Monte Carlo (as well as molecular dynamics [@confinement:berthier11]), the cavity relaxation time is larger than its bulk value and grows with decreasing $R$.
A dramatic increase of the nonswap relaxation time might suggest some kind of phase transition. Indeed a scenario involving a true phase transition has been recently described in ref. [@mosaic:cammarota11]. However, an essential ingredient of any phase transition is the thermodynamic limit. There is no true divergence at finite volume, but rather an unbounded growth of the relaxation time with volume. The transition discussed in ref. [@mosaic:cammarota11] applies to geometries where it is possible to send the system size to infinity (for example, scattered frozen particles or a sandwich geometry—see ref. [@confinement:berthier11]), in which case the relaxation time for $R\sim\xi_s$ should diverge. However, in our cavity geometry, the size is always finite, so that a phase transition does not seem the right explanation of what we see. So the nonswap scenario is hard to interpret within an RFOT context and it seems to be more compatible with a theory of the glass transition based on the idea that dynamics is facilitated by defect diffusion [@review:chandler10]: the smaller the volume, the smaller the number of defects and the slower the dynamics.
But even if the nonswap dynamics results would seem more physically relevant, a complete physical picture needs to account also for the swap results (in particular the intriguing fact of the saturation of $\tau$ at $R\sim\xi_s$) and to resolve the contradiction. This is what we attempt, in a rather speculative way, in this Section.
The hybridization between MCT and activation
--------------------------------------------
### The bulk case
To better understand what is going on in the cavity, we have to go back to the bulk. According to some theories of the glass transition [@review:biroli09], there are two relaxation channels: a nonactivated channel, well described by Mode Coupling Theory (MCT) [@rev:goetze92], which is ruled by unstable stationary points of the potential energy (saddles), and a second channel, consisting of activated barrier crossing. The first mechanism has a singularity at the MCT transition temperature $T_c$, where the MCT relaxation time diverges as a power law. On the other hand, the activated channel is insensitive to $T_c$, and its relaxation time increases in a super-Arrhenius fashion, due the the low-$T$ increase of the static correlation length, $\xi_s$.
![A schematic view of bulk relaxation. We hypothesise that there are two different channels of relaxation: i) the Mode Coupling Theory (MCT) channel, which is related to a relaxation which uses unstable stationary points (saddles) of the potential energy. The MCT dynamics has a relaxation time that diverges at $T_c$. ii) the activated barrier-crossing channel. The actual dynamics “chooses” the fastest of the two channels, so that the observed relaxation time is the lowest of the two. Below $T_c$, there is a dynamical crossover between the MCT branch to the activated branch. The crossover is exaggerated here to illustrate the point, the actual behavior is much smoother in the $T\sim T_c$ region.[]{data-label="cartoon3"}](cartoon-bulk-nonswap){width="\columnwidth"}
We make the hypothesis that the real (observed) relaxation time of the system is the lowest of the two relaxation times, because the dynamics always follows the fastest relaxation channel. We can then get an impression of what happens in Fig. \[cartoon3\]. The observed time follows the MCT branch up to close to $T_c$, where it crosses over to the activated branch, thus avoiding the MCT divergence. This hybridization between MCT and activated branches is (very roughly speaking) the origin of the dynamical crossover near $T_c$ [@review:biroli09].
Consider now what happens to this scenario when we use a swap dynamics. In general the activated relaxation time can be written as, $$\tau_\mathrm{ACT} = \tau_0 \exp(\xi_s^\psi/T) ,
\label{belladonna}$$ where $\xi_s$ is the static correlation length. The effect of swap dynamics is essentially to decrease significantly the prefactor $\tau_0$ in equation [@glassthermo:fernandez06], $$\tau_\mathrm{ACT}^\mathrm{swap} = \tau_0^\mathrm{swap} \exp(\xi_s^\psi/T) ,
\qquad \mbox{with }\tau_0^\mathrm{swap} \ll \tau_0 .
\label{bellaswap}$$ This amounts to a downward shift of the activated branch (Fig. \[cartoon4\]). Due to this, the hybridization between the two branches disappears, and the observed relaxation time does not display any significant crossover close to $T_c$. We also see that if we fix a temperature $T \gtrsim T_c$, in the nonswap case the bulk time is dominated by the MCT channel, whereas in the swap case it is dominated by the activated channel (also see Fig. 1 of ref. [@glassthermo:fernandez06], which shows how the MCT plateau seen in time correlation functions is lost with swap dynamics).
### The cavity
Let us now turn to the cavity, bearing in mind that the large $R$ value of $\tau(R)$ is nothing else than the bulk time, whose behavior we have just examined. It has been suggested that the MCT cavity relaxation time, as a function of $R$, should have a divergence at $R\sim \xi_d$, where $\xi_d$ is the [*dynamic*]{} correlation length [@mosaic:franz07]. A possible interpretation of this fact is that in a smaller cavity the frozen boundary conditions stabilize unstable saddles, thus increasing the MCT relaxation time. Below $\xi_d$ the cavity runs out of saddles and nonactivated relaxation becomes impossible. On the other hand, the activated relaxation time obeys the scenario described by eqs. and : it increases with $R$, saturating at the [*static*]{} correlation length, $\xi_s$.
![When we use a swap dynamics we are significantly lowering the prefactor of activated barrier crossing, hence shifting the activated branch downwards. As a result, there is no significant hybridization between the two branches and the resulting (observed) swap relaxation time does not detect any particular crossover close to $T_c$.[]{data-label="cartoon4"}](cartoon-bulk-swap){width="\columnwidth"}
As for the bulk, we can speculate that the observed relaxation time in the cavity will be the smallest of the two times. Let us fix a temperature slightly above $T_c$, so that the nonswap bulk relaxation is dominated by the MCT channel (Fig. \[cartoon3\]). In Fig. \[cartoon5\] we get a picture of what happens. Let us start from large values of $R$: the relaxation time follows the MCT branch, therefore giving an [*increase*]{} of $\tau(R)$ for decreasing $R$. But at some point the MCT branch crosses the activated one (and it eventually diverges at $\xi_d$), so beyond this point the dynamics sticks to the activated channel, giving rise to a maximum of $\tau(R)$. Hence, for small values of $R$ we recover a regime where $\tau(R)$ decreases for decreasing $R$.
The large $R$ regime of this nonmonotonic curve was also discussed in [@review:biroli09], where it was noted that above $T_c$ $\tau(R)$ should approach its bulk value from above. This behavior, namely a relaxation time that increases from its bulk value when decreasing $R$, is indeed what we find with nonswap dynamics, Fig. \[fig:taunonswap\]. However, in the nonswap case the increase of the relaxation time is so sharp that we struggle to follow this curve down to medium-small $R$, so we cannot access the overshooting.
What happens when we use swap dynamics? As in the bulk, by using swap we are decreasing the prefactor of activation, thus shifting the whole activated branch downwards. From Fig. \[cartoon6\] we see that this shift has the effect to weaken, or even wash out entirely, the nonmonotonic behavior of $\tau(R)$. Something similar happens by lowering the temperature (getting closer to $T_c$), because in that way we are narrowing the difference between the MCT and the activated branch (Fig. \[cartoon3\]). In the cavity, this amounts to closing the gap between the two branches at large $R$. Hence, we expect that lowering $T$ too has the effect to iron out the maximum of $\tau(R)$, eventually making it disappear [^4]. Summarizing, we expect swap dynamics to display little sign of a nonmonotonic cavity relaxation time $\tau(R)$, and to become completely monotonic at low $T$. In Fig. \[fig:overshooting\] we show a close-up of the cavity relaxation time with swap dynamics at two different temperatures: there is an overshooting of $\tau(R)$ at intermediate temperature, but it completely disappears a the lowest $T$. This expectation is therefore supported by the data.
![In the cavity, for $T\gtrsim T_c$ the hybridization between MCT and activated branch may give rise to a nonmonotonic $\tau(R)$. Starting from large cavities, the relaxation time follows the MCT branch, which has a divergence at the dynamical correlation length, $\xi_d$. In the proximity of such divergence the observed $\tau(R)$ crosses over to the activated branch, thus decreasing with decreasing $R$.[]{data-label="cartoon5"}](cartoon-cavity-nonswap){width="\columnwidth"}
![When we use swap dynamics in the cavity we shift the whole activated branch downwards, hence lowering the degree of hybridization of the two branches. In this way, the overshooting of relaxation time may be completely washed out, and $\tau(R)$ have a purely monotonously increasing behavior.[]{data-label="cartoon6"}](cartoon-cavity-swap){width="\columnwidth"}
According to this scenario (admittedly based on little evidence), in the nonswap case one should see an increase of $\tau(R)$ over its bulk value when decreasing the cavity size from large $R$ (saturation from above), whereas in the swap case (and at low $T$) the cavity relaxation time should decrease below its bulk value when decreasing the radius (saturation from below). This prediction seems to be in qualitative agreement with our numerical findings. However, there is a severe problem with this interpretation, namely the fact that nonswap dynamics at very small $R$ is stuck. Even assuming that the great increase of the cavity relaxation time that we observe in going from large $R$ down to medium $R$ (Fig. \[fig:nonswap\]) has to be identified with the large $R$ regime of a nonmonotonic $\tau(R)$ (Fig. \[cartoon5\]), the question remains: why we do not see any hint of the low $R$ regime of Fig. \[cartoon5\], where the cavity relaxation time gets smaller for small radii? It is well possible, in this scenario, that for intermediate $R$ the relaxation time is significantly larger than the bulk limit. However, for [*very*]{} small $R$ the relaxation time should drop again. Yet, we do not see this. In fact, very small cavities are [*completely stuck*]{}, as shown in Fig. \[fig:stuck\]. This phenomenon is in open disagreement with our theoretical expectation. We must address this inconsistency.
The role of boundary rearrangements
-----------------------------------
The fact that swap dynamics thermalizes a small cavity quite rapidly while nonswap dynamics remains stuck, is weird; it indicates that swapping different particles in a small volume becomes prohibitive for standard dynamics. Of course, the exchange of two particles for the standard dynamics is the result of many moves, and it is for sure a more complicated process. Yet swapping two nearby particles is definitely not a terribly collective rearrangement and it should not implicate a very large activation barrier. If it does, it means that this barrier has been made dramatically large by the amorphous boundary conditions. Why is that?
A possible explanation is that by freezing the external configuration we are preventing the surrounding system to elastically accommodate for the small rearrangements within the cavity. Although exchanging two particles is not a collective rearrangement, i.e. one in which many particles move a lot, to happen it still needs that many particles make very small movements. This phenomenon was studied in ref. [@self:prl02], where the distribution of particle displacements in moving from a local energy minimum to nearby one connected by a saddle of order $1$ was calculated. It was found that this process corresponds to [*few*]{} particles (order 2–3) moving an amount comparable to the interparticle distance and [*many*]{} particles moving very little, just to make space to the rearranging ones. Elasticity is also a central ingredient in the local elastic expansion model (also called “shoving model”) of viscous relaxation [@dynamics:dyre96]. More in general, one might argue that the whole short-time dynamics (not only elastic modes) plays a relevant role.
By freezing all the particles in the configuration external to the cavity we are inhibiting this contribution, perhaps making unnaturally large an otherwise modest barrier. Swap dynamics, on the other hand, needs not to pass through the top of a barrier to exchange two particles, and therefore is less affected by the suppression of the high-frequency response, and by the subsequent barrier’s increase. This may be the origin of the very different qualitative behavior of swap vs. nonswap dynamics observed at low $R$.
To check this last hypothesis, we suggest in the next Section a general approach to restore the short-time dynamics which is not limited to the elastic case, suggested by an alternative description of the problem in the RFOT spirit.
Frozen configuration vs. frozen state {#sec:fc-vs-vs}
=====================================
An alternative description of the over-constraining due to the boundary can be given in terms of configurations vs. states. The original aim of the amorphous boundary conditions [@mosaic:bouchaud04] was to keep the system surrounding the cavity within one fixed [*state*]{} (say $\alpha$), one of the exponentially many metastable states the supercooled liquid phase is composed of [@mosaic:kirkpatrick89]. According to this spirit, the external particles should be allowed to move enough to visit the many configurations belonging to state $\alpha$, but not enough to reach configurations that do not belong to $\alpha$. By choosing and fixing just one [*configuration*]{} within state $\alpha$, however, we are over-constraining the amorphous boundary, and this may have some side-effects on the dynamics of small rearrangements in the cavity when a standard dynamics is used.
In view of this, it seems reasonable to try to relax the constraint on the outer particles by changing the current [*frozen configuration*]{} (FC) setup, in favour of a [*frozen state*]{} (FS) one. This means that instead of completely freezing the particles outside the sphere, we let them relax subject to the condition that the overlap $q_\mathrm{ext}$ between the initial external configuration and the one at time $t$ remains at some value $\hat q_\mathrm{ext}$. The FC setup would be recovered taking simultaneously the limits $q_\mathrm{ext}\to1$ and $\ell\to0$. In this way, the external configuration is not allowed to move at all, so this amounts to a complete freezing[^5].
Of course, the choice of $\hat q_\mathrm{ext}$ is critical: with too large a value we go back to the frozen configuration case, while too small a value destroys any point-to-set correlation in the cavity. In fact, in the limit $\hat q_{\mathrm ext}= 0$ the cavity must be ergodic and the overlap must relax to zero for any value of $R$. A sensible physical choice is, $$\hat q_{\mathrm{ext}} = q_{\mathrm{EA}} ,
\label{FSbotto}$$ where $q_{\mathrm EA}$ is the self-overlap of a metastable state. In this way we ensure that the external system does not make any major structural rearrangement, and yet allows for minor movements of the particles, which can have an important elastic effect. To chose the correct value of $q_\mathrm{EA}$ we use the thermodynamic potential $V(q)$ recently discussed in [@self:prl10], whose secondary minimum indicates the value of the self-overlap $q_{\mathrm EA}$. At the temperature $T=0.246$, where we will run the FS simulations, a reasonable choice is (see Fig. 5 of ref. [@self:prl10]), $$q_{\mathrm EA} = 0.4 .$$ Of course, the final test for this choice is that the point-to-set correlation must not be lost: having switched from a FC to a FS setup will certainly imply that the infinite time limit of the overlap, $q(R)$, will be smaller at all values of $R$. What we need is $q(R)\neq 0$ at for some range $[0:R]$, in order to have a physically significant point-to-set correlation function.
Cavity dynamics with frozen state boundary conditions {#sec:cavity-dynamics-with}
=====================================================
FS simulations: technical details
---------------------------------
First, we need to make a technical, but relevant, remark. In this work, as well as in previous works [@self:prl07; @self:nphys08], the overlap is defined in such a way that it *does not* detect the exchange of particles of different size. The same definition has been adopted by other groups [@confinement:berthier11; @mode-coupling:berthier11]. However, we cannot use this definition for imposing the constraint on the external particles: an exchange of two different particles, perhaps quite far from each other, must not be allowed. Hence, the constraint must be imposed on an overlap that is sensitive to the exchange of particles of different kind (whereas we still do not distinguish the exchange of identical particles). Let us call this the [*binary*]{} overlap, defined as $$\label{eq:binary_overlap_def}
q_\mathrm{bin}(t) \equiv \frac{1}{\ell^3 \, N_i} \sum_{i \in
v} \left[ n_i^A(0)n_i^A(t) + n_i^B(0)n_i^B(t)\right] ,$$ where $n_i^X(t)$ is the number of particles of kind $X$ in box $i$. This is also the definition used in [@self:prl10] to compute the thermodynamic potential $V(q)$. In what follows we thus use $q_\mathrm{bin}$ to put the constraint on the outside particles. On the other hand, in order to compare with the previous results, we continue using the standard overlap within the cavity.
Conceptually, FS simulations are straightforward: we simply reject all moves on the external particles that violate constraint . In practice, FS simulations are much more demanding than FC ones, because now we have to update [*all*]{} particles in the system, not simply those within the cavity[^6]. For this reason we restricted our investigation of the frozen state setup to just 3 cavity sizes, $M=20, 50, 100$ particles, corresponding to $R= 1.68,
2.27, 2.88$, and to just one temperature, $T=0.246$.
FS simulation results
---------------------
The first thing we have to check is what happens to the point-to-set correlation function, i.e. to the asymptotic value of the overlap, $q(R)$, in the FS setup at this value of the temperature. To do this we run a swap BIC test in the FS setup, to be sure to get the thermalized asymptotic overlap. We report these values in Table I, where we also report the corresponding values for the standard FC setup. Recall that the effective zero of the overlap, i.e. the value it has for two uncorrelated configurations, is $q_0 = 0.062876$.
------------ ------ -------------------- --------------------
$M$ $R$ $q^\mathrm{FS}(R)$ $q^\mathrm{FC}(R)$
\[1ex\] 20 1.68 0.222 $\pm$ 0.004 0.578 $\pm$ 0.001
50 2.27 0.142 $\pm$ 0.003 0.479 $\pm$ 0.001
100 2.88 0.095 $\pm$ 0.002 0.314 $\pm$ 0.002
\[1ex\]
------------ ------ -------------------- --------------------
: Point-to-set correlation function $q(R)$: FS vs. FC
\[zoccola\]
As expected, there is a significant decrease of $q(R)$ in the FS case, due to the fact that particles in the external configuration are now partly free to move, hence lowering the constraint on the inner particles. However, in the FS case $q(R)$ is still nonzero, so that the PTS correlation function is nontrivial. We stress that the values in Table I have been obtained from a swap BIC test: the lower branch of the BIC test [*grows*]{} with time up to its asymptotic limit. We are therefore quite sure that the FS values of $q(R)$ that we report are nonzero.
Next, we turn to the time series of the overlap $q(t)$ in the FS setup, compared to the FC setup. We stress, once again, that we are using standard nonswap Montecarlo dynamics, both for FC and FS. The data are reported in Fig. \[fig:punch\] for three different values of $R$. In order to make the FS/FC comparison easier, we plot the connected overlap, i.e. the overlap subtracted by the (swap) equilibrium value, $q(R)$. The connected overlap must go to zero for infinite time.
At these values of $T$ and $R$, the FC time series (dashed lines) are completely stuck at an off-equilibrium value, so much as to make it impossible to even estimated the relaxation time. We already observed this phenomenon in Fig. \[fig:stuck\]. On the other hand, the FS time series (full lines) are quite different: the overlap does not remain stuck at any specific level; in fact, it seems to be decaying steadily towards zero. Unfortunately $2$ million Monte Carlo steps (our largest time) are not enough to directly observe the point where the connected overlap goes to zero. However, a reasonable extrapolation suggests that for all three values of $R$ the relaxation time is somewhere between $10^6$ and $10^7$ Monte Carlo steps.
![Frozen configuration (FC) vs. frozen state (FS) setup, standard nonswap dynamics. We plot the connected overlap, obtained by subtracting its equilibrium infinite time limit $q(R)$ (obtained with a swap BIC test). The asymptotic equilibrium value of the connected overlap is zero. The three values of $R$ investigated here are small, so that the FC dynamics (dashed line) is completely stuck at an out of equilibrium level. On the contrary, the FS dynamics (full line) is not stuck and, even though longer runs would be needed, it is approaching equilibrium. $T=0.246$.[]{data-label="fig:punch"}](new_qfrozenstate_portrait_BN){width="\columnwidth"}
![Frozen configuration (FC) vs. frozen state (FS) setup, standard nonswap dynamics. We plot the standard overlap (lines) and the binary overlap (symbols). These two overlaps are the same as long has no particles of different size have been swapped. On the other hand, when the (nonswap) dynamics starts swapping particles, the binary overlap gets smaller than the standard one. This never happens in the FC setup (upper panel), whereas it happens for sufficiently long times in the FS setup (lower panel). This fact explains why the FC dynamics is stuck, while the FS one is not. $R=2.88$, $T=0.246$.[]{data-label="fig:qqbin"}](new_q_vs_qbin_portrait_BN){width="\columnwidth"}
We conclude that the cavity dynamics with frozen state boundary condition no longer remains stuck at an off-equilibrium level. This result goes in the direction we expected: allowing the in-state vibrations of the external configuration unleashes some minor, but necessary, relaxation modes that are otherwise frozen in the FC setup. In particular, what happens is that in the FS case even a nonswap dynamics is able (after a while) to exchange different particles, while in the FC case this never happens. This phenomenon is shown in Fig. \[fig:qqbin\]. We report in this figure the standard cavity overlap, $q(t)$, and the binary cavity overlap, $q_\mathrm{bin}(t)$, which is sensitive to the exchange of different particles. What we see is that in the FC setup (upper panel) the two overlaps coincide up to the longest time, meaning that particles exchanges never happen[^7]. On the other hand, in the FS setup (lower panel) there is a decoupling between the two overlaps at about $5\times 10^5$ Monte Carlo steps (this decoupling is also found in FC swap dynamics —not shown— where it is naturally expected since the swap moves consist precisely in the exchange of two particles of different kind). Hence, even the nonswap dynamics is able to swap particles in the long run, and therefore to relax the cavity, provided that we confine the external system within a state, rather than a configuration.
The last, and most important, open issue is the behavior of the relaxation time as a function of $R$. We recall here the situation schematically summarized in Fig. \[cartoon5\]: the relaxation time for medium $R$ can be significantly larger that the bulk time, due to the hybridization of the MCT and activated branches. However, for small enough $R$ one should go back to a regime where $\tau$ decreases for decreasing $R$. This overshooting scenario is what happens with a local swap dynamics, Fig. \[fig:overshooting\], and our expectation was that it should also happen with a normal nonswap dynamics, provided that we use an FS setup. Is this scenario confirmed or disproved by the data in Fig. \[fig:punch\]?
Longer simulations (at least one order of magnitude longer) and several more values of $R$ and $T$ would be required to clear up unambiguously this matter. However, an unscrupulous extrapolation of the data in Fig. \[fig:punch\] suggests that the intermediate sized cavity, $R=2.27$, has the largest relaxation time, definitely closer to the right side of the $[10^6:10^7]$ window, whereas the smallest and largest cavities, $R=1.68$ and $R=2.88$, both seem to have a smaller relaxation time, closer to the $10^6$ side. In other words, the smallest cavity seems [*not*]{} to be the slowest one. If this were true, it would mean that we are exactly around the maximum of $\tau(R)$ in Fig. \[cartoon5\] and that we are starting to see a hint of the expected decrease of $\tau$ with decreasing $R$. Needless to say, we cannot push this interpretation of the data too far. Let us be content to say that current simulations with frozen state boundary conditions and nonswap dynamics do not rule out the existence of a low $R$ regime where smaller cavities have smaller relaxation times.
Further work to clear up this issue is currently in progress. Unfortunately there is no easy way to settle this. We cannot use smaller cavities, because $M=20$ particles is already about the smallest reasonable size in three dimensions. The only thing to do is to push the simulations at longer times, which is computationally very costly.
Some experimental evidence {#sec:some-exper-evid}
==========================
At the experimental level, there has been considerable interest in studying liquids in confinement conditions, in particular since nanoporous materials with well-defined pore radius have been available (see [@confinement-exp:liu89; @confinement-exp:zhang92; @confinement-exp:arndt97] and references therein) and more recently materials such as carbon nanotubes [@review:rasaiah08].
For liquids confined in nanopores the experimental glass transition temperature (as measured with differential scanning calorimetry) is *reduced* as the pore becomes smaller [@confinement-exp:zhang92], i.e. confined systems are *faster* than the bulk. However, the situation is rather more complicated, as relaxation experiments [@confinement-exp:liu89; @confinement-exp:arndt97] point to the existence of at least two regions in space, with different dynamics: a *slow* layer of molecules directly in contact with the pore walls and a *fast* region inside the cavity and far from the walls.
A particularly interesting case is reported in ref. [@confinement-exp:arndt97]: the relaxation time of salol confined in nanoporous silica glass was found with dielectric relaxation measurements (unable to distinguish the interfacial and central regions of the pore) to be larger for increasing confinement. However, after coating the pore walls with a hydrophobic lubricant (thus reducing the H bonds between salol and the pore surface) it was found that [*smaller*]{} cavities are [ *faster*]{}. In particular, they are significantly faster than the bulk. Hence, in this experimental case, once the interactions that slow down the interfacial layer were supressed, the relaxation time as a function of the radius has a qualitative behavior similar to Fig. \[fig:tau149\]. The authors of ref. [@confinement-exp:arndt97] used this to determine a cooperativity length scale.
There are intriguing similarities, as well as obvious differences, with our case. In both cases the original interaction with the cavity interface was too stiff, suppressing some relaxation channels that are not cooperative, and yet necessary to equilibrate the cavity. The strategy in [@confinement-exp:arndt97] was to lubricate the inside of the cavity, thus hindering the H bonds responsible for the artificial slowing down; our strategy was the make the surrounding system softer. In the experimental case the effect was clear: lubricated cavities are faster than unlubricated ones; smaller cavities are faster than larger cavities. In our case, we also obtain that FS cavities are faster than FC cavities; whether or not smaller cavities are faster than larger ones is unclear, but the data do not rule this out.
The differences are also relevant. In the experimental case the confined (free) system is liquid salol, and the pore is glass. Hence, even though one may say that there are amorphous boundary conditions, these are certainly not drawn from the Gibbs-Boltzmann equilibrium distribution of an external salol system. Moreover, the reasons for the original ‘stiffness’ are also different. In the experimental case it is the formation of H bonds between internal salol and the surface of the pore. In our case, the nature of the bonds between particles within the cavity and across the interface is exactly the same; however, the complete freezing of the cavity suppresses the swap, uncooperative, rearrangements useful to reach equilibrium. Accordingly, the solutions adopted are also different.
We cannot not say whether or not the similarities overcome the differences, so to make this experimental case significant to our context. We limit ourselves to register the fact that the problem of an artificial slowing down in confining geometries has already occurred in experiments and that, when solved, the cavity dynamics can change very dramatically.
Conclusions {#sec:conclusions}
===========
We have studied the dynamics of a confined cavity, using different Montecarlo algorithms and different amorphous boundary conditions. Our bare findings are:
1. FC—swap—low $T$: the cavity relaxation time $\tau$ is larger the larger $R$ and it saturates at $R\sim \xi_s$, where $\xi_s$ is the point-to-set correlation length.
2. FC—swap: in the region $R\sim \xi_s$ a colder cavity relaxes faster than a hotter cavity.
3. FC—swap: at higher $T$ the relaxation time $\tau(R)$ displays an overshooting that disappears on lowering $T$.
4. FC—nonswap: $\tau$ is larger the smaller $R$.
5. FC—nonswap: small cavities ($R< 4$) are completely stuck at an off-equilibrium level.
6. FS nonswap dynamics is significantly faster than FC nonswap dynamics; with FS small cavities are no longer stuck.
7. The FS point-to-set correlation function $q(R)$ is nonzero in the region of interest of $T$ and $R$.
8. FS—nonswap: data are compatible with a nonmonotonic $\tau(R)$; data do not rule out the possibility that in small cavities $\tau$ is smaller the smaller $R$.
We have proposed a theoretical scenario whose aim is to organize all these results into one coherent picture. Our scenario rests on two main ideas. First, depending on the values of $R$ and $T$, and on the type of dynamics, there may be an hybridization between MCT and activated relaxation channels; this hybridization, when present, gives rise to a nonmonotonic cavity relaxation time $\tau(R)$. Second, the frozen configuration setup is unsuitable to run nonswap dynamics, and in general it is not very physical, as it may give rise to an artificial dynamical freezing. We have introduced a frozen state setup, based on the idea that the amorphous boundary condition must select a certain state, not simply a certain configuration. If we trust these two ideas, then we can find an interpretation for the very diverse results we find.
Result 1 supports the concept that $\xi_s$ is the relevant scale of cooperativity in the system. According to the RFOT theory with fluctuating surface tension, the activated relaxation time is equal to the in-state relaxation time for $R\ll\xi_s$, it grows when $R$ gets across the support of the probability distribution of the rearranging sizes $P(R,\xi_s)$, and it finally saturates to its bulk value for $R\gg\xi_s$. Hence, when the cavity is larger than the scale of cooperativity relaxation factorizes, whereas when the cavity is smaller than $\xi_s$ the whole cavity must rearrange collectively. This RFOT interpretation is supported by result 2: an inversion of the relaxation time (cooler is faster) happens because a colder cavity may still be confined within just one state, thus showing only the short, in-state relaxation time, while (at the same value of $R$) a hotter cavity may be already unlocked, thus sporting the full bulk relaxation time.
The maximum displayed by the local swap $\tau(R)$ (result 3) is one piece of evidence in support of the (rather speculative) scenario described in Section VIII: the hybridization between nonactivated MCT channels and activated channels gives rise in the bulk to the crossover between MCT and activation close to $T_c$, while in the cavity it gives rise to a nonmonotonic $\tau(R)$. This hybridization implies that for large $R$ the cavity relaxation time follows the MCT branch, so that $\tau$ is larger the smaller $R$, which is in agreement with the nonswap dynamics result 4. On the other hand, switching to swap dynamics has a twofold effect: in the bulk, it eliminates the $T_c$ crossover; in the cavity, it flattens the maximum of $\tau(R)$.
We have speculated that the complete freezing out of small cavities with nonswap dynamics (result 5) is not quite physical, and we have suggested that it could be the effect of an artificial suppression of some elastic (noncooperative) relaxation modes due to the frozen configuration setup. We have proposed a practical way to implement amorphous boundary conditions with a frozen state and we have found that this setup speeds up significantly the nonswap dynamics, unlocking the small cavities (result 6). We have also checked that the point-to set correlation remains nonzero, despite a significant reduction due to the smaller degree of confinement by the external state (result 7).
Finally, we tried to understand what was the behavior of $\tau$ as a function of $R$ in the frozen state case. This is quite crucial: if we cannot find [*any*]{} regime of $R$ and $T$ where the nonswap $\tau$ is smaller for smaller $R$, then we have a problem. Our entire construction relies on the idea that for small enough $R$ the MCT branch must be gone, so that all that remains is the activated branch, and this must be faster the smaller the cavity. Our time series (Fig. \[fig:punch\]) are too short to settle this matter. But we can at least say that the data do not rule out this possibility (result 8). With a little more optimism, we can even conclude that the smallest cavity is not the slowest one, which is all we need to support our theoretical scenario.
The whole scenario still admits considerable improvements in clarity. As we have said, longer simulation with nonswap dynamics in the FS setup are needed to study carefully $\tau(R)$, and this should be done at several values of $R$ and of $T$. At the same time, FS swap simulations should be run in order to reconstruct the entire point-to-set correlation function, $q(R)$, to check whether or not it retains its essential properties. Is it still a nonexponential function at lower temperature? How does the FS correlation length $\xi_s$ compare to its FC counterpart? Work in this direction is in progress.
We thank L. Berthier, G. Biroli, J.-P. Bouchaud, C. Cammarota, L. Cugliandolo, S. Franz, J.P. Garrahan, I. Giardina, G. Gradenigo, R.L. Jack, A. Heuer, W. Kob, M. Mezard, G. Parisi, G. Tarjus, M. Wyart and F. Zamponi for several important remarks, and ECT\* and CINECA for computer time. The work of TSG was supported in part by grants from ANPCyT, CONICET, and UNLP (Argentina). PV has been partly supported through Research Contract Nos. FIS2009-12648-C03-01,FIS2008-01323 (MICINN, Spain).
[^1]: We neglect in this analysis a possible dependence of $\tau_\beta$ on $R$ due to the extended nature of the excitations related to $\beta$-relaxation [@mosaic:stevenson10].
[^2]: In fact, both surface tension and configurational entropy will fluctuate [@mosaic:dzero05]. At the practical level, though, disentangling the two effects is hard, and given that large surface tension fluctuations have been reported [@self:jstatmech09; @self:jcp09a], a generalized version of that incorporates only surface tension fluctuations seems reasonable [@self:nphys08].
[^3]: This is somewhat similar to the tests introduced by Bhatt and Young [@spin-glass:bhatt88] and later Katzgraber et al. [@spin-glass:katzgraber01] as a thermalization check in simulations of spin glasses.
[^4]: This is a general prediction of our picture: by lowering the temperature we are gradually pushing up (and therefore ruling out) the MCT branch, diminishing the hybridization of the two branches and therefore eliminating the overshooting. At very low $T$, $\tau$ should be a purely increasing function of $R$.
[^5]: We remark though that the FC results reported above are in fact obtained with *bona fide* freezing.
[^6]: Note that also in the FS setup, as in the FC one, we use a hard wall potential enclosing the particles within the cavity. In this way, particles cannot cross the surface of the cavity: whoever is in, stays in, and whoever is out, stays out. This procedure is essential in order to obtain the correct thermodynamic ensemble.
[^7]: Strictly, this means that exchanges of particles *of different kind* do not happen, but same-kind exchanges should be similarly hindered
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let $ L/K $ be a finite separable extension of local or global fields in any characteristic, let $ H_{1}, H_{2} $ be two Hopf algebras giving Hopf-Galois structures on the extension, and suppose that the actions of $ H_{1}, H_{2} $ on $ L $ commute. We show that a fractional ideal $ {\mathfrak B}$ of $ L $ is free over its associated order in $ H_{1} $ if and only if it is free over its associated order in $ H_{2} $. We also study which properties these associated orders share.'
address: |
School of Computing and Mathematics\
Keele University\
Staffordshire\
ST5 5BG\
UK
author:
- 'Paul J. Truman'
title: 'Commuting Hopf-Galois Structures on a Separable Extension'
---
Introduction and Statement of Results {#section_introduction}
=====================================
This paper is a sequel to [@PJT_Canonical]. In the introduction to that paper, we described one of the ways in which Hopf-Galois module theory generalizes the classical Galois module theory of algebraic integers. Given a finite Galois extension of local or global fields $ L/K $ with Galois group $ G $, the group algebra $ K[G] $ is a Hopf algebra, and its action on $ L $ is an example of a [*Hopf-Galois structure*]{} on the extension $ L/K $ (see [@ChTwe Definition 2.7]). However, the extension may admit a number of other Hopf-Galois structures, and each of these provides a different context in which we can ask module theoretic questions about the extension and its fractional ideals. If $ H $ is a Hopf algebra giving a Hopf-Galois structure on $ L/K $ then $ L $ is a free $ H $-module of rank $ 1 $, ([@ChTwe Proposition 2.15] and if $ {\mathfrak B}$ is a fractional ideal of $ L $ then we can define the associated order of $ {\mathfrak B}$ in $ H $: $${{\mathfrak A}_{H}}({\mathfrak B}) = \{ h \in H \mid h \cdot x \in {\mathfrak B}\mbox{ for all } x \in {\mathfrak B}\},$$ and compare the structure of $ {\mathfrak B}$ as a module over its associated orders in the various Hopf algebras. The most interesting case is $ {\mathfrak B}= {{\mathfrak O}_{L}}$, the ring of algebraic integers of $ L $, and there exist Galois extensions $ L/K $ of $ p $-adic fields for which $ {{\mathfrak O}_{L}}$ is not free over its associated order in the group algebra $ K[G] $ but is free over its associated order in some Hopf algebra giving a different Hopf-Galois structure on the extension [@By97a]. On the other hand, there exist extensions admitting multiple Hopf-Galois structures for which $ {{\mathfrak O}_{L}}$ is free over its associated order in each of the corresponding Hopf algebras [@PJT_Noether].\
\
Another way in which Hopf-Galois module theory generalizes classical Galois module theory is that a finite separable, but non-normal, extension of local or global fields may admit Hopf-Galois structures; here of course the techniques of classical Galois module theory are not available, but these Hopf-Galois structures allow us to study such extensions and their fractional ideals as described above.\
\
A theorem of Greither and Pareigis ([@Greither_Pareigis Theorem 3.1] or [@ChTwe Theorem 6.8]) enumerates and describes all of the Hopf-Galois structures admitted by a given finite separable extension of fields $ L/K $, as follows: let $ E/K $ be the Galois closure of $ L/K $, $ G={\mbox{Gal}(E/K)} $, and $ G_{L} = {\mbox{Gal}(E/L)} $. Now let $ X $ be the left coset space of $ G_{L} $ in $ G $, and $ B = {\mbox{Perm}(X)} $. Then the Hopf-Galois structures on $ L/K $ are in bijective correspondence with the regular subgroups of $ B $ normalized the image of $ G $ under the embedding $ \lambda : G \rightarrow B $ defined by $ \lambda(\sigma) gG_{L} = \sigma g G_{L} $ . (Recall that a subgroup of $ B $ is called [*regular*]{} if it has the same order as $ X $ and is transitive on $ X $.) Furthermore, the Hopf algebra corresponding to a regular subgroup $ N $ is $ E[N]^{G} $, where $ G $ acts on $ E $ as Galois automorphisms and on $ N $ by conjugation via the embedding $ \lambda $: $$\label{eqn_GP_conj}
\,^{g}\eta = \lambda(g) \eta \lambda (g^{-1}) \mbox{ for all } g \in G, \; \eta \in N.$$ Finally, such a Hopf algebra acts on $ L $ by $$\label{eqn_GP_action}
\left( \sum_{\eta \in N} c_{\eta} \eta \right) \cdot x = \sum_{\eta \in N} c_{\eta} \eta^{-1}({\overline}{1})[x] \hspace{5mm} (c_{\eta} \in E, \; x \in L),$$ where $ {\overline}{1} $ denotes the coset $ G_{L} $ in $ X $.\
\
If $ L/K $ is a Galois extension then in the notation above $ X = G $ and $ \lambda $ is the left regular embedding of $ G $ into $ {\mbox{Perm}(G)} $. In this case, one example of a regular subgroup of $ {\mbox{Perm}(G)} $ normalized by $ \lambda(G) $ is $ \rho(G) $, the image of $ G $ under the right regular embedding. Since $ \lambda(G) $ centralizes $ \rho(G) $ inside $ {\mbox{Perm}(G)} $, the Hopf-Galois structure corresponding to the regular subgroup $ \rho(G) $ has Hopf algebra isomorphic to $ K[G] $, and from equation we recover the usual action of $ K[G] $ on $ L $ (see [@ChTwe Proposition 6.10]). We call this Hopf-Galois structure the [*classical structure*]{}. If $ G $ is abelian then in fact $ \lambda(G) = \rho(G) $, but if $ G $ is nonabelian then they are distinct, and $ \lambda(G) $ is another example of a regular subgroup of $ {\mbox{Perm}(G)} $ normalized by $ \lambda(G) $, and therefore corresponds to a different Hopf-Galois structure on $ L/K $. In [@PJT_Canonical] we called this the [*canonical nonclassical structure*]{}, and denoted the corresponding Hopf algebra $ L[\lambda(G)]^{G} $ by $ H_{\lambda} $. We proved two theorems concerning the relationship between the classical structure and the canonical nonclassical structure: an element $ x \in L $ generates $ L $ as a $ K[G] $ module if and only if it generates $ L $ as an $ H_{\lambda} $-module ([@PJT_Canonical Theorem 2.1]), and, assuming that $ {{\mathfrak O}_{K}}$ is a principal ideal domain, an ambiguous ideal of $ L $ is free over its associated order in $ K[G] $ if and only if it is free over its associated order in $ H_{\lambda} $ ([@PJT_Canonical Theorem 1.1] and its correction, [@PJT_Canonical_Correction]). The proofs of these results exploited the fact that $$h \cdot g(x) = g(h \cdot x) \mbox{ for all } h \in H_{\lambda}, \; g \in G, \; x \in L$$ (see [@PJT_Canonical Lemma 3.3]). In this paper we generalize the results of [@PJT_Canonical] to pairs of Hopf-Galois structures on a given separable (but not necessarily Galois) extension whose actions on $ L $ [*commute*]{}, by which we mean that $$h_{1} \cdot ( h_{2} \cdot x ) = h_{2} \cdot ( h_{1} \cdot x ) \mbox{ for all } h_{1} \in H_{1}, \; h_{2} \in H_{2}, \; x \in L.$$ Furthermore, our results apply to all fractional ideals of $ L $ (not just to ambiguous ideals), and we can remove the hypothesis that $ {{\mathfrak O}_{K}}$ is a principal ideal domain.\
\
Our results are as follows:
\[thm\_NB\_generators\_intro\] Let $ L/K $ be a finite separable extension of fields and let $ H_{1},H_{2} $ be two Hopf algebras giving Hopf-Galois structures on $ L/K $ whose actions commute. Then an element $ x \in L $ generates $ L $ as an $ H_{1} $-module if and only if it generates $ L $ as an $ H_{2} $-module.
\[thm\_main\_thm\_intro\] Retain the notation and hypotheses of Theorem \[thm\_NB\_generators\_intro\], and assume that $ L/K $ is an extension of local or global fields (in any characteristic). Let $ {\mathfrak B}$ be a fractional ideal of $ L $, and let $ {{\mathfrak A}}_{1} $, $ {{\mathfrak A}}_{2} $ be the associated orders of $ {\mathfrak B}$ in $ H_{1} $, $ H_{2} $ respectively. Then $ {\mathfrak B}$ is a free $ {{\mathfrak A}}_{1} $-module if and only if it is a free $ {{\mathfrak A}}_{2} $-module.
If $ L/K $ is an extension of global fields then $ {\mathfrak B}$ is a locally free $ {{\mathfrak A}}_{1} $-module if and only if it is a locally free $ {{\mathfrak A}}_{2} $-module.
The proof of Theorem \[thm\_main\_thm\_intro\] does not depend on the fact that $ L $ is a field, so we may replace $ L $ with its completion at some prime $ \p $ of $ {{\mathfrak O}_{K}}$ (a Galois algebra). To say that $ {\mathfrak B}$ is locally free over $ {{\mathfrak A}}_{1} $ is to say that for each prime $ \p $ of $ {{\mathfrak O}_{K}}$, we have $ {\mathfrak B}_{\p} = \OKp \otimes_{{{\mathfrak O}_{K}}} {\mathfrak B}$ is a free $ \OKp \otimes_{{{\mathfrak O}_{K}}} {{\mathfrak A}}_{1} $-module. By Theorem \[thm\_main\_thm\_intro\] this is equivalent to saying that for each prime $ {\mathfrak B}_{\p} $ is a free $ \OKp \otimes_{{{\mathfrak O}_{K}}} {{\mathfrak A}}_{2} $-module, which is saying that $ {\mathfrak B}$ is a locally free $ {{\mathfrak A}}_{2} $-module.
In section \[section\_commuting\_structures\] we characterize Hopf-Galois structures on a given separable extension whose actions commute, using the theorem of Greither and Pareigis. In section \[section\_NBG\_module\_structure\] we generalize some of the results of [@PJT_Canonical Section 2], and use these to prove Theorems \[thm\_NB\_generators\_intro\] and \[thm\_main\_thm\_intro\]. Finally, in section \[section\_shared\_properties\] we ask, under the hypotheses of Theorem \[thm\_main\_thm\_intro\], which algebraic properties the associated orders $ {{\mathfrak A}}_{1} $ and $ {{\mathfrak A}}_{2} $ share. We show that $ {{\mathfrak A}}_{1} $ is a maximal order in $ H_{1} $ if and only if $ {{\mathfrak A}}_{2} $ is a maximal order in $ H_{2} $ (Proposition \[prop\_maximal\_orders\]), but that it is possible for $ {{\mathfrak A}}_{1} $ to be a Hopf order but for $ {{\mathfrak A}}_{2} $ not to have this property (Example \[example\_hopf\]).\
\
I would like to express my gratitude to Prof. Nigel Byott, who first suggested to me that it might be possible to generalize the results of [@PJT_Canonical] in this way.
Commuting Hopf-Galois Structures {#section_commuting_structures}
================================
Let $ L/K $ be a finite separable extension of fields. In this section we use the theorem of Greither and Pareigis to characterize Hopf-Galois structure on $ L/K $ whose actions on $ L $ commute. Although we are principally interested in extensions of local or global fields, we do not impose this hypothesis since the the results of this section are valid more generally. We retain the hypotheses and notation used in the theorem of Greither and Pareigis, as described in the introduction: $ E/K $ is the Galois closure of $ L/K $, $ G={\mbox{Gal}(E/K)} $, $ G_{L} = {\mbox{Gal}(E/L)} $, $ X = \{ gG_{L} \mid g \in G \} $, $ B = {\mbox{Perm}(X)} $, and $ \lambda : G \rightarrow B $ is defined by $ \lambda(\sigma)gG_{L} = \sigma g G_{L} $. We shall always think of a Hopf algebra produced by the theorem of Greither and Pareigis as acting via equation ; hence we may refer without ambiguity to [*the*]{} Hopf-Galois structure given by a particular Hopf algebra. In the introduction we noted that if $ L/K $ is a Galois extension with Galois group $ G $ then $ \lambda(G) $ and $ \rho(G) $ are both regular subgroups of $ B $ normalized by $ \lambda(G) $, and therefore correspond to Hopf-Galois structures on $ L/K $. (If $ G $ is abelian then these subgroups coincide, and so both yield the same Hopf-Galois structure.) We also noted that $ \lambda(G) $ centralizes $ \rho(G) $ inside $ B $; in fact $ \lambda(G) = {\mbox{Cent}_{B}\left(\rho(G)\right)} $. This relationship between $ \lambda(G) $ and $ \rho(G) $ in the Galois case is a particular example of a more general phenomenon:
\[lem\_N\_N\_prime\] If $ N $ is a regular subgroup of $ B $ which is normalized by $ \lambda(G) $, then so is $ N^{\prime} = {\mbox{Cent}_{B}\left(N\right)} $.
The subgroup $ N^{\prime} $ is explicitly constructed in [@Greither_Pareigis Lemma 2.4.2] as follows: since $ N $ is regular on $ X $, for each coset $ {\overline}{g} \in X $ there is a unique element $ \mu_{{\overline}{g}} \in N $ such that $ \mu_{{\overline}{g}}({\overline}{1})={\overline}{g} $. For each $ \eta \in N $ define $ \phi_{\eta} \in B $ by $ \phi_{\eta}({\overline}{g})=\mu_{{\overline}{g}}(\eta({\overline}{1})) $; then $ N^{\prime} = \{ \phi_{\eta} \mid \eta \in N \} $. From this it is easy to verify that $ N^{\prime} $ is a regular subgroup of $ B $. The proof that $ N^{\prime} $ is normalized by $ \lambda(G) $ appears as part of the proof of [@Greither_Pareigis Theorem 2.5]. These facts are also established in [@Kohl_4p Proposition 3.2, Corollary 3.9].
If $ N $ is a regular subgroup of $ X $ normalized by $ \lambda(G) $, and $ H $ is the Hopf algebra giving the corresponding Hopf-Galois structure, let $ N^{\prime} = {\mbox{Cent}_{B}\left(N\right)} $ and $ H^{\prime} = E[N^{\prime}]^{G} $.
We record some properties of the relationships between the regular subgroups $ N $ and $ N^{\prime} $ and between the Hopf algebras $ H $ and $ H^{\prime} $:
\[lemma\_properties\_of\_N\_opp\] Let $ N $ be a regular subgroup of $ B $ which is normalized by $ \lambda(G) $. Then:
1. We may identify $ N^{\prime} $ with $ N^{opp} $,
2. $ N^{\prime} \cong N $,
3. $ \left( N^{\prime} \right)^{\prime} = N $,
4. $ \left( H^{\prime} \right)^{\prime} = H $,
5. $ N = N^{\prime} $ if and only if $ N $ is abelian,
6. The Hopf-Galois structures given by $ H $ and $ H^{\prime} $ coincide if and only if $ H $ is commutative.
<!-- -->
1. 2. See [@Greither_Pareigis Lemma 2.4.2].
3. In the notation of the proof of Lemma \[lem\_N\_N\_prime\], the map $ \eta \mapsto \phi_{\eta^{-1}} $ is an isomorphism from $ N $ to $ N^{\prime} $.
4. See [@Kohl_4p Lemma 3.5].
5. This follows from part (iii) and the definition of $ H^{\prime} $.
6. If $ N $ is abelian then $ N \subseteq N^{\prime} $, but these groups have the same order, so in fact $ N = N^{\prime} $. For the converse, we recall from [@Kohl_4p Proposition 3.3] that $ N \cap N^{\prime} = Z(N) $, the center of $ N $, so if $ N=N^{\prime} $ then $ N $ is abelian.
7. This follows from part (v) and the definition of $ H^{\prime} $.
Since we may identify $ N^{\prime} $ with $ N^{opp} $, we may identify the group algebra $ E[N^{\prime}] $ with the opposite ring $ E[N]^{opp} $. However, we may not identify the Hopf algebra $ H^{\prime}$ with the opposite ring $ H^{opp} $, since the action of $ G $ on $ N^{opp} $ (see equation ) may not be the same as its action on $ N $. This is already apparent in the case that $ L/K $ is Galois, $ N = \rho(G) $ and $ N^{opp} = \lambda(G) $. Since $ \lambda(G) $ is the centralizer of $ \rho(G) $ in $ B $, the action of $ G $ on $ \rho(G) $ is trivial, and so $ L[\rho(G)]^{G} = L^{G}[\rho(G)] \cong K[G] $. However, the action of $ G $ on $ \lambda(G) $ is not trivial (the orbits are the conjugacy classes), and so $ L[\lambda(G)]^{G} \neq K[\lambda(G)] $. Therefore in this case we may not identify $ L[\lambda(G)]^{G} $ with $ K[\rho(G)]^{opp} $.\
\
We shall show that if $ H $ is a Hopf algebra giving a Hopf-Galois structure on $ L/K $ then the actions of $ H $ and $ H^{\prime} $ on $ L $ commute (Proposition \[prop\_actions\_commute\]) and, conversely, that if $ H_{1} $ and $ H_{2} $ are two Hopf algebras giving Hopf-Galois structure on $ L/K $ whose actions on $ L $ commute then $ H_{2} = H_{1}^{\prime} $ (Proposition \[prop\_actions\_commute\_converse\]). To do this, we recall some notation employed in the proof of the theorem of Greither and Pareigis, as detailed in [@ChTwe §6]. Let $ M = \mbox{Map}(X,E) $, and let $ \{ u_{{\overline}{g}} \mid {\overline}{g}\in X \} $ be an $ E $-basis of mutually orthogonal idempotents (where $ {\overline}{g} $ denotes the left coset $ gG_{L} \in X $). That is: $$u_{{\overline}{g}}({\overline}{\sigma}) = \delta_{{\overline}{g},{\overline}{\sigma}} \mbox{ for all } {\overline}{g},{\overline}{\sigma} \in X.$$ It can be shown [@ChTwe Theorem 6.3] that the $ E $-Hopf algebras giving Hopf-Galois structures on the extension of rings $ M / E $ are precisely the group algebras $ E[N] $ of regular subgroups $ N $ of $ {\mbox{Perm}(G)} $, where the group $ N $ acts on $ M $ by permuting the subscripts of the idempotents $ u_{{\overline}{g}} $: $$\eta \cdot u_{{\overline}{g}} = u_{\eta({\overline}{g})} \mbox{ for any } \eta \in N \mbox{ and } {\overline}{g} \in X.$$ If in addition $ N $ is normalized by $ \lambda(G) $ then the group $ G $ acts on $ E[N] $ by acting on $ E $ as Galois automorphisms and on $ N $ by conjugation via the image of the embedding $ \lambda $ into $ {\mbox{Perm}(X)} $. It also acts on $ M $ by acting on $ E $ as Galois automorphisms and on the idempotents $ u_{{\overline}{g}} $ by left translation of the subscripts. Now by Galois descent we obtain that $ E[N]^{G} $ is a $ K $-Hopf algebra giving a Hopf-Galois structure on the extension of rings $ M^{G} / K $. Note also that $ E \otimes_{K} E[N]^{G} = E[N] $ and $ E \otimes_{K} M^{G} = M $. Finally, we may identify $ L $ with the fixed ring $ M^{G} $ via the $ K $-algebra isomorphism $ L \xrightarrow{\sim} M^{G} $ defined by $$\label{eqn_L_MG_isomorphism}
x \mapsto f_{x} = \sum_{{\overline}{g} \in X} g(x) u_{{\overline}{g}} \mbox{ for all } x \in L.$$ (For $ {\overline}{g} \in X $, the element $ g(x) $ is well defined since $ g_{0}(t)=t $ for all $ g_{0} \in G_{L} $ and $ t \in L $). Thus $ E[N]^{G} $ gives a Hopf-Galois structure on $ L/K $, with the action of $ E[N]^{G} $ on $ L $ as given in equation (\[eqn\_GP\_action\]).\
\
With this notation to hand, we show that if $ H $ is a Hopf algebra giving a Hopf-Galois structure on $ L/K $ then the actions of $ H $ and $ H^{\prime} $ on $ L $ commute. We then show that any pair of Hopf algebras giving Hopf-Galois structures on $ L/K $ whose actions on $ L $ commute arises in this way.
\[prop\_actions\_commute\] Let $ H $ be a Hopf algebra giving a Hopf-Galois structure on $ L/K $, Then the actions of $ H $ and $ H^{\prime} $ on $ L $ commute.
Since $ \eta^{\prime} \eta = \eta \eta^{\prime} $ for all $ \eta \in N $ and $ \eta^{\prime} \in N^{\prime} $, we have $ z^{\prime} \cdot ( z \cdot f ) = z \cdot ( z^{\prime} \cdot f ) $ for all $ z \in E[N] $, $ z^{\prime} \in E[N^{\prime}] $ and $ f \in M $. Therefore it is certainly true that $ h^{\prime} \cdot ( h \cdot f ) = h \cdot ( h^{\prime} \cdot f ) $ for all $ h \in E[N]^{G} $, $ h^{\prime} \in E[N^{\prime}]^{G} $ and $ f \in M^{G} $, and so the actions of $ H $ and $ H^{\prime} $ on $ L $ commute.
\[prop\_actions\_commute\_converse\] If $ H_{1} = E[N_{1}]^{G} $ and $ H_{2} = E[N_{2}]^{G} $ are two Hopf algebras giving Hopf-Galois structures on the extension $ L/K $ whose actions on $ L $ commute, then we have $ N_{2} = N_{1}^{\prime} $.
Since the actions of $ H_{1}, H_{2} $ on $ L $ commute, the actions of $ H_{1}, H_{2} $ on $ M^{G} $ commute. Since $ E \otimes_{K} M^{G} = M $ and $ E \otimes_{K} H_{i} = E[N_{i}] $ for $ i=1,2 $, this implies that the actions of $ E[N_{1}], E[N_{2}] $ on $ M $ commute. Therefore for all $ \eta_{1} \in N_{1} $, $ \eta_{2} \in N_{2} $ we have $$\begin{aligned}
&& \eta_{1} \eta_{2} u_{{\overline}{g}} = \eta_{2} \eta_{1} u_{{\overline}{g}} \mbox{ for all } {\overline}{g} \in X \\
&\Rightarrow & \eta_{1}^{-1} \eta_{2}^{-1} \eta_{1} \eta_{2} {\overline}{g} ={\overline}{g} \mbox{ for all } {\overline}{g} \in X \\
&\Rightarrow & \eta_{1}^{-1} \eta_{2}^{-1} \eta_{1} \eta_{2} = 1 \\
&\Rightarrow & \eta_{1} \eta_{2} = \eta_{2} \eta_{1}. \end{aligned}$$ Therefore $ N_{2} \subset N_{1}^{\prime} $. But $ N_{2} $ and $ N_{1}^{\prime} $ have the same order, so in fact they must be equal.
Normal Basis Generators and Integral Module Structure {#section_NBG_module_structure}
=====================================================
We continue to assume that $ L/K $ is a finite separable extension of fields, and retain the notation established in section \[section\_introduction\] concerning the statement of the theorem of Greither and Pareigis, as well as the elements of its proof that we used in section \[section\_commuting\_structures\]. In this section we prove Theorems \[thm\_NB\_generators\_intro\] and \[thm\_main\_thm\_intro\], which are our main results. We require slight generalizations of two lemmas from [@PJT_Canonical]:
\[lem\_fixed\_generators\] Let $ N $ be a regular subgroup of $ {\mbox{Perm}(X)} $ normalized by $ \lambda(G) $, so that $ E[N]^{G} $ is a $ K $-Hopf algebra giving a Hopf-Galois structure on $ M^{G} / K $. An element $ f \in M^{G} $ is an $ E[N]^{G} $-generator of $ M^{G} $ if and only if it is an $ E[N] $-generator of $ M $.
This is a slight generalization of [@PJT_Canonical Lemma 2.2], with essentially the same proof.
\[lem\_transition\_matrix\] Fix orderings of the set $ X $ and the group $ N $. For $ x \in L $, the element $ f_{x} $ is an $ E[N] $-generator of $ M $ if and only if the matrix $$T_{N}(x) = ( \eta({\overline}{g})[x] )_{\eta \in N, \; {\overline}{g} \in X}$$ is nonsingular.
This is a slight generalization of [@PJT_Canonical Lemma 2.3], with essentially the same proof. We note that although the definition of the matrix $ T_{N}(x) $ depends on the orderings of $ G $ and $ N $, the question of whether it is nonsingular does not.
We now use the characterization of commuting Hopf-Galois structures from section \[section\_commuting\_structures\] to prove Theorem \[thm\_NB\_generators\_intro\]: which generalizes [@PJT_Canonical Theorem 1.2]:
By the theorem of Greither and Pareigis, $ H_{1} = E[N]^{G} $ for some regular subgroup $ N $ of $ X $ normalized by $ \lambda(G) $, and by Proposition \[prop\_actions\_commute\_converse\] we have $ H_{2} = H_{1}^{\prime} = E[N^{\prime}]^{G} $. Thus by Lemmas \[lem\_fixed\_generators\] and \[lem\_transition\_matrix\] it is sufficient to show that $ T_{N}(x) $ is nonsingular if and only if $ T_{N^{\prime}}(x) $ is nonsingular. We have: $$\begin{aligned}
\det( T_{N}(x)) & = & \det \left( ( \eta({\overline}{g})[x] )_{\eta \in N, \; {\overline}{g} \in X} \right) \\
& = & \det \left( ( \eta(\eta^{\prime}({\overline}{1_{G}}))[x] )_{\eta \in N, \; \eta^{\prime} \in N^{\prime}} \right) \mbox{ since } N^{\prime} \mbox{ is regular on } X \\
& = & \det \left( ( \eta^{\prime}(\eta({\overline}{1_{G}}))[x] )_{\eta \in N, \; \eta^{\prime} \in N^{\prime}} \right) \mbox{ since } N, N ^{\prime} \mbox{ commute inside } B \\
& = & \det \left( ( \eta^{\prime}({\overline}{g})[x] )_{\eta^{\prime} \in N^{\prime}, \; {\overline}{g} \in X} \right) \mbox{ since } N \mbox{ is regular on } X \\
& = & \det( T_{N^{\prime}}(x))\end{aligned}$$ Therefore $ T_{N}(x) $ is nonsingular if and only if $ T_{N^{\prime}}(x) $ is nonsingular, which completes the proof.
The results we have established so far are valid for any finite separable extension of fields. We now turn to questions of integral Hopf-Galois module structure in a finite separable extension of local or global fields $ L/K $. Note, though, that we make no restriction on the characteristic of $ K $. Using the notation established in section \[section\_commuting\_structures\], we will prove Theorem \[thm\_main\_thm\_intro\], which generalizes [@PJT_Canonical Theorem 1.1]. First we note that if $ H $ is a Hopf algebra giving a Hopf-Galois structure on $ L/K $ and $ {\mathfrak B}$ is a fractional ideal of $ L $ then $ {\mathfrak B}$ has an associated order in $ H $: since $ L $ is a free $ H $-module of rank $ 1 $ ([@ChTwe Proposition 2.15]), we can identify $ {\mathfrak B}$ with a full $ {{\mathfrak O}_{K}}$-lattice in $ H $, and the left multiplier ring of this lattice is an order in $ H $ [@Reiner_MO Chapter 2, §8], which we can identify with the associated order of $ {\mathfrak B}$ in $ H $. Thus the assumption in [@PJT_Canonical Theorem 1.1] that the fractional ideal $ {\mathfrak B}$ is an ambiguous ideal of $ L $ is superfluous.
Let $ H_{1}=H $; then by Proposition \[prop\_actions\_commute\_converse\] we have $ H_{2} = H^{\prime} $. Suppose that $ x \in {\mathfrak B}$ generates $ {\mathfrak B}$ as an $ {{\mathfrak A}_{H}}$-module. Then $ x $ generates $ L $ as an $ H $-module, and so by Theorem \[thm\_NB\_generators\_intro\] it generates $ L $ as an $ H^{\prime} $-module. Therefore for each $ a \in {{\mathfrak A}_{H}}$ we may define $ z_{a} \in H^{\prime} $ by $ z_{a} \cdot x = a \cdot x $. We claim that $${{\mathfrak A}_{H}}^{\prime} = \{ z_{a} \mid a \in {{\mathfrak A}_{H}}\}.$$ First we show that $ z_{a} \in {{\mathfrak A}_{H}}^{\prime} $ for each $ a \in {{\mathfrak A}_{H}}$. Let $ a \in {{\mathfrak A}_{H}}$ and $ b \in {\mathfrak B}$. Since $ {\mathfrak B}$ is a free $ {{\mathfrak A}_{H}}$-module, there exists a unique $ w \in {{\mathfrak A}_{H}}$ such that $ b = w \cdot x $. Now we have: $$\begin{aligned}
z_{a} \cdot b & = & z_{a} \cdot ( w \cdot x ) \\
& = & w \cdot ( z_{a} \cdot x ) \mbox{ since the actions of } H, H^{\prime} \mbox{ on } L \mbox{ commute}\\
& = & w \cdot ( a \cdot x ),\end{aligned}$$ and this lies in $ {\mathfrak B}$ since $ a,w \in {{\mathfrak A}_{H}}$. Therefore $ z_{a} \in {{\mathfrak A}_{H}}^{\prime} $.\
\
On the other hand, if $ z \in {{\mathfrak A}_{H}}^{\prime} $ then $ z \cdot x = z_{a} \cdot x $ for some $ a \in {{\mathfrak A}_{H}}$, and this implies that $ z = z_{a} $ because $ x $ generates $ L $ as a free $ H^{\prime} $-module. Therefore $ {{\mathfrak A}_{H}}^{\prime} = \{ z_{a} \mid a \in {{\mathfrak A}_{H}}\} $, as claimed.\
\
Since $ (H^{\prime})^{\prime} = H $, the converse statement follows by interchanging the roles of $ H, H^{\prime} $ in the argument above.
In [@PJT_Canonical], we split the proof of Theorem 1.1 into two propositions: Proposition 4.1 and Proposition 4.2. In the proofs of those propositions we assumed that certain $ {{\mathfrak O}_{K}}$ orders in certain $ K $-algebras possessed $ {{\mathfrak O}_{K}}$-bases. Although this is automatically true if $ L/K $ is an extension of local fields, it need not be true if $ L/K $ is an extension of global fields. In order for these proofs to be correct as written, we therefore issued a correction (see [@PJT_Canonical_Correction]) adding the hypothesis that $ {{\mathfrak O}_{K}}$ must be a principal ideal domain. In the proof of Theorem \[thm\_main\_thm\_intro\] above, we made no such assumptions about $ {{\mathfrak O}_{K}}$-bases, and so this hypothesis is no longer needed. In particular, [@PJT_Canonical Theorem 1.1] is valid as stated.
Shared Properties of Associated Orders {#section_shared_properties}
======================================
In this section we adopt the hypotheses of Theorem \[thm\_main\_thm\_intro\]: $ L/K $ is a finite separable extension of local or global fields, $ {\mathfrak B}$ is a fractional ideal of $ L $, $ H $ is a Hopf algebra giving a Hopf-Galois structure on $ L/K $, and $ H^{\prime} $ is the Hopf algebra whose action on $ L $ commutes with that of $ H $. We have seen that $ {\mathfrak B}$ is free over $ {{\mathfrak A}_{H}}$ if and only if it is free over $ {{\mathfrak A}_{H}}^{\prime} $, but we might wonder whether these orders share any algebraic properties. If $ K $ has characteristic zero, then the question of whether $ {{\mathfrak A}_{H}}$ is a Hopf order in $ H $ is particularly interesting, since in this case $ {{\mathfrak O}_{L}}$ is locally free over $ {{\mathfrak A}_{H}}$ by a theorem of Childs [@ChTwe Theorem 13.4]. However, there are examples of extensions for which $ {{\mathfrak A}_{H}}$ is a Hopf order but $ {{\mathfrak A}_{H}}^{\prime} $ is not:
\[example\_hopf\] Let $ p,q,r $ be prime numbers with $ r \equiv 1 \pmod{q} $ and $ p { \hspace{0mm} \nmid }qr $, and let $ L/K $ be a tamely ramified Galois extension of $ p $-adic fields whose Galois group $ G $ is isomorphic to the metacyclic group of order $ qr $. For a fixed integer $ d $ having order $ q $ modulo $ r $, we may write $$G = \langle \sigma, \tau \mid \sigma^{r} = \tau^{q} = 1, \; \tau \sigma = \sigma^{d} \tau \rangle.$$ The extension $ L/K $ is Galois with nonabelian Galois group, so it admits the classical Hopf-Galois structure, with Hopf algebra $ K[G] $, and the canonical nonclassical structure, with Hopf algebra $ H_{\lambda} = L[\lambda(G)]^{G} $. Since $ L/K $ is tamely ramified, Noether’s Theorem implies that $ {{\mathfrak O}_{L}}$ is a free module over $ {{\mathfrak O}_{K}}[G] $, which is a Hopf order in $ K[G] $. By Theorem \[thm\_main\_thm\_intro\], $ {{\mathfrak O}_{L}}$ is free over its associated order $ {{\mathfrak A}}_{\lambda} $ in $ H_{\lambda} $, but we shall show that this is not a Hopf order in $ H_{\lambda} $.\
\
By [@PJT_Noether Proposition 2.5], we have $ {{\mathfrak O}_{L}}[\lambda(G)]^{G} \subseteq {{\mathfrak A}}_{\lambda} $. We begin by showing that $ {{\mathfrak O}_{L}}[\lambda(G)]^{G} $ is not a Hopf order in $ H_{\lambda} $. By [@PJT_Integral Corollary 2.2] $ {{\mathfrak O}_{L}}[\lambda(G)]^{G} $ is a Hopf order in $ H_{\lambda} $ if and only if the inertia subgroup of $ G $ is contained in the center of $ G $. Since $ G $ is nonabelian, the extension $ L/K $ is neither unramified nor totally ramified, and so the inertia subgroup must be the unique nontrivial normal subgroup of $ G $, which is $ \langle \sigma \rangle $. But it is easy to see that $ G $ has trivial center, so the inertia subgroup is not contained in the center of $ G $, and so $ {{\mathfrak O}_{L}}[\lambda(G)]^{G} $ is not a Hopf order in $ H_{\lambda} $.\
\
Next we suppose that $ {{\mathfrak A}}_{\lambda} $ is a Hopf order in $ H_{\lambda} $ properly containing $ {{\mathfrak O}_{L}}[\lambda(G)]^{G} $. Then $ {{\mathfrak O}_{L}}\otimes_{{{\mathfrak O}_{K}}} {{\mathfrak A}}_{\lambda} $ is a Hopf order in $ L \otimes_{K} L[\lambda(G)]^{G} = L[\lambda(G)] $ properly containing $ {{\mathfrak O}_{L}}[\lambda(G)] $. But $ |G| = qr $ and $ p { \hspace{0mm} \nmid }qr $, so $ L[\lambda(G)] $ has no Hopf orders apart from $ {{\mathfrak O}_{L}}[\lambda(G)] $, by [@ChTwe Corollary 20.3]. Therefore $ {{\mathfrak A}}_{\lambda} $ is not a Hopf order in $ H_{\lambda} $.
This example illustrates that Theorem \[thm\_main\_thm\_intro\] has the potential to extend the scope of Childs’ Theorem: if $ L/K $ is a finite separable extension of local or global fields in characteristic zero and $ H $ is a Hopf algebra giving a Hopf-Galois structure on the extension then $ {{\mathfrak O}_{L}}$ is locally free over its associated orders in both $ H $ and $ H^{\prime} $ if either of these is a Hopf order.\
\
The question of whether $ {{\mathfrak A}_{H}}$ is a maximal order in $ H $ is also of interest, and this is the subject of our final result. We shall require $ H $ to be a separable $ K $ algebra, so we first note a sufficient condition for this to occur:
\[lem\_H\_separable\] Suppose that the characteristic of $ K $ does not divide $ [E:K] $. Then $ H $ is a separable $ K $-algebra.
Recall from the theorem of Greither and Pareigis that $ H = E[N]^{G} $ for some regular subgroup $ N $ of $ {\mbox{Perm}(X)} $ normalized by $ \lambda(G) $ and that we have $ E \otimes_{K} E[N]^{G} = E[N] $. Since $ N $ is regular, we have $ |N|=[L:K] $, which divides $ [E:K] $, so the characteristic of $ E $ does not divide $ |N| $, and so $ E \otimes_{K} E[N]^{G} = E[N] $ is a separable $ E $-algebra. By [@Demeyer_Ingraham Chapter II, Corollary 1.10], this implies that $ E[N]^{G} $ is a separable $ K $-algebra provided that $ K $ is a $ K $-direct summand of $ E $. The map $ E \rightarrow K $ defined by $$x \mapsto \frac{1}{[E:K]}\mbox{Tr}_{E/K}(x)$$ is surjective, $ K $-linear, and split by the inclusion $ K \hookrightarrow E $, so $ K $ indeed occurs as a $ K $-direct summand of $ E $, and so $ E[N]^{G} $ is a separable $ K $-algebra.
\[prop\_maximal\_orders\] Suppose that the characteristic of $ K $ does not divide $ [E:K] $. Let $ {\mathfrak B}$ be a fractional ideal of $ L $, and let $ {{\mathfrak A}_{H}}$, $ {{\mathfrak A}_{H}}^{\prime} $ be the associated orders of $ {\mathfrak B}$ in $ H $, $ H^{\prime} $ respectively. Then $ {{\mathfrak A}_{H}}$ is a maximal order in $ H $ if and only if $ {{\mathfrak A}_{H}}^{\prime} $ is a maximal order in $ H^{\prime} $.
Since the characteristic of $ K $ does not divide $ [E:K] $, Lemma \[lem\_H\_separable\] implies that$ H $ is a separable $ K $-algebra. Thus an $ {{\mathfrak O}_{K}}$-order $ {{\mathfrak A}}$ in $ H $ is maximal if and only if for each prime $ \p $ of $ {{\mathfrak O}_{K}}$ the order $ \OKp \otimes_{{{\mathfrak O}_{K}}} {{\mathfrak A}}$ is a maximal $ \OKp $-order in the separable $ K_{\p} $-algebra $ K_{\p} \otimes_{K} H $ [@Reiner_MO Corollaries 11.2 and 11.5]. Therefore we may reduce to the case in which $ K $ is a local field.\
\
Suppose that $ {{\mathfrak A}_{H}}^{\prime} $ is a maximal order but $ {{\mathfrak A}_{H}}$ is not. Since $ H $ is a separable $ K $-algebra, by [@Reiner_MO Corollary 10.4] $ {{\mathfrak A}_{H}}$ is properly contained in some maximal order in $ H $, say $ {\mathfrak M} $. Since $ L $ is a free $ H^{\prime} $-module of rank $ 1 $, [@Reiner_MO Theorem 18.10] implies that $ {\mathfrak B}$ is a free $ {{\mathfrak A}_{H}}^{\prime} $-module of rank $ 1 $. Let $ x \in {\mathfrak B}$ generate $ {\mathfrak B}$ as an $ {{\mathfrak A}_{H}}^{\prime} $ module; by Theorem \[thm\_main\_thm\_intro\] it also generates $ {\mathfrak B}$ as an $ {{\mathfrak A}_{H}}$-module. Now let $ \Delta = {\mathfrak M} \cdot x $. This is a full $ {{\mathfrak O}_{K}}$-lattice in $ L $ which properly contains $ {\mathfrak B}$, and it is a free $ {\mathfrak M} $-module of rank $ 1 $, generated by $ x $. Let $ {{\mathfrak A}_{H}}^{\prime}(\Delta) $ denote the associated order of $ \Delta $ in $ H^{\prime} $. Similarly to the proof of Theorem \[thm\_main\_thm\_intro\], for each $ \mu \in {\mathfrak M} $ define $ z_{\mu} \in H^{\prime} $ by $ z_{\mu} \cdot x = \mu \cdot x $; we find that $ {{\mathfrak A}_{H}}^{\prime}(\Delta) = \{ z_{\mu} \mid \mu \in {\mathfrak M} \} $. But $ {{\mathfrak A}_{H}}^{\prime}(\Delta) \supsetneq {{\mathfrak A}_{H}}^{\prime} $ since $$\Delta = {{\mathfrak A}_{H}}^{\prime}(\Delta) \cdot x \supsetneq {{\mathfrak A}_{H}}^{\prime} \cdot x = {\mathfrak B}.$$ This contradicts the assumption that $ {{\mathfrak A}_{H}}^{\prime}({\mathfrak B}) $ is a maximal order in $ H $. Therefore $ {{\mathfrak A}_{H}}$ is a maximal order in $ H $.\
\
Since $ (H^{\prime})^{\prime} = H $, the converse statement follows by interchanging the roles of $ H, H^{\prime} $ in the argument above.
[00]{}
Nigel P. Byott, *Galois structure of ideals in wildly ramified abelian $p$-extensions of a $p$-adic field, and some applications,* Journal de théorie des Nombres de Bordeaux 9, (1997): 201-219.
Lindsay N. Childs, *Taming Wild Extensions: [H]{}opf algebras and local [G]{}alois module theory,* American Mathematical Society, (2000).
Frank De Meyer and Edward Ingraham, *Separable algebras over commutative rings,* Springer, (1971).
Cornelius Greither and Bodo Pareigis, *Hopf Galois theory for separable field extensions,* Journal of Algebra 1 (1987): 239-258.
Timothy Kohl, *Groups of order 4p, twisted wreath products and Hopf-Galois theory,* Journal of Algebra 314, (2007): 42-74.
Paul J. Truman, *Towards a generalisation of [N]{}oether’s [T]{}heorem to nonclassical Hopf-Galois structures,* New York J. Math 17, (2011): 799-810.
———— *Integral Hopf-Galois structures for tame extensions,* New York J. Math 19, (2013): 647-655.
———— *Canonical nonclassical [H]{}opf-[G]{}alois module structure of nonabelian [G]{}alois extensions,* Communications in Algebra 44.3 (2016): 1119-1130.
———— *Correction to: Canonical nonclassical [H]{}opf-[G]{}alois module structure of nonabelian [G]{}alois extensions*. Communications in Algebra 45.9 (2017): 3892.
Irving Reiner, *Maximal Orders* Academic Press, (1975).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The structure of spinning particle suggested by the rotating Kerr-Newman ( black hole ) solution, super-Kerr-Newman solution and the Kerr-Sen solution to low energy string theory is considered. Main peculiarities of the Kerr spinning particle are discussed: a vortex of twisting principal null congruence, singular ring and the Kerr source representing a rotating relativistic disk of the Compton size. A few stringy structures can be found in the real and complex Kerr geometry. Low-energy string theory predicts the existence of a heterotic string placed on the sharp boundary of this disk. The obtained recently supergeneralization of the Kerr-Newman solution suggests the existence of extra axial singular line and fermionic traveling waves concentrating near these singularities.
We discuss briefly a possibility of experimental test of these predictions.
author:
- |
A. Burinskii\
Gravity Research Group, NSI Russian Academy of Sciences,\
B. Tulskaya 52, Moscow 113191, Russia, e-mail: grg@ibrae.ac.ru[^1]
date: 'October, 1999'
title: |
Structure of Spinning Particle\
Suggested by Gravity, Supergravity\
and Low Energy String Theory
---
Introduction
============
The Kerr solution is well known as a field of the rotating black hole. However, for the case of a large angular momentum $L$, $\mid a\mid = L/m \geq m$, all the horizons of the Kerr metric are absent, and the naked ring-like singularity is appeared. This naked singularity has many unpleasant manifestations and must be hidden inside a rotating disk-like source. The Kerr solution with $\mid
a\mid \gg m$ displays some remarkable features indicating a relation to the structure of the spinning elementary particles.
In the 1969 Carter [@car] observed, that if three parameters of the Kerr - Newman solution are adopted to be ($\hbar $=c=1 ) $\quad e^{2}\approx 1/137,\quad
m \approx 10^{-22},\quad a \approx 10^{22},\quad L=ma=1/2,$ then one obtains a model for the four parameters of the electron: charge, mass, spin and magnetic moment, and the giromagnetic ratio is automatically the same as that of the Dirac electron. Israel [@is] has introduced a disk-like source for the Kerr field, and it was shown by Hamity [@ham] that this source represents a rigid relativistic rotator. A model of “microgeon” with Kerr metric was suggested [@bur] and an analogy of this model with string models [@ibur]. Then a model of the Kerr-Newman source in the form of oblate spheroid was suggested [@lop]. It was shown that material of the source must have very exotic properties: null energy density and negative pressure. An attempt to explain these properties on the basis of the volume Casimir effect was given in work [@blet]. The electromagnetic properties of the material are close to those of a superconductor [@blet; @lop], that allows to consider singular ring of the Kerr source as a closed vortex string like the Nielsen-Olesen and Witten superconducting strings. Since 1992 black holes have paid attention of string theory. In 1992 the Kerr solution was generalized by Sen to low energy string theory [@sen]. It was shown that black holes can be considered as fundamental string states, and the point of view has appeared that some of black holes can be treated as elementary particles [@part]. The obtained recently super-Kerr-Newman solution [@superBH; @skn] represents a natural combination of the Kerr spinning particle and superparticle models and predicts the existence of extra axial singularity and fermionic traveling waves on the Kerr-Newman background.
Kerr singular ring
==================
The Kerr string-like singularity appears in the rotating BH solutions instead of the point-like singularity of the non-rotating BH. The simple solution possessing the Kerr singular ring was obtained by Appel in 1887 (!) [@app]. It can be considered as a Newton or Coulomb analogue to the Kerr solution. When the point-like source of the Coulomb solution $f=1/\tilde r= 1/ \sqrt{ (x-x_o)^2+(y-y_o)^2 +(z-z_o)^2}$ is shifted to a complex point of space $(x_o , y_o , z_o )\rightarrow (0,0,ia)$, the Kerr singular ring arises on the real slice of space-time. The complex equation of singularity $\tilde r=0$ represents a ring as an intersection of plane and sphere. The complex radial distance $\tilde r$ can be expressed in the oblate spheroidal coordinates $r$ and $\theta$: $\tilde r = r+i a \cos \theta$. The Kerr singular ring is a branch line of the space on two sheets: “positive” one covered by $r \geq 0 $, and “negative” one, an anti-world, covered by $ r \leq 0 $. The sheets are connected by disk $ r = 0 $ spanned by singular ring. The physical fields change signs and directions on the “negative ”sheet. Truncation of the negative sheet allows one to avoid the twosheetedness. In this case the fields will acquire a shock crossing the disk, and some material sources have to be spread on the disk surface to satisfy the field equations. The structure of electromagnetic field near the disk suggests then that the “negative” sheet of space can be considered as a mirror image of the real world in the rotating superconducting mirror.
The source of Kerr-Newman solution, like the Appel solution, can be considered from complex point of view as a “particle” propagating along a complex world-line $x^i (\tau)$ [@bkp; @c-str] parametrized by complex time $\tau$. The objects described by the complex world-lines occupy an intermediate position between particle and string. Like the string they form the two-dimensional surfaces or the world-sheets in the space-time. It was shown that the complex Kerr source may be considered as a complex hyperbolic string which requires an orbifold-like structure of the world-sheet. It induces a related orbifold-like structure of the Kerr geometry [@c-str] which is closely connected with the above mentioned twosheetedness.
Kerr congruence and disk-like source
====================================
Second remarkable peculiarity of the Kerr solution is the twisting principal null congruence (PNC) which can be considered as a vortex of null radiation. This vortex propagates via disk from negative sheet of space onto positive one forming a caustic at singular ring. PNC plays fundamental role in the structure of the Kerr geometry. The Kerr metric can be represented in the Kerr-Schild form $g_{ik} = \eta _{ik} + 2 h k_{i} k_{k}, $ where $\eta $ is metric of an auxiliary Minkowski space and $h$ is a scalar function. Vector field $ k_i(x) $ is null, $ k_i k^i=0, $ and tangent to PNC. The Kerr PNC is geodesic and shear free [@dks]. Congruences with such properties are described by the Kerr theorem via complex function $Y(x)$ representing a projective spinor coordinate $Y(x)= {\bar \Psi}^{\dot 2}/{\bar \Psi} ^{\dot 1}$. The null vector field $k_i (x)$ can be expressed in spinor form $k \sim \bar \Psi \sigma _i dx^i \Psi $.
The above complex representation of source allows one to obtain the Kerr congruence by a retarded-time construction [@bkp; @c-str]. The complex light cone with the vertex at some point $x_0$ of the complex world line $(x_i - x_{0 i})(x^i -x_0^i) = 0$ can be split into two families of null planes: “left” and “right”. In spinor form this splitting can be described as $$x_i = x_{0i} + \Psi \sigma _i \tilde \Psi ,\label{Psi}
\label{lk}$$ where “right” (or “left”) null planes can be obtained keeping $\Psi$ constant and varying $\tilde \Psi$, or keeping $\tilde \Psi$ constant and varying $ \Psi$. The rays of the twisting Kerr congruence arise as [*real slice*]{} of the “left” null planes of the complex light cones emanated from the complex world line [@bkp; @c-str].
Replacement of the negative sheet by a disk-like source at surface $r=0$ allows one to avoid twosheetedness of the Kerr space. However, there is still a small region of causality violation on positive sheet of space. By the Löpez suggestion this region has to be also covered by source [@lop]. The minimal value of $r$ covering this region is ‘classical radius’ $r_e=\frac{e^2}{2m}$. The resulting disk-like source has a thickness of order $r_e$ and its degree of oblateness is $\alpha ^{-1} \approx 137$.
Stringy suggestions
===================
In 1974, in the frame of Einstein gravity the model of microgeon with the Kerr-Newman metric was considered [@bur], where singular ring was used as a waveguide for wave excitations. It was recognized soon [@ibur] that singular ring represents in fact a string with traveling waves. Further, in dilaton gravity, the string solutions with traveling waves have paid considerable attention. The obtained by Sen generalization of the Kerr solution to low energy string theory with axion and dilaton [@sen] was analyzed in [@bsen]. It was shown that, in spite of the strong deformation of metric by dilaton (leading to a change the type of metric from type D to type I), the Kerr PNC survives in the Kerr-Sen solution and retains the properties to be geodesic and shear free. It means that the Kerr theorem and the above complex representation are valid for the Kerr-Sen solution too. It has also been obtained that the field of the Kerr-Sen solution near the Kerr singular ring is similar to the field around a fundamental heterotic string that suggested stringy interpretation of the Kerr singular ring.
Supergeneralization
===================
Description of spinning particle based only on the bosonic fields cannot be complete. On the other hand the fermionic models of spinning particles and superparticles based on Grassmann coordinates have paid considerable attention. In [@superBH; @skn] a natural way to combine the Kerr spinning particle and superparticle models was suggested leading to a non-trivial super-Kerr-Newman black hole solution.
The simplest consistent supergeneralization of Einstein gravity represents an unification of the gravitational field $g_{ik}$, with a spin 3/2 Rarita-Schwinger field $\psi _i$ . There exists the problem of triviality of supergravity solutions. Any exact solution of Einstein gravity is indeed a trivial solution of supergravity field equations with a zero field $\psi _i$. Starting from such a solution and using supertranslations, one can easily turn the gravity solution into a form containing the spin-3/2 field $\psi_i$. However, since this spin-3/2 field can be gauged away by reverse transformations such supersolutions have to be considered as [*trivial*]{}.
The hint how to avoid this triviality problem was given by complex representation of the Kerr geometry. One notes that from complex point of view the Schwarzschild and Kerr geometries are equivalent and connected by a [*trivial*]{} complex shift. The [*non-trivial*]{} twisting structure of the Kerr geometry arises as a result of the [*shifted real slice*]{} regarding the source [@bkp; @c-str].
Similarly, it is possible to turn a [*trivial*]{} super black hole solution into a [*non-trivial*]{} if one finds an analogue to the [*real slice*]{} in superspace.
The [*trivial supershift*]{} can be represented as a replacement of the complex world line by a superworldline $X^i_0(\tau)= x^i_0(\tau)-i \theta \sigma ^i \bar \zeta + i \zeta
\sigma^i \bar \theta,$ parametrized by Grassmann coordinates $\zeta, \quad \bar \zeta$, or as a corresponding coordinate supershift $x^{\prime i} = x^i + i \theta\sigma^i \bar \zeta
- i \zeta\sigma^i \bar \theta;
\qquad
\theta^{\prime}=\theta + \zeta ,\quad
{\bar\theta}^{\prime}=\bar\theta + \bar\zeta.$
Assuming that coordinates $x^i$ before the supershift are the usual c-number coordinates one sees that coordinates acquire nilpotent Grassmann contributions after supertranslations. Therefore, there appears a natural splitting of the space-time coordinates on the c-number ‘body’-part and a nilpotent part - the so called ‘soul’. The ‘body’ subspace of superspace, or B-slice, is a submanifold where the nilpotent part is equal to zero, and it is a natural analogue to the real slice in complex case.
Reproducing the real slice procedure of the Kerr geometry in superspace one obtains the condition of proportionality of the commuting spinors $\bar\Psi(x)$ determining the PNC of the Kerr geometry and anticommuting spinors $ \bar\theta$ and $\bar\zeta$, As a consequence of the B-slice and superlightcone constraints one obtains a submanifold of superspace $\theta = \theta (x),
\quad \bar \theta = \bar \theta (x).$ The initial supergauge freedom is lost now, and there appears a non-linear realization of broken supersymmetry introduced by Volkov and Akulov [@VA; @WB] and considered in N=1 supergravity by Deser and Zumino [@DZ]. It is assumed that this construction is similar to the Higgs mechanism of the usual gauge theories, and $\zeta ^\alpha (x), \quad \bar \zeta ^{\dot\alpha} (x) $ represent Goldstone fermion which can be eaten by appropriate local supertransformation with a corresponding redefinition of the tetrad and the spin-3/2 field $\psi_i$.
However, the complex character of supertranslations demands to extend this scheme to N=2 supergravity.
In this way the self-consistent super-Kerr-Newman solutions to broken N=2 supergravity coupled to Goldstone fermion field was obtained [@skn]. The solution describes the massless Dirac wave field propagating on the Kerr-Newman background along the Kerr congruence. Besides the Kerr singular ring solution contains an extra axial singularity and traveling waves propagating along the ring-like and axial singularity.
The ‘axial’ singularity represents a half-infinite line threading the Kerr singular ring and passing to ‘negative’ sheet of the Kerr geometry. The position and character of axial singularity depend on the index $n$ of elementary excitation. The case $n=-1/2$ is exclusive: there are two ‘decreasing’ singularities which are situated symmetrically at $\theta=0$ and $\theta=\pi$.
Problem of hard core
====================
The obtained supergeneralization is based on the massless Goldstone field. At present stage of investigation our knowledge regarding the origin of the Goldstone fermion is very incompleted. Analyzing the Wess-Zumino model of super-QED and some other schemes of spontaneously broken supersymmetry [@WB], one sees that it can leads to massless Goldstone fermions, at least in the region of massless fields out of the BH horizons.
However, for the known parameters of spinning particles, the angular momentum is very high, regarding the mass parameter, and the black hole horizons disappear. The resulting object is “neither black and nor hole”, and the considered above disk-like ‘hard core’ region is naked. Structure of this region represents a very important and extremely complicated problem. Among the possible field models for description this region could be mentioned the Landau-Ginzburg model, super-QED, non-abelian gauge models, Seiberg-Witten theory, as well as the recent ideas on the confinement caused by extra dimensions in the bulk/boundary ( AdS/CFT correspondence) models [@bulk]. Apparently, this problem is very far from resolution at present, and one of the most difficult points can be the concordance of the field model with the rotating disk-like bag of the Kerr geometry.
Suggestions to experimental test
================================
The predicted comparative big size of the disk-like bag looks as a serious contradiction to the traditional point of view on the structureless, point-like electron. However, the suggested by QED virtual photons surrounding electron in the region of Compton size, zitterbewegung and the vacuum zero-point fluctuations, spreading the position of electron, can be treated as some indirect evidences for the existence of an geometrical structure in the Compton region. At least, one can assume that region of virtual photons has a tendency to be very ordered with formation of the Kerr congruence and the ring-like singularity.
The modern progress in the formation of polarized beams of spinning particles suggests the possible methods for experimental test of the main predicted feature of the Kerr spinning particle - [*its highly oblated form.*]{} In particular, it could be the method proposed in [@mus] based on estimation of the cross section differences between transversely and longitudinally polarized states in proton-proton collisions. One proposes that similar experiment could be more effective for the electron-electron collisions. Another possible way of experimental test could be the analysis of the diffraction of photons on the polarized electrons.
Apparently, the strong influence of vacuum fluctuations will not allow one to observe the predicted very high oblateness of electrons. Nevertheless, one expects that an essential effect should be observed if the Kerr source model reflects the reality.
[99]{} Carter B., Phys. Rev. [**174**]{}(1968) 1559;
Israel W., Phys. Rev. [**D2**]{} (1970) 641;
Hamity V. Phys. Lett. [**A 56**]{}(1976)77;
Burinskii A., Sov. Phys. JETP [**39**]{}(1974) 193;
Ivanenko D. and Burinskii A., Izv. VUZ Fiz.[**5**]{}(1975)135; Sov. Phys. J. (USA).,nr.[**7**]{} (1978)113;
Burinskii A. Phys. Lett. [**B 216**]{}(1989)123;
López C.A., Phys. Rev. [**D30**]{} (1984) 313;
Sen A., Phys.Rev.Lett., [**69**]{}(1992)1006; Dabholkar A., Gauntlett J., Harvey J. and Waldram D., Nucl. Phys. [**B 474**]{}(1996) 85;
Sen A. Modern Phys. Lett. [**A 10**]{}(1995)2081; Nucl.Phys [**B46**]{} (Proc.Suppl)(1996)198; Burinskii A., Phys.Rev.D 57(1998)2392, hep-th/9704102;
Burinskii A. Class.Quant. Grav. [**16**]{}(1999)(Oct-Nov to appear) hep-th/9903032; E.T. Whittacker and G.N. Watson, A Course of Modern Analysis, Cambrige Univ. Press London/New York,p.400, 1969 . Burinskii A., Kerr R.P. and Perjes Z., 1995, Nonstationary Kerr Congruences, gr-qc/9501012; Burinskii A., String - like Structures in Complex Kerr Geometry, Proc. of the Fourth Hungarian Relativity Workshop Edited by R.P. Kerr and Z. Perjes, Academiai Kiado, Budapest 1994, gr-qc/9303003;
Burinskii A., Phys.Lett. [**A 185**]{}(1994)441; Burinskii A. Espec. Space Explorations, [**9 (C2)**]{}(1995) 60, Moscow, Belka, hep-th/9503094; Debney G.C., Kerr R.P., Schild A., J.Math.Phys.,[**10**]{}(1969)1842;
Burinskii A., Phys.Rev.[**D 52**]{}(1995)5826, hep-th/9504139;
Volkov D.V. and Akulov V.P. Pis’ma Zh. Eksp.Teor.Fiz. [**16**]{}621(1972);
Wess J. and Bagger J. Supersymmetry and Supergravity, Princeton, New Jersey 1983;
Deser S. and Zumino B. Phys. Rev. Lett., [**38**]{}(1977) 1433;
Musulmanbekov G. Total Cross Section Differences Between Transversely and Longitudinally Polarized States in Proton-Proton Collisions. In: Proc. of VII Workshop on High Energy Physics SPIN’97, July 1997, Dubna, Russia, Dubna 1997, p.165.
E. Witten, Adv. Theor. Math. Phys. [**2**]{} (1998) 253, hep-th/9802150; J. Maldacena, Adv. Theor. Math. Phys. [**2**]{} (1998) 231, hep-th/9711200; I.R. Klebanov and A.A. Tseytlin, hep-th/9812089; J.A. Minahan, JHEP [bf 9901]{}(1999)020, hep-th/9811156, hep-th/9902074; A.Kehagias and K. Sfetsos, Phys.Lett.[**B454**]{}(1999)270, hep-th/9902125;iibid.[**B456**]{}(1999)22, hep-th/9903109; H.Liu and A.A. Tseytlin, hep-th/9903091; R. de Mello Koch, A. Paulin-Campbell and J. P. Rodrigues, hep-th/9903029; S. Nojiri and S.D. Odintsov, hep-th/9904036, hep-th/9906216; A. Burinskii, hep-th/9908198
[^1]: Talk at the International Workshop Spin’99, Prague, 5-11 September, 1999
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present a new class of $C^\infty$-smooth finite element spaces on Cartesian grids, based on a partition of unity approach. We use these spaces to construct smooth approximations of particle fields, i.e., finite sums of weighted Dirac deltas. In order to use the spaces on general domains, we propose a fictitious domain formulation, together with a new high-order accurate stabilization. Stability, convergence, and conservation properties of the scheme are established. Numerical experiments confirm the analysis and show that the Cartesian grid-size $\sigma$ should be taken proportional to the square-root of the particle spacing $h$, resulting in significant speed-ups in vortex methods.'
author:
- 'Matthias Kirchhart [^1]'
- Shinnosuke Obi
bibliography:
- 'papers.bib'
title: A Smooth Partition of Unity Finite Element Method for Vortex Particle Regularization
---
vortex method, particle method, partition of unity finite element method, smooth shape functions, fictitious domains, Biot–Savart law
65N12, 65N15, 65N30, 65N75, 65N80, 65N85
Introduction
============
Vortex particle methods are numerical schemes for solving the incompressible Navier–Stokes equations. Instead of their formulation in the primitive variables velocity ${\mathbf{u}}$ and pressure, they make use of the equivalent formulation in terms of the vorticity ${\boldsymbol{\omega}} = \nabla\times{\mathbf{u}}$. The core idea is most easily explained in the two-dimensional, inviscid case, where the equation for the scalar vorticity $\omega$ in a bounded domain $\Omega$ then reads: $$\label{eqn:vte}
{\frac{\partial {\omega}}{\partial {t}}} + ({\mathbf{u}}\cdot\nabla)\omega = 0\qquad \mbox{in $\Omega$.}$$ Let us for the moment assume the velocity field would be known, in which case we have a linear equation. We then can discretize the initial vorticity field with *particles*: $$\label{eqn:pufem_particles}
\omega(t=0)\approx\omega_h(0) = \sum_{i=1}^{N}\Gamma_i\delta(x-x_i(0)),$$ where $\Gamma_i$ and $x_i$ denote the circulation and position of particle $i$, and $\delta$ is the Dirac delta function. Such particle fields can be seen as quadrature rules for integrating smooth functions $\varphi$ against the vorticity $\omega$ we are aiming to approximate. Let us for example assume we are given a quasi-uniform, shape-regular triangulation of the domain $\Omega$ of mesh-width $h$. If one then applies a quadrature rule of exactness degree $m$ with positive weights to each cell, one obtains a set of quadrature nodes $x_i$ with associated weights $w_i$. A particle field approximation could then be obtained by setting $\Gamma_i:= w_i\omega(x_i)$ in \[eqn:pufem\_particles\]. For such a particle field one can prove error-bounds of the form: $$\Vert\omega-\omega_h\Vert_{W^{-(m+1),2}(\Omega)}\leq Ch^{m+1}\Vert\omega\Vert_{W^{m+1,2}(\Omega)},$$ for $m+1>d/2$, where $d$ is the number of spatial dimensions. Here and throughout this text the symbol $C$ refers to a generic positive constant which is independent of the functions involved.
The reason for choosing this particular discretization is the availability of an analytic solution of \[eqn:vte\] in this case. If one modifies the particles’ positions according to ${\frac{{\mathrm d}{x_i}}{{\mathrm d}{t}}} = {\mathbf{u}}(x_i(t))$, $i = 1,\ldots,N$, the resulting approximation $\omega_h$ fulfills the vorticity equation \[eqn:vte\] *exactly*, i.e., the only error in the approximation comes from the initialization error [@cottet2000 Appendix A].
In practice, however, the velocity field is of course not known and needs to be retrieved from the vorticity. Let us for simplicity assume that the velocity would vanish at the boundaries. In this case the Helmholtz decomposition theorem tells us that the velocity can be retrieved through the Biot–Savart law without any boundary integral terms: $${\mathbf{u}} = {\mathbf{K}}\star\omega,\ {\mathbf{K}}(x) = \frac{1}{2\pi}\frac{(-x_2,x_1)^\top}{|x|^2},$$ where $\star$ denotes convolution. We have the following classical estimate due to Calderón and Zygmund [@calderon1952]: $$\label{eqn:calzyg}
\Vert{\mathbf{K}}\star\omega\Vert_{W^{1,2}(\Omega)}\leq C\Vert\omega\Vert_{L^2(\Omega)}.$$ The problem is that the particle approximation $\omega_h\not\in L^2(\Omega)$ is not smooth enough to apply this estimate; applying the Biot–Savart law to the particle field directly yields a singular velocity field. The question we try to answer in this paper is how to obtain an accurate, smooth approximation $\omega_\sigma\in L^2(\Omega)$ from the particle field $\omega_h$, where $\sigma$ refers to a smoothing length, which will be defined precisely later. This problem is called *particle regularization.* Once such a smooth approximation has been obtained, one closes the system of equations by setting ${\mathbf{u}}_\sigma :=
{\mathbf{K}}\star\omega_\sigma$ and modifying the particle’s positions according to ${\frac{{\mathrm d}{x_i}}{{\mathrm d}{t}}} = {\mathbf{u}}_\sigma(x_i(t),t)$ instead. It is this natural treatment of convection which makes vortex methods so appealing. Given an appropriate choice of $\omega_\sigma$ one can show that the resulting method is essentially free of artificial viscosity and conserves mass, circulation, linear momentum, and angular momentum, and the energy of ${\mathbf{u}}_\sigma$ exactly [@cottet2000 Section 2.6]. When extended to handle physical viscosity, this makes the method particularly attractive for flows at high Reynolds numbers, see for example the recent work by Yokota et al. [@yokota2013].
The most common approach to the regularization problem is to mollify the particle field with a certain, radially symmetric *blob-function* $\zeta_\sigma$: $\omega_\sigma := \omega_h\star\zeta_\sigma$, where $\sigma$ denotes the radius of the blob’s core [@cottet2000 Section 2.3]. Many commonly used blob-functions have infinite support, effectively extending $\omega_\sigma$ from $\Omega$ to $\mathbb{R}^d$ and blurring the domain’s boundaries. There are approaches to use blob-functions with varying shapes near boundaries [@teng1982; @lakkis2009] or to use image particles outside of the domain [@cottet2000 Section 4.5.2]. These approaches assume that the boundaries are flat and usually fail in the presence of sharp corners or kinks. Another approach is to interpret the particles’ circulations as weighted function values $\Gamma_i = w_i\omega(x_i)$, where the $w_i$ are weights from an underlying quadrature rule. While this is strictly speaking only the case during the initialization stage, the approach is then to create a triangulation of the domain using the particles’ positions as grid nodes and to use these values to construct a piece-wise linear approximation [@russo1994]. This requires a mesh to be regenerated at every time-step, which is problematic as the particle field gets distorted over time. In Vortex-in-Cell (VIC) schemes one uses interpolation formulas to obtain a grid-based approximation of the vorticity field. In the vicinity of boundaries these formulas need to be specifically adapted to the particular geometry at hand and cannot be used for arbitrary domains [@cottet2004]. In general the regularization problem causes significant difficulties, and as Cottet and Koumoutsakos point out in the introduction of their book [@cottet2000]: “To our knowledge there is no completely satisfactory solution for general geometries, in particular because of the need to regularize vortices near the boundary.”
In this work we try to address this problem with the help of a finite element formulation. The non-smooth $W^{-(m+1),2}$-nature of the particle field forces us to use shape functions that are globally $W^{m+1,2}$-smooth, which is not the case for the classical, piecewise linear elements. The partition of unity finite element method (PUFEM) by Melenk and Babuška is a generalization of the classical finite element method (FEM), which can be used to obtain such smooth spaces. Even though already mentioned in their introductory paper [@melenk1996], there seems to have been little research in this direction. Duarte et al. [@duarte2006] describe an approach which only works for certain triangulations in two dimensions.
The generation of globally smooth shape functions on general meshes in higher dimensions is a well-known, hard problem. We instead consider simple Cartesian grids, on which the construction of smooth shape functions is easier. We then apply a fictitious domain approach to deal with general geometries. This typically results in instabilities in the cut elements. Under the name *ghost penalty* Burman [@burman2010] presented an effective and accurate stabilization strategy for this problem, which has for example been successfully applied to several other flow problems with cut elements [@burman2012; @gross2016; @hansbo2013; @kirchhart2016]. We use a similar, higher-order approach inspired by Burman and Fernández [@burman2014] as well as Cattaneo et al. [@cattaneo2015] to achieve accuracy and stability of the resulting discretization.
The rest of this article is structured as follows. In \[sec:pufem\] we define and construct smooth PUFEM spaces and analyze some of their important properties. In \[sec:formulation\] we introduce a stabilized variational formulation and prove its stability and convergence. The regularization problem is then treated as a perturbation to this variational formulation. Similar to the approach with blob-functions as mentioned above, the resulting error can be split into regularization and quadrature error parts which need to be carefully balanced. The analysis will show that the smoothing parameter $\sigma$ should be taken proportional to the square-root of the particle-spacing $h$ and that this choice is in a certain sense optimal. In \[sec:experiments\] we perform numerical experiments, confirming our analysis. As a consequence of the quadratic relation between $h$ and $\sigma$, the computation of the velocity field only has a computational complexity of ${\mathcal{O}({h^{-\frac{d}{2}}})}$, enabling the use of particle numbers on desktop workstations which were previously only possible on super computers. We finish the article with concluding remarks and acknowledgements.
Smooth Partition of Unity Finite Element Spaces {#sec:pufem}
===============================================
Basic Theory
------------
We begin this subsection by defining smooth partitions of unity, similar to Melenk’s and Babuška’s theory [@melenk1996].
\[def:pou\] Let $\Omega\subset\mathbb{R}^d$ be an open set and let $\lbrace\Omega_i\rbrace$ be an open cover of $\Omega$ satisfying a pointwise overlap condition: $$\exists M\in\mathbb{N}:\ \forall x\in\Omega:\
\operatorname{card}\lbrace i\,|\,x\in\Omega_i\rbrace\leq M.$$ Let $\lbrace\varphi_i\rbrace$ be a Lipschitz partition of unity subordinate to the cover $\lbrace\Omega_i\rbrace$ satisfying $$\begin{aligned}
\operatorname{supp}\varphi_i &\subset \operatorname{clos}\Omega_i, \\
\sum_{i}\varphi_i(x) &\equiv 1\ \text{on $\Omega$},\\
\vert\varphi_i\vert_{W^{k,\infty}(\mathbb{R}^d)} &\leq
C(k)(\operatorname{diam}{\Omega_i})^{-k}\quad k\in\mathbb{N}_0,\label{eqn:phinorm}\end{aligned}$$ where the $C(k)$ are positive constants and the symbol $\vert\cdot\vert_{W^{k,p}(\mathbb{R}^d)}$ refers to the Sobolev semi-norms: $$\vert f\vert_{W^{k,p}(\mathbb{R}^d)} :=
\begin{cases}
\biggl(\sum_{|\alpha|=k}\Vert{\partial}^\alpha f\Vert_{L^p(\mathbb{R}^d)}^{p}\biggr)^{1/p} & p\in[1,\infty), \\
\max_{|\alpha|=k}\Vert{\partial}^\alpha f\Vert_{L^\infty(\mathbb{R}^d)} & p = \infty.
\end{cases}$$ Then, $\lbrace\varphi_i\rbrace$ is called a smooth partition of unity subordinate to the cover $\lbrace\Omega_i\rbrace$. The sets $\Omega_i$ are called patches.
Using these functions $\lbrace\varphi_i\rbrace$, we can define the spaces for the partition of unity finite element method (PUFEM).
\[def:pufemspace\] Let $\lbrace\varphi_i\rbrace$ be a smooth partition of unity subordinate to the open cover $\lbrace\Omega_i\rbrace$. For $P,k\in\mathbb{N}_0$, $p\in[1,\infty]$ we define polynomial enrichment spaces $V_i^P\subset W^{k,p}(\Omega\cap
\Omega_i):$ $$V_i^P := \operatorname{span}\bigl\lbrace
x^\alpha\,\bigr|\,|\alpha|\leq P
\bigr\rbrace,$$ and the PUFEM spaces $V_\sigma^P(\Omega)\subset W^{k,p}(\Omega):$ $$V_\sigma^P := \operatorname{span}\bigl\lbrace
\varphi_iv_i\,\bigr|\,v_i\in V_i^P
\bigr\rbrace$$ where $\sigma := \max_i\operatorname{diam}{\Omega_i}$ refers to the maximum patch diameter.
\[ass:bramble-hilbert\] We will assume that the shapes of the domain $\Omega$ and the patches $\lbrace\Omega_i\rbrace$ are such that we can apply the Bramble–Hilbert lemma [@brenner2008 Lemma (4.3.8)]. In particular, we will assume that for all $u\in W^{P+1,p}(\Omega\cap\Omega_i)$, $p\in[1,\infty]$, there exists a $v_i\in V_i^P$ such that: $$\label{eqn:bramble-hilbert}
\vert u-v_i\vert_{W^{k,p}(\Omega\cap\Omega_i)}
\leq
C\sigma^{P+1-k}\vert u\vert_{W^{P+1,p}(\Omega\cap\Omega_i)}\
\forall k\in\mathbb{N}_0,\ k\leq P + 1,$$ where the constant $C$ is independent of $\sigma$ and $u$.
We then have the following estimate, which is a straightforward generalization of the result of Melenk and Babuška [@melenk1996 Theorem 2.1].
\[thm:approx\] Let $V_\sigma^P(\Omega)$ be as in \[def:pufemspace\] and let \[ass:bramble-hilbert\] be fulfilled. Then for any $u\in W^{P+1,p}(\Omega)$, $p\in[1,\infty]$, there exists ${\mathcal{P}}u\in V_\sigma^P(\Omega)$ such that: $$\vert u-{\mathcal{P}}u\vert_{W^{k,p}(\Omega)}\leq C\sigma^{P+1-k}\vert u\vert_{W^{P+1,p}(\Omega)}\
\forall k\in\mathbb{N}_0,\ k\leq P + 1,$$ where the constant $C$ is independent of $\sigma$ and $u$.
Here, we will only consider the case $p\in[1,\infty)$; the proof for the case $p=\infty$ is analogous. With $v_i\in V_i^P$ as in \[ass:bramble-hilbert\] we set ${\mathcal{P}}u:=\sum_i\varphi_iv_i$. We may then write for any multi-index $\alpha$ with $|\alpha|=k$: $$\Vert{\partial}^\alpha(u-{\mathcal{P}}u)\Vert_{L^p(\Omega)}^p =
\bigl\Vert{\partial}^\alpha\sum_i(u-v_i)\varphi_i\bigr\Vert_{L^p(\Omega)}^p =
\bigl\Vert\sum_{\beta\leq\alpha}\sum_i{\binom{\alpha}{\beta}}
{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}(u-v_i)\bigr\Vert_{L^p(\Omega)}^p.$$ Considering the absolute value of the expanded derivative on the right, we obtain using Hölder’s inequality: $$\biggl|\sum_{\beta\leq\alpha}\sum_{i}\begin{pmatrix}\alpha\\\beta\end{pmatrix}{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}\bigl(u-v_i\bigr)\biggr|^p
\leq C(\alpha,p)\sum_{\beta\leq\alpha}\biggl|\sum_{i}{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}\bigl(u-v_i\bigr)\biggr|^p$$ and thus: $$\Vert{\partial}^\alpha(u-{\mathcal{P}}u)\Vert_{L^p(\Omega)}^p \leq C\sum_{\beta\leq\alpha}\
\bigl\Vert
\sum_{i}{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}(u-v_i)
\bigr\Vert_{L^p(\Omega)}^p.$$ Now, using the fact that for every point $x\in\Omega$ there are at most $M$ non-zero terms in the sum over $i$, we obtain by again using Hölder’s inequality: $$\biggl|\sum_{i}
{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}(u-v_i)
\biggr|^p
\leq C(M,p)
\sum_{i}\bigl|
{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}(u-v_i)
\bigr|^p$$ After inserting this in the previous relation we may exchange the order of summation. Using the fact that $\varphi_i \equiv 0$ outside $\Omega_i$ we obtain: $$\Vert{\partial}^\alpha(u-{\mathcal{P}}u)\Vert_{L^p(\Omega)}^p \leq C\sum_{i}\sum_{\beta\leq\alpha}\Vert{\partial}^\beta\varphi_i{\partial}^{\alpha-\beta}(u-v_i)\Vert_{L^p(\Omega\cap\Omega_i)}^p.$$ Applying \[eqn:phinorm\], \[eqn:bramble-hilbert\], and collecting all the terms yields the claim.
Construction of a Smooth Partition of Unity {#sec:smoothpart}
-------------------------------------------
In this subsection we are going to construct such a smooth partition of unity using mollification. We will make use of the following two definitions.
The function: $$\begin{split}
\zeta: \mathbb{R}\to[0,K^{-1}],\qquad x&\mapsto
\begin{cases}
0 & \mbox{if $|x|\geq \tfrac{1}{2}$,} \\
K^{-1}\exp{\bigl(-\frac{1}{1-4x^2}\bigr)} & \mbox{else,}
\end{cases}\\
K &\approx 0.221\,996\,908\,084\,039\,719,
\end{split}$$ is called Friedrichs’ mollifier in one-dimensional space. The constant $K$ was obtained numerically, such that $\Vert\zeta\Vert_{L^1(\mathbb{R})}=1$. For spatial dimensions greater than $d=1$ we define Friedrichs’ mollifier using the product: $$\zeta: \mathbb{R}^d\to[0,K^{-d}],\qquad
(x_1,\ldots,x_d)\mapsto \prod_{i=1}^{d}\zeta(x_i),$$ where under a slight abuse of notation, we reused the symbol $\zeta$.
It is well known that $\zeta\in C_0^\infty(\mathbb{R}^d)$, and thus also $\zeta\in W^{k,p}(\mathbb{R}^d),\ k\in\mathbb{N}_0,\ p\in[1,\infty]$. Furthermore, we have $\operatorname{supp}{\zeta} = [-\tfrac{1}{2},\tfrac{1}{2}]^d$, which leads to the following definition.
\[def:cartesiangrid\] Given $\sigma>0$, we define Cartesian grid points $x_i\in\mathbb{R}^d$, $i\in\mathbb{Z}^d$, $x_i := (i_1\sigma,\ldots,i_d\sigma)$. With each grid point we associate a patch $\Omega_i$ and a patch-core $\omega_i$: $$\begin{aligned}
\Omega_i &:= \bigl((i_1-1)\sigma,(i_1+1)\sigma\bigr) \times \ldots \times
\bigl((i_d-1)\sigma,(i_d+1)\sigma\bigr),\\
\omega_i &:= \bigl((i_1-\tfrac{1}{2})\sigma,(i_1+\tfrac{1}{2})\sigma\bigr) \times \ldots \times
\bigl((i_d-\tfrac{1}{2})\sigma,(i_d+\tfrac{1}{2})\sigma\bigr).\end{aligned}$$
(-3.5,-0.5) grid (3.5,2.5); (1,1) rectangle node [$Q_j$]{} (2,2); (1,1) circle node \[anchor=north east\] [$x_j$]{}; (-3,0) rectangle (-1,2); (-2.5,0.5) rectangle (-1.5,1.5); (-2,1) circle node \[anchor=north east\] [$x_i$]{}; at (-0.5,2) [$\omega_i$]{}; at (-0.5,0) [$\Omega_i$]{}; (-1.75,1.25) – (-0.75,1.95); (-1.25,0.25) – (-0.75,0.05);
An illustration of these definitions is given in \[fig:cartesian\_grid\]. It is obvious that the patches $\lbrace\Omega_i\rbrace$ form an open cover of $\mathbb{R}^d$ with $M$ from \[def:pou\] being equal to $2^d$. The patch-cores $\lbrace\omega_i\rbrace$ are pairwise disjoint and their closures form a (non-open) cover of $\mathbb{R}^d$. Using these definitions, we are now ready to construct smooth partition of unity functions $\lbrace\varphi_i\rbrace,
i\in\mathbb{Z}^d$.
For a given $\sigma>0$ and $i\in\mathbb{Z}^d$ let $\varphi_i$ be the convolution of the characteristic function $\chi_{\omega_i}$ of the patch-core $\omega_i$ with the scaled Friedrichs’ mollifier $\zeta_\sigma(x):=\sigma^{-d}\zeta(x/\sigma)$: $$\varphi_i(x):= \bigl(\chi_{\omega_i}\star\zeta_\sigma\bigr)(x)
= {\int_{\omega_i}^{} {\zeta_\sigma(x-y)}\,{\mathrm d}{y}}.$$ One then has: $$\begin{aligned}
\varphi_i\in C_0^\infty(\mathbb{R}^d) &\mbox{ and } \operatorname{supp}{\varphi_i} = \operatorname{clos}{\Omega_i},\\
\sum_{i\in\mathbb{Z}^d}\varphi_i(x) &\equiv 1\quad x\in\mathbb{R}^d,\\
|\varphi_i|_{W^{k,p}(\mathbb{R}^d)} &\leq C(k)\sigma^{d/p-k}\quad k\in\mathbb{N}_0,\ p\in[1,\infty].\end{aligned}$$
The first property directly follows from the classical properties of mollification [@adams2003 sections 2.28 and 2.29]. For the second property we immediately obtain: $$\sum_{i\in\mathbb{Z}^d}\varphi_i(x) =
\sum_{i\in\mathbb{Z}^d}{\int_{\omega_i}^{} {\zeta_\sigma(x-y)}\,{\mathrm d}{y}} =
{\int_{\mathbb{R}^d}^{} {\zeta_\sigma(x-y)}\,{\mathrm d}{y}} = 1.$$ For the last property we obtain with the help of Young’s inequality for convolutions for every multi-index $\alpha$ with $|\alpha|=k$: $$\Vert\chi_{\omega_i}\star{\partial}^\alpha\zeta_\sigma\Vert_{L^p(\mathbb{R}^d)} \leq
\Vert\chi_{\omega_i}\Vert_{L^p(\mathbb{R}^d)} \Vert{\partial}^\alpha\zeta_\sigma\Vert_{L^1(\mathbb{R}^d)} =
\sigma^{d/p-k}\Vert{\partial}^\alpha\zeta\Vert_{L^1(\mathbb{R}^d)}.$$
\[rmk:phiref\] There is no closed-form expression for the functions $\lbrace\varphi_i\rbrace$ available. However, it is important to notice that we have: $$\varphi_i(x) \equiv{\hat{\varphi}}\biggl(\frac{x-x_i}{\sigma}\biggr),\quad
{\hat{\varphi}}(x) :={\int_{(-\frac{1}{2},\frac{1}{2})^d}^{} {\zeta(x-y)}\,{\mathrm d}{y}}.$$ Furthermore, ${\hat{\varphi}}$ inherits the product structure of $\zeta$. In a computer implementation it is thus sufficient to tabulate values for ${\hat{\varphi}}$ corresponding to the case $d=1$. We can then efficiently approximate ${\hat{\varphi}}$ using, e.g., cubic Hermite splines. The graph of this function can be seen in \[fig:phi\].
\[ xlabel = [$x$]{}, ylabel = [${\hat{\varphi}}(x)$]{}, grid = [major]{}, xmin = [-1]{}, xmax = [+1]{}, ymin = [0]{}, ymax = [1]{}, \] table [phi.dat]{};
Reference Element and Inverse Estimates
---------------------------------------
In this subsection we illustrate that the smooth partition of unity constructed in \[sec:smoothpart\] leads to spaces that can be treated in a manner similar to conventional finite element spaces. In particular, we can subdivide $\mathbb{R}^d$ into *elements*: $$\label{eqn:elements}
Q_i := (i_1\sigma,(i_1+1)\sigma)\times\ldots\times(i_d\sigma,(i_d+1)\sigma),
\qquad i\in\mathbb{Z}^d.$$ Such an element is for example depicted on the right of \[fig:cartesian\_grid\]. Every $Q_i$ may be seen as the image of the *reference element* $\hat{Q}:=(0,1)^d$ under the transformation $\Phi_i: \hat{Q}\to Q_i,\
\hat{x}\mapsto x_i + \sigma\hat{x}$. In every element $Q_i$ we have a fixed set $\mathcal{J}_i$ of $2^d$ overlapping patches $\Omega_j$. Introducing: $$\mathcal{B}_j^P := \biggl\lbrace
\underbrace{\biggl(\frac{x-x_j}{\sigma}\biggr)^\alpha}_{=:g_{j,\alpha}(x)}
\,\biggr|\,
|\alpha|\leq P
\biggr\rbrace,\qquad j\in\mathbb{Z}^d$$ as bases for the enrichment spaces $V_j^P$, one quickly sees that within each element $Q_i$ the basis functions $g_{j,\alpha}$ can be expressed in terms of mapped reference functions $\hat{g}_{m,\alpha}$: $$g_{j,\alpha}(x) = \bigl(\hat{g}_{m,\alpha}\circ\Phi_i^{-1}\bigr)(x),\qquad
x\in Q_i,\ j\in\mathcal{J}_i,$$ where $m$ is the index of the node in the reference element that corresponds to $x_j$. Due to \[rmk:phiref\], the same holds true for the partition of unity functions $\varphi_j$. This allows us to infer the following classical result, which follows from a scaling argument and the norm-equivalence of finite-dimensional spaces [@brenner2008 Lemma (4.5.3)].
\[lem:investimates\] Let $V_\sigma^P(\Omega)$, $\sigma>0$, $P\in\mathbb{N}_0$ be as in \[def:pufemspace\], with patches as in \[def:cartesiangrid\]. Then, for any element $Q_i\subset\Omega$ as defined in \[eqn:elements\] that is completely contained in the domain and every $v_\sigma\in V_\sigma^P(\Omega)$ one has: $$\label{eqn:investimates}
\Vert v_\sigma\Vert_{W^{l,p}(Q_i)}\leq C\sigma^{k-l}\Vert v_\sigma\Vert_{W^{k,p}(Q_i)},\quad
p\in[1,\infty],\ k,l\in\mathbb{N}_0,\ k\leq l,$$ where the constant $C$ is independent of $\sigma$ and $i$.
Stabilized Variational Formulation {#sec:formulation}
==================================
In this section we will introduce a stabilized variational formulation with the aim of mimicking of the $L^2(\Omega)$-orthogonal projector onto $V_\sigma^P(\Omega)$. As the inverse estimates \[eqn:investimates\] are not available for elements $Q_i$ cut by the boundary ${\partial}\Omega$, we will employ a fictitious domain approach. In order to ensure coercivity of the resulting bilinear form on the entire fictitious domain, we will add a stabilization term in the cut cells. Once consistency and stability of this formulation have been established, we will model the regularization process as a perturbation to this variational problem.
Basic Definitions and Conditions
--------------------------------
We will restrict ourselves to Hilbert spaces ($p=2$), due to the rich theoretical framework available for this case. We will assume that the shape of the domain $\Omega$ is such that we may apply the Stein extension theorem [@adams2003 Section 5.24], i.e., there exists a bounded linear extension operator ${\mathcal{E}}: W^{k,2}(\Omega)\to W^{k,2}(\mathbb{R}^d)$ for any natural number $k$. We explicitly wish to include functions that do not vanish on the boundary ${\partial}\Omega$. For this reason, for any domain $\square\subset\mathbb{R}^d$, we will denote by $W^{-k,2}(\square):=
W^{k,2}(\square)^\prime$ the dual space of $W^{k,2}(\square)$. (Opposed to the convention $W^{-k,2}(\square)=W_0^{k,2}(\square)^\prime$).
We will need certain geometrical definitions. To this end, let $\sigma>0$ be arbitrary but fixed. We define the fictitious domain $\Omega_\sigma$ as the union of all elements that intersect the physical domain $\Omega$. Furthermore we define cut and uncut elements $\Omega_\sigma^\Gamma$ and $\Omega_\sigma^\circ$, respectively: $$\begin{split}
\Omega_\sigma &:= \operatorname{int}\ \bigcup\bigl\lbrace \operatorname{clos}{Q_i}\,\bigr|\,
\operatorname{meas}_d(Q_i\cap\Omega)>0\bigr\rbrace,\\
\Omega_\sigma^\Gamma &:= \operatorname{int}\ \bigcup\bigl\lbrace \operatorname{clos}{Q_i}\,\bigr|\,
Q_i\in\Omega_\sigma\wedge Q_i\not\subset\Omega\bigr\rbrace,\\
\Omega_\sigma^\circ &:= \operatorname{int}\ \bigcup\bigl\lbrace \operatorname{clos}{Q_i}\,\bigr|\,
Q_i\in\Omega_\sigma\wedge Q_i\subset\Omega\bigr\rbrace,
\end{split}$$ with $Q_i$ as in \[eqn:elements\]. Here, we write under a slight abuse of notation $Q_i\in\Omega_\sigma^\Gamma$ if $Q_i\subset\Omega_\sigma^\Gamma$. These domains obviously fulfill $\Omega_\sigma^\circ\subset\Omega\subset\Omega_\sigma$, $\Omega_\sigma = \operatorname{int}(\operatorname{clos}\Omega_\sigma^\circ\cup\operatorname{clos}\Omega_\sigma^\Gamma)$, and $\Omega_\sigma^\circ\cap\Omega_\sigma^\Gamma=\emptyset$. Two elements $Q_i$ and $Q_i'$ will be called neighbors if they share at least one node on the Cartesian grid. We will make the following somewhat technical assumption: for every $Q_i\in\Omega_\sigma^\Gamma$ there is a finite sequence of elements $(Q_i=Q_{i,1},
Q_{i,2},\ldots,Q_{i,K})\subset\Omega_\sigma^\Gamma$ with the following properties: the number $K$ is bounded independent of $\sigma$, every pair of two subsequent elements are neighbors, and $Q_{i,K}$ has a neighbor in $\Omega_\sigma^\circ$. This condition means that one can always reach uncut elements from cut elements in a bounded number of steps. For sufficiently fine Cartesian grids this condition is often fulfilled with $K=1$; if necessary it can be enforced by moving additional elements from $\Omega_\sigma^\circ$ to $\Omega_\sigma^\Gamma$.
Introduction of a Higher-order Stabilization Term
-------------------------------------------------
The basic idea of the ghost penalty method is to control the norm of cut elements by relating them to neighboring uncut elements. In the aforementioned articles [@burman2012; @hansbo2013; @kirchhart2016], for example, this is done by controlling the norms of the gradient-jumps at element boundaries. However, as our PUFEM spaces are globally smooth, they do not contain such jumps. Burman and Fernández [@burman2014] and Cattaneo et al. [@cattaneo2015] instead use the Brezzi–Pitkäranta stabilization [@brezzi1984]. We will use a higher-order variant of this idea and define the following bilinear form: $$j(u_\sigma,v_\sigma) :=
\sigma^{2(P+1)}\sum_{Q_i\in \Omega_\sigma^\Gamma}\sum_{|\alpha|=P+1}
{\int_{Q_i}^{} {({\partial}^\alpha u_\sigma)({\partial}^\alpha v_\sigma)}\,{\mathrm d}{x}},$$ such that $j(u_\sigma,u_\sigma) = \sigma^{2(P+1)}|u_\sigma|_{W^{P+1,2}
(\Omega_\sigma^\Gamma)}^2$. We then obtain the following result:
Let $u_\sigma\in V_\sigma^P(\Omega_\sigma)$. One then has with constants $c$ and $C$ independent of $\sigma$ and the position of ${\partial}\Omega$ relative to the Cartesian grid: $$\label{eqn:jestimate}
c\Vert u_\sigma\Vert^2_{L^2(\Omega_\sigma)} \leq
\Vert u_\sigma\Vert^2_{L^2(\Omega_\sigma^\circ)} + j(u_\sigma,u_\sigma) \leq
C\Vert u_\sigma\Vert^2_{L^2(\Omega_\sigma)}.$$
(0,0) rectangle node [$Q_i'$]{} (1.5,1.5); (1.5,1.5) rectangle node [$Q_i$]{} (3,3); (1.5,1.5) circle \[radius=0.1\] node \[anchor=north west\] [$x^*$]{}; (0,2.5) .. controls (2,3) and (3,2.5) .. (4,1.5) node \[above=10pt\] [${\partial}\Omega$]{};
Before moving on to the proof of this lemma, let us remark that the stabilization term is necessary. Look for example at the configuration shown in \[fig:stablem\]. The partition of unity function corresponding to the node of $Q_i$ opposite to $x^*$ vanishes on $\Omega_\sigma^\circ$. Thus its $L^2(\Omega_\sigma)$-norm cannot be controlled by looking at $\Omega_\sigma^\circ$ only, unless one adds a stabilization term.
The second inequality directly follows from the inverse inequalities \[eqn:investimates\]. For the first inequality, let us first consider the case $K=1$, i.e., a cut element $Q_i\in\Omega_\sigma^\Gamma$ and an associated uncut element $Q_i'\in\Omega_\sigma^\circ$ which share a Cartesian grid point $x^*$, as for example illustrated in \[fig:stablem\]. This configuration can be mapped to one of $P^{2^d}_2 = 4^d-2^d$ reference cases with reference elements $\hat{Q}$ and $\hat{Q}'$ using the transformation $\hat{x}=\Phi^{-1}(x):=(x-x^*)/\sigma$, such that $\hat{x}^*=0$. For an arbitrary function $v_\sigma\in V_\sigma^P(Q_i\cup Q_i')$ one obtains with $\hat{v} :=
(v_\sigma\circ\Phi)\in V_1^P(\hat{Q}\cup\hat{Q}')$: $$\Vert v_\sigma\Vert_{L^2(Q_i\cup Q_i')}^{2} =
\sigma^d\Vert\hat{v}\Vert_{L^2(\hat{Q}\cup\hat{Q}')}^2.$$ We claim that the following expression constitutes a norm on $V_1^P(\hat{Q}\cup\hat{Q}')$: $$\Vert\hat{v}\Vert_*^2 := \Vert\hat{v}\Vert_{L^2(\hat{Q}')}^2 +
\vert\hat{v}\vert_{W^{P+1,2}(\hat{Q})}^2.$$
It suffices to show that $\Vert\hat{v}\Vert_{*} = 0 \Longrightarrow\hat{v}= 0.$ From $\Vert\hat{v}\Vert_{L^2(\hat{Q}')}^2=0$ we obtain $\hat{v}\equiv 0$ on $\hat{Q}'$ and due to the global smoothness of $\hat{v}$ also ${\partial}^\alpha\hat{v}
(\hat{x}^*)= 0$ for all multi-indices $\alpha\in\mathbb{N}_0^d$. From $|\hat{v}|_{W^{P+1,2}(\hat{Q})}^2 = 0$ we obtain ${\partial}^\alpha\hat{v}\equiv 0,
|\alpha| = P + 1$ on $\hat{Q}$. Together with ${\partial}^\alpha\hat{v}(\hat{x}^*)=0$ this implies $\hat{v}\equiv 0$ on $\hat{Q}$ as well. Thus $\Vert\cdot\Vert_*$ is indeed a norm. After employing the norm-equivalence of finite-dimensional spaces, we can transform back to $Q_i\cup Q_i'$ and obtain: $$\sigma^d\Vert\hat{v}\Vert_{L^2(\hat{Q}\cup\hat{Q}')}^2 \leq
C\sigma^d\Vert\hat{v}\Vert_*^2 =
C\bigl(\Vert v_\sigma\Vert_{L^2(Q_i')}^{2} + \sigma^{2(P+1)}\vert v_\sigma\vert_{W^{P+1,2}(Q_i)}^{2}\bigr).$$ The case $K>1$ with sequences of cells $(Q_i = Q_{i,1}, Q_{i,2},\ldots, Q_{i,K} = Q_i')$ follows by induction. Now, summing over all elements and using the finite overlap condition $M=2^d$, the claim follows.
Note that the proof crucially depends on the global smoothness of the spaces $V_\sigma^P$. In particular, this stabilization does not work with the conventional finite element spaces. As an example consider the case depicted in \[fig:stablem\], and set $v_\sigma:=0$ on $Q_i'$, and $v_\sigma:=(x - x^*)\cdot y$ on $Q_i$, where $y\in\mathbb{R}^d$ is an arbitrary non-zero vector.
Stability and Convergence
-------------------------
As described before, we are aiming to mimic the $L^2(\Omega)$-orthogonal projector. To this end, we introduce the bilinear forms $a$ and $A$: $$\begin{aligned}
a:&\ L^2(\Omega_\sigma)\times L^2(\Omega_\sigma)\to \mathbb{R},\quad
&(u,v)&\mapsto{\int_{\Omega}^{} {uv}\,{\mathrm d}{x}},\\
A:&\ W^{P+1,2}(\Omega_\sigma)\times W^{P+1,2}(\Omega_\sigma)\to\mathbb{R},\quad
&(u_\sigma,v_\sigma)&\mapsto a(u_\sigma,v_\sigma) + \varepsilon j(u_\sigma,v_\sigma),
\label{eqn:bilinear_form_A}\end{aligned}$$ where $\varepsilon>0$ denotes a user-defined stabilization parameter. We define the variational problem as: given any $u$ for which the following makes sense, find $u_\sigma\in V_\sigma^P(\Omega_\sigma)$ such that: $$\label{eqn:varprob}
A(u_\sigma,v_\sigma) = {\int_{\Omega}^{} {uv_\sigma}\,{\mathrm d}{x}} \qquad
\forall v_\sigma\in V_\sigma^P(\Omega_\sigma).$$ We then obtain the following two results.
\[thm:stability\] The bilinear form $A$ from \[eqn:bilinear\_form\_A\] fulfills with a constant $C(\varepsilon)$ independent of $\sigma$, $u_\sigma$, and the position of ${\partial}\Omega$ relative to the grid: $$A(u_\sigma,u_\sigma)\geq C(\varepsilon)\Vert u_\sigma\Vert_{L^2(\Omega_\sigma)}^2\quad
\forall u_\sigma\in V_\sigma^P(\Omega_\sigma).$$
For any $u_\sigma\in V_\sigma^P(\Omega_\sigma)$ one has with the help of \[eqn:jestimate\]: $$\begin{gathered}
A(u_\sigma,u_\sigma) =
\Vert u_\sigma\Vert_{L^2(\Omega)}^2 + \varepsilon j(u_\sigma,u_\sigma) \geq \\
\Vert u_\sigma\Vert_{L^2(\Omega_\sigma^\circ)}^2 + \varepsilon j(u_\sigma,u_\sigma)
\geq C(\varepsilon)\Vert u_\sigma\Vert_{L^2(\Omega_\sigma)}^2.\end{gathered}$$
\[thm:convergence\] Let $u\in W^{P+1,2}(\Omega_\sigma)$. The solution $u_\sigma\in V_\sigma^P(\Omega_\sigma)$ of the variational problem \[eqn:varprob\] then satisfies the following error bound: $$\Vert u_\sigma - u\Vert_{L^2(\Omega_\sigma)} \leq C(\varepsilon)
\sigma^{P+1}\Vert u\Vert_{W^{P+1,2}(\Omega_\sigma)},$$ where the constant $C(\varepsilon)$ is independent of $\sigma$, $u$, and how the boundary ${\partial}\Omega$ intersects the grid.
According to \[thm:approx\], there exists ${\mathcal{P}}u\in V_\sigma^P(\Omega_\sigma)$ such that: $$\label{eqn:approx2}
\vert{\mathcal{P}}u-u\vert_{W^{k,2}(\Omega_\sigma)}\leq C\sigma^{P+1-k}
\vert u\vert_{W^{P+1,2}(\Omega_\sigma)}\qquad k\in\mathbb{N}_0,\ k\leq P + 1.$$ We may write: $$\Vert u_\sigma - u\Vert_{L^2(\Omega_\sigma)} \leq
\Vert u_\sigma - {\mathcal{P}}u\Vert_{L^2(\Omega_\sigma)} +
\Vert {\mathcal{P}}u - u\Vert_{L^2(\Omega_\sigma)}.$$ For the second term we can apply relation \[eqn:approx2\]. For the first term we obtain with \[thm:stability\] and the fact that $u_\sigma$ solves \[eqn:varprob\]: $$\begin{gathered}
\Vert {\mathcal{P}}u - u_\sigma\Vert_{L^2(\Omega_\sigma)}^2 \leq C(\varepsilon)
A({\mathcal{P}}u-u_\sigma,{\mathcal{P}}u-u_\sigma) = \\
C(\varepsilon)\biggl(\bigl({\mathcal{P}}u-u,{\mathcal{P}}u-u_\sigma\bigr)_{L^2(\Omega)} +
\varepsilon j({\mathcal{P}}u,{\mathcal{P}}u-u_\sigma)\biggr) \leq \\
C(\varepsilon)\biggl(
\Vert{\mathcal{P}}u-u\Vert_{L^2(\Omega_\sigma)}\Vert{\mathcal{P}}u-u_\sigma\Vert_{L^2(\Omega_\sigma)} +
\varepsilon j({\mathcal{P}}u,{\mathcal{P}}u)^{1/2}j({\mathcal{P}}u-u_\sigma,{\mathcal{P}}u-u_\sigma)^{1/2}
\biggr),\end{gathered}$$ where we used the Cauchy–Schwarz inequality in the last step. Noting that by the inverse estimates \[eqn:investimates\] we have: $$j({\mathcal{P}}u-u_\sigma,{\mathcal{P}}u-u_\sigma)^{1/2} \leq
C\Vert{\mathcal{P}}u-u_\sigma\Vert_{L^2(\Omega_\sigma)}$$ and together with \[eqn:approx2\]: $$j({\mathcal{P}}u,{\mathcal{P}}u)^{1/2} \leq C\sigma^{P+1}
\Vert{\mathcal{P}}u\Vert_{W^{P+1,2}(\Omega_\sigma)} \leq
C\sigma^{P+1}
\Vert u\Vert_{W^{P+1,2}(\Omega_\sigma)}.$$ After dividing both sides by $\Vert {\mathcal{P}}u - u_\sigma\Vert_{L^2(\Omega_\sigma)}$ we thus obtain: $$\label{eqn:omfg}
\Vert{\mathcal{P}}u - u_\sigma\Vert_{L^2(\Omega_\sigma)} \leq
C(\varepsilon)\biggl(\Vert{\mathcal{P}}u-u\Vert_{L^2(\Omega_\sigma)} +
\varepsilon\sigma^{P+1}\Vert u\Vert_{W^{P+1,2}(\Omega_\sigma)}\biggr).$$ Again applying \[eqn:approx2\] to the first term yields the claim.
Influence of the Quadrature Error
---------------------------------
In vortex methods we are only given a particle field, i.e., a quadrature rule for integrating smooth functions against the underlying vorticity we are aiming to approximate. Furthermore the bilinear form $A$ can usually only be computed approximately, using numerical quadrature. In this subsection we are analyzing the influence of these additional sources of error.
We will assume that the bilinear form $j$ can be computed exactly. This is justified as it is sufficient to perform computations on the reference element $\hat{Q}$, which can be done up to arbitrary precision a priori. As for the bilinear form $a$, we will assume the availability of quadrature rules $I_m$ satisfying error bounds of the following form: $$\biggl|{\int_{\Omega}^{} {f}\,{\mathrm d}{x}}-I_m(f)\biggr| \leq
Ch^{m+1}\vert f\vert_{W^{m+1,1}(\Omega)} \quad f\in W^{m+1,1}(\Omega).$$ Such error estimates typically arise from the application of quadrature rules of exactness degree $m$ and positive weights to the cells of a quasi-uniform triangulation of the domain $\Omega$ of mesh-width $h$. Note that due to the global smoothness of the PUFEM spaces, these quadrature rules *do not need to be aligned* with the Cartesian grid. We will write $a_h(u_\sigma,v_\sigma):= I_m(u_\sigma v_\sigma)$ for the resulting approximate bilinear form. For $u_\sigma,v_\sigma\in
V_\sigma^P(\Omega_\sigma)$ one then obtains with the help of Hölder’s inequality and the inverse estimates \[eqn:investimates\]: $$\begin{gathered}
\label{eqn:aerror}
|a(u_\sigma,v_\sigma)-a_h(u_\sigma,v_\sigma)| \leq
Ch^{m+1}|u_\sigma v_\sigma|_{W^{m+1,1}(\Omega_\sigma)} \leq \\
Ch^{m+1}\sum_{|\alpha|=m+1}\sum_{\beta\leq\alpha}
\begin{pmatrix}\alpha \\ \beta\end{pmatrix}
\Vert{\partial}^\beta u_\sigma {\partial}^{\alpha-\beta}
v_\sigma\Vert_{L^1(\Omega_\sigma)} \leq \\
Ch^{m+1}\sum_{|\alpha|=m+1}\sum_{\beta\leq\alpha}
\begin{pmatrix}\alpha \\ \beta\end{pmatrix}
\Vert{\partial}^\beta u_\sigma\Vert_{L^2(\Omega_\sigma)}
\Vert{\partial}^{\alpha-\beta} v_\sigma\Vert_{L^2(\Omega_\sigma)} \leq \\
Ch^{m+1}\sigma^{-(m+1)}
\Vert u_\sigma\Vert_{L^2(\Omega_\sigma)}
\Vert v_\sigma\Vert_{L^2(\Omega_\sigma)}.\end{gathered}$$ This will require us to couple $h$ and $\sigma$ through a relation like $h = \sigma^s$, for some $s>1$. We then obtain coercivity of $A_h(u_\sigma,v_\sigma):= a_h(u_\sigma,v_\sigma) +
j(u_\sigma,v_\sigma)$: $$\begin{gathered}
A_h(u_\sigma,u_\sigma) = A(u_\sigma,u_\sigma)-
\bigl(a(u_\sigma,u_\sigma)-a_h(u_\sigma,u_\sigma)\bigr) \geq \\
\bigl( C(\varepsilon) - Ch^{m+1}\sigma^{-(m+1)} \bigr)
\Vert u_\sigma\Vert^2_{L^2(\Omega_\sigma)} \geq
C(\varepsilon)\Vert u_\sigma\Vert^2_{L^2(\Omega_\sigma)},\end{gathered}$$ where the last constant $C(\varepsilon)$ is independent of $\sigma$, $u$, and the position of ${\partial}\Omega$ relative to the Cartesian grid, for $\sigma>0$ small enough, $h=\sigma^s$, $s>1$.
For the particle field $u_h$ we will assume an error bound of the following form: $$\label{eqn:particleerror}
\Vert u_h - u \Vert_{W^{-(m+1),2}(\Omega)}\leq
Ch^{m+1}\Vert u\Vert_{W^{m+1,2}(\Omega)},$$ which is the typical form arising in vortex methods [@cottet2000]. Again, the particle field does not in any way need to be aligned to the Cartesian grid. Collecting all of the previous results, we are ready to prove the main result of this article.
\[thm:perturbedprob\] Let $h=\sigma^s$, $s>1$, and denote $k:=\max\lbrace P, m\rbrace$. Let $u\in W^{k+1,2}(\Omega)$ and let the particle approximation $u_h\in W^{-(m+1),2}(\Omega)$ satisfy the error bound \[eqn:particleerror\]. Then for $\sigma>0$ small enough the solution $u_\sigma\in V_\sigma^P(\Omega_\sigma)$ of the perturbed variational problem: $$\label{eqn:perturbedprob}
A_h(u_\sigma,v_\sigma) = \langle u_h,v_\sigma\rangle\quad\forall v_\sigma\in V_\sigma^P(\Omega_\sigma)$$ satisfies the following error bound: $$\Vert u - u_\sigma\Vert_{L^2(\Omega)} \leq C(\varepsilon)
\bigl(
\sigma^{P+1} +
h^{m+1}\sigma^{-(m+1)}
\bigr)\Vert u\Vert_{W^{k+1,2}(\Omega)},$$ where the constant $C(\varepsilon)$ is independent of $\sigma$, $u$, and the position of ${\partial}\Omega$ relative to the Cartesian grid.
Let $v_\sigma\in V_\sigma^P(\Omega_\sigma)$ denote the solution of the unperturbed variational problem \[eqn:varprob\], with $u$ extended to ${\mathcal{E}}u$ by the Stein extension operator. With the help of the coercivity of $A_h$ one then obtains: $$\begin{gathered}
\Vert u_\sigma-v_\sigma\Vert_{L^2(\Omega_\sigma)}^2 \leq C(\varepsilon)
A_h(u_\sigma-v_\sigma,u_\sigma-v_\sigma) = \\
C(\varepsilon)\biggl(
\langle u_h-u,u_\sigma-v_\sigma\rangle +
A(v_\sigma,u_\sigma-v_\sigma) - A_h(v_\sigma,u_\sigma-v_\sigma)
\biggr).\end{gathered}$$ Application of the error bounds \[eqn:particleerror\] and \[eqn:aerror\] as well as the inverse estimates \[eqn:investimates\] yields: $$\begin{gathered}
\Vert u_\sigma-v_\sigma\Vert_{L^2(\Omega_\sigma)}^2 \leq C(\varepsilon)
\bigl(
h^{m+1}\sigma^{-(m+1)}\Vert u\Vert_{W^{m+1,2}(\Omega)}
\Vert u_\sigma - v_\sigma\Vert_{L^2(\Omega_\sigma)} + \\
h^{m+1}\sigma^{-(m+1)}
\Vert v_\sigma\Vert_{L^2(\Omega_\sigma)}
\Vert u_\sigma - v_\sigma\Vert_{L^2(\Omega_\sigma)}
\bigr).\end{gathered}$$ and thus: $$\Vert u_\sigma-v_\sigma\Vert_{L^2(\Omega_\sigma)} \leq
C(\varepsilon)h^{m+1}\sigma^{-(m+1)}
\bigl(
\Vert u\Vert_{W^{m+1,2}(\Omega)} +
\Vert v_\sigma\Vert_{L^2(\Omega_\sigma)}
\bigr).$$ Nothing that $\Vert v_\sigma\Vert_{L^2(\Omega_\sigma)}\leq
C(\varepsilon)\Vert u\Vert_{W^{P+1,2}(\Omega)}$ we obtain: $$\label{eqn:lol}
\Vert u_\sigma-v_\sigma\Vert_{L^2(\Omega_\sigma)} \leq
C(\varepsilon)h^{m+1}\sigma^{-(m+1)}\Vert u\Vert_{W^{k+1,2}(\Omega)}.$$ Now, by the triangle inequality: $$\Vert u_\sigma- u\Vert_{L^2(\Omega)} \leq
\Vert u_\sigma- v_\sigma\Vert_{L^2(\Omega)} +
\Vert v_\sigma-u\Vert_{L^2(\Omega)}$$ the claim follows by applying \[thm:convergence\] and the boundedness of ${\mathcal{E}}$ to the second term.
The part $\sigma^{P+1}\Vert u\Vert_{W^{k+1,2}(\Omega)}$ is called the smoothing error; $\sigma$ roughly corresponds to the blob-width in conventional vortex particle methods. The second part is called the quadrature error; choosing $s=1+\tfrac{P+1}{m+1}$ balances both terms. In the next section we will illustrate that the choice $P=m$ does not only “feel natural”, but also yields optimal results in a certain sense. In this case we obtain with $s=2$ an overall convergence rate of ${\mathcal{O}({\sigma^{P+1}})}={\mathcal{O}({h^{\frac{1}{2}(P+1)}})}$.
Optimality of the Smoothed Solution
-----------------------------------
In this subsection we will assume that we can apply the bilinear form $A$ exactly, i.e., $A_h = A$. Furthermore we assume $P=m$. We will show that the smoothed solution $u_\sigma$ then satisfies the same asymptotic error bound as $u_h$. We will need the following corollary of \[thm:convergence\].
\[thm:corollary\] The solution operator ${\mathcal{S}}$ to the problem \[eqn:varprob\] is bounded: $$\Vert{\mathcal{S}}u\Vert_{W^{P+1,2}(\Omega_\sigma)}\leq C(\varepsilon)
\Vert u\Vert_{W^{P+1,2}(\Omega_\sigma)}\qquad\forall u\in W^{P+1,2}(\Omega_\sigma).$$
One has $\Vert{\mathcal{S}}u\Vert_{W^{P+1,2}(\Omega_\sigma)}\leq
\Vert{\mathcal{S}}u - u\Vert_{W^{P+1,2}(\Omega_\sigma)} + \Vert u \Vert_{W^{P+1,2}(\Omega_\sigma)}.$ Using ${\mathcal{P}}$ from \[thm:approx\], we furthermore obtain: $\Vert\solve u - u\Vert_{W^{P+1,2}(\Omega_\sigma)} \leq
\Vert\solve u - \project u\Vert_{W^{P+1,2}(\Omega_\sigma)} +
\Vert\project u - u\Vert_{W^{P+1,2}(\Omega_\sigma)}.$ The second term can be bounded by $C\Vert u \Vert_{W^{P+1,2}(\Omega_\sigma)}$ due to the boundedness of ${\mathcal{P}}$. The first term can be bounded by first applying the inverse estimates \[eqn:investimates\] followed by estimate \[eqn:omfg\].
\[thm:optimality\] Let the conditions of \[thm:perturbedprob\] be fulfilled. Furthermore assume that $A_h = A$, $P = m$, and $s = 2$. Then the smoothed solution $u_\sigma$ fulfills: $$\Vert u_\sigma - u\Vert_{W^{-(P+1),2}(\Omega)}\leq
C(\varepsilon)h^{P+1} \Vert u\Vert_{W^{P+1,2}(\Omega)}.$$
Let $\varphi\in W^{P+1,2}(\Omega)$ be arbitrary but fixed. With ${\mathcal{P}}$ from \[thm:approx\] and the Stein extension operator ${\mathcal{E}}$ one has: $${\int_{\Omega}^{} {(u_\sigma-u)\varphi}\,{\mathrm d}{x}} =
{\int_{\Omega}^{} {(u_\sigma-u)(\varphi-{\mathcal{P}}{\mathcal{E}}\varphi)}\,{\mathrm d}{x}} +
{\int_{\Omega}^{} {(u_\sigma-u){\mathcal{P}}{\mathcal{E}}\varphi}\,{\mathrm d}{x}}.$$ For the first term we obtain with the Cauchy–Schwarz inequality, \[thm:approx\], and \[thm:perturbedprob\]: $$\begin{gathered}
{\int_{\Omega}^{} {(u_\sigma-u)(\varphi-{\mathcal{P}}{\mathcal{E}}\varphi)}\,{\mathrm d}{x}} \leq
\Vert u - u_\sigma\Vert_{L^2(\Omega)}\Vert \varphi-{\mathcal{P}}{\mathcal{E}}\varphi\Vert_{L^2(\Omega)} \\ \leq
C(\varepsilon)\sigma^{2(P+1)}\Vert u\Vert_{W^{P+1,2}(\Omega)}
\Vert\varphi\Vert_{W^{P+1,2}(\Omega)}.\end{gathered}$$ For the second term one has: $$\begin{gathered}
{\int_{\Omega}^{} {(u_\sigma-u)\mathcal{PE}\varphi}\,{\mathrm d}{x}} =
A(u_\sigma,{\mathcal{P}}{\mathcal{E}}\varphi) - {\int_{\Omega}^{} {u{\mathcal{P}}{\mathcal{E}}\varphi}\,{\mathrm d}{x}}
-\varepsilon j(u_\sigma,{\mathcal{P}}{\mathcal{E}}\varphi) = \\
\langle u_h-u,{\mathcal{P}}{\mathcal{E}}\varphi\rangle - \varepsilon j(u_\sigma,{\mathcal{P}}{\mathcal{E}}\varphi) \leq \\
C\biggl( \Vert u_h-u\Vert_{W^{-(P+1),2}(\Omega)}\Vert\varphi\Vert_{W^{P+1,2}(\Omega)} +
\varepsilon\sigma^{2(P+1)}\Vert u_\sigma\Vert_{W^{P+1,2}(\Omega_\sigma)}
\Vert\varphi\Vert_{W^{P+1,2}(\Omega)}\biggr).\end{gathered}$$ It remains to show that $\Vert u_\sigma\Vert_{W^{P+1,2}(\Omega_\sigma)} \leq
C(\varepsilon)\Vert u\Vert_{W^{P+1,2}(\Omega)}$. To see this, note that we have: $$\begin{gathered}
\Vert u_\sigma\Vert_{W^{P+1,2}(\Omega_\sigma)} \leq
\Vert{\mathcal{S}}{\mathcal{E}}u\Vert_{W^{P+1,2}(\Omega_\sigma)} +
\Vert u_\sigma - {\mathcal{S}}{\mathcal{E}}u\Vert_{W^{P+1,2}(\Omega_\sigma)} \leq\\
C(\varepsilon)\Vert u\Vert_{W^{P+1,2}(\Omega)} +
\Vert u_\sigma - {\mathcal{S}}{\mathcal{E}}u\Vert_{W^{P+1,2}(\Omega_\sigma)}.\end{gathered}$$ Applying inequality \[eqn:lol\] to the second term, collecting all the terms, and noting that $\sigma = \sqrt{h}$ yields the result.
Conservation Properties
-----------------------
In the introduction we mentioned the conservation properties of vortex methods as one of their highlights. In this section we make some brief remarks on some of these properties under the assumption that $A_h = A$. For brevity, we will focus on the two-dimensional case, but remark that all of the results we present here analogously hold in three-dimensions.
The conserved quantities circulation, linear momentum, and angular momentum are given by $I_0={\int_{\Omega}^{} {1\cdot\omega}\,{\mathrm d}{x}}$, ${\mathbf{I}}_1={\int_{\Omega}^{} {(x_2,-x_1)^\top\omega}\,{\mathrm d}{x}}$, and $I_2={\int_{\Omega}^{} {|x|^2\omega}\,{\mathrm d}{x}}$, respectively [@majda2001 Section 1.7]. Noting that the stabilization term $j$ vanishes if one of its arguments is a polynomial of total degree less than $P$, one obtains for the solution $\omega_\sigma$ of \[eqn:varprob\] with right-hand side $\omega_h$: $(\omega_\sigma,x^\alpha)_{L^2(\Omega)} = \omega_h(x^\alpha)$ for all $|\alpha|\leq P$. For $P=1$ we consequently conserve $I_0$ and ${\mathbf{I}}_1$, for $P=2$ one additionally conserves angular momentum $I_2$. This is important, because in vortex methods body forces are often computed using the relation ${\mathbf{F}} = -\rho{\frac{{\mathrm d}{{\mathbf{I}}_1}}{{\mathrm d}{t}}}$, where $\rho$ denotes the fluid’s density.
Numerical Experiments {#sec:experiments}
=====================
We have now established the necessary results to return to our original motivation. Given a particle approximation $\omega_h\in W^{-(m+1),2}(\Omega)$ of the vorticity $\omega$ that satisfies an error-bound of the form $\Vert\omega_h-\omega\Vert_{W^{-(m+1),2}(\Omega)}
\leq Ch^{m+1}\Vert\omega\Vert_{W^{m+1,2}(\Omega)}$, we want to obtain a smooth approximation $\omega_\sigma$, such that we can compute the corresponding induced velocity field using the Biot–Savart law ${\mathbf{u}}_\sigma ={\mathbf{K}}\star\omega_\sigma$. One can then use this approximate velocity field to advance $\omega_h$ in time by convecting the particles according to ${\frac{{\mathrm d}{x_i}}{{\mathrm d}{t}}}(t)={\mathbf{u}}_\sigma(x_i(t),t)$.
In \[sec:pufem\] we introduced the spaces $V_\sigma^P(\Omega)$ that can be used as test-spaces for the particle field $\omega_h$. In \[sec:formulation\] we modeled the regularization problem as a perturbation to a stabilized $L^2$-projection onto the spaces $V_\sigma^P(\Omega)$. The analysis indicated that one should choose $P=m$ and $\sigma=\sqrt{h}$, resulting in an a-priori error estimate of $\Vert\omega_\sigma-\omega\Vert_{L^2(\Omega)}\leq C\sigma^{P+1}\Vert\omega\Vert_{W^{P+1,2}(\Omega)}$. The Calderón–Zygmund inequality \[eqn:calzyg\] then tells us that one may expect $\Vert{\mathbf{u}}_\sigma-{\mathbf{u}}\Vert_{W^{1,2}(\Omega)}\leq C\sigma^{P+1}\Vert\omega\Vert_{W^{P+1,2}(\Omega)}$ for the resulting velocity field ${\mathbf{u}}_\sigma := {\mathbf{K}}\star\omega_\sigma$. This analogously holds in the three-dimensional case.
In this section we perform several numerical experiments. We will first describe the experimental setup. We then perform experiments on a scalar particle field and confirm the results of our analysis. In particular, the experiments will show that the common practice of choosing $\sigma$ proportional to $h$ instead of $\sqrt{h}$ does not lead to convergent schemes. We will then illustrate the practicality of our scheme, by approximating a vector-valued vorticity field, computing its induced velocity field, and measuring the error. We finish this section with experiments on the condition number of the resulting systems and its dependence of the stabilization parameter $\varepsilon$.
Setup {#sec:setup}
-----
We define our computational test domain as $\Omega=(-\tfrac{1}{2},\tfrac{1}{2})^3$. While this is one of the simplest cases for mesh-based methods, due to its sharp corners and edges it is one of the hardest for conventional vortex blob methods. In order to obtain quadrature rules which are not aligned to the Cartesian grid, the mesh generator Gmsh [@geuzaine2009] was used to obtain a tetrahedral mesh of the domain, consisting of 24 tetrahedra with maximum edge-length $h=1$. The quadrature rules are obtained by applying the mid-point rule to this mesh and its subsequent uniform refinements from level $l=0$ down to level $l = 8$, corresponding to $h = 2^{-8}\approx 0.004$ and $N = 402\,653\,184$ quadrature nodes.
Preliminary experiments showed good results for a stabilization parameter of $\varepsilon = 0.001$. Unless explicitly stated otherwise, we will use this value for all of our computations. We will use degree $P=1$ for the PUFEM spaces, set $\sigma := Ch^{1/s}$, and experiment on various choices of $C$ and $s=1,2$. For the integration of the bilinear form $A_h$ we use the following approach: if in a pair of basis functions one of them has cut support, we use the same quadrature rule as for the particle field. Otherwise precomputed values from the reference element $\hat{Q}$ are used. The resulting systems of equations are solved using the conjugate gradient method, where we apply a simple diagonal scaling as preconditioner. The iteration was stopped when a relative residual of $10^{-12}$ was reached. This was usually the case after less than 100 iterations, with some exceptions for coarse refinement levels $l$ and the case $C=0.5, s = 1$.
Scalar Particle Field {#sec:scalar_field}
---------------------
The common practice to choose the smoothing length $\sigma$ proportional to $h$ may in special cases be justified with the analysis of Cottet and Koumoutsakos [@cottet2000 Section 2.6]. They assume that the quadrature rules used are of infinite order, essentially corresponding to the case $m=\infty$. Such rules, however, typically only exist in very special cases, such as a cube with periodic or zero boundary conditions. To show that this approach does not work in a more general setting, we aim to approximate the following function: $$u(x) = \cos{(4\pi x_1)}\qquad
x\in (-\tfrac{1}{2},\tfrac{1}{2})^3.$$ This function *does not vanish at the boundary.* The application of conventional blob-methods would thus blur the boundaries and lead to only slowly converging schemes. We define the particle field as $u_h:=\sum_{i=1}^N w_i u(x_i)
\delta(x-x_i)$, with $\delta$ denoting the Dirac Delta, and $x_i$ and $w_i$ being the positions and weights of the mid-point quadrature rule applied to the tetrahedra of the mesh at various refinement levels $l$.
shows the error $\Vert u - u_\sigma\Vert_{L^2(\Omega)}$ for $\sigma = Ch$ for various choices of $C$ at different refinement levels. Choosing $C=0.5$ results in approximations with large errors, which do not decrease significantly under mesh refinement. The case $l=8$ was not computed due to the large memory requirements. The other curves exhibit similar behavior: in the beginning and intermediate stages the error decreases, however, only at an approximately linear, not quadratic rate. This rate further decreases and approaches zero under mesh refinement, confirming the predicted bound of the quadrature error ${\mathcal{O}({h^{m+1}\sigma^{-(m+1)}})}={\mathcal{O}({1})}$. Choosing larger values $C$ somewhat delays but does not prevent this effect, at the cost of larger errors on coarse refinement levels.
shows the corresponding error for the case $\sigma=C\sqrt{h}$. All choices of $C$ lead to convergent schemes which approach the predicted convergence rate of ${\mathcal{O}({h})}$. In our experiments, smaller choices of $C$ lead to smaller errors; however choosing $C$ too small causes larger errors in the coarser cases. In our test case a choice somewhere between $C=0.25$ and $C=0.5$ seems to be optimal.
\[ title = [$s=1$]{}, xlabel = [Refinement Level $l$]{}, ylabel = [$L^2$-Error]{}, grid = [major]{}, xmin = [0]{}, xmax = [8]{}, ymin = [1e-3]{}, ymax = [1]{}, legend entries = [$C=2$, $C=1.5$, $C=1$, $C=0.5$]{}, legend cell align = [left]{}, \] table\[x=l,y=C\_2\_00\][cosine\_s1.dat]{}; table\[x=l,y=C\_1\_50\][cosine\_s1.dat]{}; table\[x=l,y=C\_1\_00\][cosine\_s1.dat]{}; table\[x=l,y=C\_0\_50\][cosine\_s1.dat]{}; node\[below=5,pos=0.5\] [${\mathcal{O}({h})}$]{};
\[ title = [$s=2$]{}, xlabel = [Refinement Level $l$]{}, ylabel = [$L^2$-Error]{}, grid = [major]{}, xmin = [0]{}, xmax = [8]{}, ymin = [1e-3]{}, ymax = [1]{}, legend entries = [$C=1.5$, $C=1.0$, $C=0.5$, $C=0.25$]{}, legend cell align = [left]{}, \] table\[x=l,y=C\_1\_50\][cosine\_s2.dat]{}; table\[x=l,y=C\_1\_00\][cosine\_s2.dat]{}; table\[x=l,y=C\_0\_50\][cosine\_s2.dat]{}; table\[x=l,y=C\_0\_25\][cosine\_s2.dat]{}; node\[above=5,pos=0.66\] [${\mathcal{O}({h})}$]{};
Vector-valued Particle Field and Velocity Evaluation
----------------------------------------------------
In this section we show that our scheme can drastically reduce the cost of the computationally most expensive part of vortex methods, the velocity evaluation. To this end, we prescribe: $${\mathbf{u}}(x) :=
\begin{pmatrix}
x_2 \\ -x_1 \\ 0
\end{pmatrix}
\exp\biggl(-\frac{1}{1-4|x|^2}\biggr)\qquad
x\in (-\tfrac{1}{2},\tfrac{1}{2})^3.$$ This velocity field is smooth and fulfills $\nabla\cdot{\mathbf{u}}\equiv 0$. It was chosen such that it vanishes at the boundaries, so that it can be retrieved from the vorticity field ${\boldsymbol{\omega}}:=\nabla\times{\mathbf{u}}$ through the Biot–Savart law without any boundary integral terms: $${\mathbf{u}}(x) = -\frac{1}{4\pi}{\int_{\Omega}^{} {\frac{x-y}{\vert x-y\vert^3}\times{\boldsymbol{\omega}}(y)}\,{\mathrm d}{y}}.$$ Analogous to the previous section, we define the particle approximation: $${\boldsymbol{\omega}}_h :=
\sum_{i=1}^{N}w_i{\boldsymbol{\omega}}(x_i)\delta(x-x_i).$$ Experiments in the previous section suggested a choice of $\sigma = C\sqrt{h}$, with $C$ between $0.25$ and $0.5$. We consequently choose $C=0.375$ and obtain after applying the method to each component a smoothed approximation ${\boldsymbol{\omega}}_\sigma$ with an anticipated convergence rate of ${\mathcal{O}({\sigma^2})}$ in the $L^2$-norm. In order to evaluate the Biot–Savart law for this vorticity field, we chose the coarsest level $l$ such that the corresponding mesh width $2^{-l}$ is smaller than $\sigma$. We then compute the orthogonal projection of ${\boldsymbol{\omega}}_\sigma$ onto the *standard* finite element space of piecewise linear functions on that level. The Biot–Savart integral can then be computed by summing over the tetrahedra, for which Suh published analytic formulas [@suh2000]. We couple these formulas with a fast multipole method [@greengard1987; @dehnen2002] for the far-field evaluation. The resulting velocity field is approximated by taking the nodal interpolation onto the standard finite element space of piecewise *quadratics* to obtain an approximate velocity field ${\mathbf{u}}_\sigma$.
Most conventional schemes apply the fast multipole method directly to the particle field, leading to a complexity of ${\mathcal{O}({N})}={\mathcal{O}({h^{-d}})}$, with a large hidden constant. Note that in our case the method is applied to the coarser smoothed approximation, leading to a complexity of only ${\mathcal{O}({h^{-\frac{d}{2}}})}$.
\[fig:velocity\] shows the $L^2$-errors in the approximate smoothed vorticity field ${\boldsymbol{\omega}}_\sigma$ and the velocity field ${\mathbf{u}}_\sigma$. The smoothed vorticity field converges at a rate of ${\mathcal{O}({\sigma^2})}={\mathcal{O}({h})}$ as expected. With the Calderón–Zygmund inequality \[eqn:calzyg\] we obtain that the same error bound holds for the velocity in the $W^{1,2}$-norm. As the results indicate, in the $L^2$-norm the error seems to reduce by one power in $\sigma$ faster, resulting in a rate of ${\mathcal{O}({h^{1.5}})}$.
\[ title = [$C=0.375$]{}, xlabel = [Refinement Level $l$]{}, ylabel = [$L^2$-Error]{}, grid = [major]{}, xmin = [0]{}, xmax = [8]{}, ymin = [1e-6]{}, ymax = [1]{}, legend entries = [vorticity,velocity]{}, legend cell align = [left]{}, \] table\[x=l,y=vorticity\][velocity.dat]{}; table\[x=l,y=velocity \][velocity.dat]{}; node\[above=5,pos=0.5\] [${\mathcal{O}({h})}$]{}; node\[above=5,pos=0.5\] [${\mathcal{O}({h^{1.5}})}$]{};
\[ title = [$C=0.25$, $s = 2$]{}, xlabel = [Stabilization Parameter $\varepsilon$]{}, ylabel = [$\operatorname{cond}({\mathsf{D^{-1}A_h}})$]{}, grid = [major]{}, xmin = [1e-6]{}, xmax = [1]{}, ymin = [1e1]{}, ymax = [1e6]{}, xminorticks = [false]{}, legend entries = [$l=2$, $l=3$, $l=4$]{}, legend cell align = [left]{}, \] table\[x=eps,y=l2\][condition.dat]{}; table\[x=eps,y=l3\][condition.dat]{}; table\[x=eps,y=l4\][condition.dat]{};
System Condition Number {#sec:conditioning}
-----------------------
In this section we investigate the effect of the stabilization parameter $\varepsilon$ on the condition number of the system matrix. In \[sec:scalar\_field\] we observed instabilities on the coarse levels $l$ in the case $C=0.25, s=2$. We therefore chose this particular configuration for our experiments. We used the following set of functions as our basis for the spaces $V_\sigma^P(\Omega_\sigma)$: $$\mathcal{B}_\sigma^P := \biggl\lbrace
\varphi_i\biggl(\frac{x-x_i}{\sigma}\biggr)^\alpha
\,\biggl.\biggr|\,
Q_i\in\Omega_\sigma, |\alpha|\leq P
\biggr\rbrace.$$ We may assign a numbering $\mathcal{I} = \lbrace 1,\ldots,n\rbrace$ to this set, and subsequently refer to its members as $\mathcal{B}_\sigma^P\ni\psi_k$, $k\in\mathcal{I}$. We can then define the system matrix ${\mathsf{A_h}}\in\mathbb{R}^{n\times n}$ via the relation: $${\mathsf{e_k}}^\top{\mathsf{A_h}}{\mathsf{e_l}} = A_h(\psi_k,\psi_l) = a_h(\psi_k,\psi_l) + \varepsilon j(\psi_k,\psi_l),
\qquad\forall k,l\in\mathcal{I},$$ where ${\mathsf{e_k}}\in\mathbb{R}^n$ refers to the $k$-th Cartesian basis vector, and the approximate bilinear form $a_h$ is defined as described in the numerical setup (\[sec:setup\]). We are then interested in the condition number of the diagonally scaled matrix ${\mathsf{D^{-1}A_h}}$, where ${\mathsf{D}}:=\operatorname{diag}{\mathsf{A_h}}$.
shows the condition number of ${\mathsf{D^{-1}A_h}}$ for various refinements levels $l$ as a function of $\varepsilon$. In the case $l=2$ the quadrature error is so large that the resulting matrix ${\mathsf{A_h}}$ ceased being positive definite for $\varepsilon = 10^{-3}$, and is even singular for $\varepsilon = 0$. This explains the large error observed in \[sec:scalar\_field\] for this case. But even then a sufficiently large choice of $\varepsilon$ results in a well conditioned system. For the finer refinement levels a choice of $\varepsilon$ between $10^{-3}$ and $10^{-1}$ seems to be optimal and reduces the matrix’ condition number below 100. The effect becomes slightly less pronounced with increasing $l$. We can thus conclude that for such a choice of $\varepsilon$ the stabilization removes the ill-conditioning of the system, especially in the presence of moderate quadrature errors.
Conclusions and Outlook
=======================
We have presented a new method to tackle the particle regularization problem, based on a stabilized fictitious domain formulation with smooth shape-functions. Our approach enjoys all the benefits of the conventional blob-methods: the resulting smoothed approximations are $C^\infty$ functions and conserve all moments up to order $P$. On top of that, our approach can accurately handle general geometries. The evaluation of the smoothed approximations is cheap and straightforward and does not require a summation over all particles as in the case of blob functions. The fact that we can only achieve a convergence rate of $\mathcal{O}(h^{\frac{m+1}{2}})$ as opposed to $\mathcal{O}(h^{m+1})$ might seem disappointing, but is intrinsic to the smoothing problem at hand. This can be illustrated in a simple one-dimensional example: given an interval of length $h$ on the real line, the $m$-node Gaussian quadrature rule will have an error bound of $\mathcal{O}(h^{2m})$. With $m$ function values, however, we can only construct an interpolation polynomial of degree $m-1$, having the halved error bound $\mathcal{O}(h^m)$. shows that the smoothed approximation is essentially just as accurate as the particle field. This also means that it can be used to reinitialize overly distorted particle fields. Furthermore, this means that the smoothed vorticity field has much greater length-scales than the particle spacing.
As a consequence the velocity evaluation – usually the most expensive part of vortex methods – can be drastically sped up. In our numerical experiments we gave an example of a simple mesh-based scheme for this, which was chosen because of its simplicity. A disadvantage of this approach is that the resulting velocity approximation ceases being divergence-free. For future research it would be interesting to make use of the fact that in three-dimensional space the Biot–Savart law is of the form ${\mathbf{u}}=\nabla\times(G\star{\boldsymbol{\omega}})$, where $G(x)= (4\pi|x|)^{-1}$ denotes the fundamental solution of the Laplacian. The “curl spaces” $\nabla\times\bigl(V_\sigma^P(\Omega)\bigr)^3$ would thus be a more natural choice for approximating the velocity, while also being divergence-free in the strong, pointwise sense.
It is not clear whether the exact variational formulation \[eqn:varprob\] is actually unstable without stabilization. A result by Reusken [reusken2008]{} indicates that a rescaling might be sufficient to achieve a stable formulation. On the other hand, this result assumes that the bilinear form $a$ can be computed exactly. The experiments of \[sec:conditioning\] suggest that stabilization is especially beneficial in the presence of quadrature errors.
Our current approach uses a uniform grid size $\sigma$ and a fixed polynomial degree $P$. The partition of unity approach, however, is general enough to be extended to adaptive grids and varying polynomial degrees. For the future, experiments with $\sigma$-, $P$-, or $\sigma P$-adaptive schemes are another interesting field for further research.
We believe the stabilized fictitious domain approach with smooth shape functions is not only useful for vortex particle regularization but also for other problems which require higher degrees of smoothness, such as certain problems from linear elasticity like the Kirchhoff–Love thin-plate theory.
The source code of the software used to obtain the results of this article can be obtained from the authors upon request.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to thank the reviewers for their fruitful comments and suggestions. The first author would also like to express his gratitude to Sven Gro[ß]{}, Arnold Reusken, and all the members of the DROPS team at the Lehrstuhl für Numerische Mathematik (LNM) at RWTH Aachen University, Germany, with whom he previously worked on two-phase flows and the ghost penalty stabilization. Without their support and the knowledge received during that time, this work would not have been possible.
Last but not least, the first author receives the MEXT scholarship of the Japanese Ministry of Education and was supported by the Keio Leading Edge Laboratory of Science and Technology (no grant numbers allotted). Without their financial support this research would have been impossible to conduct.
[^1]: Department of Mechanical Engineering, Keio University, Japan. (, )
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
The class of metrizable spaces $M$ with the following approximation property is introduced and investigated: $M\in AP(n,0)$ if for every ${\varepsilon}>0$ and a map $g\colon{\mathbb I}^n\to M$ there exists a $0$-dimensional map $g'\colon{\mathbb I}^n\to M$ which is ${\varepsilon}$-homotopic to $g$. It is shown that this class has very nice properties. For example, if $M_i\in
AP(n_i,0)$, $i=1,2$, then $M_1\times M_2\in AP(n_1+n_2,0)$. Moreover, $M\in AP(n,0)$ if and only if each point of $M$ has a local base of neighborhoods $U$ with $U\in AP(n,0)$. Using the properties of $AP(n,0)$-spaces, we generalize some results of Levin and Kato-Matsuhashi concerning the existence of residual sets of $n$-dimensional Lelek maps.
address:
- 'Department of Mechanics and Mathematics, Ivan Franko Lviv National University (Ukraine) and Instytut Matematyki, Akademia Świȩtokrzyska w Kielcach (Poland)'
- 'Department of Computer Science and Mathematics, Nipissing University, 100 College Drive, P.O. Box 5002, North Bay, ON, P1B 8L7, Canada'
author:
- Taras Banakh
- Vesko Valov
title: Approximation by light maps and parametric Lelek maps
---
[^1]
Introduction
============
All spaces in the paper are assumed to be metrizable and all maps continuous. By $C(X,M)$ we denote all maps from $X$ into $M$. Unless stated otherwise, all function spaces are endowed with the source limitation topology.
One of the important properties of the $n$-dimensional cube ${\mathbb I}^n$ or the Euclidean space $\mathbb{R}^n$, widely exploited in dimension theory, is that any map from an $n$-dimensional compactum $X$ into ${\mathbb I}^n$ (resp., $\mathbb{R}^n$) can be approximated by maps with $0$-dimensional fibers. The aim of this article is to introduce and investigate the class of spaces having that property. More precisely, we say that a space $M$ has the [*$(n,0)$-approximation property*]{} (br., $M\in AP(n,0))$ if for every ${\varepsilon}>0$ and a map $g\colon{\mathbb I}^n\to M$ there exists a $0$-dimensional map $g'\colon{\mathbb I}^n\to M$ which is ${\varepsilon}$-homotopic to $g$. Here, $g'$ is ${\varepsilon}$-homotopic to $g$ means that there is an ${\varepsilon}$-small homotopy $h\colon{\mathbb I}^n\times{\mathbb I}\to M$ connecting $g$ and $g'$. It is easily seen that this definition doesn’t depend on the metric generating the topology of $M$. If $M$ is $\LC[n]$, then $M\in AP(n,0))$ if and only if for every ${\varepsilon}>0$ and $g\in C({\mathbb I}^n,M)$ there exists a $0$-dimensional map $g'$ which is ${\varepsilon}$-close to $g$. We show that the class of $AP(n,0)$-spaces is quite large and it has very nice properties. For example, if $M_i\in AP(n_i,0)$, $i=1,2$, then $M_1\times M_2\in AP(n_1+n_2,0)$. We also prove that $M\in AP(n,0)$ if and only if each point of $M$ has a local base of neighborhoods $U$ with $U\in AP(n,0)$. Moreover, every path-connected compactum $M\in AP(n,0)$ is a $(V^n)$-continuum in the sense of P. Alexandroff [@ps]. In particular, according to [@hv], any such $M$ is a Cantor $n$-manifold, as well as a strong Cantor $n$-manifold in the since of Hadžiivanov [@nh]. All complete $\LC[0]$-spaces without isolated points are $AP(1,0)$, as well as, every manifold modeled on the $n$-dimensional Menger cube or the $n$-dimensional Nöbeling space has the $AP(n,0)$-property.
The class of $AP(n,0)$-spaces is very natural for obtaining results about approximation by dimensionally restricted maps. We present such a result about $n$-dimensional Lelek maps. Recall that a map $g\colon X\to M$ between compact spaces is said to be an [*$n$-dimensional Lelek map*]{} [@l] if the union of all non-trivial continua contained in the fibers of $g$ is of dimension $\leq n$. For convenience, if $n\leq 0$, by an $n$-dimensional Lelek map we simply mean a $0$-dimensional map. Every $n$-dimensional Lelek map $g$ between compacta is at most $n$-dimensional because the components of each fiber of $g$ is at most $n$-dimensional (we say that $g$ is $n$-dimensional if all fibers of $g$ are at most $n$-dimensional). Lelek [@le] constructed such a map from ${\mathbb I}^{n+1}$ onto a dendrite. Levin [@l] improved this result by showing that the space $C(X,{\mathbb I}^n)$ of all maps from $X$ into ${\mathbb I}^n$ contains a dense $G_{\delta}$-subset consisting of $(m-n)$-dimensional Lelek maps for any compactum $X$ with $\dim
X=m\leq n$. Recently Kato and Matsuhashi [@km] established a version of the Levin result with ${\mathbb I}^n$ replaced by more general class of spaces. This class consists of complete metric $ANR$-spaces having a piecewise embedding dimension $\geq n$.
In the present paper we generalize the result of Kato-Matsuhashi in two directions. First, we establish a parametric version of their theorem and second, we show that this parametric version holds for $AP(n,0)$-spaces (see Section 6 where it is shown that the class of $AP(n,0)$-spaces contains properly the class of $ANR$’s with piecewise embedding dimensions $\geq n$). Here is our result about parametric Lelek maps (a more general version is established in Section 3; moreover we also present a version of Theorem 1.1 when $M$ has a property weaker than $AP(n,0)$, see Theorem 4.5).
\[lelek\] Let $f\colon X\to Y$ be a perfect map with ${\dim_\triangle}(f)\leq m$, where $X$ and $Y$ are metric spaces. If $M\in AP(n,0)$ is completely metrizable, then there exists a $G_\delta$-set $\mathcal{H}\subset
C(X,M)$ such that every simplicially factorizable map in $C(X,M)$ is homotopically approximated by maps from $\mathcal{H}$ and for every $g\in\mathcal{H}$ and $y\in Y$ the restriction $g|f^{-1}(y)$ is an $(m-n)$-dimensional Lelek map.
Let us explain the notions in Theorem 1.1. A map $g\in C(X,M)$ is homotopically approximated by maps from $\mathcal{H}$ means that for every function ${\varepsilon}\in
C(X,(0,1])$ there exists $g'\in\mathcal{H}$ which is ${\varepsilon}$-homotopic to $g$. Here, the maps $g$ and $g'$ are said to be ${\varepsilon}$-homotopic, if there is a homotopy $h\colon X\times{\mathbb I}\to M$ connecting $g$ and $g'$ such that each set $h(\{x\}\times{\mathbb I})$ has a diameter $<{\varepsilon}(x)$, $x\in X$.
The function space $C(X,M)$ appearing in this theorem is endowed with the source limitation topology whose neighborhood base at a given function $f\in C(X,M)$ consists of the sets $$B_\varrho(f,{\varepsilon})=\{g\in C(X,M):\varrho(g,f)<{\varepsilon}\},$$ where $\varrho$ is a fixed compatible metric on $M$ and ${\varepsilon}:X\to(0,1]$ runs over continuous positive functions on $X$. The symbol $\varrho(f,g)<{\varepsilon}$ means that $\varrho(f(x),g(x))<{\varepsilon}(x)$ for all $x\in X$. Since $X$ is metrizable, the source limitation topology doesn’t depend on the metric $\rho$ [@nk] and it has the Baire property provided $M$ is completely metrizable [@munkers].
We say that a map $g\colon X\to M$ is simplicially factorizable [@bv] if there exists a simplicial complex $L$ and two maps $g_1\colon X\to L$ and $g_2\colon L\to M$ such that $g=g_2\circ
g_1$. In each of the following cases the set of simplicially factorizable maps is dense in $C(X,M)$ (see [@bv Proposition 4]): (i) $M$ is an $ANR$; (ii) $\dim X\leq k$ and $M$ is $\LC[k-1]$; (iii) $X$ is a $C$-space and $M$ is locally contractible.
The dimension ${\dim_\triangle}(f)$ was defined in [@bv]: ${\dim_\triangle}(f)$ of a map $f:X\to Y$ is equal to the smallest cardinal number $\tau$ for which there is a map $g:X\to \mathbb I^\tau$ such that the diagonal product $f\Delta g:X\to Y\times \mathbb I^\tau$ is a $0$-dimensional map. For any perfect map $f:X\to Y$ between metric spaces we have: (i) $\dim(f)\leq {\dim_\triangle}(f)$; (ii) ${\dim_\triangle}(f)=\dim(f)$ if $Y$ is a $C$-space, see [@bp:96] and [@tv]; (iii) ${\dim_\triangle}(f)\leq\dim
(f)+1$ if the spaces $X,Y$ are compact, see [@lev].
Since every metric space admitting a perfect finite-dimensional map onto a $C$-space is also a $C$-space [@hy], Theorem 1.1 implies the following (here, $\mathcal{H}$ is homotopically dense in $C(X,M)$ if every $g\in C(X,M)$ is homotopically approximated by maps from $\mathcal{H}$):
Let $f:X\to Y$ be a perfect $m$-dimensional map between metric spaces with $Y$ being a $C$-space. If $M\in A(n,0)$ is completely metrizable, then in each of the following cases there exists a homotopically dense $G_{\delta}$-subset $\mathcal{H}\subset C(X,M)$ consisting of maps $g$ such that $g|f^{-1}(y)$ is an $(m-n)$-dimensional Lelek map for every $y\in Y$:
- $M$ is locally contractible;
- $X$ is finite dimensional and $M\in\LC[k-1]$ for $k=\dim X$.
The paper is organized as follows. In Section 2 some preliminary results about $AE(n,0)$-spaces are presented. In Section 3 we establish a generalized version of Theorem 1.1. Section 4 is devoted to almost $AP(n,0)$-spaces. The final two sections contain some interesting properties of $AP(n,0)$-spaces. In Section 5 we prove that $AP(n,0)$-property has a local nature, as well as, $M_1\times
M_2\in AP(n_1+n_2,0)$ provided each $M_i\in AP(n_i,0)$ is completely metrizable. It is also established in this section that every path connected compactum is a $(V^n)$-continuum. Section 6 is devoted to the interplay of $AP(n,0)$-property and the general position properties $m-\overline{DD}^{\{n,k\}}$-properties. In particular, it is shown that every point of a locally path-connected $AP(n,0)$-space $X$ is a homological $Z_{n-1}$-point in $X$. Another result from this section states that every completely metrizable space possessing the disjoint $(n-1)$-disks property is an $AP(n,0)$-space.
Preliminary results about $AP(n,0)$-spaces
==========================================
In this section we established some preliminary results on $AP(n,0)$-spaces which are going to be used in next sections.
Suppose $M$ is a metric space and $\varepsilon >0$. We write $d_n(M)<\varepsilon$ if $M$ can be covered by an open family $\gamma$ such that ${\mathrm{diam}}U<{\varepsilon}$ for all $U\in\gamma$ and $ord(\gamma)\leq n+1$ (the last inequality means that at most $n+1$ elements of $\gamma$ can have a common point). It is easily seen that if $F\subset M$ is closed and $d_n(F)<\varepsilon$, then $F$ is covered by an open in $M$ family $\gamma$ of mesh $<\varepsilon$ and $ord\leq n+1$. We also agree to denote by $cov(M)$ the family of all open covers of $M$. Let us begin with the following technical lemma.
Let $Z=A\cup B$ be a compactum, where $A$ and $B$ are closed subsets of $Z$. Suppose $C\subset Z$ is closed such that $\dim C\cap A\leq
0$ and $d_0(C\cap B)<\varepsilon$. Then $d_0(C)<\varepsilon$.
Since $d_0(C\cap B)<{\varepsilon}$, there exists a disjoint open family $\gamma=\{W_1,..,W_k\}$ in $B$ of ${\mathrm{mesh}}<{\varepsilon}$. We extend every $W_i$ to an open set $\tilde{W_i}$ in $Z$ such that $\tilde{W_i}\cap
B=W_i$ and $\tilde{\gamma}=\{\tilde{W_i}:i=1,..,k\}$ is a disjoint family of ${\mathrm{mesh}}<\varepsilon$ (this can be done because $B$ is closed in $Z$). Observe that $C_1=C\backslash\bigcup_{i=1}^{i=k}\tilde{W_i}$ is a closed subset of $C\cap A$ disjoint from $B$. Since $\dim C\cap A\leq 0$, there exists a clopen set $C_2$ in $C\cap A$ disjoint from $B$ and containing $C_1$. Obviously, $C_2$ is clopen in $C$. Moreover, $\dim
C_2\leq 0$. Hence, there exists a cover $\omega=\{V_j:j=1,..,m\}\in
cov(C_2)$ consisting of clopen subsets of $C_2$ with ${\mathrm{diam}}V_j<{\varepsilon}$ for every $j$. Then $\omega_1=\omega\cup\{\tilde{W_i}\cap
C\backslash C_2\}$ is a disjoint open cover of $C$ and $mesh(\omega_1) <{\varepsilon}$. So, $d_0(C)<{\varepsilon}$.
The proof of next lemma is extracted from [@bv Theorem 4].
Let $G\subset C(X,M)$, where $(M,\varrho)$ is a complete metric space. Suppose $(U(i)_{i\geq 1}$ is a sequence of open subsets of $C(X,M)$ such that
- for any $g\in G$, $i\geq 1$ and any function $\eta\in
C(X,(0,1])$ there exists $g_i\in B_\varrho(g,\eta)\cap U(i)\cap G$ which is $\eta$-homotopic to $g$.
Then, for any $g\in G$ and ${\varepsilon}\colon X\to (0,1]$ there exists $g'\in\bigcap_{i=1}^{\infty} U(i)$ and an ${\varepsilon}$-homotopy connecting $g$ and $g'$. Moreover, $g'|A=g_0|A$ for some $g_0\in C(X,M)$ and $A\subset X$ provided $g_i|A=g_0|A$ for all $i$.
For fixed $g\in G$ and ${\varepsilon}\in C(X,(0,1])$ let $g_0=g$ and ${\varepsilon}_0={\varepsilon}/3$. We shall construct by induction a sequence $(g_i:X\to
M)_{i\geq 1}\subset G$, a sequence $({\varepsilon}_i)_{i\geq 1}$ of positive functions, and a sequence $(H_i:X\times[0,1]\to M)_{i\geq 1}$ of ${\varepsilon}_{i-1}$-homotopies satisfying the conditions:
- $H_{i+1}(x,0)=g_{i}(x)$ and $H_{i+1}(x,1)=g_{i+1}(x)$ for every $x\in X$;
- $g_{i+1}\in B_\varrho(g_i,{\varepsilon}_i)\cap U(i+1)\cap G$;
- ${\varepsilon}_{i+1}\leq{\varepsilon}_i/2$;
- $B_\varrho(g_{i+1},3{\varepsilon}_{i+1})\subset
U(i+1)$.
Assume that, for some $i$, we have already constructed maps $g_1,\dots,g_i$, positive numbers ${\varepsilon}_1,\dots,{\varepsilon}_i$, and homotopies $H_1,\dots,H_i$ satisfying the above conditions. According to the hypotheses, there exists a map $g_{i+1}\in B_\varrho(g_i,{\varepsilon}_i)\cap
U(i+1)\cap G$ such that $g_{i+1}$ is ${\varepsilon}_i$-homotopic to $g_i$. Let $H_{i+1}:X\times[0,1]\to M$ be an ${\varepsilon}_i$-homotopy connecting the maps $g_{i}$ and $g_{i+1}$. Since the set $U(i+1)$ is open in $C(X,M)$, there is a positive function ${\varepsilon}_{i+1}\leq{\varepsilon}_i/2$ such that $B_\varrho(g_{i+1},3{\varepsilon}_{i+1})\subset U(i+1)$. This completes the inductive step.
It follows from the construction that the function sequence $(g_i)_{i\geq 1}$ converges uniformly to some continuous function $g':X\to M$. Obviously, $\varrho(g',g_i)\leq \sum_{j=i}^\infty{\varepsilon}_j\leq
2{\varepsilon}_i$ for every $i$. Hence, according to the choice of the sequences $({\varepsilon}_i)$ and $(g_i)$, $g'\in U(i)$ for every $i\geq 1$. So, $g'\in\bigcap_{i=1}^{\infty} U(i)$. Moreover, the ${\varepsilon}_{i-1}$-homotopies $H_i$ compose an ${\varepsilon}$-homotopy $H:X\times[0,1]\to M$ $$H(x,t)=\begin{cases} \displaystyle
H_i\big(x,2^i(t-1+\frac1{2^{i-1}})\big)&\mbox{if $\displaystyle t\in
[1-\frac1{2^{i-1}},1-\frac1{2^{i}}]$,
$i\geq 1$};\\
g'(x)&\mbox{if $t=1$}.
\end{cases}$$ connecting $g_0=g$ and $g'$.
If follows from our construction that $g'|A=g_0|A$ if $g_i|A=g_0|A$ for all $i\geq 1$.
Let $M$ be a completely metrizable space with the $AP(n,0)$-property and $X$ be an $n$-dimensional compactum. Then $$\mathcal E=\{g\in
C(X,M): \dim g\leq 0\}$$ is a $G_{\delta}$-subset of $C(X,M)$ such that every simplicially factorizable map is homotopically approximated by maps from $\mathcal E$.
We fix a metric on $X$ and a complete metric $\varrho$ on $M$. For every ${\varepsilon}>0$ let $C(X,M;{\varepsilon})$ be the set all maps $g\in C(X,M)$ with $d_0(g^{-1}(a))<\varepsilon$ for every $a\in M$. It is easily seen that $C(X,M;{\varepsilon})$ is open in $C(X,M)$ and $\mathcal
E=\bigcap_{i=1}^{\infty}C(X,M;1/i)$. So, $\mathcal E$ is a $G_{\delta}$-subset of $C(X,M)$.
[*Claim. For every simplicially factorizable map $h\in C(X,M)$ and positive numbers $\eta$ and $\delta$ there exists a simplicially factorizable map $h'\in C(X,M;\eta)$ which is $\delta$-homotopic to $h$*]{}.
To prove the claim, fix a simplicially factorizable map $h\in
C(X,M)$ and positive numbers $\delta$, $\eta$. Then there exists a simplicial complex $L$ and maps $q_1\colon X\to L$, $q_2\colon L\to
M$ with $h=q_2\circ q_1$. We can assume that $L$ is finite and $q_2(L)\subset\cup\gamma$, where $\gamma=\{B_{\varrho}(y,\delta/4):
y\in g(X)\}$. Because $L$ is an $ANR$ and $\dim X\leq n$, there is a polyhedron $K$ with $\dim K\leq n$ and two maps $f\colon X\to K$, $\alpha: K\to L$ such that $f$ is an $\eta$-map and $q_1$ and $\alpha\circ f$ are $q_2^{-1}(\gamma)$-homotopic. So, $h$ and $q_2\circ\alpha\circ f$ are $\delta/2$-homotopic. Moreover, there is a cover $\omega\in cov(K)$ such that ${\mathrm{diam}}f^{-1}(W)<\eta$ for every $W\in\omega$. Let $\theta$ be a Lebesgue number of $\omega$. It remains to find a map $h^*\colon K\to M$ which is $\delta/2$-homotopic to $q_2\circ\alpha$ and $h^*\in C(K,M;\theta)$ (then $h'=h^*\circ f$ would be a simplicially factorizable map $\delta$-homotopic to $h$ and $h'\in C(X,M;\eta)$).
To find such a map $h^*\colon K\to M$, let $\{\sigma_1,..,\sigma_m\}$ be an enumeration of the simplexes of $K$ and $K_i=\bigcup_{j=1}^{j=i}\sigma_j$. We are going to construct by induction maps $h_i\colon K\to M$, $i=0,..,m$, satisfying the following conditions:
- $h_0=q_2\circ\alpha$;
- $h_i|K_i$ belongs to $C(K_i,M;\theta)$, $1\leq i\leq m$;
- $h_i$ and $h_{i+1}$ are $(\delta/2m)$-homotopic, $i=0,..,m-1$.
Let ${\mathcal V}=\{B_{\varrho}(y,\delta/4m): y\in M\}\in cov(M)$ and assume that $h_i$ has already been constructed. Since $h_i|K_i\in
C(K_i,M;\theta)$, every fiber $h_i^{-1}(y)\cap K_i$ of $h_i|K_i$, $y\in h_i(K_1)$, is covered by a finite open and disjoint family $\Omega(y)$ in $K_i$ with ${\mathrm{mesh}}(\Omega(y))<\theta$. Using that $h_i|K_i$ is a perfect map, we find a cover ${\mathcal V}_i\in cov(M)$ such that ${\mathcal V}_i$ is a star-refinement of ${\mathcal V}$ and if ${St}(y,{\mathcal V}_i)\cap
h_i(K_i)\neq\emptyset$ for some $y\in M$, then there is $z\in
h_i(K_i)$ with $h_i^{-1}({St}(y,{\mathcal V}_i))\cap K_i\subset\Omega(z)$. Since $M$ has the $AP(n,0)$-property, there exists a $0$-dimensional map $p_i\colon\sigma_{i+1}\to M$ which is ${\mathcal V}_i$-homotopic to $h_i|\sigma_{i+1}$. By the Homotopy Extension Theorem, $p_i$ can be extended to a map $h_{i+1}\colon K\to M$ being ${\mathcal V}_i$-homotopic to $h_i$. Then $h_{i+1}$ is $(\delta/2m)$-homotopic to $h_i$. To show that $h_{i+1}|K_{i+1}$ belongs to $C(K_{i+1},M;\theta)$, we observe that $h_{i+1}^{-1}(y)\cap K_{i+1}=\big(h_{i+1}^{-1}(y)\cap
K_i\big)\cup\big(h_{i+1}^{-1}(y)\cap\sigma_{i+1}\big)$, $y\in M$. According to our construction, we have $h_i\big(h_{i+1}^{-1}(y)\cap
K_i\big)\subset{St}(y,{\mathcal V}_i)$. Hence, $h_{i+1}^{-1}(y)\cap K_i$ is contained in $\Omega(z)$ for some $z\in h_i(K_i)$. Therefore, $d_0\big(h_{i+1}^{-1}(y)\cap K_{i}\big)<\theta$. Since $h_{i+1}^{-1}(y)\cap\sigma_{i+1}$ is $0$-dimensional, Lemma 2.1 implies that $d_0(h_{i+1}^{-1}(y)\cap K_{i+1})<\theta$. Obviously the map $h^*=h_m$ is $\delta/2$-homotopic to $q_2\circ\alpha$ and $h^*\in
C(K,M;\theta)$. This completes the proof of the claim.
To finish the proof of the proposition, we apply Lemma 2.2 with $G$ being the set of all simplicially factorizable maps from $C(X,M)$ and $U(i)=C(X,M;1/i)$, $i\geq 1$ (we can apply Lemma 2.2 because of the claim). Hence, for every simplicially factorizable map $g\in
C(X,M)$ and a positive number ${\varepsilon}$ there exists a map $g'\in\bigcap_{i=1}^\infty U(i)$ which is ${\varepsilon}$-homotopic to $g$. Finally, we observe that $\bigcap_{i=1}^\infty U(i)$ consists of $0$-dimensional maps.
\[n-1\] Let $M$ be a completely metrizable space with the $AP(n,0)$-property, $X$ a compactum and $Z$ an $F_{\sigma}$-subset of $X$ with $\dim Z\leq n-1$. Then there exists a $G_{\delta}$-subset $\mathcal H\subset C(X,M)$ with the following properties:
- $Z$ is contained in the union of trivial components of the fibers of $g$ for all $g\in\mathcal H$;
- for every simplicially factorizable map $g\in C(X,M)$ and every ${\varepsilon}>0$ there exists $g'\in\mathcal H$ which is ${\varepsilon}$-homotopic to $g$.
We represent $Z$ as the union of an increasing sequence $(Z_i)_{i\geq 1}$ with all $Z_i$ being closed in $Z$. For every ${\varepsilon}>0$ let $$\mathcal H(Z_i,{\varepsilon})=\{g\in C(X,M): F(g,{\varepsilon})\cap Z_i=\emptyset\},$$ where $$F(g,{\varepsilon})=\cup\{C: \mbox{$C$ is a component
of a fiber of $g$ with ${\mathrm{diam}}C\geq{\varepsilon}$}\}.$$ It is easily seen that each $\mathcal H(Z_i,{\varepsilon})$ is open in $C(X,M)$. So, the set $\mathcal
H=\bigcap_{i=1}^\infty\mathcal H(Z_i,1/i)$ is $G_{\delta}$ and $Z$ is contained in the union of trivial components of the fibers of $g$ for every $g\in\mathcal H$.
To prove the second item of our proposition, we first consider the particular case when $X$ is a polyhedron and $Z$ is a (compact) subpolyhedron of $X$.
[*Claim $1$. Suppose that, in addition to hypotheses of Proposition $2.4$, $X$ is a polyhedron and $Z$ is a subpolyhedron of $X$. Then, for every $g_0\in C(X,M)$ and $\delta>0$ there exists $g\in\mathcal H=\bigcap_{i=1}^\infty\mathcal H(Z,1/i)$ which is $\delta$-homotopic to $g_0$.*]{}
The proof of Claim 1 is a slight modification of the proof of [@km Theorem 2.2]. Let $g_0\in C(X,M)$ and $\delta>0$. We take an open neighborhood $W$ of $Z$ in $X$, a retraction $r\colon W\to
Z$ and a function $\alpha\colon X\to{\mathbb I}$ such that $\alpha^{-1}(0)=Z$, $\alpha^{-1}(1)=X\backslash W$. Since $X$ is a compact $ANR$, we can choose $W$ so small that $g_0|W$ is $\delta/2$-homotopic to $(g_0\circ r)|W$. Next, denote by $\pi\colon
Z\times{\mathbb I}\to Z$ the projection and consider the map $\varphi\colon
W\to Z\times{\mathbb I}$, $\varphi (x)=(r(x),\alpha(x))$. Obviously, $(g_0\circ r)|W=(g_0\circ \pi\circ\varphi)|W$. So, $(g_0\circ
\pi\circ\varphi)|W$ is $\delta/2$-homotopic to $g_0|W$. Because $Z$ is a polyhedron, so is $Z\times{\mathbb I}$. Hence, every map from $C(Z\times{\mathbb I},M)$ is simplicially factorizable. Moreover, $\dim
Z\times{\mathbb I}\leq n$ and $M\in AP(n,0)$. Therefore, by Proposition 2.3, there exists a $0$-dimensional map $h\colon Z\times{\mathbb I}\to M$ which is $\delta/2$-homotopic to the map $g_0\circ\pi$. Then $(h\circ\varphi)|W$ is $\delta/2$-homotopic to $(g_0\circ
\pi\circ\varphi)|W$. Consequently, $(h\circ\varphi)|W$ is $\delta$-homotopic to $g_0|W$. By the Homotopy Extension Theorem, there exits a map $g\in C(X,M)$ such that $g$ is $\delta$-homotopic to $g_0$ and $g|\overline{U}=(h\circ\varphi)|\overline{U}$, where $U$ is an open neighborhood of $Z$ in $X$ with $\overline{U}\subset
W$.
To finish the proof of Claim 1, it remains to show that $g\in\mathcal H$. Let $C$ be a subcontinuum of $g^{-1}(y)$ for some $y\in M$ and let $Z\cap C\neq\emptyset$. We are going to prove that $C\subset Z$. Otherwise, there would be a subcontinuum $C'\subset
C\cap U$ such that $C'\cap Z\neq\emptyset$ and $C'\backslash
Z\neq\emptyset$. Then $g(C')=h(\varphi(C'))=y$ and, according to the definition of $\alpha$, $\varphi (C')$ is a non-degenerate continuum in $h^{-1}(y)$. Since $h$ is $0$-dimensional, this is a contradiction. Hence, $C\subset Z$. Using again that $\dim(h)\leq
0$, we conclude that $\varphi(C)$ is a point. On the other hand $\varphi(C)=C\times\{0\}\subset Z\times{\mathbb I}$. Therefore, $C$ should be a trivial continuum.
Now, consider the general case of Proposition 2.4.
[*Claim $2$. Let $g_0\in C(X,M)$ be a simplicially factorizable map and $\delta$, $\eta$ positive numbers. Then for any $i$ there exists a simplicially factorizable map $g\in\mathcal H(Z_i,\eta)$ which is $\delta$-homotopic to $g_0$.*]{}
Since $g_0$ is simplicially factorizable, there exist a finite simplicial complex $L$ and maps $q_1\colon X\to L$, $q_2\colon L\to
M$ with $g_0=q_2\circ q_1$. Let ${\mathcal V}=\{B_{\varrho}(y,\delta/4): y\in
M\}\in cov(M)$ and ${\mathcal W}=q_2^{-1}({\mathcal V})\in cov(L)$. Next, choose a finite cover ${\mathcal U}\in cov(X)$ of $X$ with ${\mathrm{mesh}}({\mathcal U})<\eta$ such that:
- at most $n$ elements of the family $\gamma=\{U\in{\mathcal U}: U\cap
Z_i\neq\emptyset\}$ can have a common point;
- there exists a map $h\colon\mathcal N({\mathcal U})\to L$ such that $h\circ f_{{\mathcal U}}$ is ${\mathcal W}$-homotopic to $q_1$, where $\mathcal N({\mathcal U})$ is the nerve of ${\mathcal U}$ and $f_{{\mathcal U}}\colon X\to\mathcal N({\mathcal U})$ is the natural map.
Then $q_2\circ h\circ f_{{\mathcal U}}$ is $\delta/2$-homotopic to $g_0$ and the subpolyhedron $K$ of $\mathcal N({\mathcal U})$ generated by the family $\gamma$ is of dimension $\leq n-1$. So, according to Claim 1, there exists a map $g_1\colon\mathcal N({\mathcal U})\to M$ such that $g_1$ is $\delta/2$-homotopic to $q_2\circ h$ and $K$ is contained in the union of trivial components of the fibers of $g_1$. Since all fibers of $f_{{\mathcal U}}$ are of diameter $<\eta$ and $f_{{\mathcal U}}(Z_i)\subset K$, the map $g=g_1\circ
f_{{\mathcal U}}$ belongs to $\mathcal H(Z_i,\eta)$. Moreover, $g$ is obviously simplicially factorizable and $\delta$-homotopic to $g_0$. This completes the proof of Claim 2.
Finally, the proof of Proposition 2.4 follows from Lemma 2.2 (with $G$ being the set of all simplicially factorizable maps from $C(X,M)$ and $U(i)=\mathcal H(Z_i,1/i)$ for every $i\geq 1$) and Claim 2.
Parametric Lelek maps
=====================
In this section we are going to prove Theorem 1.1. Everywhere in this section we suppose that $(M,\varrho)$ is a given complete metric space and $\mu=\{W_\nu:\nu\in\Lambda\}$, $\mu_1=\{G_\nu:\nu\in\Lambda\}$ are locally finite open covers of $M$ such that
1. $\overline{G}_\nu\subset W_\nu$ and $W_\nu\in AP(n_\nu,0)$ with $0\leq n_\nu$ for every $\nu\in\Lambda$.
Obviously, Theorem 1.1 follows directly from next theorem.
Let $f\colon X\to Y$ be a perfect map between metrizable spaces with $\dim\triangle(f)\leq m$. Suppose $(M,\varrho)$ is a complete metric space and $\mu$, $\mu_1$ two locally finite open covers of $M$ satisfying condition $(*)$. Then there is a $G_{\delta}$-set $\mathcal{H}\subset C(X,M)$ such that any simplicially factorizable map in $C(X,M)$ can be homotopically approximated by maps from $\mathcal{H}$ and every $g\in\mathcal{H}$ has the following property: for any $y\in Y$ and $\nu\in\Lambda$ the restriction $g|\big(f^{-1}(y)\cap g^{-1}(\overline{G}_\nu)\big)$ is an $(m-n_{\nu})$-dimensional Lelek map from $f^{-1}(y)\cap
g^{-1}(\overline{G}_\nu)$ into $\overline{G}_\nu$.
The proof of Theorem 3.1 consists of few propositions. We can assume that $n_\nu\leq m$ for all $\nu\in\Lambda$. For every $y\in Y$, ${\varepsilon}>0$, $\nu\in\Lambda$ and $g\in C(X,M)$ we denote by $F_\nu(g,{\varepsilon},y)$ the union of all continua $C$ of diameter $\geq{\varepsilon}$ such that $C\subset f^{-1}(y)\cap g^{-1}(z_C)$ for some $z_C\in\overline{G}_\nu$. Note that each $F_\nu(g,{\varepsilon},y)$ is compact as a closed subset of $f^{-1}(y)$. Then, for a fixed $\eta>0$, let $\mathcal{H}(y,{\varepsilon},\eta)$ be the set of all $g\in C(X,M)$ such that $d_{m-n_\nu}\big(F_\nu(g,{\varepsilon},y)\big)<\eta$ for every $\nu\in\Lambda$. The last inequality means that $F_\nu(g,{\varepsilon},y)$ can be covered by an open family $\gamma$ in $X$ such that ${\mathrm{mesh}}(\gamma)<\eta$ and no more $m-n_\nu+1$ elements of $\gamma$ have a common point. If $V\subset Y$, then $\mathcal{H}(V,{\varepsilon},\eta)$ denotes the intersection of all $\mathcal{H}(y,{\varepsilon},\eta)$, $y\in V$.
For every $y\in Y$ and every $g\in \mathcal{H}(y,{\varepsilon},\eta)$ there exists a neighborhood $V_y$ of $y$ in $Y$ and $\delta_y>0$ such that if $y'\in V_y$ and $\varrho(h(x),g(x))<\delta_y$ for all $x\in
f^{-1}(y')$, then $h\in\mathcal{H}(y',{\varepsilon},\eta)$.
Since $\mu_1$ is locally finite and $f^{-1}(y)$ is compact, for every $y\in Y$ and $g\in C(X,M)$ there exists a neighborhood $O_g(y)$ of $g(f^{-1}(y))$ such that the family $\Lambda_g(y)=\{\nu\in\Lambda: O_g(y)\cap G_\nu\neq\emptyset\}$ is finite.
Assume the lemma is not true for some $y\in Y$ and $g\in
\mathcal{H}(y,{\varepsilon},\eta)$. Then there exists a local base of neighborhoods $V_i$ of $y$, points $y_i\in V_i$ and functions $g_i\in C(X,M)$ such that $g_i|f^{-1}(y_i)$ is $1/i$-close to $g|f^{-1}(y_i)$ but $g_i\not\in\mathcal{H}(y_i,{\varepsilon},\eta)$. It is easily seen that for some $k$ and all $i\geq k$ we have $g_i(f^{-1}(y_i))\subset O_g(y)$ . Consequently, for every $i\geq k$ there exists $\nu(i)\in\Lambda_g(y)$ such that $d_{m-n_{\nu(i)}}\big(F_{\nu(i)}(g_i,{\varepsilon},y_i)\big)\geq\eta$ (because $g_i\not\in\mathcal{H}(y_i,{\varepsilon},\eta)$). So, each $F_{\nu(i)}(g_i,{\varepsilon},y_i)$ doesn’t have any open cover of ${\mathrm{mesh}}<\eta$ and order $\leq m-n_{\nu(i)}+1$. Since the family $\Lambda_g(y)$ is finite, there exists $\nu(0)\in\Lambda_g(y)$ with $\nu(i)=\nu(0)$ for infinitely many $i$. With out loss of generality, we may suppose that $\nu(i)=\nu(0)$ for all $i\geq k$. This implies that $g_i(f^{-1}(y_i))\cap\overline{G}\neq\emptyset$, $i\geq k$, where $G=G_{\nu(0)}$. Consequently, $g(f^{-1}(y))\cap\overline{G}\neq\emptyset$. Since $g\in
\mathcal{H}(y,{\varepsilon},\eta)$, $F_{\nu(0)}(g,{\varepsilon},y)$ can be covered by an open family $\gamma$ in $X$ of order $\leq m-n_{\nu(0)}+1$ and ${\mathrm{mesh}}(\gamma)<\eta$. Let $U=\cup\gamma$. To obtain a contradiction, it suffice to show that $F_{\nu(i)}(g_i,{\varepsilon},y_i)\subset U$ for some $i\geq k$. Indeed, otherwise for every $i\geq k$ there would exist points $x_i\in F_{\nu(i)}(g_i,{\varepsilon},y_i)\backslash U$, $z_i\in\overline{G}$, and a component $C_i$ of $f^{-1}(y_i)\cap
g_i^{-1}(z_i)$ containing $x_i$ with ${\mathrm{diam}}C_i\geq{\varepsilon}$. Using that $P=f^{-1}(\{y_i\}_{i=k}^\infty\cup\{y\})$ is a compactum, we can suppose that $\{x_i\}_{i=k}^\infty$ converges to a point $x_0\in
f^{-1}(y)$, $\{z_i\}_{i=k}^\infty$ converges to a point $z_0\in\overline{G}$ and $\{C_i\}_{i=k}^{\infty}$ (considered as a sequence in the space of all closed subsets of $P$ equipped with the Vietoris topology) converges to a closed set $C\subset f^{-1}(y)\cap
g^{-1}(z_0))$. It is easily seen that $C$ is connected and ${\mathrm{diam}}C\geq{\varepsilon}$. Hence, $C\subset F_{\nu(0)}(g,{\varepsilon},y)\subset U$. So, $x_i\in
C_i\subset U$ for some $i$, a contradiction.
Now, we are in a position to show that the sets $\mathcal{H}(Y,{\varepsilon},\eta)$ are open in $C(X,M)$.
\[open\] For any closed set $F\subset Y$ and any ${\varepsilon}, \eta>0$, the set $\mathcal{H}(F,{\varepsilon},\eta)$ is open in $C(X,M)$.
Let $g_0\in\mathcal{H}(F,{\varepsilon},\eta)$. Then, by Lemma 3.2, for every $y\in F$ there exist a neighborhood $V_y$ and a positive $\delta_y\leq 1$ such that $h\in\mathcal{H}(y',{\varepsilon},\eta)$ provided $y'\in V_y$ and $h|f^{-1}(y')$ is $\delta_y$-close to $g_0|f^{-1}(y')$. The family $\{V_y\cap Y:y\in F\}$ can be supposed to be locally finite in $F$. Consider the set-valued lower semi-continuous map $\varphi\colon F\to (0,1]$, $\varphi(y)=\cup\{(0,\delta_z]:y\in V_z\}$. By [@rs Theorem 6.2, p.116], $\varphi$ admits a continuous selection $\beta\colon
F\to (0,1]$. Let $\overline{\beta}:Y\to (0,1]$ be a continuous extension of $\beta$ and $\alpha=\overline{\beta}\circ f$. It remains only to show that if $g\in C(X,M)$ with $\varrho\big(g_0(x),g(x)\big)<\alpha(x)$ for all $x\in X$, then $g\in\mathcal{H}(F,{\varepsilon},\eta)$. So, we take such a $g$ and fix $y\in
F$. Then there exists $z\in F$ with $y\in V_{z}$ and $\alpha(x)\leq\delta_{z}$ for all $x\in f^{-1}(y)$. Hence, $\varrho\big(g(x),g_0(x)\big)<\delta_z$ for each $x\in f^{-1}(y)$. According to the choice of $V_z$ and $\delta_z$, $g\in\mathcal{H}(y,{\varepsilon},\eta)$. Therefore, $\mathcal{H}(F,{\varepsilon},\eta)$ is open in $C(X,M)$.
To prove Theorem 3.1 it suffices to show that if $g\in
C(X,M)$ is a simplicially factorizable map and $\delta\in
C(X,(0,1])$, then for any ${\varepsilon},\eta>0$ there exists a simplicially factorizable map $g_{{\varepsilon}\eta}\in\mathcal{H}(Y,{\varepsilon},\eta)$ which is $\delta$-homotopic to $g$. Indeed, since any set $\mathcal{H}(Y,{\varepsilon},\eta)$ is open, Lemma 2.2 would imply that every simplicially factorizable map is homotopically approximated by simplicially factorizable maps from $\mathcal{H}=\bigcap_{i,j=1}^{\infty}\mathcal{H}(Y,1/i,1/j)$. But for every $g\in\mathcal H$, $y\in Y$, $\nu\in\Lambda$ and $i\ge1$ we have $d_{m-n_\nu}\big(F_{\nu}(g,1/i,y)\big)=0$. So, $\dim F_{\nu}(g,1/i,y)\leq m-n_\nu$. Then $\dim
F_\nu(g,y)\leq m-n_\nu$, where $F_{\nu}(g,y)=\bigcup_{i=1}^{\infty}F_\nu(g,1/i,y)$ . On the other hand, $F_\nu(g,y)$ is the union of all non-trivial continua contained in the fibers of $g|g^{-1}(\overline{G}_\nu)\cap
f^{-1}(y)$. Therefore, $\mathcal{H}$ consists of maps $g$ such that $g|g^{-1}(\overline{G}_\nu)\cap f^{-1}(y)\colon
g^{-1}(\overline{G}_\nu)\cap f^{-1}(y)\to \overline{G}_\nu$ is $(m-n_\nu)$-dimensional Lelek map for every $\nu\in\Lambda$ and $y\in Y$.
Next proposition shows that all simplicially factorizable maps in $C(X,M)$ can be homotopically approximated by simplicially factorizable maps from $\mathcal{H}(Y,{\varepsilon},\eta)$ provided the set $\mathcal{H}_p(L)$ is homotopically dense in $C(N,M)$ , where $p\colon N\to L$ is any perfect $m$-dimensional $PL$-map between two simplicial complexes $N, L$ equipped with the $CW$-topology and $\mathcal{H}_p(L)$ is the set of all maps $q\colon N\to M$ such that $q|q^{-1}(\overline{G}_\nu)\cap p^{-1}(z)\colon
q^{-1}(\overline{G}_\nu)\cap p^{-1}(z)\to \overline{G}_\nu$ is an $(m-n_\nu)$-dimensional Lelek map for every $\nu\in\Lambda$ and $z\in L$. Recall that $p\colon N\to L$ is a $PL$-map (resp., a simplicial map) if $p$ maps every simplex $\sigma$ of $N$ into (resp., onto) some simplex of $L$ and $p$ is linear on $\sigma$.
Let $X$, $Y$, $f$ and $M$ satisfy the hypotheses of Theorem $3.1$. Suppose the set $\mathcal{H}_p(L)$ is homotopically dense in $C(N,M)$ for any perfect $m$-dimensional $PL$-map $p\colon N\to L$. Then for any simplicially factorizable map $g\in C(X,M)$ and any $\delta\in C(X,(0,1])$, ${\varepsilon},\eta>0$ there exists a simplicially factorizable $h\in\mathcal{H}(Y,{\varepsilon},\eta)$ such that $h$ is $\delta$-homotopic to $g$.
For fixed $\delta\in C(X,(0,1])$ and a simplicially factorizable map $g\in C(X,M)$ we are going to find a simplicially factorizable $h\in\mathcal{H}(Y,{\varepsilon},\eta)$ such that $\varrho(g(x),h(x))<\delta(x)$ for all $x\in X$, where ${\varepsilon}$ and $\eta$ are arbitrary positive reals. Since $g$ is simplicially factorizable, there exists a simplicial complex $D$ and maps $g_D\colon X\to D$, $g^D\colon D\to M$ with $g=g^D\circ g_D$. The metric $\varrho$ induces a continuous pseudometric $\varrho_D$ on $D$, $\varrho_D(x,y)=\varrho(g^D(x),g^D(y))$. By [@ca] and [@si], $D$ being a stratifiable $ANR$ is a neighborhood retract of a locally convex space. Hence, we can apply [@bv Lemma 8.1] to find an open cover ${\mathcal U}$ of $X$ satisfying the following condition: if $\alpha\colon X\to K$ is a ${\mathcal U}$-map into a paracompact space $K$ (i.e., $\alpha^{-1}(\omega)$ refines ${\mathcal U}$ for some $\omega\in cov(K)$), then there exists a map $q'\colon G\to D$, where $G$ is an open neighborhood of $\overline{\alpha(X)}$ in $K$, such that $g_D$ and $q'\circ\alpha$ are $\delta/2$-homotopic with respect to the metric $\varrho_D$. Let ${\mathcal U}_1$ be an open cover of $X$ refining ${\mathcal U}$ with ${\mathrm{mesh}}{\mathcal U}_1<\min\{{\varepsilon},\eta\}$ and $\inf\delta(U)>0$ for all $U\in{\mathcal U}_1$.
Next, according to [@bv Theorem 6], there exists an open cover ${\mathcal V}$ of $Y$ such that: for any ${\mathcal V}$-map $\beta\colon Y\to L$ into a simplicial complex $L$ we can find an ${\mathcal U}_1$-map $\alpha\colon X\to
K$ into a simplicial complex $K$ and a perfect $m$-dimensional $PL$-map $p\colon K\to L$ with $\beta\circ f=p\circ\alpha$. We can assume that ${\mathcal V}$ is locally finite. Take $L$ to be the nerve of the cover ${\mathcal V}$ and $\beta\colon Y\to L$ the corresponding natural map. Then there are a simplicial complex $K$ and maps $p$ and $\alpha$ satisfying the above conditions. Hence, the following diagram is commutative: $$\begin{CD}
X@>{\alpha}>>K\cr @V{f}VV @VV{p}V\cr Y@>{\beta}>>L\cr
\end{CD}$$ Since $K$ is paracompact, the choice of the cover ${\mathcal U}$ guarantees the existence of a map $q_D\colon G\to D$, where $G\subset K$ is an open neighborhood of $\overline{\alpha(X)}$, such that $g_D$ and $h_D=q_D\circ\alpha$ are $\delta/2$-homotopic with respect to $\varrho_D$. Then, according to the definition of $\varrho_D$, $h'=g^D\circ q_D\circ\alpha$ is $\delta/2$-homotopic to $g$ with respect to $\varrho$. Replacing the triangulation of $K$ by a suitable subdivision, we may additionally assume that no simplex of $K$ meets both $\overline{\alpha(X)}$ and $K\backslash G$. So, the union $N$ of all simplexes $\sigma\in K$ with $\sigma\cap\overline{\alpha(X)}\neq\emptyset$ is a subcomplex of $K$ and $N\subset G$. Moreover, since $N$ is closed in $K$, $p\colon
N\to L$ is a perfect $m$-dimensional $PL$-map. Therefore, we have the following commutative diagram, where $q=g^D\circ q_D$:
(120,95)(-100,0) (30,10)[$L$]{} (0,30)[$Y$]{} (12,28)[(3,-2)[18]{}]{} (14,14)[$\beta$]{} (1,70)[$X$]{} (5,66)[(0,-1)[25]{}]{} (-1,53)[$f$]{} (11,73)[(1,0)[45]{}]{} (30,77)[$h'$]{} (12,68)[(3,-2)[18]{}]{} (15,56)[$\alpha$]{} (31,50)[$N$]{} (35,46)[(0,-1)[25]{}]{} (37,33)[$p$]{} (46,58)[(4,3)[13]{}]{} (44,64)[$q$]{} (60,70)[$M$]{}
Now, we shall construct a continuous function $\delta_1:N\to(0,1]$ with $\delta_1\circ\alpha\leq\delta$. Since $\alpha$ is a ${\mathcal U}_1$-map, there is an open cover ${\mathcal V}_1$ of $N$ such that the cover $\alpha^{-1}({\mathcal V}_1)=\{\alpha^{-1}(V):V\in{\mathcal V}_1\}$ refines ${\mathcal U}_1$. Because $\inf\delta(U)>0$ for any $U\in{\mathcal U}_1$, $\inf\delta(\alpha^{-1}(V))>0$ for any $V\in{\mathcal V}_1$. We can assume that ${\mathcal V}_1$ is locally finite and consider the lower semi-continuous set-valued map $\varphi\colon N\to (0,1]$ defined by $\varphi
(z)=\cup\{(0,\inf\delta(\alpha^{-1}(V))]:z\in V\in{\mathcal V}_1\}$. Then, by [@rs Theorem 6.2, p. 116], $\varphi$ admits a continuous selection $\delta_1\colon N\to (0,1]$. Obviously, $\delta_1(z)\leq\inf\delta(\alpha^{-1}(z))$ for all $z\in N$. Hence, $\delta_1\circ\alpha\leq\delta$.
Since, according to our assumption, $\mathcal{H}_p(L)$ is homotopically dense in $C(N,M)$, there exists a map $q_1\in\mathcal{H}_p(L)$ such that $q_1$ is $\delta_1/2$-homotopic to $q$. Let $h=q_1\circ\alpha$. Then $h$ and $q\circ \alpha$ are $\delta/2$-homotopic because $\delta_1\circ\alpha\leq\delta$. On the other hand, $q\circ\alpha=h'$ is $\delta/2$-homotopic to $g$. Hence, $g$ and $h$ are $\delta$-homotopic. Moreover, $h$ is obviously simplicially factorizable.
It remains to show that $h\in\mathcal{H}(Y,{\varepsilon},\eta)$. To this end, we fix $y\in Y$ and $\nu\in\Lambda$, and consider the set $F_\nu(h,{\varepsilon},y)$. Recall that $F_\nu(h,{\varepsilon},y)$ is the union of all continua of diameter $\geq{\varepsilon}$ such that $C\subset f^{-1}(y)\cap
h^{-1}(a_C)$ for some $a_C\in\overline{G}_\nu$. For any such continuum $C$ we have $\alpha (C)\subset p^{-1}(\beta(y))\cap
q_1^{-1}(a_C)$. Since the diameters of all fibers of $\alpha$ are $<{\varepsilon}$ (recall that $\alpha$ is an ${\mathcal U}_1$-map), $\alpha(C)$ is a non-trivial continuum in $p^{-1}(\beta(y))\cap q_1^{-1}(a_C)$. Therefore, $\alpha(C)\subset F_\nu(q_1,\beta(y))$, where $F_\nu(q_1,\beta(y))$ denotes the union of all non-trivial continua which are contained in the fibers of the restriction $q_{y\nu}=q_1|\big(p^{-1}(\beta(y))\cap
q_1^{-1}(\overline{G}_\nu)\big)$. Actually, we proved that $\alpha\big(F_\nu(h,{\varepsilon},y)\big)\subset F_\nu(q_1,\beta(y))$. Since $q_1\in\mathcal{H}_p(L)$, $q_{y\nu}\colon p^{-1}(\beta(y))\cap
q_1^{-1}(\overline{G}_\nu)\to\overline{G}_\nu$ is $(m-n_\nu)$-dimensional Lelek map. Consequently, $\dim
F_\nu(q_1,\beta(y))\leq m-n_\nu$. So, there exists an open cover $\gamma$ of $F_\nu(q_1,\beta(y))$ of order $\leq m-n_\nu+1$ (such a cover $\gamma$ exists because $F_\nu(q_1,\beta(y))$ is metrizable as a subset of the metrizable compactum $p^{-1}(\beta(y)$). We can suppose that $\gamma$ is so small that $\alpha^{-1}(\gamma)$ refines ${\mathcal U}_1$. But ${\mathrm{mesh}}({\mathcal U}_1)<\eta$. Consequently, $\alpha^{-1}(\gamma)\cap F_\nu(h,{\varepsilon},y)$ is an open cover of $F_\nu(h,{\varepsilon},y)$ of order $\leq m-n_\nu+1$ and mesh $<\eta$. This means that $d_{m-n\nu}\big(F_\nu(h,{\varepsilon},y)\big)<\eta$. Therefore, we found a simplicially factorizable map $h\in\mathcal{H}(Y,{\varepsilon},\eta)$ which is $\delta$-homotopic to $g$.
In next two lemmas we suppose that $p\colon N\to L$ is an $m$-dimensional $PL$-map between finite simplicial complexes. As everywhere in this section, $(M,\varrho)$ is a complete metric space possessing two locally finite open covers $\mu=\{W_\nu:\nu\in\Lambda\}$ and $\mu_1=\{G_\nu:\nu\in\Lambda\}$ such that $\overline{G}_\nu\subset W_\nu$, $W_\nu\in AP(n_\nu,0)$ and $0\leq n_\nu\leq m$ for every $\nu\in\Lambda$. For given ${\varepsilon},\eta>0$ and $y\in L$ we denote by $\mathcal{H}_p(y,{\varepsilon},\eta)$ the set of $g\in C(N,M)$ such that $d_{m-n_\nu}\big(F_\nu(g,{\varepsilon},y)\big)<\eta$ for every $\nu\in\Lambda$. Here, $F_\nu(g,{\varepsilon},y)$ is the union of all continua $C\subset
p^{-1}(y)\cap g^{-1}(z_C)$ with $z_C\in\overline{G}_\nu$ and ${\mathrm{diam}}C\geq{\varepsilon}$. If $B\subset L$, then $\mathcal{H}_p(B,{\varepsilon},\eta)$ stands for the intersection of all $\mathcal{H}_p(y,{\varepsilon},\eta)$, $y\in B$.
Let $B\subset L$ be a subcomplex of $L$ and $g_0\in C(N,M)$ be a map such that $g_0\in\mathcal{H}_p(B)=\bigcap_{i\geq 1}\mathcal{H}_p(B,1/i,1/i)$. Then, for every ${\varepsilon}>0$, $q\geq 1$ and $g\in C(N,M)$ with $g|p^{-1}(B)=g_0|p^{-1}(B)$ there exists a map $g_q\in C(N,M)$ extending $g_0|p^{-1}(B)$ such that $g_q$ is ${\varepsilon}$-homotopic to $g$ and $g_q\in\mathcal{H}_p(y,1/q,1/q)$ for all $y\in L$.
We fix $q$ and $g\in C(N,M)$ with $g|p^{-1}(B)=g_0|p^{-1}(B)$. Then, by Lemma 3.2, for every $y\in B$ there exists a neighborhood $V_y$ in $L$ and $\delta_y>0$ such that any $h\in
C(N,M)$ belongs to $\mathcal{H}_p(V_y,1/q,1/q)$ provided $h|p^{-1}(V_y)$ is $\delta_y$-close to $g|p^{-1}(V_y)$ (we can apply Lemma 3.2 because $g|p^{-1}(B)=g_0|p^{-1}(B)$ yields $g\in\mathcal{H}_p(y,1/q,1/q)$ for every $y\in B$). Let $\{V_{y(i)}\}_{i\leq s}$ be a finite subfamily of $\{V_y:y\in B\}$ covering $B$ and $V=\bigcup_{1\leq i\leq s}V_{y(i)}$.
Since $\mu$ is locally finite in $M$, $g(N)$ meets only finitely many $W_j=W_{\nu(j)}$, $j=1,..,k$. For any $j\leq k$ let $P_j=g^{-1}(\overline{G}_{\nu(j)})$ and $U_j^1, U_j^2$ be open subsets of $N$ such that $P_j\subset
U_j^1\subset\overline{U}_j^1\subset
U_j^2\subset\overline{U}_j^2\subset g^{-1}(W_j)$. We also choose ${\varepsilon}_0>0$ satisfying the following condition:
- If $h\in C(N,M)$ with $\varrho\big(g(x),h(x)\big)<{\varepsilon}_0$ for every $x\in N$, then $\{\nu\in\Lambda:h(N)\cap W_\nu\neq\emptyset\}$ is contained in $\{\nu(j):j=1,..,k\}$, and for all $j$ we have $h^{-1}(\overline{G}_{\nu(j)})\subset U_j^1$ and $h(\overline{U}_j^2)\subset W_j$.
Let $\delta_0=\min\{{\varepsilon},{\varepsilon}_0,\delta_{y(i)}:i\leq s\}$. Considering suitable subdivisions of $N$ and $L$, we can suppose that $p$ is a simplicial map and the following conditions hold:
- Every simplex $\sigma\in L$ intersecting the set $L\backslash V$ does not meet $B$;
- Every simplex $\tau\in N$ intersecting the set $\overline{U}_j^1$ does not meet $N\backslash U_j^2$, $j=1,..,k$.
Let $L_0=\cup\{\sigma\in L:\sigma\cap L\backslash V\neq\emptyset\}$ and $N_j=\cup\{\tau\in N:\tau\cap\overline{U}_j^1\neq\emptyset\}$, $1\leq j\leq k$. Obviously $L_0$ is a subpolyhedron of $L$ disjoint from $B$ and containing $L\backslash V$. Moreover, each $N_j$ is a subpolyhedron of $N$ such that $\overline{U}_j^1\subset N_j\subset
U_j^2$. Now, for every $j\leq k$ consider the map $p_j=p|N_j\colon
N_j\to p(N_j)$. Since $p$ is $m$-dimensional, by [@bp:96] or [@ht:85], there exists an $(n_{\nu(j)}-1)$-dimensional sigma-compact set $Z_j=\bigcup_{i\geq 1}Z_{ij}\subset N_j$ such that each $Z_{ij}$ is compact and $\dim p_j^{-1}(y)\backslash Z_j\leq m-n_{\nu(j)}$ for all $y\in p(N_j)$. Denote $$T_{ij}=\{h\in C(N,M):F^j(h,1/q)\cap Z_{ij}=\emptyset\},$$ where $F^j(h,1/q)$ is the union of all continua $C\subset N_j$ of diameter $\geq 1/q$ which are contained in fibers of the map $h|N_j$. It is easily seen that $T_{ij}$ are open in $C(N,M)$.
[*Claim. For any $h\in B_\varrho(g,\delta_0)$, $\eta>0$ and $i,j\geq 1$ there exists $h_{ij}\in T_{ij}\cap B_\varrho(h,\eta)\cap
B_\varrho(g,\delta_0)$ which is $\eta$-homotopic to $h$.*]{}
Indeed, $h\in B_\varrho(g,\delta_0)$ and $\delta_0\leq{\varepsilon}_0$ imply that $h(N_j)\subset W_j$ and $h(N)\cap W_\nu=\emptyset$ if $\nu\neq\nu(j)$ for all $j$. Let $\eta_1=\min\{\eta,\delta_0-\varrho\big(g(x),h(x)\big):x\in N\}$. Since $h|N_j$ is simplicially factorizable (as a map whose domain is a polyhedron), $W_j\in AP(n_{\nu(j)},0)$ and $Z_{ij}$ is a compact subset of $N_j$ with $\dim Z_{ij}\leq n_{\nu(j)}-1$, according to Proposition 2.4, there is a map $h'\colon N_j\to W_j$ which is $\eta_1$-homotopic to $h|N_j$ and the union of all non-trivial components of the fibers of $h'$ is disjoint from $Z_{ij}$. By the Homotopy Extension Theorem, $h'$ admits an extension $h_{ij}\in
C(N,M)$ with $h_{ij}$ being $\eta_1$-homotopic to $h$. Obviously, $h_{ij}\in T_{ij}$ and $h_{ij}\in B_\varrho(g,\delta_0)$.
The above claim allows us to apply Lemma 2.2 for the set $B_\varrho(g,\delta_0)$ and the sequence $\{T_{ij}\}_{i,j\geq 1}$ to obtain a map $h_1\in C(N,M)$ such that $h_1\in\bigcap_{i,j=1}^{\infty}T_{ij}$ and $h_1$ is $\delta_0$-homotopic to $g$. Let $h_2\colon p^{-1}(L_0\cup
B)\to M$ be defined by $h_2(x)=g_0(x)$ if $x\in p^{-1}(B)$ and $h_2(x)=h_1(x)$ if $x\in p^{-1}(L_0)$. Obviously, $h_2$ is $\delta_0$-homotopic to $g$. Since $L_0\cup B$ is a subpolyhedron of $L$ and $p$ is a simplicial map, $p^{-1}(L_0\cup B)$ is a subpolyhedron of $N$. Then, by the Homotopy Extension Theorem, there exists a map $g_q\in C(N,M)$ extending $h_2$ with $g_q$ being $\delta_0$-homotopic to $g$.
It remains only to show that $g_q\in\mathcal{H}_p(y,1/q,1/q)$ for all $y\in L$. This is true if $y\in V$. Indeed, then $y$ belongs to some $V_{y(i)}$. Since $g_q|V_{y(i)}$ is $\delta_0$-close to $g|V_{y(i)}$ and $\delta_0\leq\delta_{y(i)}$, $g_q\in\mathcal{H}_p(y,1/q,1/q)$ according to the choice of $V_{y(i)}$ and $\delta_{y(i)}$. If $y\in L_0$, then $g_q|p^{-1}(y)=h_1|p^{-1}(y)$. Since $h_1$ is $\delta_0$-close to $g$ and $\delta_0\leq{\varepsilon}_0$, $h_1^{-1}(\overline{G}_{\nu(j)})\subset
U_j^1\subset N_j$ for every $j\leq k$. So, $p^{-1}(y)\cap
g_q^{-1}(\overline{G}_{\nu(j)})=p^{-1}(y)\cap
h_1^{-1}(\overline{G}_{\nu(j)})=p_j^{-1}(y)\cap
g_q^{-1}(\overline{G}_{\nu(j)})$, $j\leq k$. On the other hand, $h_1\in\bigcap_{i,j=1}^{\infty}T_{ij}$ implies that every restriction $h_1|h_1^{-1}(\overline{G}_{\nu(j)})\colon
h_1^{-1}(\overline{G}_{\nu(j)})\to\overline{G}_{\nu(j)}$ has the following property: the union of all non-trivial components of the fibers of $h_1|h_1^{-1}(\overline{G}_{\nu(j)})$ is contained in $N_j\backslash Z_j$. Hence, the union of all non-trivial components of the fibers of $g_q|p^{-1}(y)\cap g_q^{-1}(\overline{G}_{\nu(j)})$ is contained in $p_j^{-1}(y)\backslash Z_j$. Since $\dim
p_j^{-1}(y)\backslash Z_j\leq m-n_{\nu(j)}$, every $g_q|p^{-1}(y)\cap g_q^{-1}(\overline{G}_{\nu(j)})$ is an $(m-n_{\nu(j)})$-dimensional Lelek map. But $g_q(N)$ doesn’t meet any $\overline{G}_\nu$ except for $\nu\in\{\nu(j):j=1,..,k\}$. Therefore, $g_q\in\mathcal{H}_p(y,1/i,1/i)$ for every $i\geq 1$.
Let $B\subset L$ be a subcomplex of $L$ and $g_0\in C(N,M)$ be a map such that $g_0\in\mathcal{H}_p(B)=\bigcap_{i\geq
1}\mathcal{H}_p(B,1/i,1/i)$. Then, for every $\delta>0$ there exists $\overline{g}_0\in\mathcal{H}_p(L)$ which is $\delta$-homotopic to $g_0$ and $\overline{g}_0|p^{-1}(B)=g_0|p^{-1}(B)$
Each set $\mathcal{H}_p(L,1/i,1/i)$ is open in $C(N,M)$ according to Proposition 3.3. So, by Lemma 3.5, we can apply Lemma 2.2 (with $U(i)$ being in our case $\mathcal{H}_p(L,1/i,1/i)$ and $A=p^{-1}(B)$) to find a map $\overline{g}_0\in\bigcap_{i\geq
1}\mathcal{H}_p(L,1/i,1/i)$ which is $\delta$-homotopic to $g_0$ and $\overline{g}_0|p^{-1}(B)=g_0|p^{-1}(B)$. Finally, $\overline{g}_0\in\mathcal{H}_p(L)$ because $\mathcal{H}_p(L)=\bigcap_{i\geq 1}\mathcal{H}_p(L,1/i,1/i)$.
Next proposition completes the proof of Theorem 3.1. We suppose that $p\colon N\to L$ is a perfect $m$-dimensional $PL$-map between simplicial complexes, $(M,\varrho)$ a complete metric space and $\mu=\{W_\nu:\nu\in\Lambda\}$, $\mu_1=\{G_\nu:\nu\in\Lambda\}$ are locally finite open covers of $M$ with $\overline{G}_\nu\subset
W_\nu$ and $W_\nu\in AP(n_\nu,0)$ for every $\nu\in\Lambda$, where $n_\nu\leq m$ are integers. If $B\subset A\subset L$, denote by $\mathcal{H}_p(A,B)$ the set of the maps $g\in C(p^{-1}(A),M)$ such that $g|g^{-1}(\overline{G}_\nu)\cap p^{-1}(y)\colon
g^{-1}(\overline{G}_\nu)\cap p^{-1}(y)\to \overline{G}_\nu$ is $(m-n_\nu)$-dimensional Lelek map for every $\nu\in\Lambda$ and $y\in B$. When $A=B$, we write $\mathcal{H}_p(B)$ instead of $\mathcal{H}_p(A,B)$.
Let $p\colon N\to L$ and $(M,\varrho)$ be as above. Then the set $\mathcal{H}_p(L)$ is homotopically dense in $C(N,M)$.
As usual, the simplicial complexes $N$ and $L$ are equipped with the $CW$-topology. But when consider a diameter of any subset of $N$ we mean the diameter with respect to the standard metric generating the metric topology of $N$. According to the notations in this section, for every sets $B\subset A\subset L$ and ${\varepsilon},\eta>0$, let $\mathcal{H}_p(A,B,{\varepsilon},\eta)$ be the set of all $g\in C(p^{-1}(A),M)$ such that any $d_{m-n_\nu}\big(F_\nu(g,{\varepsilon},y)\big)<\eta$ for all $y\in B$ and $\nu\in\Lambda$. Although the domain of $g$ is the set $A$ (not the whole space $N$), we use the same notation $F_\nu(g,{\varepsilon},y)$ to denote the union of all continua $C\subset
p^{-1}(y)\cap g^{-1}(z_C)$ with $z_C\in\overline{G}_\nu$ and ${\mathrm{diam}}C\geq{\varepsilon}$. Let us also denote by $\mathcal{H}_p(A,B)$ the set of the maps $g\in C(p^{-1}(A),M)$ such that $g|g^{-1}(\overline{G}_\nu)\cap
p^{-1}(y)\colon g^{-1}(\overline{G}_\nu)\cap p^{-1}(y)\to
\overline{G}_\nu$ is an $(m-n_\nu)$-dimensional Lelek map for every $\nu\in\Lambda$ and $y\in B$. This means that $\dim F_\nu(g,y)\leq
m-n_\nu$ for any $y\in B$ and $\nu\in\Lambda$, where $F_\nu(g,y)=\bigcup_{i=1}^{\infty}F_\nu(g,1/i,y)$. It is easily seen that $\mathcal{H}_p(A,B)=\bigcap_{i\geq
1}\mathcal{H}_{p}(A,B,1/i,1/i)$.
Now, let us finish the proof of Proposition 3.7. Fix $g\in C(N,M)$ and $\delta\in C(N,(0,1])$. We are going to find $h\in\mathcal{H}_p(L)$ which is $\delta$-homotopic to $g$. To this end, let $L^{(i)}$, $i\geq 0$, denote the $i$-dimensional skeleton of $L$ and $L^{(-1)}=\emptyset$. We put $h_{-1}=g$ and construct inductively a sequence $(h_i:N\to M)_{i\geq 0}$ of maps such that
- $h_{i}|p^{-1}(L^{(i-1)})=h_{i-1}|p^{-1}(L^{(i-1)})$;
- $\displaystyle h_{i}$ is $\displaystyle\frac{\delta}{2^{i+2}}$-homotopic to $h_{i-1}$;
- $h_i\in\mathcal{H}_p(L,L^{(i)})$.
Assuming that the map $h_{i-1}:N\to M$ has been constructed, consider the complement $L^{(i)}\setminus L^{(i-1)}=\sqcup_{j\in
J_i}\overset{\circ}\sigma_j$, which is the discrete union of open $i$-dimensional simplexes. Since $h_{i-1}|\sigma_j$ belongs to $\mathcal{H}_p(\sigma_j,\sigma^{(i-1)}_j)$ for any simplex $\sigma_j\in L^{(i)}$, we can apply Lemma 3.6 to find a map $g_j:p^{-1}(\sigma_j)\to M$ such that
- $g_j$ coincides with $h_{i-1}$ on the set $p^{-1}(\sigma^{(i-1)}_j)$;
- $g_j$ is $\displaystyle\frac{\delta}{2^{i+2}}$-homotopic to $h_{i-1}$;
- $g_j\in\mathcal{H}_p(\sigma_j,\sigma_j)$.
Next, define a map $q_i:p^{-1}(L^{(i)})\to M$ by the formula $$q_i(x)=\begin{cases} h_{i-1}(x)&\mbox{if $x\in
p^{-1}(L^{(i-1)})$;}\\ g_j(x)&\mbox{if $x\in p^{-1}(\sigma_j)$.}
\end{cases}$$ It can be shown that $q_i$ is $\displaystyle\frac{\delta}{2^{i+2}}$-homotopic to $h_{i-1}|p^{-1}(L^{(i)})$. Since $p^{-1}(L^{(i)})$ is a subpolyhedron of $N$, we can apply the Homotopy Extension Theorem to find a continuous extension $h_i:N\to M$ of the map $q_i$ which is $\displaystyle\frac{\delta}{2^{i+2}}$-homotopic to $h_{i-1}$. Moreover, $h_i\in\mathcal{H}_p(L,L^{(i)})$ because $h_{i-1}\in\mathcal{H}_p(L,L^{(i-1)})$ and $g_j\in\mathcal{H}_p(\sigma_j,\sigma_j)$ for any $j$. This completes the inductive step.
Then the limit map $h=\lim_{i\to\infty}h_i:N\to M$ is well-defined, continuous and $\delta$-homotopic to $g$ (the last two properties of $h$ hold because $h$ has this properties for any simplex from $N$ and because of the definition of the $CW$-topology on $N$). Finally, since $h|p^{-1}(L^{(i)})=h_i|p^{-1}(L^{(i)})$ and $h_i\in\mathcal{H}_p(L,L^{(i)})$ for every $i$, $h\in\mathcal{H}_p(L)$.
Almost $AE(n,0)$-spaces
=======================
We already observed that if $M$ is an $\LC[n]$-space, then $M\in
AP(n,0)$ if and only if $M$ has the following property:
- for every map $g\in C({\mathbb I}^n,M)$ and every ${\varepsilon}>0$ there exists a $0$-dimensional map $g'\in C({\mathbb I}^n,M)$ which is ${\varepsilon}$-close to $g$.
Any space having the above property will be referred as [*almost $AP(n,0)$*]{}. Obviously, every $\LC[n-1]$ almost $AP(n,0)$-space has the $AP(n-1,0)$-property.
We are going to establish an analogue of Theorem 3.1 for almost $AP(n,0)$-spaces.
\[anr\] Every complete $\LC[n-1]$-space $M$ admits a complete metric $\varrho$ generating its topology and satisfying the following condition: If $Z$ is an $n$-dimensional space, $A\subset Z$ its closed set and $h\colon Z\to M$, then for every function $\alpha:Z\to(0,1]$ and every map $g\colon A\to M$ with $\varrho(g(z),h(z))<\alpha(z)/8$ for all $z\in A$ there exists a map $\bar{g}\colon Z\to M$ extending $g$ such that $\varrho(\bar{g}(z),h(z))<\alpha(z)$ for all $z\in Z$.
We embed $M$ in a Banach space $E$ as a closed subset. Since the Hilbert cube is the image of the $n$-dimensional Menger compactum under an $n$-invertible map [@ad:84], we can find a metric space $E(n)$ with $\dim E(n)\leq n$ and a perfect $n$-invertible surjection $p\colon E(n)\to E$. Here, $p$ is $n$-invertible means that every map from at most $n$-dimensional space into $E$ can be lifted to a map into $E(n)$.
Since $M\in\LC[n-1]$, there exist a neighborhood $W$ of $M$ in $E$ and a map $q\colon p^{-1}(W)\to M$ extending the restriction $p|p^{-1}(M)$. For every open $U\subset M$ let $T(U)=W\backslash
p(q^{-1}(M\backslash U))$. Obviously, $T(U)\subset W$ is open, $T(U)\cap M=U$ and $q(p^{-1}(T(U)))=U$. Let $\mathcal{T}$ be the collection of all pairs $(U,V)$ of open sets in $M$ such that $\overline{conv(V)}\subset T(U)$, where $\overline{conv(V)}$ is the closed convex hull of $V$ in $E$. Now, consider the family $$\mathcal{T}=\{(U,V):U,V\hbox{~}\mbox{are open
in}\hbox{~}M\hbox{~}\mbox{and}\hbox{~}\overline{conv(V)}\subset
T(U)\}.$$ The family $\mathcal{T}$ has the following properties: (i) for any $z\in M$ and its neighborhood $U$ in $M$ there is a neighborhood $V\subset U$ of $z$ with $(U,V)\in\mathcal{T}$; (ii) for any $(U,V)\in\mathcal{T}$ and open sets $U',V'\subset M$ we have $(U',V')\in\mathcal{T}$ provided $U\subset U'$ and $V'\subset V$. By [@dm:97 Proposition 2.3], there exists a complete metric $\varrho$ on $M$ such that for every $z\in M$ and $r\in (0,1)$ the pair of open balls $(B_{\varrho}(z,r),(B_{\varrho}(z,r/8))$ belongs to $\mathcal{T}$.
Suppose we are given an $n$-dimensional space $Z$, its closed subset $A\subset Z$ and two maps $h\colon Z\to M$ and $g\colon A\to M$ such that $\varrho(g(z),h(z))<\alpha(z)/8$ for all $z\in A$, where $\alpha\in C(Z,(0,1])$. Consider the set-valued map $\phi\colon Z\to E$, $\phi(z)=g(z)$ if $z\in A$ and $\phi(z)=\overline{conv(B_{\varrho}(h(z),\alpha(z)/8))}$ if $z\not\in A$. Then $\phi$ is lower semi-continuous and has closed and convex values in $E$. So, by the Michael convex-valued selection theorem [@em:56], $\phi$ has a continuous selection $g_1$. Next, we lift $g_1$ to a map $g_2\colon Z\to E(n)$. According to the definition of $\mathcal{T}$, every $p^{-1}\big(\overline{conv(B_{\varrho}(h(z),\alpha(z)/8))}\big)$ is contained in $q^{-1}\big(B_{\varrho}(h(z),\alpha(z))\big)$. Hence, $\overline{g}=q\circ g_2$ is the required extension of $g$.
We also need the following lemma whose proof is similar to that one of Lemma 2.2.
Let $G\subset C(X,M)$, where $(M,\varrho)$ is a complete metric space. Suppose $(U(i)_{i\geq 1}$ is a sequence of open subsets of $C(X,M)$ such that
- for any $g\in G$, $i\geq 1$ and any function $\eta\in
C(X,(0,1])$ there exists $g_i\in B_\varrho(g,\eta)\cap U(i)\cap G$ which is $\eta$-close to $g$.
Then, for any $g\in G$ and ${\varepsilon}\colon X\to (0,1]$ there exists $g'\in\bigcap_{i=1}^{\infty} U(i)$ which is and ${\varepsilon}$-close to $g$. If, in addition, all $g_i$ can be chosen such that $g_i|A=g_0|A$ for some $g_0\in C(X,M)$ and $A\subset X$, then $g'|A=g_0|A$.
Next two propositions are analogues of Proposition 2.3 and Proposition 2.4.
\[n-1\] Let $M$ be a complete $\LC[n-1]$ almost $AP(n,0)$-space. Then for every $n$-dimensional compactum $X$ there exists a dense $G_{\delta}$-subset $\mathcal H\subset C(X,M)$ of $0$-dimensional maps.
First of all, let us note that since $\dim X\leq n$ and $M$ is $\LC[n-1]$, the set of all simplicially factorizable maps from $C(X,M)$ is dense in $C(X,M)$. Analyzing the proof of Proposition 2.3 and using Lemma 4.2 instead of Lemma 2.2, one can see that it suffices to establish the following claim:
[*Claim. If $K$ is a finite $n$-dimensional polyhedron, then every map $g\colon K\to M$ can be approximated by a $0$-dimensional map $g'\colon K\to M$.*]{}
Since the sets $C(K,M;\eta)$ consisting of maps $g\in C(K,M)$ with $d_0(g^{-1}(g(x))<\eta$, $\eta>0$, are open in $C(K,M)$, according to Lemma 4.2, it is enough to show that for every $g\in C(K,M)$ and $\eta>0$ there exists $g'\in C(K,M;\eta)$ which is $\eta$-close to $g$. To this end, we equipped $M$ with a complete metric satisfying the hypotheses of Lemma 4.1 and fix $0<\eta\leq 1$ and $g\in
C(K,M)$. Since $M\in\LC[n-1]$ and $M$ is almost $AP(n,0)$, $M\in
AP(n-1,0)$. So, there exists a $0$-dimensional map $h\colon
K^{(n-1)}\to M$ which is $\eta/16$-close to $g|K^{(n-1)}$. Here, $K^{(n-1)}$ is the $(n-1)$-dimensional skeleton of $K$. By Lemma 4.1, $h$ can be extended to a map $\bar{h}\in C(K,M)$ with $\varrho(\bar{h}(x),g(x))<\eta/2$ for all $x\in K$. Let $\sigma_j$, $j=1,..,k$, be all $n$-dimensional simplexes of $K$. Then $K\backslash K^{(n-1)}$ is a disjoint union of the open simplexes $\overset{\circ}\sigma_j$. Let $K_i=K^{(n-1)}\bigcup_{j=1}^{j=i}\sigma_j$, $i=1,..,k$ We are going to construct by induction maps $h_i\colon K\to M$, $i=0,..,k$, satisfying the following conditions:
- $h_0=\bar{h}$;
- $h_i|K_i$ belongs to $C(K_i,M;\eta)$, $1\leq i\leq k$;
- $h_i$ and $h_{i+1}$ are $\displaystyle (\eta/2k)$-close, $i=0,..,k-1$.
Assume that $h_i$ has already been constructed. Since $h_i|K_i$ belongs to $C(K_i,M;\eta)$, every fiber $h_i^{-1}(y)\cap K_i$ of $h_i|K_i$, $y\in h_i(K_1)$, is covered by a finite open and disjoint family $\Omega(y)$ in $K_i$ with ${\mathrm{mesh}}(\Omega(y))<\eta$. Using that $h_i(K_i)$ is compact, we find $\displaystyle 0<\delta_i<\eta/2k$ such that if $\varrho(y,h_i(K_i))<\delta_i$ for some $y\in M$, then there is $z\in
h_i(K_i)$ with $h_i^{-1}(B_\varrho(y,\delta_i))\cap K_i\subset\Omega(z)$. Since $M$ is almost $AP(n,0)$, there exists a $0$-dimensional map $p_i\colon\sigma_{i+1}\to M$ which is $\delta_i/8$-close to $h_i|\sigma_{i+1}$. By Lemma 4.1, $p_i$ can be extended to a map $h_{i+1}\colon K\to M$ being $\delta_i$-close to $h_i$. To show that $h_{i+1}|K_{i+1}$ belongs to $C(K_{i+1},M;\eta)$, we observe that $h_{i+1}^{-1}(y)\cap K_{i+1}=\big(h_{i+1}^{-1}(y)\cap
K_i\big)\cup\big(h_{i+1}^{-1}(y)\cap\sigma_{i+1}\big)$, $y\in M$. According to our construction, we have $h_i\big(h_{i+1}^{-1}(y)\cap
K_i\big)\subset B_\varrho(y,\delta_i)\cap h_i(K_i)$. Hence, $h_{i+1}^{-1}(y)\cap K_i\subset h_i^{-1}(B_\varrho(y,\delta_i))\cap K_i
\subset\Omega(z)$ for some $z\in h_i(K_i)$. Therefore, $d_0\big(h_{i+1}^{-1}(y)\cap K_{i}\big)<\eta$. Since $h_{i+1}^{-1}(y)\cap\sigma_{i+1}$ is $0$-dimensional, Lemma 2.1 implies that $d_0(h_{i+1}^{-1}(y)\cap K_{i+1})<\eta$. Obviously $h_k\in
C(K,M;\eta)$ and $h_k$ is $\eta/2$-close to $\bar{h}$. Hence, $g'=h_k$ is $\eta$-close to $g$. This completes the proof of the claim.
Let $M$ be a complete $\LC[n-1]$ almost $AP(n,0)$-space. Then for every $n$-dimensional compactum $X$ and its $F_{\sigma}$-subset $Z$ with $\dim Z\leq n-1$ there exists a dense $G_{\delta}$-subset $\mathcal H\subset C(X,M)$ of maps $g$ such that $Z$ is contained in the union of trivial components of the fibers of $g$.
Following the proof of Proposition 2.4 and using Lemma 4.2 and Proposition 4.3 instead of Lemma 2.2 and Proposition 2.3, respectively, it suffices to prove the following analogue of Claim 1 from Proposition 2.4.
[*Claim. Suppose $X$ is a finite $n$-dimensional polyhedron and $Z$ a subpolyhedron of $X$ with $\dim Z\leq n-1$. Then for every $g_0\in C(X,M)$ and $\delta>0$ there exists $g\in\mathcal{H}=\bigcap_{i=1}^{\infty}\mathcal{H}(Z,1/i)$ which is $\delta$-close to $g_0$.*]{}
Here, $\mathcal{H}(Z,1/i)$ is the set of all $g\in C(X,M)$ with $F(g,1/i)\cap Z=\emptyset$. We follow the proof of Claim 1, Proposition 2.4 and use the same notations. The first difference is that we take $W$ to be a neighborhood of $Z$ in $X$ such that $g_0|W$ and $(g_0\circ r)|W$ are $\delta/16$-close. Then, $(g_0\circ\pi\circ\varphi)|W$ is $\delta/16$-close to $g_0|W$. Next, we use by Proposition 4.3 to choose a $0$-dimensional map $h\colon Z\times{\mathbb I}\to M$ which is $\delta/16$-close to $g_0\circ\pi$. So, $(h\circ\varphi)|W$ is $\delta/8$-close to $g_0|W$. Finally, take a neighborhood $U$ of $Z$ in $X$ with $\overline{U}\subset W$, and use Lemma 4.1 to find an extension $g\in C(X,M)$ of $(h\circ\varphi)|\overline{U}$ with $g$ being $\delta$-close to $g_0$. Then $g\in\mathcal{H}$.
We establish now an analogue of Theorem 3.1 for almost $AP(n,0)$-spaces.
Let $f\colon X\to Y$ be a perfect map between metrizable spaces with $\dim X\leq n$ and $M$ is a complete $\LC[n-1]$-space. Suppose $\mu=\{W_\nu:\nu\in\Lambda\}$ and $\mu_1=\{G_\nu:\nu\in\Lambda\}$ are locally finite open covers of $M$ such that $\overline{G}_\nu\subset W_\nu$ and each $W_\nu$ is almost $AP(m_\nu,0)$ for every $\nu\in\Lambda$. Then there is a dense $G_{\delta}$-set $\mathcal{H}\subset C(X,M)$ of maps $g$ such that for any $y\in Y$ and $\nu\in\Lambda$ the restriction $g|\big(f^{-1}(y)\cap g^{-1}(\overline{G}_\nu)\big)$ is an $(n-m_{\nu})$-dimensional Lelek map.
Following the notations and the proof of Theorem 3.1, we can assume that $m_\nu\leq n$ for all $\nu$. Let $\mathcal{H}(y,{\varepsilon},\eta)$ denote the set of all $g\in C(X,M)$ such that $d_{n-m_\nu}\big(F_\nu(g,{\varepsilon},y)\big)<\eta$ for every $\nu\in\Lambda$, where $y\in Y$ and ${\varepsilon},\eta>0$ are fixed. Moreover, for any $F\subset Y$, let $\mathcal{H}(F,{\varepsilon},\eta)$ be the intersection of all $\mathcal{H}(y,{\varepsilon},\eta)$, $y\in F$. One can establish a lemma analogical to Lemma 3.2. Then, as in Proposition 3.3, we can show that any $\mathcal{H}(F,{\varepsilon},\eta)$ is open in $C(X,M)$, where $F\subset Y$ is closed.
Observe that $\dim\triangle(f)\leq m$ because $X$, as a space of dimension $\leq n$, admits a uniformly $0$-dimensional map into ${\mathbb I}^n$, see [@mk]. Moreover, we can assume that the simplicial complex $K$ in Proposition 3.4 is $n$-dimensional. So, we can apply Proposition 3.4 in the present situation to show that any $\mathcal{H}(Y,{\varepsilon},\eta)$ is dense in $C(X,M)$ provided for any perfect $PL$ map $p\colon N\to L$ with $\dim N\leq n$ the set $\mathcal{H}_p(L)$ is dense in $C(N,M)$. Here, $\mathcal{H}_p(L)$ is the set of all maps $q\colon N\to M$ such that $q|q^{-1}(\overline{G}_\nu)\cap p^{-1}(z)$ is an $(n-m_\nu)$-dimensional Lelek map for every $\nu\in\Lambda$ and $z\in L$. Hence, it remains to show that $\mathcal{H}_p(L)$ is dense in $C(N,M)$ for any $PL$-map $p\colon N\to L$ with $\dim N\leq n$. And this follows from the proof of Lemma 3.5, Lemma 3.6 and Proposition 3.7 with the only difference that now we replace the application of Lemma 2.2, Proposition 2.3 and Proposition 2.4 by Lemma 4.2, Proposition 4.3 and Proposition 4.4, respectively.
Let $M$ be a complete $\LC[n-1]$-space such that each point $z\in M$ has a neighborhood which is almost $AP(n,0)$. Then, for every perfect map $f\colon X\to Y$ between metric spaces with $\dim X\leq
n$, there is a dense $G_{\delta}$-set $\mathcal{H}\subset C(X,M)$ consisting of maps $g$ such that every restriction $g|f^{-1}(y)$, $y\in Y$, is a $0$-dimensional map.
Any manifold $M$ modeled on the $n$-dimensional Menger cube has the $AP(n,0)$-property, see Corollary 6.5 below. So, Theorem 1.1 holds for such a space $M$. But Theorem 1.1 does not provide any information about the density of the set $\mathcal{H}$ in $C(X,M)$ except that every simplicially factorizable map in $C(X,M)$ can be approximated by maps from $\mathcal{H}$. Next proposition shows that, in this special case, the set $\mathcal{H}$ is dense in $C(X,M)$ with respect to the uniform convergence topology.
Let $f\colon X\to Y$ be a perfect map between metrizable spaces with $\dim\triangle(f)\leq m$ and $M$ be a manifold modeled on the $n$-dimensional Menger cube. Then there is a $G_{\delta}$-set $\mathcal{H}\subset C(X,M)$ dense in $C(X,M)$ with respect to the uniform convergence topology such that for any $g\in\mathcal{H}$ and $y\in Y$ the restriction $g|f^{-1}(y)$ is an $(m-n)$-dimensional Lelek map.
Let $\mathcal{H}$ be the set from the proof of Theorem 3.1. To show that $\mathcal{H}$ is dense in $C(X,M)$ equipped with the uniform convergence topology, we used an idea from the proof of [@km Corollary 2.8]. According to [@be], for any ${\varepsilon}>0$ there exists an $n$-dimensional polyhedron $P\subset M$ of piecewise embedding dimension $n$ and maps $u\colon M\to P$ and $v\colon P\to
M$ such that $u$ is a retraction, $v$ is $0$-dimensional, $v\circ u$ is ${\varepsilon}/2$-close to the identity $id_M$. Since every $ANR$ of piecewise embedding dimension $n$ has the $AP(n,0)$-property (see [@km Propsition 2.1]), according to Theorem 3.1, for every $g\in C(X,M)$ there is $g'\colon X\to P$ such that $g'$ is $\delta$-close to $u\circ g$ and $g'|f^{-1}(y)$ is an $(m-n)$-dimensional Lelek map for all $y\in Y$. Here $\delta>0$ is chosen such that $dist(v(x),v(y))<{\varepsilon}/2$ for any $x,y\in P$ which are $\delta$-close. Then $v\circ g'$ is ${\varepsilon}$-close to $g$ and since $v$ is $0$-dimensional, $v\circ g'\in\mathcal{H}$.
Properties of $AP(n,0)$-spaces
==============================
In this section we investigate the class of $AP(n,0)$-spaces.
Let $K$ be a polyhedron of dimension $\leq n$ and $L\subset K$ a subpolyhedron. Suppose $(X,\varrho)$ is a complete metric space possessing the $AP(n,0)$-property and $g_0\in C(K,X)$ with $\dim
g_0^{-1}(g_0(x))\leq 0$ for all $x\in L$. Then for every $\delta>0$ there exists a $0$-dimensional map $g\colon K\to X$ which is $\delta$-homotopic to $g_0$ and $g|L=g_0|L$.
We already observed in Section 2 that all sets $C(K,X,{\varepsilon})$, ${\varepsilon}>0$, consisting of maps $h\in C(K,X)$ with $d_0(h^{-1}(h(x)))<{\varepsilon}$ for every $x\in K$ are open in $C(K,X)$. Since every map from $C(K,X)$ is simplicially factorizable, $C(K,X,{\varepsilon})$ are homotopically dense in $C(K,X)$ according to Proposition 2.3.
[*Claim $1$. For every $x\in L$ and $j\geq 1$ there exist ${\varepsilon}_x>0$ and a neighborhood $U_x$ of $x$ in $K$ satisfying the following condition: If $h\in C(K,X)$ and $Z\subset K$ with $\varrho\big(g_0(y),h(y)\big)<{\varepsilon}_x$ for all $y\in Z\cup\overline{U}_x$, then $d_0(h^{-1}(h(y))\cap Z)< 1/j$ for any $y\in\overline{U}_x$.*]{}
The proof of this claim is similar to the proof of Lemma 3.2.
We fix $j$ and $\eta>0$. Choose finitely many points $x(i)\in L$, $i\leq k$, positive reals ${\varepsilon}_{x(i)}$ and neighborhoods $U(x_i)$ in $K$ satisfying the hypotheses of Claim 1 such that $L\subset
U=\bigcup_{i\leq k}U(x_i)$. Taking a smaller neighborhood, if necessarily, we can suppose that $\varrho\big(g_0(x),h(x)\big)<\eta_1/2$ for all $x\in\overline{U}$, where $\eta_1=\min\{\eta,{\varepsilon}_{x(i)}:i\leq k\}$. Consider a triangulation $T$ of $K$ such that $\sigma\in T$ and $\sigma\cap
K\backslash U\neq\emptyset$ imply $\sigma\cap L=\emptyset$. Now, let $N$ be the subpolyhedron of $K$ given by $N=\cup\{\sigma\in
T:\sigma\cap K\backslash U\neq\emptyset\}$. Obviously, $N$ and $L$ are disjoint. Since $X\in AP(n,0)$ and $\dim N\leq n$, by Proposition 2.3, there exists a $0$-dimensional map $g_N\in C(N,X)$ which is $\eta_1/2$-homotopic to $h|N$ (we can apply Proposition 2.3 because $h|N$ is simplicially factorizable as a map with a polyhedral domain). The map $h'\colon N\cup L\to X$, $h'|N=g_N$ and $h'|L=g_0|L$, is $\eta_1/2$-homotopic to $h|(N\cup L)$. So, by the Homotopy Extension Theorem, $h'$ admits an extension $h_j\colon K\to
X$ which is $\eta_1/2$-homotopic to $h$. It remains only to show that $d_0(h_j^{-1}(h_j(x)))< 1/j$ for any $x\in K$. To this end, observe that $\varrho\big(g_0(x),h_j(x)\big)<\eta_1$ for all $x\in
\overline{U}$. Because $K=N\cup\overline{U}$, for every $x\in K$ we have $h_j^{-1}(h_j(x))=\big(h_j^{-1}(h_j(x))\cap
N\big)\cup\big(h_j^{-1}(h_j(x))\cap\overline{U}\big)$. Since $h_j|N=g_N$ and $g_N$ is $0$-dimensional, $\dim h_j^{-1}(h_j(x))\cap
N\leq 0$. On the other hand, $h_j^{-1}(h_j(x))\cap\overline{U}=h_j^{-1}(h_j(y))\cap\overline{U}$ for some $y\in\overline{U}$. So, there exists $m\leq k$ with $y\in\overline{U}_{x(m)}$. Since $h_j|\overline{U}$ is $\eta_1$-close to $g_0|\overline{U}$, $\varrho\big(g_0(z),h_j(z)\big)<{\varepsilon}_{x(m)}$ for all $z\in
\overline{U}$. Hence, according to the choice of $U_{x(m)}$ and ${\varepsilon}_{x(m)}$, $d_0\big(h_j^{-1}(h_j(y))\cap\overline{U})\big)<1/j$. Therefore, $d_0\big(h_j^{-1}(h_j(x))\cap\overline{U})\big)<1/j$. Finally, by Lemma 2.1, $d_0\big(h_j^{-1}(h_j(x))\big)<1/j$. This completes the proof of Claim 2.
We are in a position to complete the proof of Lemma 5.1. Because of Claim 2, we can apply Lemma 2.2 (with $G$ being in our case the set $\{h\in C(K,X):h|L=g_0|L\}$ and $U(j)=C(K,X,1/j)$) to obtain a map $g\in\bigcap_{j=1}^{\infty}C(K,X,1/j)$ such that $g|L=g_0|L$ and $g$ is $\delta$-homotopic to $g_0$. Obviously, $g\in\bigcap_{j=1}^{\infty}C(K,X,1/j)$ yields $\dim g\leq 0$.
For any space $X$ we have:
1. If $X$ has the $AP(n,0)$-property, then every open subset of $X$ also has this property.
2. If $X$ is completely metrizable, then $X$ has the $AP(n,0)$-property if and only if it admits a cover by open subsets with that property.
To prove the first item, suppose $U$ is an open subset of $X\in
AP(n,0)$, ${\mathcal U}\in cov(U)$ and $g\in C({\mathbb I}^n,U)$. Let ${\mathcal U}'={\mathcal U}\cup\{X\backslash g({\mathbb I}^n)\}$. Since $X\in AP(n,0)$, there is a $0$-dimensional map $g'\in C({\mathbb I}^n,X)$ such that $g'$ is ${\mathcal U}'$-homotopic to $g$. Then $g'({\mathbb I}^n)\subset U$ and there exists a ${\mathcal U}$-homotopy $h\colon{\mathbb I}^n\to U$ joining $g$ and $g'$.
According to Michael’s theorem on local properties [@mi], the second item will be established if we show that: (i) A space has the $AP(n,0)$-property provided $X$ is a discrete sum of spaces with the same property; (ii) A completely metrizable space has the $AP(n,0)$-property provided it is a union of two open subspaces with this property. Condition (i) trivially follows from the definition. Let us check condition (ii).
Suppose $X$ is a completely metrizable space and $X=X_0\cup X_1$ is the union of two open subspaces $X_1,X_2\subset X$ with the $AP(n,0)$-property. Fix an open cover ${\mathcal U}$ of $X$ and a map $g:\mathbb I^n\to X$ and choose two open sets $W_1,W_2\subset X$ such that $X=W_1\cup W_2$ and $W_i\subset \overline{W}_i\subset X_i$ for $i\in\{1,2\}$. Next, find a complete metric $\varrho$ on $X$ such that
- $\varrho(X\setminus W_1,X\setminus W_2)\geq 1$;
- $B_\varrho(\overline{W}_i,1)\subset X_i$ for $i\in\{1,2\}$ and
- each set of diameter $<1$ in $(X,\varrho)$ lies in some $U\in{\mathcal U}$.
Let $V=V_1\cap V_2$, where $V_i=B_\rho(\overline{W}_i,1/2)$, $i=1,2$, and choose a triangulation $T$ of $\mathbb I^n$ such that for any simplex $\sigma\in T$ we have ${\mathrm{diam}}g(\sigma)<1/4$. Now, consider the polyhedra $$\begin{aligned} K_i&=\cup\{\sigma\in T:g(\sigma)\cap
\overline{W}_i\neq\emptyset\},\hbox{} i\in\{1,2\}
\end{aligned}$$ and
$$\begin{aligned} L_2&=\cup\{\sigma\in T:g(\sigma)\cap
\overline{V}_2\neq\emptyset\}.
\end{aligned}$$
Obviously, $g(K_i)\subset V_i\subset\overline{V}_i\subset X_i$, $i=1,2$. So, $g(K_0)\subset V$, where $K_0=K_1\cap K_2$. Moreover, we have $$K_2\subset g^{-1}(V_2)\subset\overline{g^{-1}(V_2)}\subset
L_2\subset g^{-1}(X_2)$$ and $$K_0\subset g^{-1}(V)\subset L_2.$$ Choose now positive $\delta\leq\min\{\varrho\big(g(L_2),X\backslash
X_2\big),\varrho\big(g(K_1),X\backslash X_1\big),1/2\}$ such that $$h^{-1}(h(K_0))\subset g^{-1}(V)$$ for any $h\in C({\mathbb I}^n,X)$ which is $\delta$-close to $g$.
Since $X_2\in AP(n,0)$ and $L_2$ is a polyhedron of dimension $\leq
n$, by Proposition 2.3, there exists a $0$-dimensional map $g_2\colon L_2\to X_2$ which is $\delta$-homotopic to $g|L_2$. Next, by the Homotopy Extension Theorem, $g_2$ can be extended to a map $\overline{g}_2\in C({\mathbb I}^n,X)$ $\delta$-homotopic to $g$. According to the choice of $\delta$, $\overline{g}_2(L_2)\subset X_2$, $\overline{g}_2(K_1)\subset X_1$ and $(\overline{g}_2)^{-1}(\overline{g}_2(K_0))\subset g^{-1}(V)\subset
L_2$. Hence, the restriction $g_0=\overline{g}_2|K_1$ is a map from $K_1$ into $X_1$ and, for every $x\in K_0$, we have $g_0^{-1}(g_0(x))\subset g_2^{-1}(g_2(x))$. The last inclusion implies that $\dim g_0^{-1}(g_0(x))\leq 0$ for all $x\in K_0$ because $g_2$ is $0$-dimensional. Since $X_1\in AP(n,0)$, we can apply Lemma 5.1 (for the polyhedra $K_0\subset K_1$ and the map $g_0$) to obtain a $0$-dimensional map $g_1\in C(K_1,X_1)$ such that $g_1|K_0=g_0$ and $g_1$ is $1/2$-homotopic to $g_0$. Finally, consider the map $g_{12}\in C({\mathbb I}^n,X)$ defined by $g_{12}|K_1=g_1$ and $g_{12}|K_2=g_2|K_2$. Since both $g_1$ and $g_2$ are $0$-dimensional, so is $g_{12}$. Moreover, $g_{12}$ is $1$-homotopic to $g$ because $\delta\leq 1/2$. So, according to the choice of $\varrho$, $g_{12}$ is ${\mathcal U}$-homotopic to $g$.
If $X$ and $Y$ are complete metric spaces such that $X\in AP(n,0)$ and $Y\in AP(m,0)$, then $X\times Y\in
AP(n+m,0)$.
Let $\varrho_X$ and $\varrho_Y$ be complete metrics on $X$ and $Y$, respectively. We fix $g\in C({\mathbb I}^{n+m},X\times Y)$ and consider the complete metric $\varrho=\max\{\varrho_X,\varrho_Y\}$ on $X\times
Y$. It suffices to show that for every $\varepsilon>0$ there exists a $0$-dimensional map $h\in C({\mathbb I}^{n+m},X\times Y)$ which is $\varepsilon$-homotopic to $g$. Let $g_X=\pi_X\circ g$ and $g_Y=\pi_Y\circ g$, where $\pi_X$ and $\pi_Y$ are the projections from $X\times Y$ onto $X$ and $Y$, respectively. We represent ${\mathbb I}^{n+m}$ as the union $\bigcup_{i=1}^{n+m+1}A_i$ with each $A_i$ being a $0$-dimensional $G_{\delta}$-subset of ${\mathbb I}^{n+m}$. Then $Z_X={\mathbb I}^{n+m}\backslash\bigcup_{i=n+1}^{n+m+1}A_i$ is an $F_{\sigma}$-subset of ${\mathbb I}^{n+m}$ which is contained in $\bigcup_{i=1}^{n}A_i$. So, $\dim Z_X\leq n-1$. Since $X\in AE(n,0)$ and every map from $C({\mathbb I}^{n+m},X)$ is simplicially factorizable, Proposition 2.4 yields the existence of a map $h_X\in
C({\mathbb I}^{n+m},X)$ such that $h_X$ is $\varepsilon$-homotopic to $g_X$ and $F(h_X)=\bigcup_{j=1}^\infty F(h_X,1/j)\subset{\mathbb I}^{n+m}\backslash
Z_X$. Obviously, $F(h_X)$ is an $F_{\sigma}$-set in ${\mathbb I}^{n+m}$ with $ F(h_X)\subset\bigcup_{i=n+1}^{n+m+1}A_i$. Hence, $Z_Y=F(h_X)\backslash A_{n+m+1}$ is also an $F_{\sigma}$-set in ${\mathbb I}^{n+m}$ with $\dim F_Y\leq m-1$. Now, since $Y\in AE(m,0)$, we may apply again Proposition 2.4 to obtain a map $h_Y\in
C({\mathbb I}^{n+m},Y)$ which is $\varepsilon$-homotopic to $g_Y$ and $F(h_Y)=\bigcup_{j=1}^\infty F(h_Y,1/j)\subset{\mathbb I}^{n+m}\backslash
Z_Y$. Then the diagonal product $h=h_X\triangle h_Y\colon{\mathbb I}^{n+m}\to
X\times Y$ is $\varepsilon$-homotopic to $g$. It remains only to show that $h$ is $0$-dimensional. If $C$ is a non-trivial component of a fiber of $h$, then $C\subset F(h_X)\cap F(h_Y)\subset
F(h_X)\backslash Z_Y\subset A_{m+n+1}$. Since $A_{m+n+1}$ is 0-dimensional, $C$ should be a point. Therefore, all components of the fibers of $h$ are trivial, i.e. $\dim h=0$.
Finally, we are going to show that every arc-wise connected $AP(n,0)$-compactum is a continuum $(V^n)$ in the sense of P. Alexandroff. Recall that a compact metric space $(M,\varrho)$ is a $(V^n)$-continuum [@ps] if for any pair of disjoint closed subsets $A$ and $B$ of $M$ both having non-empty interiors there exists ${\varepsilon}>0$ such that $d_{n-2}(C)>{\varepsilon}$ for every partition $C$ in $M$ between $A$ and $B$. It is easily seen that this is a topological property, i.e., it doesn’t depend on the metric $\varrho$. Obviously, every continuum $(V^n)$ is a Cantor $n$-manifold (a compactum which is not disconnected by any $(n-2)$-dimensional closed subset). Moreover, any $(V^n)$ continuum has a stronger property [@hv]: it cannot be decomposed into a countable union of proper closed subsets $F_i$ with $\dim F_i\cap
F_j\leq n-2$. The compacta with the last property are called strong Cantor $n$-manifolds [@nh].
Every path-connected compactum $M\in AP(n,0)$ is a continuum $(V^n)$.
Let $A$ and $B$ be disjoint closed subsets of $M$ with non-empty interior and $\varrho$ be a metric on $M$. Since $M$ is path-connected, we can choose a map $g\colon{\mathbb I}^n\to M$ such that $g({\mathbb I}^n)\cap Int(A)\neq\emptyset\neq g({\mathbb I}^n)\cap Int(B)$. Then there exists a $0$-dimensional map $g_1\colon{\mathbb I}^n\to M$ which is so close to $g$ that $g_1({\mathbb I}^n)$ meets both $Int(A)$ and $Int(B)$. Thus, $A_1=g_1^{-1}(A)$ and $B_1=g_1^{-1}(B)$ are disjoint closed subsets of ${\mathbb I}^n$ with non-empty interiors. Since ${\mathbb I}^n$ is a continuum $(V^n)$ [@ps], there exists ${\varepsilon}>0$ such that $d_{n-2}(C_1)>{\varepsilon}$ for every partition $C_1\subset{\mathbb I}^n$ between $A_1$ and $B_1$. Because $g_1$ is $0$-dimensional, every $y\in M$ has a neighborhood $W_y$ such that $g_1^{-1}(W_y)$ splits into a finite disjoint family of open subsets of ${\mathbb I}^n$ each of diameter $<{\varepsilon}$. Let $\delta$ be the Lebesgue number of the cover $\{W_y:y\in M\}$. Then $d_{n-2}(C)\geq\delta$ for any partition $C\subset M$ between $A$ and $B$. Indeed, otherwise there would be a partition $C$ and an open family $\gamma_C$ in $M$ of order $\leq n+1$ such that $\gamma_C$ covers $C$ and ${\mathrm{diam}}(W)<\delta$ for every $W\in\gamma_C$. Hence, $g_1^{-1}(\gamma_C)$ is an open family in ${\mathbb I}^n$ of order $\leq n+1$ and covering $g_1^{-1}(C)$. Moreover, each $g_1^{-1}(W)$, $W\in\gamma_C$, splits into a finite disjoint open family consisting of sets with diameter $<{\varepsilon}$. Therefore, $d_{n-2}(g_1^{-1}(C))<{\varepsilon}$. This is a contradiction because $g_1^{-1}(C)$ is a partition in ${\mathbb I}^n$ between $A_1$ and $B_1$.
$AP(n,0)$-spaces and general position properties
================================================
The parametric general position properties were introduces in [@bv]. We say that a space $M$ has the $m\mbox{-}\overline{DD}^{\{n,k\}}$-property, where $m,n,k\geq 0$ are integers or $\infty$, if for any open cover ${\mathcal U}$ of $M$ and any two maps $f:\mathbb I^m\times \mathbb I^n\to M$, $g:\mathbb I^m\times
\mathbb I^k\to M$ there exist maps $f':\mathbb I^m\times \mathbb
I^n\to M$, $g':\mathbb I^m\times \mathbb I^k\to M$ which are ${\mathcal U}$-homotopic to $f$ and $g$, respectively, and $f'(\{z\}\times
\mathbb I^n)\times g'(\{z\}\times \mathbb I^k\}=\emptyset$ for all $z\in \mathbb I^m$. It is clear that this is exactly the well known disjoint $n$-disks property provided $m=0$, $n=k$ and $M$ is $\LC[n]$. When $m=0$, we simply write $\overline{DD}^{\{n,k\}}$ instead of $0\mbox{-}\overline{DD}^{\{n,k\}}$.
Let $M$ be completely metrizable having the $\overline{DD}^{\{n,k\}}$-property. Suppose $X$ is a compactum and $A,B\subset X$ closed disjoint subsets with $\dim A\leq n$ and $\dim B\leq k$. Then every simplicially factorizable map $g\colon
X\to M$ can be homotopically approximated by maps $g'\in C(X,M)$ such that $g'(A)\cap g'(B)=\emptyset$.
Let $g\in C(X,M)$ be simplicially factorizable and $\delta>0$. As in Claim 2 from the proof of Proposition 2.4, we can find a finite open cover ${\mathcal U}$ of $X$ such that:
- at most $n+1$ elements of the family $\gamma_A=\{U\in{\mathcal U}: U\cap
A\neq\emptyset\}$ intersect;
- at most $k+1$ elements of the family $\gamma_B=\{U\in{\mathcal U}: U\cap
B\neq\emptyset\}$ intersect;
- $\cup\gamma_A\cap \cup\gamma_B=\emptyset$;
- there exists a map $h\colon\mathcal{N}({\mathcal U})\to M$ such that $h\circ f_{{\mathcal U}}$ is $\delta/2$-homotopic to $g$, where $\mathcal{N}({\mathcal U})$ is the nerve of ${\mathcal U}$ and $f_{{\mathcal U}}\colon
X\to\mathcal{N}({\mathcal U})$ is the natural map.
Let $K_A$ and $K_B$ be the subpolyhedra of $\mathcal{N}({\mathcal U})$ generated by the families $\gamma_A$ and $\gamma_B$, respectively. Then $\dim K_A\leq n$, $\dim K_B\leq k$ and $K_A\cap K_B=\emptyset$. For any simplexes $\sigma\in K_A$ and $\tau\in K_B$ let $$G(\sigma,\tau)=\{p\in C(\mathcal{N}({\mathcal U}),M): p(\sigma)\cap
p(\tau)=\emptyset\}.$$ Obviously, each $G(\sigma,\tau)$ is open in $C(\mathcal{N}({\mathcal U}),M)$. Using $M\in\overline{DD}^{\{n,k\}}$ and the Homotopy Extension Theorem, one can show that all $G(\sigma,\tau)$ are homotopically dense in $C(\mathcal{N}({\mathcal U}),M)$. So, by Lemma 2.2, there exists $h'\in\bigcap\{G(\sigma,\tau):\sigma\in K_A,\tau\in
K_B\}$ which is $\delta/2$-homotopic to $h$. Then $g'=h'\circ
f_{{\mathcal U}}$ is $\delta$-homotopic to $g$ and $h'(A)\cap
h'(B)=\emptyset$.
For a space $X$ we have:
1. If $X$ is $\LC[0]$, then $X$ has the $\overline{DD}^{\{0,0\}}$-property if and only if $X$ has no isolated point;
2. If $X$ is completely metrizable and $X\in\overline{DD}^{\{0,0\}}$, then $X$ has the $AP(1,0)$-property.
For the first item, see [@bv Proposiion 7]. To proof the second item, we fix countably many closed $0$-dimensional subsets $P_i$ of ${\mathbb I}$ consisting of irrational numbers such that $\dim{\mathbb I}\backslash\bigcup_{i\geq 1}P_i=0$. For every rational $t\in
Q$ and $i\geq 1$ let $U_i(t)=\{g\in C({\mathbb I},M): g(t)\not\in g(P_i)\}$. By Lemma 6.1, each $U_i(t)$ is homotopically dense in $C({\mathbb I},M)$. On the other hand, obviously all $U_i(t)$ are open in $C({\mathbb I},M)$. Therefore, by Lemma 2.2, the set $U=\bigcap\{U_i(t):t\in Q, i\geq 1\}$ is homotopically dense in $C({\mathbb I},X)$. Moreover, for every $g\in U$ and $x\in{\mathbb I}$ we have the following: if $g^{-1}(g(x))\cap Q\neq\emptyset$, then $g^{-1}(g(x))\subset{\mathbb I}\backslash\bigcup_{i\geq 1}P_i$; if $g^{-1}(g(x))\cap Q=\emptyset$, then $g^{-1}(g(x))\subset{\mathbb I}\backslash Q$. Hence, $U$ consists of $0$-dimensional maps.
Next corollary of Propositions 5.3 and 6.2 provides more examples of $AP(n,0)$-spaces.
If $M_i$, $i=1,..n$, are completely metrizable $\LC[0]$-spaces without isolated points, then $\prod_{i=1}^{i=n}M_i$ has the $AP(n,0)$-property.
Here is a characterization of the $\overline{DD}^{\{n,k\}}$-property.
A completely metrizable space $M$ has the $\overline{DD}^{\{n,k\}}$-property, where $n\leq k\leq\infty$, if and only if $M$ satisfies the following condition $(n,k)$:
- If $X$ is a compactum and $A\subset B\subset X$ its $\sigma$-compact subsets with $\dim A\leq n$ and $\dim B\leq k$, then any simplicially factorizable map $g\in C(X,M)$ can be homotopically approximated by maps $h\in C(X,M)$ such that $h^{-1}(h(x))\cap B=x$ for all $x\in A$.
Suppose $M\in\overline{DD}^{\{n,k\}}$ and let $A\subset B\subset X$ be two $\sigma$-compact subsets of a compactum $X$ with $\dim A\leq
n$ and $\dim B\leq k$. Then $A=\bigcup_{p\geq 1}A_p$ and $B=\bigcup_{m\geq 1}B_m$, where $A_p$ and $B_m$ are compact sets of dimension $\dim A_p\leq n$ and $\dim B_m\leq k$. For every $p,i\geq
1$ let $\omega_i(p)=\{A_{pj}(i):j=1,2,..,s(p,i)\}$ be a family of closed subsets of $A_p$ such that $\omega_i(p)$ covers $A_p$ and ${\mathrm{mesh}}(\omega_i(p))<1/i$. We also choose sequences $\{B_{mq}(p,i,j)\}_{q=1}^{\infty}$ of closed sets $B_{mq}(p,i,j)\subset B_m$ with $B_m\backslash
A_{pj}(i)=\bigcup_{q=1}^{\infty}B_{mq}(p,i,j)$, where $p,i\geq 1$ and $j=1,2,..,s(p,i)$. Then the sets $$G(p,i,j,m,q)=\{g\in C(X,M): g(A_{pj}(i))\cap g(B_{mq}(p,i,j))=\emptyset\}$$ are open in $C(X,M)$ and the intersection $G$ of all $G(p,i,j,m,q)$ consists of maps $g$ with $g^{-1}(g(x))\cap B=x$ for any $x\in A$. Hence, according to Lemma 2.2, every simplicially factorizable map is homotopically approximated by maps from $G$ provided the following is true: For any simplicially factorizable map $g\in C(X,M)$, ${\varepsilon}>0$ and $(p,i,j,m,q)$ with $p,i,m,q\geq 1$ and $1\leq j\leq s(i,p)$, there exists a simplicially factorizable map $g(p,i,j,m,q)\in G(p,i,j,m,q)$ which is ${\varepsilon}$-homotopic to $g$. Since $M\in\overline{DD}^{\{n,k\}}$ and any couple $A_{pj}(i)$, $B_{mq}(p,i,j)$ consists of disjoint closed sets in $X$ of dimension $\leq n$ and $\leq k$, respectively, the last statement follows from Lemma 6.1. Hence $M$ satisfies condition $(n,k)$.
Suppose now that $M$ satisfies condition $(n,k)$. To show that $M\in\overline{DD}^{\{n,k\}}$, let $f\colon{\mathbb I}^n\to M$ and $g\colon{\mathbb I}^k\to M$ be two maps. If $k<\infty$, we denote by $X$ the disjoint sum ${\mathbb I}^n\biguplus{\mathbb I}^k$ and consider the map $h\colon X\to
M$, $h|{\mathbb I}^n=f$ and $h|{\mathbb I}^k=g$. Since $h$ is simplicially factorizable (as a map with a polyhedral domain) and $M$ has the $(n,k)$-property, for every ${\varepsilon}>0$ there is a map $h_1\in C(X,M)$ such that $h_1^{-1}(h_1(x))=x$ for all $x\in{\mathbb I}^n$ and $h_1$ is ${\varepsilon}$-homotopic to $h$. Then the maps $f_1=h_1|{\mathbb I}^n$ and $g_1=h_1|{\mathbb I}^k$ are ${\varepsilon}$-homotopic to $f$ and $g$, respectively, and $f_1({\mathbb I}^n)\cap g_1({\mathbb I}^k)=\emptyset$. So, $M\in\overline{DD}^{\{n,k\}}$. If $n<k=\infty$, we choose $k(1)<\infty$ such that $n\leq k(1)$ and the map $g'=g\circ r_{k(1)}$ is ${\varepsilon}/2$-homotopic to $g$, where $r_{k(1)}\colon{\mathbb I}^{\infty}\to{\mathbb I}^{k(1)}$ is the retraction of ${\mathbb I}^{\infty}$ onto ${\mathbb I}^{k(1)}\subset{\mathbb I}^{\infty}$ defined by $r_{k(1)}((x_1,x_2,..))=(x_1,x_2,..,x_{k(1)},0,0,..)$. Then the map $h\colon{\mathbb I}^n\biguplus{\mathbb I}^{\infty}\to M$, $h|{\mathbb I}^n=f$ and $h|{\mathbb I}^{\infty}=g'$, is simplicially factorizable. Hence, as in the previous case, we can use the $(n,\infty)$-property of $M$ to show that $M\in\overline{DD}^{\{n,\infty\}}$. If $n=k=\infty$, we homotopically approximate both $f$ and $g$ by maps $f'=f\circ
r_{n(1)}$ and $g'=g\circ r_{k(1)}$, respectively, and proceed as in the first case.
Every completely metrizable space $\displaystyle
M\in\overline{DD}^{\{n-1,n-1\}}$ has the $AP(n,0)$-property. In particular, every manifold modeled on the $n$-dimensional Menger cube or the $n$-dimensional Nöbeling space has the $AP(n,0)$-property.
Let ${\mathbb I}^n=A\cup B$, where $B={\mathbb I}^n\backslash A$ is $0$-dimensional and $A=\bigcup_{i\geq 1}A_i$ is $\sigma$-compact with $\{A_i\}$ being a sequence of closed $(n-1)$-dimensional subsets of ${\mathbb I}^n$. Since $C({\mathbb I}^n,M)$ consists of simplicially factorizable maps, by Proposition 6.4, every map in $C({\mathbb I}^n,M)$ is homotopically approximated by maps $g\in C({\mathbb I}^n,M)$ such that $g^{-1}(g(x))\cap
A=x$ for all $x\in A$. Let us show that any such $g$ is $0$-dimensional. Indeed, since for every $x\in{\mathbb I}^n$ the intersection $g^{-1}(g(x))\cap A$ can have at most one point, $g^{-1}(g(x))=(g^{-1}(g(x))\cap A)\cup (g^{-1}(g(x))\cap B)$ is $0$-dimensional.
The second part of the corollary follows from the fact that any $n$-dimensional Menger manifold, as well as any manifold modeled on the $n$-dimensional Nöbeling space, is a complete $\LC[n-1]$-space with the disjoint $n$-disks property [@be]. So, any such a space has the $\overline{DD}^{\{n-1,n-1\}}$-property.
The following notions were introduced in [@bv] and [@bck]. Let A be a closed subset of a space $X$. We say that $A$ is:
- a (homotopical) $Z_n$-set in $X$ if for any an open cover ${\mathcal U}$ of $X$ and a map $f:{\mathbb I}^n\to X$ there is a map $g:{\mathbb I}^n\to X$ such that $g({\mathbb I}^n)\cap A=\emptyset$ and $g$ is ${\mathcal U}$-near (${\mathcal U}$-homotopic) to $f$;
- homological $Z_n$-set in $X$ if $H_k(U,U\setminus A)=0$ for all open sets $U\subset
X$ and all $k<n+1$.
Here, $H_{*}(U,U\setminus A)=0$ are the relative singular homology groups with integer coefficients. Each homotopical $Z_n$-set in $X$ is both a $Z_n$-set and a homological $Z_n$-set in $X$. The converse is not always true.
\[t1\] Let $X$ be a locally path-connected $AP(n,0)$-space. Then every $x\in X$ is a homological $Z_{n-1}$-point in $X$.
By [@bck Corollary 8.4], it suffices to check that $(x,0)$ is a homotopical $Z_n$-point in $X\times[-1,1]$ for every $x\in X$. Given a map $f:{\mathbb I}^n\to X\times[-1,1]$ we are going to homotopically approximate $f$ by a map $g:{\mathbb I}^n\to X\times[-1,1]$ with $(x,0)\notin g({\mathbb I}^n)$. Let $f=(f_1,f_2)$ where $f_1:{\mathbb I}^n\to X$ and $f_2:{\mathbb I}^n\to[-1,1]$ are the components of $f$. Since $X\in
AP(n,0)$, there is a 0-dimensional map $g_1:{\mathbb I}^n\to X$ that approximates $f_1$. Because $Z=g^{-1}_1(x)$ is a $0$-dimensional subset of ${\mathbb I}^n$ and $\{0\}$ is a $Z_0$-set in $[-1,1]$, the map $f_2$ can be approximated by a map $g_2:{\mathbb I}^n\to[-1,1]$ such that $0\notin f_2(Z)$. Then for the map $g=(g_1,g_2):{\mathbb I}^n\to
X\times[-1,1]$ we have $(x,0)\notin g({\mathbb I}^n)$.
\[p1\] If $X\in AP(n,0)$ and $Y\in \infty\mbox{-}\overline{DD}^{\{0,0\}}$, then $X\times Y\in \overline{DD}^{\{n,n\}}$.
Let $f=(f_X,f_Y):{\mathbb I}^n\oplus{\mathbb I}^n\to X\times Y$ be a given map. Since $X\in AP(n,0)$, we homotopically approximate the first component $f_X$ by a 0-dimensional map $g_X:{\mathbb I}^n\oplus{\mathbb I}^n\to X$. Next, use $\infty\mbox{-}\overline{DD}^{\{0,0\}}$-property of $Y$ to homotopically approximate the second component $f_Y$ by a map $g_Y:{\mathbb I}^n\oplus{\mathbb I}^n\to Y$ that is injective on the fibers of $g_X$. Then the map $g=(g_X,g_Y):{\mathbb I}^n\oplus{\mathbb I}^n\to X\times Y$ is injective, witnessing that $X\times Y$ has the $\overline{DD}^{\{n,n\}}$-property.
We say that a space $X$ has the $AP(\infty,0)$-property if $X\in
AP(n,0)$ for every $n\geq 1$.
\[c1\] If $X,Y$ are two locally path-connected spaces both possessing the $AP(\infty,0)$-property and $Y$ contains a dense set of homotopical $Z_2$-points, then $X\times Y$ has the $\overline{DD}^{\{\infty,\infty\}}$-property.
By Theorem \[t1\], each point of $Y$ is a homological $Z_\infty$-point. Taking into account that $Y$ has a dense set of homotopical $Z_2$-points and applying Theorem 26(4,5) of [@bv], we conclude that the space $Y\in\infty\mbox{-}\overline{DD}^{\{0,0\}}$. Finally, by Proposition \[p1\], $X\times Y\in
\overline{DD}^{\{\infty,\infty\}}$.
\[c2\] If $X,Y$ are two compact $AR$’s both possessing the $AP(\infty,0)$-property, then $X\times Y\times{\mathbb I}$ is homeomorphic to the Hilbert cube.
Theorem \[t1\] implies that every $y\in Y$ is a homological $Z_\infty$-point in $Y$. Then, according to [@bv Theorem 26(5)], $Y\in\overline{DD}^{\{1,1\}}$. Consequently, by [@bv Theorem 10(5)], each point of $Y$ is a homotopical $Z_1$-point in $Y$. Now, applying the Multiplication Formula for homological $Z$-sets [@bv Theorem 17(1)]), we conclude that each point of $Y\times{\mathbb I}$ is a homotopical $Z_2$-point. Therefore, Corollary \[c1\] yields that the product $X\times (Y\times{\mathbb I})$ has the $\overline{DD}^{\{\infty,\infty\}}$-property. Being a compact $AR$, this product is homeomorphic to the Hilbert cube according to the Toruńczyk characterization theorem of $Q$-manifolds.
In light of the preceding corollary, it is interesting to remark that there exists a compact space $X\in AR\cap AP(\infty,0)$ which is not homeomorphic to the Hilbert cube. To present such an example, we first prove that the $AP(n,0)$-property is preserved by a special type of maps.
A map $\pi:X\to Y$ is called an elementary cell-like map if:
1. $\pi$ is a fine homotopy equivalence, i.e., for every open cover ${\mathcal U}$ of $Y$ there exists a map $s:Y\to X$ with $\pi\circ s$ being ${\mathcal U}$-homotopic to the identity of $Y$;
2. the non-degeneracy set $N_\pi=\{y\in Y:|\pi^{-1}(y)|\ne 1\}$ is at most countable;
3. each fiber $\pi^{-1}(y)$, $y\in N_\pi$, is an arc.
Let $\pi:X\to Y$ be an elementary cell-like map between metric spaces such that $X$ is complete and every $x\in \pi^{-1}(N_\pi)$ is a homotopical $Z_n$-point in $X$. Then $Y\in AP(n,0)$ provided $X\in AP(n,0)$.
Suppose $X\in AP(n,0)$. To prove that $Y\in AP(n,0)$, fix an open cover $\gamma$ of $Y$ and a map $f:{\mathbb I}^n\to Y$. Let $\gamma_1$ be an open cover of $Y$ which is star-refinement of $\gamma$. Since $\pi$ is a fine homotopy equivalence, there is a map $s:Y\to X$ such that $\pi\circ s$ is $\gamma_1$-homotopic to the identity map of $Y$. Let $D=\{x_i:i\geq 1\}\subset f^{-1}(N_\pi)$ be a sequence such that $f^{-1}(y)\backslash D$ is $0$-dimensional for every $y\in N_f$. Then the sets $W_i=\{g\in C({\mathbb I}^n,X): x_i\notin g({\mathbb I}^n)\}$ are open in $C({\mathbb I}^n,X)$ equipped with the uniform convergence topology. Moreover, each $W_i$ is homotopically dense in $C({\mathbb I}^n,X)$ because $x_i$ is a homotopical $Z_n$-point in $X$. We also consider the sets $U_i=\{g\in C({\mathbb I}^n,X): d_0(g^{-1}(x))<1/i\hbox{~}\mbox{for
all}\hbox{~}x\in X\}$. Since $X\in AP(n,0)$ and all maps from $C({\mathbb I}^n,X)$ are simplicially factorizable, it follows from Proposition 2.3 that any $U_i$ is open and homotopically dense in $C({\mathbb I}^n,X)$. This easily implies that the intersections $V_i=W_i\cap
U_i$ are open and homotopically dense in $C({\mathbb I}^n,X)$. Consequently, by Lemma 2.2 (with $G$ being $C({\mathbb I}^n,X)$), there exists a map $g\in\bigcap_{i=1}^{\infty}V_i$ which is $\pi^{-1}(\gamma_1)$-homotopic to $s\circ f$. Obviously, $g$ is $0$-dimensional, $g({\mathbb I}^n)\subset X\backslash D$ and $f_1=\pi\circ
g:I^n\to Y$ is $\gamma_1$-homotopic to $\pi\circ s\circ f$. Then $f_1$ is $\gamma$-homotopic to $f$ because $\pi\circ s\circ f$ is $\gamma_1$-homotopic to $f$. Hence, $Y\in AP(n,0)$.
Singh [@sin] constructed an elementary cell-like map $f:Q\to X$ from the Hilbert cube $Q$ onto a compact $X\in AR$ such that $X\times{\mathbb I}$ is homeomorphic to $Q$ but $X$ contains no proper $ANR$-subspace of dimension $\ge 2$. By the preceding proposition, Singh’s space has the $AP(\infty,0)$-property. Thus we have:
There is a compact $X\in AR$ with $X\in AP(\infty,0)$ such that $X\times{\mathbb I}$ is homeomorphic to the Hilbert cube but $X$ contains no proper $ANR$-subspace of dimension $\ge 2$.
Now, we consider the spaces with piecewise embedding dimension $n$. According to [@km], a map $h\colon P\to M$ from a finite polyhedron $P$ is said to be a piecewise embedding if there is a triangulation $T$ of $P$ such that $h$ embeds each simplex $\sigma\in T$ and $h(P)$ is an $ANR$. For a space $M$, the piecewise embedding dimension $ped(M)$ is the maximum $k$ such that for any ${\varepsilon}>0$ and any map $g\colon P\to M$ from a finite polyhedron $P$ with $\dim P\leq k$ there exists a piecewise embedding $g'\colon
P\to M$ which is ${\varepsilon}$-close to $g$. If $y\in M$, then $ped_y(M)$ is the maximum of all $ped(U)$, where $U$ is a neighborhood of $y$ in $M$. Obviously, $ped(M)\leq\min\{ped_y(M):y\in m\}$.
As we noted, by [@km Propsition 2.1], every complete $ANR$-space $M$ with $ped(M)\geq n$ has the $AP(n,0)$-property. But there are even compact $ANR$’s having the $AP(n,0)$-property with $ped\leq n-1$. For example, according to Corollary 6.3, any product $M=\prod_{i=1}^{i=n}M_i$ of dendrites with dense set of endpoints is an $AP(n,0)$. On the other hand, by [@ys Theorem 3.4], $ped(M)\leq n-1$. Next proposition also shows that the property $ped=n$ is quite restrictive.
If $(M,\varrho)$ is a complete space and $ped_y(M)=n+1$, then $y$ is a $Z_{n}$-point in $M$.
Since $ped_y(M)=n+1$, there exists a neighborhood $U_y$ in $M$ with $ped(U_y)=n+1$. It is easily seen that if $y$ is a $Z_{n}$-point in $U_y$, it is also a $Z_{n}$-point in $M$. So, we can suppose that $U_y=M$, and let ${\varepsilon}>0$ and $g\in C({\mathbb I}^{n},M)$. We identify ${\mathbb I}^n$ with the set $\{(x_1,..,x_{n+1})\in{\mathbb I}^{n+1}:x_{n+1}=0\}$ and let $\pi:{\mathbb I}^{n+1}\to{\mathbb I}^n$ be the projection. Since $ped(M)=n+1$, there exists a triangulation $T$ of ${\mathbb I}^{n+1}$ and a map $g_1\in
C({\mathbb I}^{n+1},M)$ which is ${\varepsilon}/2$-close to $g\circ\pi$ and $g_1|\sigma$ is an embedding for all $\sigma\in T$. Hence, $g_1^{-1}(y)$ consists of finitely many points. Moreover, there is $\delta>0$ such that for any $\delta$-close points $x',x''$ in ${\mathbb I}^{n+1}$ we have $\varrho(g_1(x'),g_1(x''))<{\varepsilon}/2$. Since each $x\in{\mathbb I}^{n+1}$ is a $Z_n$-point in ${\mathbb I}^{n+1}$, $g_1^{-1}(y)$ is a $Z_n$-set in ${\mathbb I}^{n+1}$. Hence, there exists a map $h\colon{\mathbb I}^n\to{\mathbb I}^{n+1}$ such that $h$ is $\delta$-close to $\displaystyle id_{{\mathbb I}^{n}}$ and $h({\mathbb I}^n)\subset{\mathbb I}^{n+1}\backslash g_1^{-1}(y)$. Then $g_1\circ h\in
C({\mathbb I}^n,M)$ is ${\varepsilon}$-close to $g$ and $y\not\in g_1\circ h({\mathbb I}^n)$.
Finally, let us complete the paper with the following question:
Let $X\in AR$ be a compact space such that $X\in\overline{DD}^{\{2,2\}}\cap AP(\infty,0)$. Is $X$ homeomorphic to the Hilbert cube?
[999]{}
P. Alexandroff, *Die Kontinua $(V^p)$ - eine Verscharfung der Cantorschen Mannigfaltigkeiten*, Monasth. Math. **61** (1957), 67–76.
T. Banakh and V. Valov, *Parametric general position properties and embedding of $n$-dimensional maps into trivial bundles*, preprint.
T. Banakh, R. Cauty and A. Karassev, *On homotopical and homological $Z_n$-sets*, preprint.
M. Bestvina, *Characterizing $k$-dimensional universal Menger compacta*, Mem. Amer. Math. Soc. **380(71)** (1988).
R. Cauty, *Convexité topologique et prolongement des fonctions continues*, Compos. Math. **27** (1973), 233–273.
T. Dobrowolski and W. Marciszewski, *Rays and the fixed point property in noncompact spaces*, Tsukuba J. Math. **21** (1997), no. 1, 97–112.
A. Dranishnikov, *Absolute extensors in dimension $n$ and $n$-soft mappings raising dimension*, Uspekhi Mat. Nauk **39** (1984), 55–95 (in Russian).
R. Engelking, *Theory of dimensions: Finite and Infinite*, Heldermann Verlag, Lemgo (1995).
N. Hadžiivanov, *An $n$-dimensional cube cannot be decomposed into a countable union of proper closed subsets whose pairwise intersections are of dimension no greater than $n-2$*, Dokl. Akad. Nauk SSSSR **195** (1970), 43–45 (Russian). English translation: Soviet Math. Dokl. **11** (1970), 1428–1430.
N. Hadžiivanov and V. Todorov, *On non-Euclidean manifolds*, C.R. Acad. Bulgare Sci. **33, 4** (1980), 449–452 (Russian).
Y. Hattori and K. Yamada, *Closed pre-images of $C$-spaces*, Math. Japon. **34, 4** (1989), 555–561.
H. Kato and E. Matsuhashi, *Lelek maps and $n$-dimensional maps from compacta to polyhedra*, Topology and Appl. **153, 8** (2006), 1241–1248.
M. Katetov, *On the dimension of non-separable spaces. I*, Czech. Math. Journ. **2** (1952), 333–368 (in Russian).
N. Krikorian, *A note concerning the fine topology on function spaces*, Compos. Math. **21** (1969), 343–348.
A. Lelek, *On mappings that change dimension of spheres*, Coloq. Math. **10** (1963), 45–48.
M. Levin, *Bing maps and finite-dimensional maps*, Fund. Math. **151, 1** (1996), 47–52.
M. Levin, *Certain finite-dimensional maps and their application to hyperspaces*, Israel J. Math. **105** (1998), 257–262.
E. Michael, *Local properties of topological spaces*, Duke Math. J. **21** (1954), 163–171.
E. Michael, *Continuous selections I*, Ann. of Math. **63** (1956), 361–382.
J. Munkers, *Topology* (Prentice Hall, Englewood Cliffs, NY, 1975).
B. Pasynkov, *On geometry of continuous maps of finite-dimensional compact metric spaces*, Trudy Mat. Inst. Steklova **212** (1996), 147–172 (Russian); English translation: Proc. Steklov Inst. Math. **212, 1** (1996), 138–162.
D. Repovš and P. Semenov, *Continuous selections of multivalued mappings*, Math. and its Appl. **455**, Kluwer, Dordrecht (1998).
S. Singh, *Exotic $ANR$s via null decompositions of Hilbert cube manifolds*, Fund. Math. **125** (1985), 175–183.
O. Sipachëva, *On a class of free locally convex spaces*, Mat. Sb. **194, 3** (2003), 25–52 (Russian); English translation: Sb. Math. **194, 3-4** (2003), 333–360.
Y. Sternfeld, *Mappings in dendrites and dimension*, Houston J. Math. **19** (1993), 483–497.
H. Torunczyk, *Finite-to-one restrictions of continuous functions*, Fund. Math. **125** (1985), 237–249.
M. Tuncali and V. Valov, *On dimensionally restricted maps*, Fund. Math. **175, 1**, (2002), 35–52.
[^1]: The second author was partially supported by NSERC Grant 261914-03.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
In this paper, I give a pedagogical discussion of the GSI anomaly. Using two different formulations, namely the intuitive Quantum Field Theory language of the second quantized picture as well as the language of amplitudes, I clear up the analogies and differences between the GSI anomaly and other processes (the Double Slit experiment using photons, $e^+ e^- \rightarrow \mu^+ \mu^-$ scattering, and charged pion decay). In both formulations, the conclusion is reached that the decay rate measured at GSI cannot oscillate if only Standard Model physics is involved and the initial hydrogen-like ion is no coherent superposition of more than one state (in case there is no new, yet unknown, mechanism at work). Furthermore, a discussion of the Quantum Beat phenomenon will be given, which is often assumed to be able to cause the observed oscillations. This is, however, not possible for a splitting in the final state only.\
Keywords: GSI anomaly, neutrino oscillations, quantum theory\
PACS: 14.60.Pq, 23.40.-s
author:
- Alexander Merle
bibliography:
- './GSI-DSlit-refs.bib'
title: 'Why a splitting in the final state cannot explain the GSI-Oscillations'
---
\[sec:intro\] Introduction
==========================
In the last months, a measurement of the lifetime of several highly charged ions with respect to electron capture (EC) decays at GSI Darmstadt [@Litvinov:2008rk] has caused a lot of discussion: Instead of seeing only the exponential decay law, a superimposed oscillation has been observed. The cause of this phenomenon, often referred to as [*Darmstadt oscillations*]{} or [*GSI anomaly*]{}, is not yet clear and a huge debate arose whether it could be related to neutrino mixing [@Ivanov:2008sd; @Ivanov:2008nb; @Ivanov:2008zzc; @Faber:2008tu; @Lipkin:2008ai; @Ivanov:2008xw; @Lipkin:2008in; @Kleinert:2008ps; @Walker:2008zzb; @Lipkin:2009zy; @Ivanov:2009rc; @Ivanov:2009en; @Ivanov:2009kt; @Ivanov:2009ku; @Lipkin:2009ge], or not [@Kienert:2008nz; @Giunti:2008ex; @Giunti:2008im; @Giunti:2008eb; @Burkhardt:2008ek; @Peshkin:2008vk; @Peshkin:2008qz; @Gal:2008sw; @Cohen:2008qb; @Giunti:2008db; @Giunti:2009ds; @Flambaum:2009di]. Alternative attempts for an explanation involve spin-rotation coupling [@Lambiase:2008ki; @Pavlichenkov:2008tm; @Faber:2009mg; @Faestermann:2009tj], the interference of the final states [@Isakov:2009yr], or hyperfine excitation [@Winckler:2009jm]. From the experimental side, two test-experiments (with, however, different systematics [@Litvinov:2008hf]) have been performed [@Vetter:2008ne; @Faestermann:2008jt].\
Several times in this discussion, the analogy of the GSI-experiment to the famous historical Double Slit experiment using, e.g., photons [@Young] has been drawn [@Lipkin:2008ai; @Giunti:2008im; @Giunti:2008db], which also led to lively debates at several meetings [@meetings]. In this article, I show that the intuitive Quantum Field Theory (QFT) formulation of the problem always leads to the correct result. As there is still a lot of discussion in part of the community, it might be useful to give one more detailed explanation of the Quantum Mechanics (QM) involved. This can be done best by presenting easy and familiar examples that are not necessarily directly related to the GSI anomaly but do involve the same logical steps and are not under dispute. To do this, I start with the superposition principle and discuss the Double Slit experiment with photons, $e^+ e^- \rightarrow \mu^+ \mu^-$ scattering, and the experiment performed at GSI. Afterwards, the language of amplitudes is used to further justify the QFT-treatment by carefully considering several cases, where $\pi^+$ decay serves as additional example before the considerations are applied to the GSI-experiment, too. Furthermore, the so-called [*Quantum Beats*]{} [@Chow:1975zz] will be discussed, a well-known phenomenon that could indeed cause oscillatory decay rates. This has been argued to cause the observation at GSI in several places (see, e.g., Refs. [@Ivanov:2008sd; @Ivanov:2009rc; @Ivanov:2009en]). It will, however, be shown that this cannot cause the observed behavior if a splitting is only present in the final state.\
In the course of the paper, we will see that all three languages naturally lead to the same result, namely that a splitting in the final state cannot explain the GSI anomaly. Depending on the field of the reader, one or the other part may be clearer, but in the end it turns out that the intuitive QFT picture is correct and in perfect agreement with the results obtained using probability amplitudes or the language of Quantum Beats, which are just different formulations of the same basic principles.
\[sec:disctinction\] The Quantum Field Theory formulation of the problem
========================================================================
Category Double Slit $e^+ e^-\rightarrow \mu^+ \mu^-$ GSI experiment
---------- ----------------------------- ---------------------------------- ----------------------------------------------------
1 No slit-monitoring at all $e^+ e^-$-collider N/A
2A Monitoring & read out N/A GSI-like experiment with more kinematical accuracy
2B Monitoring without read out N/A Actual GSI-experiment
The starting point for the discussion is the superposition principle in QM. One common formulation is [@Peskin:1995ev]: “When a process can happen in alternative ways, we [*add*]{} the amplitudes for each possible way.” The problem in the interpretation arises in the term “alternative ways”, because it is not a priori clear what the word “way” actually means, as well as in the word “process”, which exhibits similar ambiguities.\
Let us use the following terminology: [*Process*]{} means a reaction with a well-defined initial and final state, whereas [*way*]{} is a particular intermediate state of a process. E.g., the scattering reaction $e^+ e^- \rightarrow \mu^+ \mu^-$ is one single process, no matter by which way ($\gamma$-, $Z^0$-, or $H^0$-exchange at tree-level in the Standard Model (SM) of elementary particles) it is mediated. $Z^0 \rightarrow \nu_e \overline{\nu}_e$ and $Z^0 \rightarrow \nu_\mu \overline{\nu}_\mu$ are, however, two distinct processes.\
Using this terminology, the superposition principle can be formulated in the following way:
1. If different ways lead from the same initial to the same final state in one particular process, then one has to add the respective partial amplitudes to obtain the total amplitude. The absolute square of this total amplitude is then proportional to the probability of the process to happen ([*coherent summation*]{}).
2. If a reaction leads to physically distinct final states, then one has to add the probabilities for the different processes ([*incoherent summation*]{}).
If a certain situation belongs to category 1, an interference pattern will be visible (or oscillations, in case the interfering terms have different phases as functions of time), while if it belongs to category 2, there will be no interference. The remaining question is at which point the measurement comes in. This can be trivially said for point 2: Either the experimental apparatus is sufficiently good to distinguish between the different final states (2A) – then no summation whatsoever is necessary simply because one can divide the data set into two (or more), one for each of the different final states. If this is not the case (2B), the experiment will be able to lead to either of the final states, but one would not know which one had been the actual result – then one would simply have to add the probabilities for the different final states to occur in order to obtain the total probability.\
What if we do such a measurement for category 1? If we can indeed distinguish several ways that a process can happen, then this has to be done by some measurement. Since this measurement then has selected one particular way, we have actually transformed a situation belonging to category 1 into a situation of category 2. However, then there would be no terms to interfere with – the interference would have been “killed”.\
Let us now turn to Table \[tab:class\], that illustrates how our three examples fit into the categories 1, 2A, and 2B. These three cases will be discussed one by one in the following.
\[sec:dslit\] The Double Slit experiment with photons
-----------------------------------------------------
This is the “classical” situation of an experiment that reveals the nature of QM. It has first been performed by Thomas Young [@Young] and has later been the major example to illustrate the laws of QM. Its basic procedure is the following: Light emitted coherently by some source (e.g. a laser) hits a wall with two slits, both with widths comparable to the wavelength of the light. If it hits a screen behind the wall, one will observe an interference pattern, as characteristic for wave-like objects (category 1). There is, however, the interpretation of light as photons, i.e., quanta of a well-defined energy. Naturally, one could ask which path such a photon has taken, meaning through which of the two slits it has travelled. The amazing observation is that, as soon as one can resolve this by monitoring the slits accurately enough, the interference pattern will vanish, no matter if one actually reads out the information of the monitoring (2A), or not (2B). The reason is that, regardless of using the information or not, the measurement itself has disturbed the QM process in a way that the interference pattern is destroyed [@feynman].\
The key point is that one cannot even say that the photon takes only one way: In the QM-formulation, amplitudes are added (and not probabilities), and hence the photon does not take one way or the other (and we simply sum over the results), but it rather has a total amplitude that includes a partial amplitude to take way 1 as well as another partial amplitude way 2. By taking the absolute square of this sum of amplitudes, interference terms appear.\
A QFT-formulation involving elementary fields only would be much more complicated: One would sum over the amplitudes for the photon to interact with each electron and each quark in the matter the slits are made of, after having propagated to this particular particle and before further propagating to a certain point on the screen. Of course, by using an effective formulation of the theory, one can find a much more economical description and the easiest one is to simply comprise all possible interactions into two amplitudes, one for going through the first and one for going through the second slit.\
Let us go back to this effective formulation: If there is monitoring, one actually “kills” one of these two amplitudes, the other one remains, and the interference is destroyed. Whenever there has been such a measurement, the interference will vanishe. As we will see, the question is [*if*]{} in a certain situation a measurement has been performed (or is implicitly included in the process considered), no matter if the corresponding information is read out, or not.
\[sec:eemumu\] $e^+ e^- \rightarrow \mu^+ \mu^-$ scattering at a collider
-------------------------------------------------------------------------
{height="2.3cm"}
Let us now consider the scattering of $e^+ e^-$ to form a pair of muons. This is, differently from the Double Slit experiment, a fundamental process where only a small number of elementary particles is involved. If one wants to calculate the scattering probability, the amplitude for the process is again decisive. In the SM, there are only three possibilities for this process to happen at tree-level and in all three of them the $e^+ e^-$ pair annihilates to some intermediate (virtual) boson which in the end decays again, but this time into a $\mu^+ \mu^-$ pair. The intermediate particle can either be a photon, a $Z^0$-boson, or a Higgs scalar, see Fig. \[fig:eemm\].\
Here, we have three different ways to form the process. The difference to the Double Slit experiment, however, is that these three ways cannot be separated easily. In a real collider-experiment we are not able to say that the reaction $e^+ e^- \rightarrow \mu^+ \mu^-$ has taken place by the exchange of, e.g., a photon only, but it will always be the sum of the three diagrams (and a lot more, in case we include higher orders). Hence, this process will always fall into category 1 and interference terms will appear.\
This is an easy and familiar example for the appearance of interference terms in a real experiment. In the next paragraph it will be shown exactly what is different in the case of the GSI-experiment.
\[sec:GSI\] The GSI experiment
------------------------------
![The Feynman diagram for the GSI experiment.[]{data-label="fig:feyn-GSI"}](f2.eps){height="3cm"}
The remaining question is what the situation looks like for the GSI-experiment. Even though the QFT-calculation of what happens is pretty straightforward, fitting everything in the language used above might be a bit more subtle. We will, however, see in Sec. 3.2 that the formulation in terms of amplitudes additionally justifies the result obtained here.\
Let us at first consider the Feynman diagram of the process involved in the GSI-experiment in Fig. \[fig:feyn-GSI\] [@Kienert:2008nz]: Here, in the absence of extreme kinematics (meaning that the $Q$-value of the reaction is large enough so that all neutrino mass eigenstates can be emitted), the neutrino is produced as electron neutrino. What happens to this neutrino? Since it is not detected, it escapes to infinity in the view of QFT (in the picture of second quantization). Physically, it loses its coherence after some propagation distance and travels as a unique mass eigenstate.\
The key point is the following: Since the neutrino will not interact before it loses its coherence, it must be asymptotically a mass eigenstate. This can be shown easily: The coherence length of a (relativistic) neutrino is given by [@Giunti:1991sx] $$L_{\rm coh}=2\sqrt{2} \sigma_x \cdot \frac{2p^2}{(\Delta m^2)_\odot},
\label{eq:coh-length}$$ where $\sigma_x$ is the size of the neutrino wave packet, $p$ is the momentum of the mean value neutrino momentum in the limit $m_\nu=0$, and $(\Delta m^2)_\odot=7.67\cdot 10^{-5}~{\rm eV}^2$ [@GonzalezGarcia:2007ib] is the solar neutrino mass square difference as known from neutrino oscillation experiments. The question is how to obtain an estimate for $\sigma_x$: If the nucleus was inside a lattice, one could estimate a width like the typical interatomic distance, $\sigma_x\sim 1$Å, which would lead to $L_{\rm coh}\sim 2\cdot 10^8$ m. Of course, this precision cannot be reached in the GSI experiment. However, at least during the electron cooling [@Steck], the nucleus will be localized to some precision. Since the velocity of the nucleus is known, this information could in principle be extrapolated for each run. A fair estimate would then be the average distance between two electrons in the cooling process, which is roughly given by $1/\sqrt[3]{n}\sim 0.1$mm, where $n$ is the electron density [@Bosch_ESR]. This leads to a more realistic coherence length of $L_{\rm coh}\sim 2\cdot 10^{14}$ m. The pessimistic case, where $\sigma_x$ is taken to be the approximate diameter $108.36~{\rm m}/\pi$ [@ESR] of the Experimental Storage Ring (ESR) produces $L_{\rm coh}\sim 6\cdot 10^{19}$ m. The mean free path of a neutrino in our galaxy, however, is roughly $1\cdot 10^{40}$ m (for an assumed matter density in the Milky Way of $1\cdot 10^{-23}~{\rm g/cm}^3$), so the assumption that the neutrino does not interact before losing its coherence is completely safe.\
Even if we do not know in which of the three mass eigenstates the neutrino actually is, we know that it has to be in one of them. This knowledge is somehow obtained “a posteriori”, since the mass eigenstate only reveals its identity after some propagation. But, by conservation of energy and momentum, one could treat the process as if the kinematical selection had already been present at the production point of the neutrino. This “measurement” is enforced by the physical conservation laws.\
An analogous reasoning is given by Feynman and Hibbs [@feynman], using the example of neutron scattering: Neutrons prepared to have all spin up scatter on a crystal. If one of the scattered neutrons turns out to have spin down, one knows by angular momentum conservation that it must have been scattered by a certain nucleus. In principle, by noting down the spin state of every nucleus in the crystal before and after the measurement, one could find the corresponding scattering partner of the neutron without disturbing it. No matter if this would be difficult practically, by a physical conservation law one knows that a particular scattering must have been present, even if the corresponding nucleus is not “read out”. Accordingly, the corresponding interference vanishes and the neutrons that have spin down after the scattering come out diffusely in all directions.\
This can also be formulated in the language of wave packets: We have complete 4-momentum conservation for each single component (which is a plane wave!) of the wave packets, but if we consider the whole wave packet, its central momentum does not have to be conserved [@Beuthe:2001rc; @Giunti:2003ax]. However, all the different components can produce both possible neutrino mass eigenstates, but for a certain kinematical configuration of parent and daughter components only one of the mass eigenstates will actually be produced.\
The rest is easy: If the GSI experiment had infinite kinematical precision, one could read out which of the mass eigenstates has been produced and it would clearly fall into category 2A. Since, however, this information is not read out but could in principle have been obtained (e.g. by detecting the escaping neutrino), the GSI experiment falls into category 2B and one has to sum over probabilities. This logic works because we know that the neutrino is, after some propagation, no superposition of mass eigenstates anymore, but just one particular eigenstate with a completely fixed mass.\
A viewpoint closer to the amplitude formulation would be: If the neutrino finally interacts, it has to “decide” which mass eigenstate it has, even if it was a superposition of several mass eigenstates before. This is then equivalent to the image of having produced one particular mass eigenstate from the beginning on.
\[sec:amps\] Amplitudes - probably the easiest language to use
==============================================================
In this section, I use time-dependent amplitudes for the different basis states to describe another example, namely charged pion decay, which I compare then to neutrino oscillations (with referring to the actual situation in the GSI-experiment). The logical steps needed to understand the familiar example of pion decay are exactly the same as the ones needed to understand what is going on at GSI. This description is clear enough to account for very different situations and allows for an easy and nearly intuitive understanding of the various cases. Furthermore, it yields an a posteriori justification of the view used in the preceding section.
\[sec:pion\] Charged pion decay
-------------------------------
It is well-known that a charged pion (e.g. $\pi^+$) can decay into either a positron in combination with an electron neutrino, or into the corresponding pair of $\mu$-like particles. Let us consider the case of a pure (and normalized) initial state pion $|\pi^+\rangle$. As this state evolves with time (and is not monitored), it will become a coherent superposition of the parent-state, as well as all possible daughter states: $$|\pi^+ (t)\rangle=\mathcal{A}_\pi (t) |\pi^+\rangle+ \mathcal{A}_\mu (t) |\mu^+ \nu_\mu\rangle+ \mathcal{A}_e (t)|e^+ \nu_e\rangle,
\label{eq:pion_1}$$ where all time-dependence is inside the partial amplitudes $\mathcal{A}_i$. Of course, this state has to be normalized correctly: $$|\mathcal{A}_\pi (t)|^2 + |\mathcal{A}_\mu (t)|^2 + |\mathcal{A}_e (t)|^2=1,
\label{eq:pion_2}$$ with $\mathcal{A}_\pi (0)=1$ and $\mathcal{A}_\mu (0)=\mathcal{A}_e (0)=0$. One can understand Eq. in the following way: The state at time $t$ is a coherent superposition of the basis states $\{ |\pi^+\rangle, |\mu^+ \nu_\mu\rangle, |e^+ \nu_e\rangle \}$ with time-dependent coefficients. Note that the basis states are orthogonal. The outcome of a certain measurement is some state $|\Psi\rangle$: All that a detector does is projecting on just this state $|\Psi\rangle$. Of course, different detectors will in general be described by projections on different $|\Psi\rangle$’s, which is a reflection of the influence of the process of measurement on the measurement itself. If one wants to know the probability for measuring that particular state, one has to calculate it according to the standard formula, $$P(\Psi)=|\langle \Psi|\pi^+ (t) \rangle|^2.
\label{eq:pion_3}$$ The question is what $|\Psi\rangle$ looks like. To make that clear, let us discuss several cases:
- The (trivial) case is that there has been no detection at all: Then we have gained no information. This means that the projected state is just the time-evolved state itself (we do not know anything except for the time passed since the experiment has started), and we get $$|\langle \Psi|\pi^+ (t) \rangle|^2=|\langle \pi^+ (t)|\pi^+ (t) \rangle|^2=1.
\label{eq:pion_4}$$ This result is trivial, since the probability for anything to happen must be equal to 1.
- The next situation is when our experimental apparatus can give us the information that the pion has decayed, but we do not know the final state. Then, it can be either $|\mu^+ \nu_\mu\rangle$ or $|e^+ \nu_e\rangle$ and we remain with a superposition of these two states. The only information that we have gained is that the amplitude for the initial pion to be still there is now zero, $\mathcal{A}_\pi=0$ in Eq. . Then, the properly normalized state $|\Psi\rangle$ is $$| \Psi\rangle = \frac{\mathcal{A}_\mu (t) |\mu^+ \nu_\mu\rangle+ \mathcal{A}_e (t)|e^+ \nu_e\rangle}{\sqrt{|\mathcal{A}_\mu (t)|^2+|\mathcal{A}_e (t)|^2}}.
\label{eq:pion_5}$$ The absolute value square of the corresponding projection is $$|\langle \Psi|\pi^+ (t) \rangle|^2=|\mathcal{A}_\mu (t)|^2+|\mathcal{A}_e (t)|^2,
\label{eq:pion_6}$$ and if there is any oscillatory phase in the amplitudes, $\mathcal{A}_k (t)= \mathcal{\tilde A}_k (t) e^{i\omega_k t}$, it will have no effect due to the absolute values.
- What if we know that the initial pion is still present? This sets $\mathcal{A}_\mu (t)=\mathcal{A}_e (t)=0$, and $|\Psi\rangle$ is just $\mathcal{A}_\pi (t) |\pi^+\rangle /\sqrt{|\mathcal{A}_\pi (t)|^2}$. The projection gives $$|\langle \Psi|\pi^+ (t) \rangle|^2=|\mathcal{A}_\pi (t)|^2,
\label{eq:pion_7}$$ which again does not oscillate.
- If one particular final state, let us say $|e^+ \nu_e\rangle$, is detected, then $\mathcal{A}_\pi (t)=\mathcal{A}_\mu (t)=0$ and we get another term free of oscillations: $$|\langle \Psi|\pi^+ (t) \rangle|^2=|\mathcal{A}_e (t)|^2.
\label{eq:pion_8}$$
The question remains when we do get oscillations at all. The answer is: It depends on what our detector measures. If, e.g., the detector measures not exactly the state $|\mu^+ \nu_\mu\rangle$ or $|e^+ \nu_e\rangle$, but instead some (hypothetical) superposition (e.g., some quantum number which is not yet known, under which neither $\mu^+$ nor $e^+$ is an eigenstate, but some superposition of them), then one could measure the following (correctly normalized!) state: $$|\Psi\rangle=\frac{1}{\sqrt{2}} \left(|\mu^+ \nu_\mu\rangle + |e^+ \nu_e\rangle \right).
\label{eq:pion_9}$$ The squared overlap is $$|\langle \Psi|\pi^+ (t) \rangle|^2=\frac{1}{2}\left[ |\mathcal{A}_\mu (t)|^2+|\mathcal{A}_e (t)|^2 + 2 \Re \left( \mathcal{A}_\mu^* (t) \mathcal{A}_e (t) \right) \right],\nonumber
$$ where the $2 \Re \left( \mathcal{A}_\mu^* (t) \mathcal{A}_e (t) \right)$-piece will, in general, lead to oscillatory terms. What has been done differently than before? This time, we have done more than simply killing one or more amplitudes in Eq. , and this is the cause of oscillations: Whenever we are in a situation, in which the state playing the role of $|\Psi\rangle$ in Eq. is physical, the corresponding projection will yield oscillatory terms. As we will see in a moment, this is exactly what happens in neutrino oscillations.
\[sec:neutrino\] Neutrino oscillations and the GSI-experiment
-------------------------------------------------------------
Let us now turn to neutrino oscillations. Here, as we will see, a state like $|\Psi\rangle$ in Eq. can indeed be physical in some situations. To draw a clean analogy to the experiment done at GSI, we consider a hydrogen-like ion as initial state $|M\rangle$ that can decay to the state $|D \nu_e\rangle$ via electron capture. Since there was an electron in the initial state, we know that the amplitude for producing the mass eigenstate $|\nu_i\rangle$ is just $U_{ei}$. If there is no relative phase between the two mass eigenstates, the neutrino produced in the decay is exactly the particle that we call [*electron neutrino*]{}. In any case, due to different kinematics, the two mass eigenstates will in general develop different phases in the time-evolution. This means that, in spite of the mixing matrix elements $U_{ei}$ being time-independent, there will be a phase between the two neutrino mass eigenstates. Completely analogous to Eq. , the time-evolution of the initial state will be given by: $$| M(t)\rangle = \mathcal{A}_M (t) |M\rangle + U_{e1} \mathcal{A}_1 (t) |D \nu_1\rangle+U_{e2} \mathcal{A}_2 (t) |D \nu_2\rangle,
\label{eq:neutrino_1}$$ with $|\mathcal{A}_M (t)|^2+|U_{e1}\mathcal{A}_1 (t)|^2+|U_{e2}\mathcal{A}_2 (t)|^2=1$ and $\mathcal{A}_M (0)=1$. We can immediately look at different cases:
- The parent ion is seen in the experiment: This kills all daughter amplitudes, $\mathcal{A}_{1,2} (t)=0$. The only remaining amplitude is $\mathcal{A}_M (t)$, very similar to Eq. . With the proper normalization for $|\Psi\rangle$ one gets no oscillation again: $$|\langle \Psi | M(t)\rangle |^2 = |\mathcal{A}_M (t)|^2
\label{eq:neutrino_2}$$
- The next case corresponds to the GSI-experiment: One sees only the decay, but cannot tell which of the two neutrino mass eigenstates has been produced. This leads to $\mathcal{A}_M (t)=0$ and one has to perform a projection on the state $$|\Psi \rangle = \frac{U_{e1}\mathcal{A}_1 (t) |D \nu_1\rangle+U_{e2}\mathcal{A}_2 (t)|D \nu_2\rangle}{\sqrt{|U_{e1}\mathcal{A}_1 (t)|^2+|U_{e2}\mathcal{A}_2 (t)|^2}}.
\label{eq:neutrino_3}$$ Doing this with $|M(t)\rangle$ from Eq. yields $$\begin{aligned}
&& |\langle \Psi | M(t)\rangle |^2= \left| \frac{|U_{e1}\mathcal{A}_1 (t)|^2 \cdot 1 +|U_{e2}\mathcal{A}_2 (t)|^2 \cdot 1}{\sqrt{|U_{e1}\mathcal{A}_1 (t)|^2+|U_{e2}\mathcal{A}_2 (t)|^2}} \right|^2=\nonumber\\
&& =|U_{e1}\mathcal{A}_1 (t)|^2+|U_{e2}\mathcal{A}_2 (t)|^2,
\label{eq:neutrino_4}\end{aligned}$$ which exhibits no oscillations, but is rather an incoherent sum over probabilities. This result is the justification of the intuitive treatment in Sec. 2.3: The elementary QM-discussion using probability amplitudes gives us just the correct prescription for how to sum up the amplitudes for the final states.
- The GSI-experiment with infinite kinematical precision: In this case, one could actually distinguish the states $|D \nu_1\rangle$ and $|D \nu_2\rangle$. If one knows that $|D \nu_1\rangle$ is produced (e.g., by having very precise information about the kinematics), one will again have no oscillation, $$|\langle \Psi | M(t)\rangle |^2 = |\mathcal{A}_1 (t) U_{e1}|^2,
\label{eq:neutrino_5}$$ just as in Eq. .
These are in principle all cases that can appear. One can, however, have a closer look at the realistic situation in the GSI-experiment. Let us re-consider Eq. : In reality, the parent ion will be described by a wave packet with a finite size or, equivalently, a finite spreading in momentum space, due to the Heisenberg uncertainty relation. If this wave-packet is broad enough that each component can equivalently decay into $|D \nu_1\rangle$ or $|D \nu_2\rangle$, then both of the corresponding amplitudes will actually have the same phase ($\mathcal{A}_1 (t)=\mathcal{A}_2 (t)$), since they have the same energy, and one can write Eq. as $$| M(t)\rangle = \mathcal{A}_M (t) |M\rangle + \mathcal{A} (t) \underbrace{\left[ U_{e1} |D \nu_1\rangle+U_{e2} |D \nu_2\rangle \right]}_{=|D \nu_e\rangle}.
\label{eq:neutrino_6}$$ Since the knowledge of the momentum of the parent ion is not accurate enough at the GSI-experiment to make a distinction between both final states $|D \nu_k\rangle$, this is a realistic situation. Of course, this does not at all change the above argumentation, since the final state $|\Psi\rangle$ will experience the same modification. The neutrino produced is an electron-neutrino, as to be expected.\
The question remains, why some authors come to the conclusion that there should be oscillations? The answer is simple: If the correspondence between time-evolved initial state and detected state is wrong, then oscillations may appear. As example, we will consider the situation that the kinematics of the parent and daughter are fixed so tightly, that indeed the production amplitudes for $|D \nu_1\rangle$ and $|D \nu_2\rangle$ are not equal. This would correspond to an extremely narrow wave packet in momentum space. Let us, e.g, have in mind the extreme case when by kinematics only the production of $\nu_1$ is possible. This is no problem in principle and we would be used to it if neutrinos had higher masses, so that the $Q$-value of the capture was only sufficient to produce the lightest neutrino mass eigenstates. If only the disappearance of the parent is seen, the corresponding state $|\Psi\rangle$, which is detected, is given by Eq. (with $\mathcal{A}_2 (t)=0$ in the extreme case, but anyway with $\mathcal{A}_1 (t)\neq \mathcal{A}_2 (t)$). The corresponding neutrino is, however, no electron-neutrino anymore (which would be $U_{e1} |\nu_1\rangle+U_{e2} |\nu_2\rangle$, with the same phase for both states)! Indeed this is no surprise at all, since the kinematics in the situation considered is so tight that it changes the neutrino state which is emitted. This is a clear consequence of quantum mechanics, since for obtaining the necessary pre-knowledge (namely the very accurate information about the kinematics), one has to do a measurement that is precise enough to have an impact on the QM state.\
If one consideres the state from Eq. as being the one emitted but then projects onto an electron neutrino state, oscillations will appear: $$\begin{aligned}
&& |\langle D,\nu_e|M(t)\rangle|^2= |(U_{e1}^* \langle D \nu_1 |+U_{e2}^* \langle D \nu_2 |) \cdot \nonumber\\
&& \cdot(\mathcal{A}_M (t) |M\rangle + U_{e1} \mathcal{A}_1 (t) |D \nu_1\rangle+U_{e2} \mathcal{A}_2 (t) |D \nu_2\rangle)|^2=\nonumber\\
&& =|\mathcal{A}_1 (t)|^2+|\mathcal{A}_2 (t)|^2+2\Re (\mathcal{A}_1 (t) \mathcal{A}_2^* (t)).
\label{eq:neutrino_7}\end{aligned}$$ This is, however, wrong: One has not used all the information that could in principle have been obtained! But Nature does not care about if one uses information or not, so this treatment does simply not correspond to what has happened in the actual experiment. The oscillations, however, only arise due to the incorrect projection, and have no physical meaning.\
The remaining question to obtain a complete understanding of the situation is if the neutrino that is emitted in the GSI-experiment oscillates. The answer is yes, of course. But to see that, we will have to modify our formalism a bit. Knowing that an electron neutrino has been emitted corresponds to $\mathcal{A}_M (t)=0$ in Eq. , and the remaining (normalized) state is: $$| \Psi\rangle = \frac{\mathcal{A} (t)}{|\mathcal{A} (t)|}\left[ U_{e1} |D \nu_1\rangle+U_{e2} |D \nu_2\rangle \right].
\label{eq:neutrino_8}$$ Re-phasing this state and measuring the time from $t$ on gives as initial state: $$| \Psi\rangle =U_{e1} |D \nu_1\rangle+U_{e2} |D \nu_2\rangle.
\label{eq:neutrino_9}$$ This is the state which will undergo some evolution in time according to $$| \Psi (t')\rangle =\mathcal{A}'_1(t') U_{e1} |D \nu_1\rangle+\mathcal{A}'_2(t')U_{e2} |D \nu_2\rangle,
\label{eq:neutrino_10}$$ with $|\mathcal{A}'_1(t') U_{e1}|^2+|\mathcal{A}'_2(t') U_{e2}|^2=1$ and $\mathcal{A}'_1(0)=\mathcal{A}'_2(0)=1$. If we ask what happens to this neutrino if it is detected after some macroscopic distance, it is necessary to take into account what has happend to the daughter nucleus that has been produced together with the neutrino, due to entanglement. The daughter nucleus, which is accurately described by a wave packet, is detected, but not with sufficient kinematical accuracy to distinguish the different components $|D\rangle$ of the wave packet. The effect of such a non-measurement is studied most easily in the density matrix formalism. The density matrix $\rho'$ corresponding to Eq. is given by $$\begin{aligned}
&& |\Psi (t')\rangle \langle \Psi(t')| =|\mathcal{B}_1(t')|^2 |D\rangle |\nu_1\rangle \langle \nu_1| \langle D|+\nonumber\\
&& +|\mathcal{B}_2(t')|^2 |D\rangle |\nu_2\rangle \langle \nu_2| \langle D|+\nonumber\\
&&+[\mathcal{B}_1(t') \mathcal{B}^*_2(t') |D\rangle |\nu_1\rangle \langle \nu_2| \langle D|+h.c.],
\label{eq:neutrino_11}\end{aligned}$$ where $\mathcal{B}_k(t')=\mathcal{A}'_k(t') U_{ek}$. If the exact kinematics of the daughter is not measured, then one has to calculate the trace over the corresponding states. It gives $$\begin{aligned}
&&\rho\equiv \int dD \langle D|\rho'|D\rangle=|\mathcal{B}_1(t')|^2 |\nu_1\rangle \langle \nu_1| +\label{eq:neutrino_12}\\
&&+|\mathcal{B}_2(t')|^2 |\nu_2\rangle \langle \nu_2|+ (\mathcal{B}_1(t') \mathcal{B}^*_2(t') |\nu_1\rangle \langle \nu_2| +h.c.).\nonumber\end{aligned}$$ If we want to know the probability to detect, e.g., a muon neutrino, $|\nu_\mu\rangle=U_{\mu 1} |\nu_1\rangle+ U_{\mu 2}|\nu_2\rangle$, the corresponding projection operator is given by $$\mathcal{P}_\mu= |\nu_\mu\rangle \langle \nu_\mu|,
\label{eq:neutrino_13}$$ and the probability to detect this state is $$P_\mu={\rm Tr}(\mathcal{P}_\mu \rho)=\langle \nu_1|\mathcal{P}_\mu \rho|\nu_1\rangle+\langle \nu_2|\mathcal{P}_\mu \rho|\nu_2\rangle.
\label{eq:neutrino_14}$$ Note, however, that the neutrino states $|\nu_{1,2}\rangle$ will always be orthogonal, since they correspond to eigenstates of different masses (like an electron is in that sense orthogonal to a muon). The result is $$\begin{aligned}
&& P_\mu= |U_{\mu 1}|^2 |\mathcal{B}_1(t')|^2 + |U_{\mu 2}|^2 |\mathcal{B}_2(t')|^2+\nonumber\\
&& +[U_{\mu 1} U_{\mu 2}^* \mathcal{B}^*_1(t') \mathcal{B}_2(t') +c.c.],
\label{eq:neutrino_15}\end{aligned}$$ whose second line contains oscillatory contributions. These oscillation are indeed physical: Eq. is a description of a detector that is sensitive to $\nu_\mu$’s only. If it could not distinguish different neutrino flavours, the oscillation would vanish again.
\[sec:beats\] Quantum Beats
===========================
![Type I (left) and type II (right) of the Quantum Beats settings.[]{data-label="fig:QBs"}](QBboth.eps){height="5.2cm"}
The last point to discuss are the so-called Quantum Beats (QBs) [@Chow:1975zz]. This phenomenon is known from Quantum Optics and has often been mentioned as possible explanation for the GSI anomaly. As we will see, the corresponding language can be equally used to describe the GSI-experiment and (of course) yields the same result as already obtained. Still, it is also useful to consider the experiment from this point of view in order not to be misled by claims that erroneously make QBs arising from a splitting in the final state responsible for the observation at GSI.\
Normally, one considers atomic levels for this discussion, and we will stick to that here for illustrative purposes and give the relation to the GSI-experiment at the end of each section. This way also easily clarifies the analogies to the Quantum Optics formulation.
\[sec:beats\_typeI\] Single atom of type I
------------------------------------------
Let us start with the classic example of QBs, namely an atom in a coherent superposition of three states $|a\rangle$, $|b\rangle$, and $|c\rangle$, where the first two states are above and closely spaced compared to $|c\rangle$. This setting is drawn on the left panel of Fig. \[fig:QBs\] and is referred to as “type I”. First note that the three levels correspond to different (but fixed) eigenvalues of the energy and are hence orthogonal vectors in Hilbert space. This is not at all changed by an energy uncertainty which, however, makes it possible to have a coherent superposition of the three states. Initially, we assume the atom to be in such a superposition of these states, but having emitted no photon yet. Accordingly, the photon state can only be the vacuum $|0\rangle_\gamma$. Then, the initial state of this system can be written as $$|\Psi(0)\rangle=\mathcal{A}_0|a\rangle |0\rangle_\gamma+\mathcal{B}_0|b\rangle |0\rangle_\gamma+\mathcal{C}_0|c\rangle |0\rangle_\gamma,
\label{eq:QBs_1}$$ where $|\mathcal{A}_0|^2+|\mathcal{B}_0|^2+|\mathcal{C}_0|^2=1$. If this system undergoes a time-evolution, the lower state might be populated by de-excitation of the upper ones, which is done by photon emission. If the state $|1_x\rangle_\gamma=a^\dagger_x |0\rangle_\gamma$ is assumed to describe a state with one photon of frequency $\om_x$, then the state at time $t$ can be written as $$\begin{aligned}
&& |\Psi(t)\rangle=\mathcal{A}(t)|a\rangle |0\rangle_\gamma+\mathcal{B}(t)|b\rangle |0\rangle_\gamma+\mathcal{C}(t)|c\rangle |0\rangle_\gamma+\nonumber\\
&& + \mathcal{C}_1(t)|c\rangle |1_{ac}\rangle_\gamma + \mathcal{C}_2(t)|c\rangle |1_{bc}\rangle_\gamma,
\label{eq:QBs_2}\end{aligned}$$ where $\mathcal{A}(0)=\mathcal{A}_0$, $\mathcal{B}(0)=\mathcal{B}_0$, $\mathcal{C}(0)=\mathcal{C}_0$, $\mathcal{C}_{1,2}(0)=0$, and $|\mathcal{A}(t)|^2+|\mathcal{B}(t)|^2+|\mathcal{C}(t)|^2+|\mathcal{C}_1(t)|^2+|\mathcal{C}_2(t)|^2=1$. Under the assumption that all levels are equally populated, the radiated intensity will be proportional to $\langle \Psi(t)| {{\mathbf{E}}}^2({{\mathbf{0}}},t) |\Psi(t)\rangle$, where $${{\mathbf{E}}}({{\mathbf{x}}},t)=\sum_{{{\mathbf{k}}},\lambda} \eps_{{{\mathbf{k}}},\lambda} \left( a_{{{\mathbf{k}}},\lambda} e^{-ikx} + a_{{{\mathbf{k}}},\lambda}^\dagger e^{+ikx} \right)
\label{eq:QBs_3}$$ is the electric field operator and $\eps_{{{\mathbf{k}}},\lambda}$ is the electric field per photon of momentum ${{\mathbf{k}}}$ and polarization $\lambda$. Note that the creation and annihilation operators have only one non-trivial commutation relation, namely $[a_{{{\mathbf{k}}},\lambda},a_{{{\mathbf{k}}}',\lambda'}^\dagger]=\delta_{{{\mathbf{k}}},{{\mathbf{k}}}'} \delta_{\lambda,\lambda'}$. In our case we obtain effectively: $$\begin{aligned}
&&{{\mathbf{E}}}({{\mathbf{0}}},t)^2=\eps_{ac}^2 (1+2a_{ac}^\dagger a_{ac})+\label{eq:QBs_4}\\
&&+\eps_{bc}^2 (1+2a_{bc}^\dagger a_{bc})+ 2 \eps_{ac} \eps_{bc} (a_{ac}^\dagger a_{bc} e^{i\Delta t} + a_{bc}^\dagger a_{ac} e^{-i\Delta t}),\nonumber\end{aligned}$$ where $\Delta=\om_{ac}-\om_{bc}$. Here, we have already used that terms like, e.g., $a_{ac}^2$ give no contribution with $|\Psi\rangle$ from Eq. . Remember now, that the atomic states are orthonormal. This means that one can, e.g., combine a term proportional to $\langle b|$ in $\langle \Psi(t)|$ only with the corresponding term $|b\rangle$ in $|\Psi(t)\rangle$. The corresponding combination of amplitudes $|\mathcal{B}(t)|^2$ does, however, not oscillate, since any phase will be killed by the absolute value. This is also true for every term involving one of the constant parts of Eq. : E.g. the term proportional to $\mathcal{C}^*(t) \mathcal{C}_1(t)$ can involve a factor $${_\gamma} \langle 0| a_{ac}^\dagger a_{ac} a_{ac}^\dagger |0\rangle_\gamma= 0,
\label{eq:QBs_5}$$ because of $a_{ac}^\dagger$ acting on the left. There are, however, remaining oscillatory terms such as $\mathcal{C}_1^*(t) \mathcal{C}_2(t) e^{i\Delta t}$, which is proportional to $${_\gamma} \langle 0|a_{ac} a_{ac}^\dagger a_{bc} a_{bc}^\dagger |0\rangle_\gamma= {_\gamma} \langle 0|(1+a_{ac}^\dagger a_{ac}) (1+a_{bc}^\dagger a_{bc}) |0\rangle_\gamma=1.\nonumber
$$ These terms cause the Quantum Beats for a type I atom. Actually, one could have expected this result intuitively: Both of the coherently excited upper levels can decay into [*the same*]{} state $|c\rangle$ via the emission of a photon. Hence, one cannot in any way determine the photon energy without measuring it directly. Without such a measurement, interference terms will appear.
How is the situation for the GSI-experiment? In this case one simply has to replace the photon by the neutrino. As explained in Ref. [@Kienert:2008nz] for instance, a splitting in the initial state could lead to an oscillatory behavior. This splitting, however, would have to be tiny, $\sim 10^{-15}$ eV, a value which can hardly be explained. Furthermore, there exists preliminary data on the lifetimes of $^{142}{\rm Pm}^{60+}$ with respect to $\beta^+$-decay that shows no oscillatory behavior [@Ivanov:2009rc]. An initial splitting in the nucleus would lead to an oscillatory rate in this case, too. Accordingly, if such a splitting is present in the initial state, it could be in the levels of the single bound electron, since this would then affect EC-decays while leaving $\beta^+$-decays untouched.
\[sec:beats\_typeII\] Single atom of type II
--------------------------------------------
We can study a similar setting, namely an atom of type II, shown on the right panel of Fig. \[fig:QBs\]. The corresponding initial state would again be described by Eq. , but its time-evolution would now look like $$\begin{aligned}
&& |\Psi(t)\rangle=\mathcal{A}(t)|a\rangle |0\rangle_\gamma+\mathcal{B}(t)|b\rangle |0\rangle_\gamma+\mathcal{C}(t)|c\rangle |0\rangle_\gamma+\nonumber\\
&& + \mathcal{B}'(t)|b\rangle |1_{ab}\rangle_\gamma + \mathcal{C}'(t)|c\rangle |1_{ac}\rangle_\gamma,
\label{eq:QBs_7}\end{aligned}$$ where $\mathcal{A}(0)=\mathcal{A}_0$, $\mathcal{B}(0)=\mathcal{B}_0$, $\mathcal{C}(0)=\mathcal{C}_0$, $\mathcal{B}'(0)=0$, $\mathcal{C}'(0)=0$, and $|\mathcal{A}(t)|^2+|\mathcal{B}(t)|^2+|\mathcal{C}(t)|^2+|\mathcal{B}'1(t)|^2+|\mathcal{C}'(t)|^2=1$. The square of the electric field has again the form of Eq. , just with $bc \rightarrow ab$. Due to the orthogonality of the atomic states, there are not too many combinations which are possible:
- 0-photon state coupled with itself:\
If we take, e.g., the term $|\mathcal{A}(t)|^2$, it does not oscillate anyway. Hence, only the time-dependent parts in Eq. (with $bc \rightarrow ab$) could lead to oscillations. But they are proportional to $${_\gamma} \langle 0| a_{ac}^\dagger a_{ab} |0\rangle_\gamma={_\gamma} \langle 0| a_{ab}^\dagger a_{ac} |0\rangle_\gamma=0.\nonumber$$
- 1-photon state coupled with itself:\
$|\mathcal{B}'(t)|^2$ does not oscillate, too, and the time-dependent terms from the electric field yield $$\begin{aligned}
&& {_\gamma} \langle 1_{ab}| a_{ac}^\dagger a_{ab} |1_{ab}\rangle_\gamma={_\gamma} \langle 0| a_{ab} a_{ac}^\dagger a_{ab} a_{ab}^\dagger|0\rangle_\gamma=0\ {\rm and}\nonumber\\
&& {_\gamma} \langle 1_{ab}| a_{ab}^\dagger a_{ac} |1_{ab}\rangle_\gamma={_\gamma} \langle 0| a_{ab} a_{ab}^\dagger a_{ac} a_{ab}^\dagger|0\rangle_\gamma=0, \nonumber\end{aligned}$$ which follows immediately from the action of $a_{ac}^\dagger$ to the left and of $a_{ac}$ to the right, respectively.
- 0-photon state coupled with 1-photon state:\
This is the only possibility, which is left. If we take for instance the term $\mathcal{B}^*(t) \mathcal{B}'(t)$, this will oscillate anyway, so we will also have to check the constant terms in Eq. . The ones proportional to $1$ are naturally zero, ${_\gamma} \langle 0|1_{ab}\rangle_\gamma={_\gamma} \langle 0| a_{ab}^\dagger |0\rangle_\gamma=0.$ The other terms are $$\begin{aligned}
&& {_\gamma} \langle 0| \underbrace{a_{ac}^\dagger}_{0\leftarrow} a_{ac} |1_{ab}\rangle_\gamma=0,\ {_\gamma} \langle 0| \underbrace{a_{ab}^\dagger}_{0\leftarrow} a_{ab} |1_{ab}\rangle_\gamma=0,\label{eq:QBs_8}\\
&& {_\gamma} \langle 0| \underbrace{a_{ac}^\dagger}_{0\leftarrow} a_{ab} |1_{ab}\rangle_\gamma=0,\ {\rm and}\ {_\gamma} \langle 0| \underbrace{a_{ab}^\dagger}_{0\leftarrow} a_{ac} |1_{ab}\rangle_\gamma=0,\nonumber\end{aligned}$$ where the action of the operators to give zero is always indicated by the arrow. The argumentation is analogous for the complex conjugated term.
Hence, there can be no Quantum Beats for a single atom of type II! The intuitive reason is that, by waiting long enough, one could reach an accuracy in energy that is good enough to distinguish the possible final states $|b\rangle$ and $|c\rangle$. This would then be a way to determine the energy of the emitted photon without disturbing it.
To give an analogous argument for the GSI-experiment, one has to turn the comparison given in Sec. 4.1 round and replace the [*atom by the neutrino*]{} and the [*photon by the ion*]{}. The reason is that what is claimed to interfere in this situation is the neutrino states themselves (see, e.g., Ref. [@Ivanov:2008sd]). This neutrino is not expected to interact, before losing its coherence (cf. Sec. 2.3). However, once it interacts, it has to decide for a certain mass eigenstate. By monitoring this interaction, it would in principle be no problem to determine the neutrino’s mass (e.g., by exploiting the spatial separation of the mass eigenstates far away from the source) and from this one could easily reconstruct the kinematics of the daughter ion in the GSI-experiment. Accordingly, no QBs are to be expected in this situation.
\[sec:beats\_twotypeII\] Two atoms of type II
---------------------------------------------
On the other hand, there is a situation in which we can expect QBs even for atoms of type II, namely if we have two of them. If these two atoms are separated by a distance which is smaller than the wavelength of the emitted photons, there is no way to resolve their separation in space and we have to write down a combined initial state for both atoms, 1 and 2: $$\begin{aligned}
&& |\Psi(0)\rangle=\mathcal{A}_0 |a\rangle_1 |a\rangle_2 |0\rangle_\gamma+\mathcal{B}_0 |b\rangle_1 |b\rangle_2 |0\rangle_\gamma+\mathcal{C}_0 |c\rangle_1 |c\rangle_2 |0\rangle_\gamma\nonumber\\
&& +\mathcal{D}_{1,0} |a\rangle_1 |b\rangle_2 |0\rangle_\gamma+\mathcal{D}_{2,0} |b\rangle_1 |a\rangle_2 |0\rangle_\gamma+\mathcal{E}_{1,0} |a\rangle_1 |c\rangle_2 |0\rangle_\gamma+\nonumber\\
&& +\mathcal{E}_{2,0} |c\rangle_1 |a\rangle_2 |0\rangle_\gamma+\mathcal{F}_{1,0} |b\rangle_1 |c\rangle_2 |0\rangle_\gamma+\mathcal{F}_{2,0} |c\rangle_1 |b\rangle_2 |0\rangle_\gamma.\nonumber\end{aligned}$$ The corresponding time-evolution $|\Psi(t)\rangle$ looks a bit complicated: $$\begin{aligned}
&& \mathcal{A}(t) |a\rangle_1 |a\rangle_2 |0\rangle_\gamma
+\mathcal{B}(t) |b\rangle_1 |b\rangle_2 |0\rangle_\gamma
+\mathcal{C}(t) |c\rangle_1 |c\rangle_2 |0\rangle_\gamma+\nonumber\\
&& +\mathcal{D}_1(t) |a\rangle_1 |b\rangle_2 |0\rangle_\gamma
+\mathcal{D}_2(t) |b\rangle_1 |a\rangle_2 |0\rangle_\gamma+\nonumber\\
&& +\mathcal{E}_1(t) |a\rangle_1 |c\rangle_2 |0\rangle_\gamma
+\mathcal{E}_2(t) |c\rangle_1 |a\rangle_2 |0\rangle_\gamma+\nonumber\\
&& +\mathcal{F}_1(t) |b\rangle_1 |c\rangle_2 |0\rangle_\gamma
+\mathcal{F}_2(t) |c\rangle_1 |b\rangle_2 |0\rangle_\gamma+\nonumber\\
&& +\mathcal{G}_1(t) |b\rangle_1 |a\rangle_2 |1_{ab}\rangle_\gamma
+\mathcal{G}_2(t) |a\rangle_1 |b\rangle_2 |1_{ab}\rangle_\gamma+\nonumber\\
&& +\mathcal{H}_1(t) |c\rangle_1 |a\rangle_2 |1_{ac}\rangle_\gamma
+\mathcal{H}_2(t) |a\rangle_1 |c\rangle_2 |1_{ac}\rangle_\gamma+\nonumber\\
&& +\mathcal{I}_1(t) |b\rangle_1 |b\rangle_2 |1_{ab}\rangle_\gamma
+\mathcal{I}_2(t) |c\rangle_1 |c\rangle_2 |1_{ac}\rangle_\gamma+\nonumber\\
&& +\mathcal{J}_1(t) |b\rangle_1 |c\rangle_2 |1_{ab}\rangle_\gamma
+\mathcal{J}_2(t) |c\rangle_1 |b\rangle_2 |1_{ab}\rangle_\gamma+\nonumber\\
&& +\mathcal{K}_1(t) |b\rangle_1 |c\rangle_2 |1_{ac}\rangle_\gamma
+\mathcal{K}_2(t) |c\rangle_1 |b\rangle_2 |1_{ac}\rangle_\gamma.
\label{eq:QBs_9}\end{aligned}$$ One oscillatory term would then be, e.g., $\mathcal{J}_1^* \mathcal{K}_1 e^{-i\Delta t}$, which is proportional to $$\begin{aligned}
&& {_\gamma} \langle 1_{ab}| a_{ab}^\dagger a_{ac} |1_{ac}\rangle_\gamma={_\gamma} \langle 0| a_{ab} a_{ab}^\dagger a_{ac} a_{ac}^\dagger |0\rangle_\gamma=\label{eq:QBs_10}\\
&& ={_\gamma} \langle 0| (1+\underbrace{a_{ab}^\dagger}_{0\leftarrow} a_{ab}) (1+a_{ac}^\dagger \underbrace{a_{ac}}_{\rightarrow 0}) |0\rangle_\gamma={_\gamma} \langle 0|0\rangle_\gamma=1.\nonumber\end{aligned}$$ If the spatial separation is less than the photon wavelength, one cannot determine the photon energy, because one does not know which atom has emitted the radiation. Accordingly, we expect QBs.
For the GSI-case, this possibility has to be taken into account, because even for runs with one single EC only, there can have been more ions in the ring that were lost or decayed via $\beta^+$. In this case (comparing the neutrino again with the photon), one has to replace the wavelength of the photon by the de Broglie wavelength of the neutrino. The neutrino energy should be of the same order as the $Q$-value of the EC-reaction, which is roughly 1 MeV [@Litvinov:2008rk]. The corresponding wavelength is, however, $\lambda=\frac{2\pi\hbar c}{Ec}\sim 10^{-12}$ m, while the average distance between two ions should be of the order of the storage ring [@Steck:1996me], which is roughly 100 m [@ESR]. Hence, this possibility is excluded for the GSI-experiment.
\[sec:conclusions\] Conclusions
===============================
A comparison of the GSI-experiment with several other processes (the Double Slit experiment with photons, $e^+ e^- \rightarrow \mu^+ \mu^-$ scattering, and charged pion decay) has been given. By using the language of QFT as well as the intuitive formulation with probability amplitudes, I have shown that the situation at GSI cannot lead to any oscillation of the decay rate, if the correct treatment is chosen and no additional assumptions (as, e.g., a splitting in the initial state) are taken into account. Also the frequently mentioned possibility of Quantum Beats of the final state cannot explain the observed oscillations, at least not in the standard picture. Hopefully this article will contribute to the clarification of the physical situation in the experiment that has been performed at GSI.
\[sec:ack\] Acknowledgements {#secack-acknowledgements .unnumbered}
============================
I would like to thank A. Blum and M. Lindner for useful discussions, as well as K. Blaum and especially J. Kopp for carefully reading the manuscript and giving valuable comments. I am furthermore grateful to the referee of this paper for making valuable suggestions that have improved the manuscript considerably. This work has been supported by Deutsche Forschungsgemeinschaft in the Sonderforschungsbereich Transregio 27.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Recently, a novel family of biologically plausible online algorithms for reducing the dimensionality of streaming data has been derived from the similarity matching principle. In these algorithms, the number of output dimensions can be determined adaptively by thresholding the singular values of the input data matrix. However, setting such threshold requires knowing the magnitude of the desired singular values in advance. Here we propose online algorithms where the threshold is self-calibrating based on the singular values computed from the existing observations. To derive these algorithms from the similarity matching cost function we propose novel regularizers. As before, these online algorithms can be implemented by Hebbian/anti-Hebbian neural networks in which the learning rule depends on the chosen regularizer. We demonstrate both mathematically and via simulation the effectiveness of these online algorithms in various settings.'
author:
- 'Yuansi Chen[^1]'
- 'Cengiz Pehlevan[^2]'
- 'Dmitri B. Chklovskii [^3]'
title: '**Self-calibrating Neural Networks for Dimensionality Reduction** '
---
Introduction
============
Dimensionality reduction plays an important role in both artificial and natural signal processing systems. In man-made data analysis pipelines it denoises the input, simplifies further processing and identifies important features. Dimensionality reduction algorithms have been developed for both the offline setting where the whole dataset is available to the algorithm from the outset and the online setting where data are streamed one sample at a time [@van2009dimensionality; @yan2006effective]. In the brain, online dimensionality reduction takes place, for example, in early processing of streamed sensory inputs as evidenced by a high ratio of input to output nerve fiber counts [@hyvarinen2009natural]. Therefore, dimensionality reduction algorithms may help model neuronal circuits and observations from neuroscience may inspire the future development of artificial signal processing systems.
Recently, a novel principled approach to online dimensionality reduction in neuronal circuits has been developed [@pehlevan2015hebbian; @hu2014hebbian; @pehlevan2014hebbian; @pehlevan2015normative]. This approach is based on the principle of similarity matching: the similarity of outputs must match the similarity of inputs under certain constraints. Mathematically, pairwise similarity is quantified by the inner products of the data vectors and matching is enforced by the classical multidimensional scaling (CMDS) cost function. Dimensionality is reduced by constraining the number of output degrees of freedom, either explicitly or by adding a regularization term.
To formulate similarity matching mathematically, we represent each centered input data sample received at time $t$ by a column vector $\mathbf{x}_t \in \mathbb{R}^n$ and the corresponding output by a column vector, $\mathbf{y}_t \in \mathbb{R}^k$. We concatenate the input vectors into an $n\times T$ input data matrix $\mathbf{X} = [\mathbf{x}_1, ..., \mathbf{x}_T]$ and the output vectors into an $k\times T$ output data matrix $\mathbf{Y} = [\mathbf{y}_1, ..., \mathbf{y}_T]$. Then we match all pairwise similarities by minimizing the summed squared differences between all pairwise similarities (known as the CMDS cost function, [@carroll1972idioscal; @mardia1980multivariate; @cox2000multidimensional]): $$\label{eq_CMDS_cost}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top \mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \right\|_F^2$$ To avoid the trivial solution, $\mathbf{Y} = \mathbf{X}$, we restrict the number of degrees of freedom in the output by setting $k<n$. Then the solution to the minimization problem (\[eq\_CMDS\_cost\]) is the projection of the input data on the $k$-dimensional principal subspace of $\mathbf{X}^\top\mathbf{X}$ [@mardia1980multivariate; @cox2000multidimensional], i.e. the subspace spanned by the eigenvectors corresponding to the top $k$ eigenvalues of the input similarity matrix $\mathbf{X}^\top\mathbf{X}$.
In [@pehlevan2015hebbian; @hu2014hebbian; @pehlevan2014hebbian], the similarity matching objective (\[eq\_CMDS\_cost\]) was optimized in the online setting, where input data vectors arrive sequentially, one at a time, and the corresponding output is computed prior to the arrival of the next input. Remarkably, the derived algorithm can be implemented by a single-layer neural network (Figure \[fig:network\], left), where the components of input (output) data vectors are represented by the activity of input (output) neurons at the corresponding time. The algorithm proceeds in alternating steps where the neuronal dynamics computes the output $\mathbf{y}_T$ and the synaptic weights are updated according to local learning rules, meaning that a weight update of each synapse depends on the activity of only its pre- and postsynaptic neurons. The family of similarity matching neural networks [@pehlevan2015hebbian; @hu2014hebbian; @pehlevan2014hebbian; @pehlevan2015normative] is unique among dimensionality reduction networks in combining biological plausibility with the derivation from a principled cost function.
In real-world signal-processing applications and neuroscience context the desired number of output dimensions is often unknown to the algorithm a priori and varies with time because of input non-stationarity. Because the number of output dimensions in the neural circuit solution of (1) is the number of output neurons it cannot be adjusted quickly.
To circumvent this problem, we proposed to penalize the rank of the output by adding a regularizer $R_{T\mathbf{Y}}=\alpha T{\mathrm Tr}(\mathbf{Y}^\top \mathbf{Y})= \alpha T\left \| \mathbf{Y}^\top \mathbf{Y} \right\|_*$ where $\left \| .\right \|_*$ is a nuclear norm of a matrix known to be a convex relaxation of matrix rank. From the regularized cost function we derived adaptive algorithms [@pehlevan2015normative] in which the number of output dimensions is given by the number of input singular values exceeding a threshold that depends on the parameter $\alpha$. Because singular values scale with the number of time steps $T$ the threshold scales with $T$. However, choosing a regularization parameter is hard because it requires knowing the exact scale of input singular values in advance. Furthermore, a scale-dependent threshold is not adaptive to non-stationary inputs with changing singular values.
In this paper, we introduce [*self-calibrating*]{} regularizers for the cost function (\[eq\_CMDS\_cost\]) which do not depend on time explicitly and are designed to automatically adjust to the variation in singular values of the input. Specifically, we propose $R_{\mathbf{XY}}= \alpha {\mathrm Tr}(\mathbf{X}^\top \mathbf{X}){\mathrm Tr}(\mathbf{Y}^\top \mathbf{Y})$ and $R_{\mathbf{YY}}= \alpha ({\mathrm Tr}(\mathbf{Y}^\top \mathbf{Y}))^2$. We solve the cost function with these regularizers in both offline (Section II) and online (Section IV) settings. These two online algorithms also map onto a single-layer neuronal network but, importantly, with different learning rules. In Section III, we mathemathically illustrate the difference among the three regularizers in a very simple data generation scenario. In Section V, we propose the corresponding algorithms for the non-stationary input scenario by introducing discounting or forgetting.
Adaptive dimension reduction in the offline setting
===================================================
In this section, we first summarize previous derivation of adaptive dimensionality reduction from cost function (\[eq\_CMDS\_cost\]) with a scale-dependent regularizer, $R_{T{\mathbf{Y}}}$, [@pehlevan2015normative] and discuss its potential shortcomings. Then we introduce two self-calibrating adaptive dimension reduction methods, involving solving cost function (\[eq\_CMDS\_cost\]) offline with two new regularizers $R_{\mathbf{XY}}$ and $R_{\mathbf{YY}}$.
Scale-dependent regularizer, $R_{T{\mathbf{Y}}}$
------------------------------------------------
In order to adaptively choose the output dimension, [@pehlevan2015normative] proposed to modify the objective function (\[eq\_CMDS\_cost\]) by adding a scale-dependent regularizer: $$\label{eq_offline_TY}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \right\|_F^2 + 2\alpha T {\mathrm Tr} ( \mathbf{Y}^\top\mathbf{Y} ),$$ with $\alpha \geq 0$. Such a regularizer corresponds to the trace norm of $\mathbf{Y}^\top\mathbf{Y}$ which is a convex relaxation of rank. Its impact on the solution can be better understood by rewriting the cost as a full square which has the same $\mathbf{Y}$-dependent terms: $$\label{eq_offline_TY_variant}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} - \alpha T \mathbf{I}_T\right\|_F^2,$$ where $\mathbf{I}_T$ is a $T\times T$ identity matrix.
The optimal output matrix $\mathbf{Y}$ is a projection of the input data $\mathbf{X}$ onto its principal subspace [@pehlevan2015normative], with soft-thresholding on the input singular values. Indeed, suppose the eigen-decomposition of $\mathbf{X}^\top\mathbf{X}$ is $\mathbf{X}^\top\mathbf{X} = \mathbf{V}^X \mathbf{\Lambda}^X {\mathbf{V}^X}^\top$, where $\mathbf{\Lambda}^X = \mathrm{diag}(\lambda_1^X, ..., \lambda_T^X)$ with $\lambda_1^X \geq ... \geq \lambda_T^X \geq 0$ are ordered eigenvalues of $\mathbf{X}^\top{\bf X}$. Then the solution to the offline problem (\[eq\_offline\_TY\]) is $$\label{eq_sol_offline_TY}
\mathbf{Y} = \mathbf{U}_k \mathrm{ST}_k(\mathbf{\Lambda}^X, \alpha T)^{1/2} {\mathbf{V}_k^X}^\top,$$ where $$\mathrm{ST}_k({\mathbf{\Lambda}}^X,\alpha T) = {\rm diag} \left(\mathrm{ST}\left(\lambda^X_{1},\alpha T\right),\ldots,\mathrm{ST}\left(\lambda^X_{k},\alpha T \right)\right),$$ $\mathrm{ST}$ is the soft-thresholding function, $\mathrm{ST}(a,b) = \max(a-b,0)$. $\mathbf{V}_k^X$ consists of the columns of $\mathbf{V}^X$ corresponding to the top $k$ eigenvalues, i.e. $\mathbf{V}_k^X = [\mathbf{v}_1^X, ..., \mathbf{v}_k^X]$ and $\mathbf{U}_k$ is any $k \times k$ orthogonal matrix, i.e. $\mathbf{U}_k \in \mathcal{O}(k)$.
Equation (4) shows that the regularization coefficient $\alpha$ sets the threshold on the eigenvalues of input covariance. Input modes with eigenvalues above $\alpha T$ are included in the output, albeit with a eigenvalue shrunk by $\alpha T$. Modes below $\alpha T$ are rejected by setting corresponding output singular values to zero. The scaling of regularization coefficient with time, $T$, ensures that the threshold occupies the same relative position in the spectrum of eigenvalues, which grow linearly with time for a stationary signal.
The algorithm can separate signal from the noise if the signal eigenvalues are greater than the noise eigenvalues and $\alpha$ is set in between. However, setting such value of $\alpha$ requires knowing the variance of the input signal and noise. In the offline setting, $\alpha$ can be computed from the data. However, if $\alpha$ has to be chosen a priori, e.g. in the online setting, choosing the value that suits various inputs with different signal variance may be difficult. In particular, when the noise variance of one possible input exceeds the signal variance of another input, a universal value of $\alpha$ does not exist. Is it possible to regularize the problem so that the regularization coefficient is chosen only once and applies universally to inputs of arbitrary variance?
Input-output regularizer, $R_{\mathbf{XY}}$
-------------------------------------------
A regularizer that applies a relative, rather than absolute, threshold to input singular values would be able to deal with various input setting. Rather than using the threshold depending on time $T$ explicitly, we set the threshold value proportional to the sum of eigenvalues of $\mathbf{X}^\top\mathbf{X}$. Formally, this leads to the following optimization problem with input-output regularizer $R_{\mathbf{XY}}$: $$\label{eq_offline_XY}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y}\right\|_F^2+ 2\alpha \mathrm{Tr}(\mathbf{X}^\top\mathbf{X})\mathrm{Tr}(\mathbf{Y}^\top\mathbf{Y})$$ which is equivalent to: $$\label{eq_offline_XY_variant}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} - \alpha \mathrm{Tr}(\mathbf{X}^\top\mathbf{X}) \mathbf{I}_T\right\|_F^2$$ In the offline setting, the change relative to previous method is minor because we can always have a good choice of $\alpha$ to make the former formulation similar to the later after observing the whole $\mathbf{X}$. As a consequence, the offline solution of Eq. (\[eq\_offline\_XY\]) is very similar to that of the previous method. $\mathbf{Y}$ is a projection of the input data onto its principal subspace with a different eigenvalue cutoff based on the input eigenvalue sum. In turn, the coefficient, $\alpha$, sets the threshold relative to the sum of eigenvalues of $\mathrm{Tr}(\mathbf{X}^\top\mathbf{X})$: $$\label{eq_sol_offline_XY}
\mathbf{Y} = \mathbf{U}_k \mathrm{ST}_k(\mathbf{\Lambda}^X, \alpha \mathrm{Tr}(\mathbf{X}^\top\mathbf{X}) )^{1/2} {\mathbf{V}_k^X}^\top.$$ Even though this solution looks very similar to the scale-dependent adaptive dimension reduction’s offline solution (\[eq\_sol\_offline\_TY\]), as we will see in Section IV, the fact that the thresholding depends on the input singular values will allow corresponding online algorithms to calibrate to various input statistics.
Squared-output regularizer, $R_{\mathbf{YY}}$
---------------------------------------------
An alternative way to apply a relative threshold to input singular values is to deploy a regularization proportional to the sum of eigenvalues of $\mathbf{Y}^T\mathbf{Y}$. When doing dimension reduction, the sum of eigenvalues of $\mathbf{Y}^T\mathbf{Y}$ is reflective of the sum of top eigenvalues of $\mathbf{X}^T\mathbf{X}$. This reasoning leads us to the following optimization problem with squared-output regularizer $R_{\mathbf{YY}}$: $$\label{eq_offline_XY}
\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \right\|_F^2 + \alpha [\mathrm{Tr}(\mathbf{Y}^\top\mathbf{Y})]^2.$$
This optimization problem is not as simple as in the previous case. Yet, the optimal output $\mathbf{Y}$ is still a projection of input data but with an adaptively thresholded singular values of $\mathbf{X}$: $$\begin{aligned}
\label{eq_sol_offline_YY}
\mathbf{Y} = \mathbf{U}_k {(\mathbf{D}_k^Y)}^{1/2} {\mathbf{V}_k^X}^\top,\end{aligned}$$ where $\mathbf{D}_k^Y = \mathrm{diag}(\lambda_1^Y, ..., \lambda_p^Y, 0, ..., 0)$ and $$\begin{aligned}
\mathbf{D}_p^Y = (\mathbf{I}_p - \frac{\alpha}{1+\alpha p} \mathbf{1}_p\mathbf{1}_p^\top) \mathbf{\Lambda}_p^X.\end{aligned}$$ The interger $p$ decides how many singular values of the input $\mathbf{X}$ are cut off. It is chosen as the largest integer in $\{1, ..., k\}$ such that all diagonal elements of $\mathbf{D}_p^Y$ are nonnegative.
More details about the derivation of this closed-form offline solution can be found in Appendix A. Intuitively, the amount of shrinkage still depends approximately on the sum of input eigenvalues but the sum is computed only on the top $p$ eigenvalues. Similar to the input-output regularizer, the regularization coefficient $\alpha$ sets the threshold relatively to the input statistics.
Three methods on a simple case: two sets of degenerate eigenvalues with a gap
=============================================================================
To gain intuition about the three similarity matching algorithms, we compare their offline solutions for the input covariance with only two sets of degenerate eigenvalues. Suppose that the eigenvalues of the normalized input similarity matrix $\frac 1T \mathbf{X}^T\mathbf{X}$ are $$(\underbrace{a,...,a}_{n_1}, \underbrace{b,...,b}_{n_2})$$ with $a>b$ and $n_1+n_2 = n$. This kind of scenario models a situation when signal and noise eigenvalues are separated by a gap.
We ask when the output similarity matrix keeps track of the $n_1$ signal eigenmodes but rejects the $n_2$ noise eigenmodes. We first derive, for each method, the range of $\alpha$’s achieving this goal.
1. For the scale-dependent regularizer $R_{T\mathbf{Y}}$, according to Eq. (\[eq\_sol\_offline\_TY\]), it is sufficient and necessary to choose the regularization coefficient $\alpha$ between $a$ and $b$. Note that this regularization coefficient $\alpha$, unlike the following two, depends on the absolute scale of the noise level $b$.
2. For the input-output regularizer $R_{\mathbf{XY}}$, $\text{Tr}(\frac{1}{T}\mathbf{X}^\top\mathbf{X}) = n_1 a + n_2 b$. According to Eq. (\[eq\_sol\_offline\_XY\]), it is sufficient and necessary to choose $\alpha$ such that $$\begin{aligned}
\frac{a}{n_1 a + n_2 b} \geq \alpha \geq \frac{b}{n_1 a + n_2 b}.\end{aligned}$$
3. For the squared-output regularizer $R_{\mathbf{YY}}$, we are aiming at choosing $p = n_1$. According to Eq. (\[eq\_sol\_offline\_YY\]), it is sufficient and necessary to choose $\alpha$ such that the $n_1+1$st output similarity matrix’s eigenvalue is non-positive for $p = n_1 + 1$. This results in $$\begin{aligned}
\alpha \geq \frac{b}{(a-b)n_1}\end{aligned}$$
Table \[tab:summary\_reg\] summarizes the different ranges of regularization coefficients $\alpha$ with which three methods can keep track of the first $n_1$ eigenmodes and the resulting output similarity matrix’s top eigenvalue.
Regularizer Choice of $\alpha$ Output top eigenvalue
---------------------------------------------------- -------------------------------------------------------------------- -------------------------------
1\. Scale-dependent Regularizer, $R_{\mathbf{TY}}$ $a \geq \alpha \geq b$ $a - \alpha$
2\. Input-output Regularizer, $R_{\mathbf{XY}}$ $\frac{a}{n_1 a + n_2 b} \geq \alpha \geq \frac{b}{n_1 a + n_2 b}$ $a - \alpha (n_1 a + n_2 b)$
3\. Squared-output Regularizer, $R_{\mathbf{YY}}$ $\alpha \geq \frac{b}{(a-b)n_1}$ $a (\frac{1}{1+\alpha n_1})$
\[tab:summary\_reg\]
To illustrate the difference between the known scale-dependent regularizer and the two newly proposed regularizers we compute for what fraction of various pairs of $a$ and $b$ (see Appendix) each algorithm achieves the goal for values of $\alpha$ from $0$ to $\infty$. Fig. \[fig:signal2noise\] shows the fraction of $a$ and $b$ pairs for which the signal is transmitted vs. the fraction for which the noise is transmitted as $\alpha$ varies along each curve. The curve corresponding to the scale-dependent regularizer does not reach the point $(0, 1)$ in Fig. \[fig:signal2noise\] indicating that no value of $\alpha$ achieves the goal for all pairs of $a$ and $b$. Yet, the input-output and the squared-output regularizers pass through the point $(0,1)$ indicating that universal values of $\alpha$ exist for which these algorithms transmit all signal while discarding all noise.
![Fraction of $a$ and $b$ pairs for which all signal is transmitted vs. fraction of $a$ and $b$ paris for which all noise is transmitted. Each curve is computed for $\alpha$ varying from $0$ to $\infty$ for each of the three regularizers. []{data-label="fig:signal2noise"}](signal2noise.png){width="0.8\columnwidth"}
Nex, to illustrate the difference between the input-output and the squared-output regularizer we plot a phase diagram (Fig. \[fig:phase\_diag\]) illustrating the range of parameters where each algorithm transmits all signal and rejects all noise. Specifically, Fig. \[fig:phase\_diag\] shows the range of regularization coefficient $\alpha$ for different noise-to-signal ratio $b/a$ of the input. The range of $\alpha$ for which the squared-output regularizer achieves the goal is much larger than that for the input-output regularizer indicating its robustness.
![Colored regions indicate the range of the regularization coefficient $\alpha$ for which all signal and no noise is transmitted. []{data-label="fig:phase_diag"}](phase_diag.png){width="0.8\columnwidth"}
{width="90.00000%"}
Online adaptive dimensionality reduction with Hebbian/anti-Hebbian neural networks
==================================================================================
In this section, we first formulate online versions of the dimensionality reduction optimization problems presented in Section II. Then we derive corresponding online algorithms and map them onto the dynamics of neural networks with biologically plausible local learning rules. Our derivations follow [@pehlevan2015normative]. At time $T$, the algorithm minimizes the cost depending on the previous inputs and outputs up to time $T$ with respect to $\mathbf{y}_T$, while keeping the previous $\mathbf{y}_t$ fixed. $$\label{eq_online_CMDS_cost}
\min_{\mathbf{y}_T} \left\| \mathbf{X}^\top \mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \right\|_F^2$$ In the unregularized formulation, the output dimensionality is determined by the column dimensition of $\mathbf{Y}$, $k$. Online adaptive dimension reduction methods adaptively choose the output dimensionality based on the trade-off of the CMDS cost with regularizers.
Scale-dependent regularizer, $R_{T{\mathbf{Y}}}$
------------------------------------------------
Consider the following optimization problem in the online setting: $$\begin{aligned}
\label{eq_online_TY}
\mathbf{y}_T \leftarrow \arg\min_{\substack{\mathbf{y}_T}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} - \alpha T \mathbf{I}_T \right\|_F^2.\end{aligned}$$ By expanding the squared norm and keeping only the terms that depend on $y_T$, the problem is equivalent to: $$\begin{aligned}
\mathbf{y}_T \leftarrow \arg\min_{\substack{\mathbf{y}_T}} & \phantom{+} [ -4\mathbf{x}_T^\top \left(\sum_{t=1}^{T-1} \mathbf{x}_t\mathbf{y}_t^\top\right) \mathbf{y}_T , \nonumber \\
& + 2\mathbf{y}_T^\top (\sum_{t=1}^{T-1}\mathbf{y}_t\mathbf{y}_t^\top+ \alpha T\mathbf{I}_m) \mathbf{y}_T, \nonumber \\
& - 2 \|\mathbf{x}_T\|^2 \|\mathbf{y}_T\|^2 + \|\mathbf{y}_T\|^4 ].\end{aligned}$$ In the large-$T$ limit, the last two terms can be ignored since the first two terms of order $T$ dominates. The remaining cost function is a positive quadratic form in $\mathbf{y}_T$ and we could find the minimum by solving the system of linear equations: $$\begin{aligned}
\label{fp}
\left (\sum_{t=1}^{T-1} \mathbf{y}_t\mathbf{y}_t^\top + \alpha T \mathbf{I}_m \right) \mathbf{y}_T = \left( \sum_{t=1}^{T-1} \mathbf{y}_t\mathbf{x}_t^\top\right)\mathbf{x}_T.\end{aligned}$$ Out of various ways to solve Eq. , we choose the weighted Jacobi iteration because it leads to an algorithm implementable by a biologically plausible neural network [@pehlevan2015hebbian]. In this algorithm, $$\label{eq_jacobi}
\mathbf{y}_T \leftarrow (1-\eta) \mathbf{y}_T + \eta (\mathbf{W}_T^{YX} \mathbf{x}_T - \mathbf{W}_T^{YY}\mathbf{y}_T),$$ where $\eta$ is the weight parameter, and $\mathbf{W}_T^{YX}$ and $\mathbf{W}_T^{YY}$ are normalized input-output and output-output covariances: $$\begin{aligned}
\label{TYweights}
W_{T, ik}^{YX} &= \frac{\sum_{t=1}^{T-1} y_{t,i}x_{t,k}}{\alpha T + \sum_{t=1}^{T-1} y_{t,i}^2}, \nonumber \\
W_{T, ij\neq i}^{YY} &= \frac{\sum_{t=1}^{T-1} y_{t,i}y_{t,j}}{\alpha T + \sum_{t=1}^{T-1} y_{t,i}^2}, \nonumber \\
W_{T, ii}^{YY} &= 0.\end{aligned}$$ When the Jacobi iteration converges to a fixed point, it obtains the solution of the quadratic program.
Such algorithm can be implemented by the dynamics of neural activity in a single-layer network. $\mathbf{W}_T^{YX}$ and $\mathbf{W}_T^{YY}$ represent the weights of feedforward $(\mathbf{x}_t \rightarrow \mathbf{y}_t)$ and lateral $(\mathbf{y}_t \rightarrow \mathbf{y}_t)$ synaptic connections. At each time step $T$, we first iterate until convergence, then update the weights online according to the following learning rules: $$\begin{aligned}
\label{eq_online_update_TY}
\mu_{T+1, i} &\leftarrow \mu_{T, i} + \alpha + y_{T,i}^2 \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}} \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha+y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$ where we introduce scalar variables $\mu_{T,i}$ representing cumulative activity of neuron $i$ up to time $T-1$. The left figure of Fig. \[fig:network\] illustrates this network implementation.
Input-output regularizer, $R_{\mathbf{XY}}$
-------------------------------------------
The online optimization problem with input-output regularizer is similar to the previous one. At every time step, the amount of thresholding is given by the cumulative sum of input eigenvalues. After slight modifications, we arrive at the following neural network algorithm. At time step $T$, $\mathbf{y}_T$ is iterated until convergence based on the following Jacobi iteration $$\begin{aligned}
\label{eq_jacobi_XY}
\mathbf{y}_T \leftarrow (1-\eta) \mathbf{y}_T + \eta (\mathbf{W}_T^{YX} \mathbf{x}_T - \mathbf{W}_T^{YY} \mathbf{y}_T).\end{aligned}$$ The online synaptic updates are: $$\begin{aligned}
\label{eq_online_XY}
\mu_{T+1, i} &\leftarrow \mu_{T, i} + \alpha \|\mathbf{x}_T\|^2 + y_{T,i}^2 \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha \|\mathbf{x}_T\|^2+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}} \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha \|\mathbf{x}_T\|^2 +y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$ This differs from Eq. in that a new scalar variable $\|\mathbf{x}_T\|^2$ is needed that sums up the current input amplitude across all input neurons. Biologically, in an Hebbian/anti-Hebbian neural network, such summation could be implemented via extracellular space or glia (Fig. \[fig:network\], middle).
Squared-output regularizer, $R_{\mathbf{YY}}$ {#sec_online_YY}
---------------------------------------------
Finally, we consider the following online optimization problem: $$\begin{aligned}
\label{eq_online_YY}
\mathbf{y}_T \leftarrow \arg\min_{\substack{\mathbf{y}_T}} \|\mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \|_F^2 + \alpha \text{Tr}(\mathbf{Y}^\top\mathbf{Y})^2.\end{aligned}$$ In the large-T limit, $\sum_{t=1}^{T-1} \|\mathbf{y}_t\|^2 \gg \|\mathbf{y}_T\|^2$, we could instead solve a simplified version to the problem: $$\begin{aligned}
\label{eq_online_update_YY}
\mathbf{y}_T &\leftarrow \arg\min_{\substack{\mathbf{y}_T}} \|\mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y}\|_F^2 \nonumber \\
&\qquad\qquad+ 2 \alpha \text{Tr}(\mathbf{Y}_{T-1}^\top\mathbf{Y}_{T-1}) \text{Tr}(\mathbf{Y}^\top\mathbf{Y}).\end{aligned}$$ After this simplification, the update is similar to the previous input-output regularizer except that the learning rate depends on the norm of current output vector, $\|\mathbf{y}_T\|^2$. Since this is a scalar, such summation can be easily implemented in biology using summation in extracellular space or glia (Fig. \[fig:network\], right). At time step $T$ , $\mathbf{y}_T$ the neural network dynamics iterates until convergence of the Jacobi iteration . After convergence, synaptic weights are updated online as follows, $$\begin{aligned}
\label{eq_online_YY}
\mu_{T+1, i} &\leftarrow \mu_{T, i} + \alpha \|\mathbf{y}_T\|^2 + y_{T,i}^2 \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha \|\mathbf{y}_T\|^2+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}} \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha \|\mathbf{y}_T\|^2 +y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$
Online algorithms for non-stationary statistics
===============================================
The online algorithms we proposed assume that the input has stationary statistics. A truly adaptive algorithm, in addition to self-calibrating the number of dimensions to transmit, should be able to adapt to temporal statistics changes. To address this issue, we introduce discounting into the cost function which reduces the contribution of older data samples [@pehlevan2015hebbian; @yang1995projection], or equivalently “forgets" them, in order to react to changes in input statistics.
Scale-dependent regularizer, $R_{T\mathbf{Y}}$
----------------------------------------------
Following [@pehlevan2015hebbian], we discount past inputs, ${\bf x}_t$, and past outputs, ${\bf y}_t$, with $\beta^{T-t}$, where $0\leq\beta\leq 1$. With such discounting, the effective time scale of forgetting is $-1/\ln \beta$. This procedure leads to a modified online cost function, $$\begin{aligned}
\label{eq_online_forget_TY}
\mathbf{y}_T \leftarrow \arg\min_{\substack{\mathbf{y}_T}} \left\| B \mathbf{X}^\top\mathbf{X} B - B \mathbf{Y}^\top\mathbf{Y} B - \alpha \mathrm{Tr}(B^2) \right\|_F^2\end{aligned}$$ where $B$ is a diagonal matrix with $(\beta^0, \beta^1, \beta^2, ..., \beta^{T-1})$ on the diagonal. The term $\mathrm{Tr}(B^2)$ takes the place of the time variable $T$ in the original online formulation \[eq\_online\_TY\]. When $\beta$ is $1$, reduces to . To derive an online algorithm, we follow the same steps as before. By keeping only the terms that depend on current output $\mathbf{y}_T$, we again arrive at a quadratic function of $\mathbf{y}_T$, which is solved as in by a weighted Jacobi iteration. The online synaptic learning rules get modified: $$\begin{aligned}
\label{eq_rule_forget_TY}
\mu_{T+1, i} &\leftarrow \beta^2 \mu_{T, i} + \alpha + y_{T,i}^2, \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}}, \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha+y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$ The difference from the non-discounted learning rules (Eq. ) is in how $\mu_{T, i}$ gets updated. The $\beta^2$ decay in $\mu_{T, i}$ update prevents $\mu_{T, i}$ from growing indefinitely. Consequently, the learning rate, $1/\mu_{T+1, i}$ does not steadily decrease with $T$, but saturates, allowing the synaptic weights to react to changes in input statistics.
Input-output regularizer, $R_{\mathbf{XY}}$
-------------------------------------------
We can implement forgetting in this case again by discounting past inputs and outputs: $$\begin{aligned}
\mathbf{y}_T \leftarrow \arg\min_{\substack{\mathbf{y}_T}} \left\| B \mathbf{X}^\top\mathbf{X} B - B \mathbf{Y}^\top\mathbf{Y} B - \alpha \mathrm{Tr}(B \mathbf{X}^\top \mathbf{X} B) \right\|_F^2.\end{aligned}$$ Following the same steps as before, a weighted Jacobi iteration is still deployed. The following learning rules can be derived: $$\begin{aligned}
\label{eq_rule_forget_XY}
\mu_{T+1, i} &\leftarrow \beta^2 \mu_{T, i} + \alpha \|\mathbf{x}_T\|^2 + y_{T,i}^2, \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha \|\mathbf{x}_T\|^2+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}}, \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha \|\mathbf{x}_T\|^2 +y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$ Again, $\beta^2$ decay in $\mu_{T,i}$ update allows the network to react to non-stationarity.
Squared-output regularizer, $R_{\mathbf{YY}}$ {#squared-output-regularizer-r_mathbfyy-1}
---------------------------------------------
Discounting past inputs and outputs, we arrive at the online cost function: $$\begin{aligned}
\mathbf{y}_T &\leftarrow \arg\min_{\substack{\mathbf{y}_T}} \left\| B \mathbf{X}^\top\mathbf{X} B - B \mathbf{Y}^\top\mathbf{Y} B\right\|_F^2 \nonumber \\
&\qquad\qquad + \alpha \mathrm{Tr}(B \mathbf{Y}^\top \mathbf{Y} B)^2\end{aligned}$$ Following the same steps as before, a weighted Jacobi iteration is still deployed. The following learning rules can be derived: $$\begin{aligned}
\label{eq_rule_forget_YY}
\mu_{T+1, i} &\leftarrow \beta^2 \mu_{T, i} + \alpha \|\mathbf{y}_T\|^2 + y_{T,i}^2, \nonumber \\
W_{T+1,ij}^{YX} &\leftarrow W_{T,ij}^{YX} + \frac{(y_{T,i} x_{T,j} - (\alpha \|\mathbf{y}_T\|^2+y_{T,i}^2) W_{T,ij}^{YX})}{\mu_{T+1,i}}, \nonumber \\
W_{T+1,ij\neq i}^{YY} &\leftarrow W_{T,ij}^{YY} + \frac{(y_{T,i} y_{T,j} - (\alpha \|\mathbf{y}_T\|^2 +y_{T,i}^2) W_{T,ij}^{YY})}{\mu_{T+1,i}}.\end{aligned}$$
Experiments
===========
We evaluate the performance of three online algorithms on a synthetic dataset. We first generate a $n=64$ dimensional colored Gaussian process with a specified covariance matrix. In this covariance matrix, the eigenvalues, $\lambda_{1..4} = \{6, 5, 4, 2\}$ and the remaining $\lambda_{5..60}$ are chose uniformly from the interval $[0, 0.2]$. Correlations are introduced in the covariance matrix by generating random orthonormal eigenvectors. We set $\eta$ to be $0.1$. Synaptic weight matrices were initialized randomly.
Initially, we choose different $\alpha$’s for three algorithms such that the thresholding has the approximately the same effect on the original data: the output keeps track of the top three principal components, while discarding the rest principal components. Additionally, the top three eigenvalues of the input similarity matrix are soft-thresholded by $2$.
Stationary input
----------------
As shown in Fig.\[fig:eigen\_track\], with appropriately chosen $\alpha$, all three online algorithms are able to keep track of the top three input eigenvalues correctly.
{width="73.00000%"}
To quantify the performance of these algorithms more precisely than looking at individual eigenvalues, we use two different metrics. The first metric, eigenvalue error, measures the deviation of output covariance eigenvalues from their optimal offline values derived in Section II. The eigenvalue error at iteration $T$ is calculated by summing squared differences between the eigenvalues of $\frac{1}{T} Y^\top Y$. The second metric, subspace error, quantifies the deviation of the learned subspace from the input principal subspace. The exact formula for the subspace error metric can be found in [@pehlevan2015normative]. Fig \[fig:eigen\_subspace\_error\] shows that three algorithms perform similarly in terms of these two metrics. Both errors for each algorithm decrease as a function of iterations $T$.
![Left: eigenvalue error as a function of iteration for three algorithms; Right: subspace error as a function of iteration for three algorithms.[]{data-label="fig:eigen_subspace_error"}](eigen_error.png "fig:"){width="100.00000%"} \[fig:eigen\_error\]
![Left: eigenvalue error as a function of iteration for three algorithms; Right: subspace error as a function of iteration for three algorithms.[]{data-label="fig:eigen_subspace_error"}](subspace_error.png "fig:"){width="100.00000%"} \[fig:subspace\_error\]
Non-stationary input
--------------------
We evaluate the performance of three online algorithms with forgetting on non-stationary input. The non-stationary input we use here has a sudden change of input statistics. We first use the original data generation process for $1000$ iterations. $\alpha$ is chosen for each algorithm such that the top three principal components are retained and the rest are discarded. Then we change the input data generation by multiplying the eigenvalues of $\mathbf{X}^\top\mathbf{X}$ by $2$, in order to see whether the algorithms can still keep track of only the top three principal components. Finally at $6000$ iteration, we change back to the original statistics. Since the input statistics changes over time, $\frac{1}{T}\mathbf{X}^\top\mathbf{X}$ is not reflective of the eigenvalues our online algorithms are tracking at time $T$. Thus we the eigenvalues over a short period $T_0$ of data before $T$.
For the first $1000$ iterations, all three online algorithms keep track of the top three principal components (See Fig. \[fig:sudden2\]). The fourth output singular value (fourth red line) is kept zero all the time, while the top three singular values (top three red lines) are above zero. At $1000$ iteration, there is sudden change of input data generation. The fourth output singular value for input-output regularizer and squared-output regularizer remains zero, however, the fourth output singular value for scale-dependent regularizer becomes larger than zero (See Fig. \[fig:sudden2\]). Scale-dependent regularizer now has an output with effective dimension four rather than three. The other two are doing a better job in keeping track of only three principal components.
When the input data generation is changed back to the original one at iteration $6000$, because of the forgetting mechanism we introduced, all three regularizers are able to keep track of the top three principal components like during the first $1000$ iterations.
{width="73.00000%"}
Conclusion
==========
We have introduced online dimensionality reduction algortihms with self-calibrating regularizers. Unlike the scale-dependent adaptive dimensionality reduction algorithm [@pehlevan2015normative], these self-calibrating algorithms are designed to automatically adjust to the variation in singular values of the input. As a consequence, they may be more appropriate for modeling neuronal circuits or any related artificial signal processing systems.
APPENDIX {#Appendix .unnumbered}
========
Proof: offline solution of squared-output regularizer {#Appendix:proof}
-----------------------------------------------------
First, we cite a lemma from [@pehlevan2015normative]:
[**Lemma 1**]{}: Let $\mathbf{\Lambda} = \mathrm{diag}(\lambda_1,...,\lambda_p)$, where $\lambda_1 \geq ... \geq \lambda_p$ are real numbers, and let $\hat{\mathbf{\Lambda}} = \mathrm{diag}(\hat{\lambda}_1,...,\hat{\lambda}_p)$, where $\hat{\lambda}_1 \geq ... \geq \hat{\lambda}_p$ are real numbers. Then, $$\max_{\mathbf{O}\in O(p)} \mathrm{Tr}(\mathbf{\Lambda}\mathbf{O}\hat{\mathbf{\Lambda}}\mathbf{O}^\top) = \mathrm{Tr}(\mathbf{\Lambda}\hat{\mathbf{\Lambda}})$$ where $O(p)$ is the set of $p\times p$ orthogonal matrices.
This lemma states that identity belongs to the optimal orthogonal transformations for diagonal matrix alignment. A complete proof of the lemma can be found in [@pehlevan2015normative].
Offline adaptive soft-thresholding optimization problem has the following form: $$\min_{\mathbf{Y}} \left\| \mathbf{X}^\top\mathbf{X} - \mathbf{Y}^\top\mathbf{Y} \right\|_F^2 + \alpha [\text{Tr}(\mathbf{Y}^\top\mathbf{Y})]^2.$$ Suppose an eigen-decomposition of $\mathbf{Y}^\top\mathbf{Y}$ is $\mathbf{Y}^\top\mathbf{Y} = \mathbf{V}^Y\mathbf{\Lambda}^Y{\mathbf{V}^Y}^\top$. Since the Frobenius norm is invariant to multiplication of unitary matrices, we could multiply on left ${\mathbf{V}^X}^\top$ and on right ${\mathbf{V}^X}$ to obtain an equivalent objective $$\min_{\mathbf{\Lambda}^Y \geq 0, \mathbf{G}\in O(T)} \left\| \mathbf{\Lambda}^X - \mathbf{G}\mathbf{\Lambda}^Y\mathbf{G}^\top \right\|_F^2 + \alpha [\text{Tr}(\mathbf{\Lambda}^Y)]^2.$$ According to [**Lemma 1**]{}, we conclude that $\mathbf{G}$ could be take $\mathbf{I}_T$ at optimum. Observing that $\text{Tr}(\mathbf{\Lambda}^Y)$ can be written as a linear transform of its diagonal elements. Then the remaining optimization on the diagonal matrix $\mathbf{\Lambda}^Y$ could be written as $$\min_{\mathbf{d}^Y \geq 0} \left\| \mathbf{d}^X - \mathbf{d}^Y \right\|^2 + \alpha (\mathbf{1}^\top\mathbf{d}^Y)^2.$$ where $\mathbf{d}^X$ and $\mathbf{d}^Y$ are diagonals of $\mathbf{\Lambda}^X$ and $\mathbf{\Lambda}^Y$ respectively. We have an extra constraint that less than $n$ coordinates of $\mathbf{d}^Y$ could be nonzero. The problem could be written equivalently, $$\min_{\mathbf{d}_n^Y \geq 0} {\mathbf{d}_n^Y}^\top (\mathbf{I}_n + \alpha \mathbf{1}_n\mathbf{1}_n^T) \mathbf{d}_n^Y - 2 {\mathbf{d}_n^X}^\top \mathbf{d}_n^Y.$$ where $\mathbf{d}_n^Y$ are the first $n$ elements of $\mathbf{d}^Y$.
This is a nonnegative least squares problem (NNLS), which has the general form $$\min_{x\geq 0} \left\| \mathbf{b} - \mathbf{A}\mathbf{x}\right\|^2.$$ For general form of $\mathbf{A}$, this NNLS does not allow for closed form solution. In general, it is solved by an active-set type optimization algorithm [@nocedal2006numerical], and the number of iterations in the worse case could be exponential on the input dimension. In our case, $\mathbf{A} = \mathbf{I}_n + \alpha \mathbf{1}_n\mathbf{1}_n^\top$ and it almost allows for an closed form solution.
1. Since $\mathbf{A}$ is a diagonal matrix plus a constant matrix, when the values of $\mathbf{d}^X$ is ordered, the values of $\mathbf{d}^Y$ is also ordered.
2. Once the support of $\mathbf{d}^Y$ is known, the problem is a unconstrained positive definite quadratic program, which always allows for a closed form solution.
3. Combining (1) and (2), the support of the solution is always the first $p$ elements. It is sufficient to try $n$ different supports and find the best feasible solution.
Now suppose that we have found that the support of solution is of size $p$, we could obtain a closed form solution of the offline problem. Given the support, the NNLS problem is equivalent to the unconstrained quadratic problem. $$\min_{\mathbf{d}_p^Y} {\mathbf{d}_p^Y}^\top (\mathbf{I}_p + \alpha \mathbf{1}_p\mathbf{1}_p^T) \mathbf{d}_p^Y - 2 {\mathbf{d}_p^X}^\top \mathbf{d}_p^Y.$$ Solving the unconstrained quadratic problem, we obtain $$\mathbf{d}_p^Y = (\mathbf{I}_p - \frac{\alpha}{1+\alpha p} \mathbf{1}_p\mathbf{1}_p^\top) \mathbf{d}_p^X,$$ since $\mathbf{A}$ is invertible with inverse $\mathbf{I}_p - \frac{\alpha}{1+\alpha p} \mathbf{1}_p\mathbf{1}_p^\top$.
Performance of three regularizers in various signal and noise setups {#Appendix:performance}
--------------------------------------------------------------------
The set of signal and noise setups $(a, b)$ corresponds to the following $5050$ cases $$S = \{ (a, b) | a \geq b, (a, b) \in \{0.01, 0.02, ..., 0.99, 1.00\}^2\}.$$
ACKNOWLEDGMENT {#acknowledgment .unnumbered}
==============
We would like to thank Anirvan Sengupta for discussions.
[10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
L. Van Der Maaten, E. Postma, and J. Van den Herik, “Dimensionality reduction: a comparative review,” *J Mach Learn Res*, vol. 10, pp. 66–71, 2009.
J. Yan, B. Zhang, N. Liu, S. Yan, Q. Cheng, W. Fan, Q. Yang, W. Xi, and Z. Chen, “Effective and efficient dimensionality reduction for large-scale and streaming data preprocessing,” *IEEE transactions on Knowledge and Data Engineering*, vol. 18, no. 3, pp. 320–333, 2006.
A. Hyv[ä]{}rinen, J. Hurri, and P. O. Hoyer, *Natural Image Statistics: A Probabilistic Approach to Early Computational Vision.*1em plus 0.5em minus 0.4emSpringer Science & Business Media, 2009, vol. 39.
C. Pehlevan, T. Hu, and D. B. Chklovskii, “A hebbian/anti-hebbian neural network for linear subspace learning: A derivation from multidimensional scaling of streaming data,” *Neural computation*, vol. 27, pp. 1461–1495, 2015.
T. Hu, C. Pehlevan, and D. B. Chklovskii, “A hebbian/anti-hebbian network for online sparse dictionary learning derived from symmetric matrix factorization,” in *2014 48th Asilomar Conference on Signals, Systems and Computers*.1em plus 0.5em minus 0.4emIEEE, 2014, pp. 613–619.
C. Pehlevan and D. B. Chklovskii, “A hebbian/anti-hebbian network derived from online non-negative matrix factorization can cluster and discover sparse features,” in *2014 48th Asilomar Conference on Signals, Systems and Computers*.1em plus 0.5em minus 0.4emIEEE, 2014, pp. 769–775.
C. Pehlevan and D. Chklovskii, “A normative theory of adaptive dimensionality reduction in neural networks,” in *Advances in Neural Information Processing Systems*, 2015, pp. 2269–2277.
J. Carroll and J. Chang, “Idioscal (individual differences in orientation scaling): A generalization of indscal allowing idiosyncratic reference systems as well as an analytic approximation to indscal,” in *Psychometric meeting, Princeton, NJ*, 1972.
K. Mardia, J. Kent, and J. Bibby, *Multivariate analysis*.1em plus 0.5em minus 0.4emAcademic press, 1980.
T. Cox and M. Cox, *Multidimensional scaling*.1em plus 0.5em minus 0.4emCRC Press, 2000.
B. Yang, “Projection approximation subspace tracking,” *IEEE Transactions on Signal Processing*, vol. 43, no. 1, pp. 95–107, 1995.
J. Nocedal and S. Wright, *Numerical optimization*.1em plus 0.5em minus 0.4emSpringer Science & Business Media, 2006.
[^1]: yuansi.chen@berkeley.edu
[^2]: cpehlevan@simonsfoundation.org
[^3]: mitya@simonsfoundation.org
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We built a multi-wavelength dataset for galaxies from the Local Volume [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} Survey (LVHIS), which comprises 82 galaxies. We also select a sub-sample of ten large galaxies for investigating properties in the galactic outskirts. The LVHIS sample covers nearly four orders of magnitude in stellar mass and two orders of magnitude in [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass fraction ($\fHI$). The radial distribution of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas with respect to the stellar disc is correlated with $\fHI$ but with a large scatter. We confirm the previously found correlations between the total [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass and star formation rate (SFR), and between [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities and SFR surface densities beyond R$_{25}$. However, the former correlation becomes much weaker when the average surface densities rather than total mass or rate are considered, and the latter correlation also becomes much weaker when the effect of stellar mass is removed or controlled. Hence the link between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} is intrinsically weak in these regions, consistent with what was found on kpc scales in the galactic inner regions. We find a strong correlation between the SFR surface density and the stellar mass surface density, which is consistent with the star formation models where the gas is in quasi-equilibrium with the mid-plane pressure. We find no evidence for [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps to be linked with decreasing star forming efficiencies.'
author:
- |
Jing Wang$^{1,2}$[^1], Bärbel S. Koribalski$^{1}$[^2], Tom H. Jarrett$^3$, Peter Kamphuis$^4$, Zhao-Yu Li$^{5,6}$,\
Luis C. Ho$^{2,7}$, Tobias Westmeier$^8$, Li Shao$^1$, Claudia del P. Lagos$^{8,9}$, O. Ivy Wong$^{8,9}$,\
Paolo Serra$^{10}$, Lister Staveley-Smith$^{8,9}$, Gyula Józsa$^{11,12,13}$, Thijs van der Hulst$^{14}$\
Á.R. López-Sánchez$^{15,16}$\
$^1$Australia Telescope National Facility, CSIRO Astronomy and Space Science, PO box 76, Epping, NSW 1710, Australia\
$^2$Kavli Institute for Astronomy and Astrophysics, Peking University, Beijing 100871, China\
$^3$Department of Astronomy, University of Cape Town, Private Bag X3, Rondebosch, 7701, South Africa\
$^4$National Centre for Radio Astrophysics, TIFR, Ganeshkhind, Pune 411007, India\
$^5$Key Laboratory for Research in Galaxies and Cosmology, Shanghai Astronomical Observatory, Chinese Academy of Sciences,\
80 Nandan Road, Shanghai 200030, China\
$^6$College of Astronomy and Space Sciences, University of Chinese Academy of Sciences, 19A Yuquan Road, Beijing 100049, China\
$^7$Department of Astronomy, School of Physics, Peking University, Beijing 100871, China\
$^8$International Centre for Radio Astronomy Research (ICRAR), M468, University of Western Australia, 35 Stirling Hwy,\
Crawley, WA 6009, Australia\
$^9$ARC Centre of Excellence for All-Sky Astrophysics (CAASTRO), Australia\
$^{10}$INAF - Osservatorio Astronomico di Cagliari, Via della Scienza 5, I-09047 Selargius (CA), Italy\
$^{11}$SKA South Africa Radio Astronomy Research Group, 3rd Floor, The Park, Park Road, Pinelands, 7405, South Africa\
$^{12}$Rhodes University, Department of Physics and Electronics, Rhodes Centre for Radio Astronomy Techniques & Technologies,\
PO Box 94, Grahamstown, 6140, South Africa\
$^{13}$Argelander-Institut für Astronomie, Auf dem Hügel 71, D-53121 Bonn, Germany\
$^{14}$University of Groningen, Kapteyn Astronomical Institute, Landleven 12, 9747 AD, Groningen, The Netherlands\
$^{15}$Australian Astronomical Observatory, PO Box 915, North Ryde, NSW 1670, Australia\
$^{16}$Department of Physics and Astronomy, Macquarie University, NSW 2109, Australia\
bibliography:
- 'LVHISmw.bib'
date: 'Accepted 2017 ???? ?? Received 2017 ???? ??; in original form 2017 January'
title: 'The Local Volume HI Survey: star formation properties'
---
interstellar medium, spiral galaxies, star formation
Introduction {#sec:introduction}
============
How galaxies evolve in their stellar population is a key question in galactic astronomy, and the properties of atomic hydrogen [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas contain key information to that question. The star formation rates (SFR) and stellar masses of star forming galaxies are observed to correlate with each other and this relationship is often referred to as the star formation ‘main sequence’ [MS, @Noeske07; @Elbaz07; @Dutton10]. The star formation quenching or quenched galaxies lie below the sequence with a wide range of star formation rates [@Whitaker12]. The deviation to the median SFR of the MS at a fixed stellar mass strongly depends on the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas fraction of galaxies [@Saintonge16]. Hence the quenching of star formation is accompanied by removal or exhaustion of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas. This is because [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} forms the initial reservoir from which molecular gas and subsequently, stars are formed. In order to better understand galaxy quenching, we first need to understand how [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} fuels star formation, i.e. the relation between [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} and SFR in star-forming galaxies.
Several large single-dish [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surveys have been carried out in the past decade. The shallow blind surveys, ALFALFA [Arecibo Legacy Fast ALFA, @Giovanelli05] and HIPASS [HI Parkes All-Sky Survey, @Meyer04; @Koribalski04], and the targeted, optically selected, GASS survey [GALEX Arecibo SDSS Survey, @Catinella10] have significantly advanced our understanding of the SFR-[[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} relations on the global scales. Among the most interesting findings are, 1) SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass are tightly correlated for the ALFALFA sample whose galaxies are mostly on the star-forming sequence [@Huang12]; 2) the star forming efficiencies (SFE$=$SFR$/\mHI$) in the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-selected ALFALFA sample are on average three times lower than those in the optically selected GASS sample, which includes both star-forming and transitional galaxies [@Huang12; @Schiminovich10]; 3) there is a weak correlation between SFE and stellar mass in the ALFALFA sample [@Huang12], which was not found within the GASS sample [@Schiminovich10], possibly due to the small range in stellar masses for GASS and$\slash$or the different sample selections; 4) there is a weak correlation between SFE and optical surface brightness for galaxies selected from the HIPASS sample [@Wong16].
The strong correlation between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass found on global scales seems to be at odds with the lack of correlation between surface densities of SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} found on sub-kpc scales in the inner parts of galaxies [@Bigiel08]. The latter is contrasted by the tight scaling relation between surface densities of SFR and molecular gas [@Bigiel08], and the explanation is that the molecular gas clouds rather than the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas directly form stars. This difference in [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR relations on kpc and global scales might be caused by the difference in scales, but can also possibly indicate different physics working on these two scales. Especially we need to keep in mind that a large fraction of the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas in [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-rich galaxies is distributed beyond the stellar disc where the SFE is so much lower than that in the inner discs that it would take several times the Hubble time to deplete the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas there [@Bigiel10].
On the other hand, the global [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR correlation is present on kpc-scales in the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} dominated regions, i.e., in irregular galaxies [@Roychowdhury14] and the outskirts of spiral and early-type galaxies [@Koribalski09; @Bigiel10; @Yim16; @Yildiz16]. However there is large scatter ($\sim$1.5 dex) in these relations [@Roychowdhury14; @Bigiel10]. @Hunter98 showed that in irregular galaxies the SFR radial profiles actually follow the shape of stellar mass profiles better than that of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} profiles. This makes us wonder whether the same behaviour can be found in the outskirts of spiral galaxies. A possibility that could explain these trends is that both the SFR and the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities are correlated with the stellar mass surface densities in the outer regions of galaxies. The mid-plane pressure model has been commonly used to explain the link between mass surface densities and SFR [@Leroy08; @Ostriker10 e.g.], however one significant limitation is that it doesn’t explain the stochastic behaviour of star formation in these low gas density regions [@Gerola80; @Teich16].
Mass surface density is not the only possible parameter that regulates SFR. Yim & van der Hulst (2016) found that disturbances from the external environment seem to hardly affect the radial [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} SFE of galaxies. On the other hand, the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} disc internal structures may affect SFE in the galactic outskirt. As demonstrated nicely in @vanderKruit07, [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps usually start near the optical truncation radius where the optical discs abruptly and steeply turn fainter [also see @Jozsa07b; @GarciaRuiz02; @Briggs90]. One possible explanation is that the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas in a warp has an external origin and was accreted with an angular momentum mis-aligned with that of the main disc. Conversely, the existence of an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp may affect the radial distribution of SFR, which is adopted in some theoretical models to explain the optical truncation radius. For example, in the model of @SanchezBlazquez09 the gas near a warp would have low volume densities due to the warp, and hence could not efficiently form stars. As a result a large fraction of stars beyond the warp radius were not formed in situ but rather migrated from the inner part of the galaxy disc. These theories add complexities to the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR relationship in the galactic outskirts. However, the influence of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps on localised SF has never been investigated statistically with observational data.
We have reviewed the apparently different behaviours of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR relations on different scales and in different regions. However, to understand them in a consistent picture has been difficult, largely due to the lack of proper datasets. The single-dish [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data spatially resolve very few galaxies while [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} interferometric samples usually lack homogeneity in galaxy selection. Another common problem with HI interferometry is that extended emission of objects with large angular scales (typically $>$ 15 arcmin) is not properly measured because there are no baselines short enough to sample such angular scales. This affects the surface densities of all extended emission, especially on the outer parts. The Local Volume [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} Survey (LVHIS[^3], Koribalski et al. in prep, Paper I hereafter) tends to overcome these difficulties. It is close to a volume limited [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} sample of galaxies beyond the Milky Way, built on the single-dish blind survey HIPASS [@Meyer04], by targeting all detections with a declination less than or similar to $-30$ degree and a distance less than 8 Mpc from the HIPASS catalogs [@Meyer04; @Koribalski04]. Special care has been taken to address the missing flux problem by using compact array configurations. We also combine the interferometric data with single-dish imaging mosaics [from the Parkes Galactic All Sky Survey, GASS, @McClure-Griffiths09] to produce the HI image for the galaxy NGC 300, for which ATCA data alone is not sufficient to sample the extended HI structures [see Paper I and @Stanimirovic02 for a description of such techniques]. LVHIS hence provides us a suitable laboratory for coherently understanding the different [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR relations.
This paper presents the multi-wavelength dataset for the LVHIS sample. As a first application of the dataset it investigates [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-SFR relations on global scales and in the outskirts, with the goal of reconciling the seeming contradictions of these relations from the relation observed in galactic inner regions.
The paper is organised as follows. In Section \[sec:sample\] we introduce the data used in this paper, including a brief description about the LVHIS sample and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data. In Section \[sec:analysis\] we analyse the multi-wavelength data for the whole LVHIS sample while in Section \[sec:analysis2\] we focus on a sub-sample of ten large LVHIS galaxies. A comparison of LVHIS with other [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surveys can be found in Section \[sec:scale\_relation\] compares LVHIS. In Section \[sec:dist\_HI\] we study the dependence of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} radial distributions on stellar properties. In Section \[sec:glob\_SFE\] we use resolved [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} properties to explore causes for the global star formation relations. In Section \[sec:loc\_SFE\] we investigate relations between SFR, [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} and stellar surface densities in the galactic outskirts. In Section \[sec:warp\_SFR\] we search for possible links between the onset of an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp and the radial change of SFE. In Section \[sec:summary\] we summarise the paper.
Throughout this paper, we assume a Kroupa IMF [initial mass function, @Kroupa01], and a $\Lambda$CDM cosmology with $\Omega_{m}=0.3$, $\Omega_{lambda}=0.7$ and $h=0.7$. Any quantities and trends taken from the literature are converted to be consistent with these assumptions before used in this paper. Because this paper extensively discusses correlations, we define here that a strong correlation should have Pearson correlation coefficient $\rho>$0.8, a significant correlation should have $\rho$ between 0.6 and 0.8, a weak$\slash$moderate correlation should have $\rho$ between 0.2 and 0.6, while no correlation corresponds to $\rho<$0.2. Errors for correlation coefficients are calculated through bootstrapping.
Data {#sec:data}
====
Sample and ATCA HI data {#sec:sample}
-----------------------
The Local Volume [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} Survey (LVHIS, Koribalski et al. in prep, Paper I) consists of a complete sample of 82 nearby galaxies ($D < 10$ Mpc) with a declination $\delta \lesssim -30{\ensuremath{^\circ}}$ selected from HIPASS [@Meyer04; @Koribalski04].
Using the Australia Telescope Compact Array (ATCA), sensitive [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} spectral line and 20-cm radio continuum data were obtained for all LVHIS galaxies. Large galaxies like M83, NGC 3621, Circinus and others were mosaiced to ensure that the extended [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} emission of the outer disc was not missed. Paper I presents the LVHIS [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} galaxy atlas (cubes and moment maps), based on the ‘natural’ weighted [*uv*]{}-data, and focused on the large-scale [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} emission in and around the selected galaxies. Typical angular resolution is B$_{\rm maj} \ga$40 (ie. $\sim$1 kpc at $D$ = 5 Mpc), where B$_{\rm maj}$ is the major axis of the synthesized beam. The typical r.m.s. sensitivity is $\sim$1.5 mJybeam$^{-1}$ per 4 km$/$s channel, which corresponds to a 5 r.m.s column density limit of 0.46$\times 10^{20}$ cm$^{-2}$ assuming a 20 km$/$s line width. This paper makes use of the primary beam corrected cubes and moment 0 images to derive [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} related properties. Results based on the LVHIS dataset have been presented in a series of publications, including @Koribalski09 [@vanEymeren10; @LopezSanchez12; @Kirby12; @Kamphuis15; @Johnson15] and @Wang16. A parallel study of the 20-cm radio continuum emission of the sample will be presented in Shao et al. (in prep). We take R$_{25}$ (the semi-major axis of the 25 mag arcsec$^{-2}$ isophote in the $B$-band) from the surface photometry catalogue of the ESO-Uppsala galaxies [@Lauberts89], and the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface density radial profiles and R$_{\rm HI}$ (the semi-major axis for the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} 1 $\mspc$ isophote) from @Wang16.
The LVHIS galaxy names along with some of their key properties derived throughout this paper are listed in Table \[tab:galprop\]. The table columns are as follows: (1) LVHIS ID, (2) HIPASS name, (3) galaxyÊname, (4) star formation rate (${\rm SFR = SFR_{FUV} + SFR_{W4}}$; Section \[sec:estimate\_sfr\]),Ê(5) stellar mass ($M_*$) derived from WISE luminosities (Section \[sec:estimate\_mst\]), (6) HI fraction ($\fHI=\mHI/M_*$), (7) ratio of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} to optical radii, (8) the averaged stellar surface densities within the W1 effective radius ($\Sigma_{\rm *,eff}$, Section \[sec:irphot\]), (9) the averaged SFR surface density within R$_{25}$ ($\Sigma_{\rm SFR,r<R25}$, Section \[sec:best\_res\]), (10) the averaged [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities within R$_{25}$, (11) the averaged [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities within 3.2 r$_{\rm s}$ of the W1 band (Section \[sec:irphot\]).ÊProperties in Columns 7, 10 and 11 are obtained only when the relevant radii are resolved in [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} (${\rm >b_{maj}}$). The tables presented in this paper have the same order as in Paper I and Shao et al. (in prep) and we also use ‘ID’ to denote the index (starting from 1). Figure \[fig:HIcontour\] shows an example multi-wavelength image of one of the 82 LVHIS galaxies.
From the LVHIS sample we also select a sub-sample of large galaxies for detailed multi-wavelength analysis (marked with [*l*]{} in Table \[tab:galprop\]). These, in total ten galaxies, are all the LVHIS galaxies that have an optical diameter, $D_{25}$, at least four times larger than the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} angular resolution and have GALEX FUV data available (Section \[sec:data\_uv\]). They include seven galaxies with Hubble types ranging between Sc and Sd, two Sm galaxies and one irregular dwarf galaxy. An atlas of these galaxies can be found in Appendix \[sec:appendix\_figure\]. The [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} discs are 15 to 73 times larger than the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} angular resolution. We obtain multi-wavelength radial profiles and tilted-ring [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} models for this subsample in Section \[sec:analysis2\].
We note that for analysis based on this sub-sample, we use the natural weighted [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} image from THINGS [The HI Nearby Galaxy Survey @Walter08] instead of the ATCA [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} image for the galaxy NGC 7793. This is because for NGC 7793, the THINGS [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data have much better resolution ($\sim$15.6 acsec) than the ATCA [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data ($\sim$364 arcses) while they do not significantly miss flux (its [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} flux 12% less than the HIPASS [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} flux, comparable to the 1$\sigma=15\%$ difference between ATCA and HIPASS fluxes for LVHIS, see Paper I) due to the lack of short baselines.
{width="14cm"}
GALEX {#sec:data_uv}
-----
The Galaxy Evolution Explorer [@Martin05 GALEX] provides near-ultraviolet (NUV, effective wavelength $\lambda\sim1528\AA$) and far-ultraviolet (FUV, $\lambda\sim2271\AA$) images over a large fraction of the sky. Each image has a pixel size of 15, and the typical full width half maximum (FWHM) of the PSF (point spread function) is 50 (FUV) and 55 (NUV). The provided magnitude zero-points are calibrated in the AB mag-system. Of the 82 LVHIS galaxies, 56 are covered by the GALEX imaging survey (data release 6+7) in the NUV band and 46 in the FUV band. The remaining 26 galaxies are missing because GALEX avoided the sky near the Galactic Plane.
The depth of the available GALEX images varies substantially: 22 galaxies have relatively deep images from the Medium Imaging Survey (MIS), with typical exposure times of $t_{\rm exp}>1500~s$ and a depth of 22.6 mag in FUV, 9 have ultra deep images from the Deep Imaging Survey (DIS), with $t_{\rm exp} > 10,000~s$ and a depth of 24.2 mag in FUV, while 15 have short exposures from the All-Sky Imaging Survey (AIS), with $t_{\rm exp}\sim100~s$ and a depth of 20 mag in FUV.
WISE
----
The Wide-field Infrared Survey Explorer [@Wright10 WISE] has mapped the whole sky in the 3.4, 4.6, 12 and 22 $\mu$m MIR (mid-infrared) bands, commonly referred to as W1, W2, W3 and W4. We make use of specially constructed WISE mosaics produced with ICORE @Masci13 [Image Co-addition with opitional Resolution Enhancement] as described in @Cluver14. These mosaics notably improve the angular resolution by $\sim$30% from the ALLWISE images [@Cutri12] published in the WISE public release archive, which were optimised for detecting point sources. They have a pixel size of 1 and a resolution of 59, 65, 70 and 124 for the four bands, respectively. The typical sky rms (1$\sigma$) is [@Jarrett12]. The magnitude zero-points are calibrated in the Vega mag system.
Other data {#sec:otherdata}
----------
Past studies have provided photometric data for several LVHIS galaxies. A comprehensive overview of the complimentary datasets is given in Table 1 of Paper I. Here we review a few datasets that are used for analysis or comparison in this paper.
As part of the Local Volume Legacy (LVL) project [@Kennicutt08; @Dale09], @Lee11 obtained asymptotic fluxes (a measure of the total flux, see Section \[sec:uvphot\]) in the GALEX bands for 390 local volume galaxies including 53 LVHIS galaxies. We will use this dataset to validate our GALEX flux measurements (Section \[sec:uvphot\]).
The Carnegie-Irvine Galaxy Survey [@Ho11 CGS] provides $B$, $V$, $R$ and $I$-band images for the 605 optically brightest galaxies in the southern sky, including 22 LVHIS galaxies. The field of view for each image is $8\farcm9 \times 8\farcm9$, and the typical seeing is 1. The details are described in @Ho11 and @Li11. The survey also provides a comprehensive set of photometric measurements including total fluxes, sizes, and radial surface brightness (SB) profiles. We will use this dataset to compare with our stellar mass ($M_*$) estimate for the high-mass galaxies (Section \[sec:estimate\]).
@Kirby12 obtained deep $H$-band (1.65$\mu$m) images with the 3.9-m Anglo Australian Telescope (AAT) and performed photometry for 57 galaxies in the Local Volume. @Young14 followed up another 40 dwarf galaxies. The field of view for each image is $7\farcm7 \times 7\farcm7$, and the typical seeing is 13. The two samples combined include 37 LVHIS galaxies. We will use this dataset to compare with our stellar mass ($M_*$) estimate for the low-mass galaxies (Section \[sec:estimate\]). The SINGG [Survey for Ionization in Neutral Gas Galaxies, @Meurer06] obtained H$\alpha$ and $R$-band images for 468 galaxies randomly selected from the HIPASS catalogues to uniformly sample the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass function. Its sister survey, SUNGG [Survey of Ultraviolet emission in Neutral-Gas Galaxies, @Wong16] used GALEX to image a subset of SINGG galaxies in the FUV and NUV bands. SINGG (SUNGG) have 19 (6) galaxies in overlap with LVHIS. We will compare our SFR and $M_*$ estimates with those obtained in the SINGG studies (Section \[sec:estimate\]). @Lee09 estimated SFRs for 300 local galaxies using data from the 11 Mpc H$\alpha$ and Ultraviolet Galaxy Survey (11HUGS). There are 29 galaxies in common for 11HUGS and LVHIS. We will compare our SFR estimates with those from @Lee09 for validation (Section \[sec:estimate\]).
Analysis for LVHIS galaxies {#sec:analysis}
===========================
In this section, we describe how we compute the total and radial distribution of fluxes from the WISE and GALEX images. We convert these measurements into physical properties and compare them to literature data.
GALEX photometry {#sec:uvphot}
----------------
We use the same pipeline described briefly in @Overzier11 to obtain photometric measurements for the LVHIS galaxies. Here we describe each step of the procedure:
1. Global background estimate and neighbour masking. We use SExtractor [@Bertin96] to produce a segmentation map for each image, where the pixels for all detected sources are labelled. We then exclude the pixels belonging to detected sources and calculate a global background value. Next we manually set this background value as input and run SExtractor again. Finally, we use the new segmentation map to mask sources around our galaxies.
2. Foreground star removal. For galaxies with available FUV images, we convolve these to match the slightly poorer resolution of the NUV images. Pixels associated with foreground stars are identified as having flux ratios NUV$/$FUV $> 5$ and NUV signal-to-noise (S/N) $>2$. For galaxies without FUV images, we inspect the NUV images and manually identify pixels containing bright point sources [^4]. We replace the pixels associated with foreground stars with pixels on the other side of the galaxy centre, or the average of pixels on the galaxy isophote, depending on whether or not the pixels on the other side are also associated with foreground stars. This step makes a clean image that will also be used in the resolution matched photometric analysis (Section \[sec:phot\_conv\]).
3. Radial profiles and residual background removal. We measure the radial profile of surface brightness (SB) along elliptical rings with a width of at least 1 pixel ($\sim1/3$ of the PSF size). For large galaxies, the spacing is increased to sample the radial profile with at most 50 data points. The ellipticity and position angle of the rings are fixed at the optical values of the 25 mag arcsec$^{-2}$ isophote. The SB is calculated as the 3-sigma clipped mean of the distribution of the pixel values in each elliptical ring. We identify the radius where the profile flattens within the noise, and use the annulus between one and 1.4 times this radius around the galaxy centre to estimate a local residual of the background. We remove this residual background and cut the radial profile at a S$/$N of 1.25 (equivalent to an SB error of 0.87 mag).
4. Growth curve and total flux. We measure the growth curve (GC) of flux enclosed by elliptical apertures out to 1.25 times the last radial point of the SB radial profile. We use the GC to obtain the asymptotic flux. Following the method outlined in @Lee11, we extrapolate the relation between the derivative of the GC and the GC, and the intercept is taken as the asymptotic flux. The error of the asymptotic flux is a combination of the photometric error and error from the extrapolation. We further add a zero-point error of 0.05 mag for FUV and 0.027 mag for NUV [@Morrissey07]. The asymptotic flux is taken as the total flux for GALEX bands in this paper.
Figure \[fig:uvphot\_example\] shows the intermediate products of this pipeline for two galaxies, one is bright and with deep imaging and the other is faint and with shallow imaging. We are able to reliably measure total fluxes (with an error less than 0.5 mag) for 44 and 54 galaxies in the FUV and NUV bands, respectively. We present these photometric measurements in Table \[tab:uvphot\]).
There are alternative methods to obtain the total fluxes. As adopted by @Huang12, fitting models (e.g. exponential functions) to SB radial profiles and extrapolating helps recover faint fluxes lower than the noise level, which is especially useful for low SB dwarf galaxies. Although the majority of the LVHIS sample are dwarf galaxies, we have chosen to use the asymptotic fluxes instead of the model fluxes. This is because nearly half of the LVHIS sample only have shallow AIS images (see Section \[sec:data\_uv\]) while @Huang12 selected galaxies with GALEX images of at least MIS-depth. Model fitting could be unreliable on our shallow images. To demonstrate the reason for our choice, in Figure \[fig:fluxes\_asy\_model\], we test the stabilities of the two methods on images of different depths. We take the DIS FUV image (with $t_{exp} = 12987.6~s$) of NGC 300 and pretend to observe it again with shorter $t_{\rm exp}$, i.e. by adding Poissonian noise to the counts. It is clear that when $t_{\rm exp}<1000~s$ (typical for AIS), model fluxes have large uncertainties and systematic uncertainties while asymptotic fluxes are much more stable.
As an external validation, we compare our asymptotic flux measurements with those from @Lee11 for the galaxies included in both studies (53 in total). GALEX magnitudes before data release 6 as presented in @Lee11 are 0.045 mag brighter in NUV and 0.033 mag brighter in FUV as compared to the most recent releases, due to a less accurate zero-point calibration. We have corrected for this effect before the comparison. As can be seen in Figure \[fig:LVHIS\_LVL\_flux\], the LVHIS and LVL measurements of FUV fluxes agree reasonably well, with a median value and scatter of 0.085 and 0.089 mag for the differences between the two. We have reviewed our photometry steps for the outliers in the comparison, and make sure that there are no obvious errors in our processing.
{width="13cm"} {width="4.5cm"}
{width="13.cm"} {width="4.5cm"}
![Stabilitiy of flux measurements at different image depths. Shallow images are simulated based on the deep GALEX FUV image of NGC 300 (see Section \[sec:uvphot\]). Asymptotic and model magnitudes measured from shallow simulated images (output) are compared to the magnitude measured from the original deep image (input). []{data-label="fig:fluxes_asy_model"}](plot/fluxes_asy_model.pdf){width="8cm"}
{width="7cm"}
\[fig:LVHIS\_LVL\_flux\]
WISE photometry {#sec:irphot}
---------------
We use the same pipeline outlined in detail in @Jarrett17 [upgraded upon the pipeline of Jarrett et al. 2012 and Jarrett et al. 2013] to perform photometric measurements on the WISE images. Here we briefly summarise the major steps of the procedure. For each image, the pipeline first removes nearby sources and produces a clean image. Then it obtains the aperture flux and SB radial profile within the 1 $\sigma$ (sky rms) isophote around the galaxy centre. A double S[' e]{}rsic model is fitted to the SB profile, and the final total flux is calculated as the sum of the aperture flux and the extrapolation of the S[' e]{}rsic model beyond the 1 $\sigma$ isophote out to three times the scale-length of the disc (the second component of the double S[' e]{}rsic model). Calibration and spectral shape related corrections are applied to the magnitudes as described in @Jarrett17. Structural parameters like position angle and axis ratio b$/$a (minor-to-major axis ratio) of the outmost isophote, effective radius r$_e$ (radius that enclose half of the total flux), effective surface brightness (the averaged surface brightness within r$_e$), and concentration indices R$_{75}/$R$_{25}$ (the ratio between radii that enclose 75% and 25% of the total fluxes) are also derived.
Figure \[fig:irphot\_example\] shows the intermediate products of this pipeline for two galaxies, NGC 300 and ESO 381-G020. We are able to reliably measure total fluxes (with errors less than 0.5 mag) for 81, 76, 65 and 57 galaxies in the W1–4 bands. We are unable to measure W1 fluxes for only one galaxy, J1131-31, because a foreground star outshines the whole galaxy. We further impose the W1 aperture on the W2 image for the five galaxies detected in W1 but not in W2, and obtain a rough measure of the W2 flux by summing all pixels within the aperture. As described in @Jarrett17, we obtain the rest-frame W1-4 fluxes based on these photometric fluxes. Especially, we remove the stellar continuum to obtain the W3 fluxes of the 11.3 $\mu$m PAH (polycyclic aromatic hydrocarbon) and the W4 fluxes originated from warm and cold small dust grains. These rest-frame and stellar continuum removed fluxes are obtained through comparing the WISE spectral energy distribution (SED) of each galaxy with extragalactic population SED templates built in @Brown14 based on the SWIRE$\slash$GRASIL models [@Polletta06; @Polletta07; @Silva98]. These corrected fluxes are more directly linked to physical quantities (e.g. $M_*$ and SFR) than the photometric fluxes.
We present these WISE rest frame luminosities and structural parameters in Table \[tab:irphot\].
{width="12.cm"} {width="5.5cm"}
{width="12.cm"} {width="5.5cm"}
Estimate of $M_*$ and SFR {#sec:estimate}
-------------------------
### Estimating $M_*$ {#sec:estimate_mst}
Using the method of @Jarrett12, we calculate the W1 band stellar mass-to-light ratio based on the W1-W2 colour to account for the stellar population dependence. Similar to the treatment in @Jarrett17, we put a floor/ceiling limits on the W1-W2 colour, -0.05 to 0.2 mag, the colour range where the @Jarrett12 formula was calibrated. It also minimises AGN contamination and the systematics caused by the less sensitive W2 band. Then we calculate $M_*$ for 81 LVHIS galaxies based on W1 luminosities.
The @Jarrett12 method was calibrated using nearby bright spiral galaxies, so there might be a worry that it fails for low-mass irregular galaxies. Hence we compare our WISE-based $M_*$ with the $M_*$ derived from optical and $H$-band data in Figure \[fig:comp\_lgm\]. The CGS $M_*$ is estimated with the $B-R$ colour dependent $R$-band mass-to-light ratios based on the formula in @Bell03. The $H$-band based $M_*$ is taken from @Kirby12 and @Young14 and was estimated with a fixed mass-to-light ratio of 0.9. Finally, the SINGG $M_*$ is estimated with a $R$-band luminosity dependent R-band mass-to-light ratio as discussed in @Wong16. The median and scatter of the differences in $M_*$ are –0.04 and 0.23 dex respectively for the comparison between WISE and CGS (19 galaxies), –0.12 and 0.20 dex respectively for the comparison between WISE and $H$-band (22 galaxies), and 0.06 and 0.17 dex for the comparison between WISE and SINGG (13 galaxies). The scatter is consistent with the typical errors of $M_*$-to-light ratios with and without accounting for dependence on stellar populations [@Bell03]. As can be seen from Figure \[fig:comp\_lgm\], CGS contains the massive LVHIS galaxies, the $H$-band survey focusses on dwarf galaxies, while SINGG connects the two populations. The WISE-based $M_*$ estimates in this paper are in excellent agreement with the $M_*$ estimates from optical and near-infrared luminosities for both dwarf and massive galaxies.
### Estimating SFR {#sec:estimate_sfr}
We estimate the dust un-attenuated part of the total SFR based on FUV luminosities (SFR$_{\rm FUV}$), and recover the dust attenuated part based on W4 luminosities (SFR$_{\rm W4}$). This method is widely adopted in the literature [@Calzetti13; @Hao11] and based on the fact that FUV fluxes are primarily contributed by the massive B stars which have a typical age of 100 Myr, and part of the FUV fluxes are absorbed and re-emitted by cool and warm dust in the far and mid-infrared, while W4-band traces the warm dust with a temperature $\sim$150 K.
We calculate SFR$_{\rm FUV}$ with the equation from @Calzetti13 [their table 1.2], which is derived from stellar population models (dust-free). We calculate SFR$_{\rm W4}$ with equation 2 from @Jarrett13, which is derived from the infrared SED models of starburst galaxies (SFR fully attenuated). Then total . Because FUV data are not always available while W4 fluxes are not always detectable for all the galaxies, we use the W4 estimated SFR as a lower limit when only W4 data is available, and assume the dust attenuation to be zero when we get no detection from the W4 data.
We also caution that while we have tried our best to remove the contribution from stellar continuum (Section \[sec:irphot\]), W4 band as an SFR indicator has uncertainties for being a far extrapolation of the dominant bolometric emission that arises from cool dust in the far-infrared, and it has risks of being contaminated by the warm and dusty TP-AGB populations. Hence for validation we compare in Figure \[fig:comp\_sfr\] our SFR with SFR derived in other ways (see Section \[sec:otherdata\]). The 11HUGS (FUV+TIR) SFR [from @Lee09] was estimated in a similar way as in this paper, but the more accurate attenuation tracer, total infrared luminosities (TIR) instead of 22 $\mu$m were used to indicate the dust attenuated SFR. The 11HUGS (H$\alpha$) SFR [also from @Lee09] was derived with the H$\alpha$ luminosities with a dust attenuation correction based on the Balmer decrement H$\alpha/$H$\beta$. The SINGG (H$\alpha$) SFR is estimated from the H$\alpha$ luminosities with a dust attenuation correction determined by the R-band absolute magnitude [@Meurer06].
The LVHIS SFRs agree well with these three datasets, with a median difference of -0.08$\pm$0.09, 0.07$\pm$0.36 and -0.02$\pm$0.17 dex from the 11HUGS (FUV+TIR), 11HUGS (H$\alpha$) and SINGG (H$\alpha$) SFRs. When SFR$<$0.01 $\ms$ yr$^{-1}$, 11HUGS (H$\alpha$) SFRs tend to be systematically lower than our SFRs, which might be caused by insufficient samplings of the massive end of the initial mass function (IMF), or by a different IMF shape at the massive end in the low gas surface density regions [@Meurer06; @Lee09; @LopezSanchez15]. On the other hand, there is no systematic offset between LVHIS and SINGG (H$\alpha$) SFRs at the low SFR end.
![Comparison of WISE-based $M_*$ (this work) with the estimates from other published data (see Section \[sec:estimate\_mst\]). The dashed line marks the one-to-one relation. Typical error bars are shown in the left-top corner. Measurements from different datasets have been adjusted to have the same luminosity distance and @Kroupa01 IMF before comparison. []{data-label="fig:comp_lgm"}](plot/comp_lgm.pdf){width="8cm"}
![Comparison of FUV+W4 based SFR (this work) with the estimates from other published data (see Section \[sec:estimate\_sfr\]). The dashed line marks the one-to-one relation. Typical error bars are shown in the left-top corner. Measurements from different datasets have been adjusted to have the same luminosity distance and @Kroupa01 IMF before comparison. []{data-label="fig:comp_sfr"}](plot/comp_lgsfr.pdf){width="8cm"}
\[tab:galprop\]
-------- --------------- ----------------- ---------------------- --------------- ----------------- ----------------------------- --------------------------------- ----------------------------------- ------------------------------- ---------------------------------
[ID]{} [HIPASS ID]{} [Galaxy Name]{} SFR [log $M_*$]{} [$\log \fHI$]{} [${\rm R_{HI}/R_{25}}^r$]{} [$\log \Sigma_{{\rm *,eff}}$]{} [$\log \Sigma_{\rm SFR,r<R25}$]{} [$\Sigma_{\rm HI,r<R25}^r$]{} [$\Sigma_{\rm HI,r<3.2rs}^r$]{}
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
1 J0008-34 ESO349-G031 0.0006$\pm$0.0001 7.3 -0.22 - 6.81 -2.92 - 0.35
2 J0026-41 ESO294-G010 0.0001$\pm$0.0001 6.7 - - 7.36 -3.16 - -
3 J0015-32 ESO410-G005 0.0001$\pm$0.0000 6.2 -0.32 - 6.72 -3.67 - 0.35
4$^l$ J0015-39 NGC55 0.1896$\pm$0.0220 9.4 -0.03 1.26 8.01 -3.14 6.53 3.39
5$^l$ J0054-37 NGC300 0.1598$\pm$0.0210 9.5 -0.21 2.30 7.91 -2.86 5.44 4.67
6$^l$ J0047-25 NGC253 3.8612$\pm$0.5582 10.4 -0.99 0.82 8.84 -2.38 2.94 7.13
7 J0047-20 NGC247 0.1374$\pm$0.0206 9.7 -0.42 1.33 7.92 -3.37 4.74 1.70
8 J0135-41 NGC625 0.0455$\pm$0.0051 8.7 -0.75 0.77 8.00 -2.97 1.83 2.29
9 J0145-43 ESO245-G005 0.0249$\pm$0.0037 8.3 0.27 2.60 7.15 -2.74 7.73 2.65
10 J0150-44 ESO245-G007 0.0000$\pm$0.0000 6.8 -1.85 - 6.73 -3.45 - 0.03
11$^l$ J0237-61 ESO115-G021 0.0133$\pm$0.0020 8.1 0.76 2.25 7.28 -3.48 5.09 3.72
12 J0256-54 ESO154-G023 0.0366$\pm$0.0055 8.7 0.27 - 7.32 - - -
13 J0258-49 ESO199-G007 0.0015$\pm$0.0004 7.0$^h$ 0.23 - - -3.02 - 0.62
14$^l$ J0317-66 NGC1313 0.5164$\pm$0.0642 9.5 -0.20 1.65 7.83 -2.40 8.45 8.31
15 J0320-52 NGC1311 0.0128$\pm$0.0019 8.2 -0.29 1.18 7.69 -3.15 3.80 2.22
16 J0321-66 AM0319-662 0.0001$\pm$0.0001 7.1 -1.14 - - -4.35 - -
17 J0333-50 IC1959 0.0261$\pm$0.0036 8.5 -0.19 1.55 8.02 -3.00 5.08 4.37
18 J0454-53 NGC1705 0.0481$\pm$0.0069 8.2 -0.37 - 8.15 -2.06 - 2.33
19 J0457-42 ESO252-IG001 - 7.3 0.76 2.11 6.77 - - 2.95
20 J0605-33 ESO364-GQ029 - 7.6 0.90 2.14 6.46 - - 2.57
21 J0607-34 AM0605-341 - 7.9 0.18 - 7.80 - - -
22 J0610-34 NGC2188 $>$0.0211$\pm$0.0036 8.9 -0.28 0.91 7.92 $>$-3.71 3.91 1.66
23 J0615-57 ESO121-G020 0.0033$\pm$0.0006 7.2 0.58 2.92 6.96 -2.66 - 1.38
24 J0639-40 ESO308-G022 - 7.8 -0.04 4.54 7.01 - - -
25 J0705-58 AM0704-582 - 6.2 2.09 - 6.76 - - -
26 J0731-68 ESO059-G001 - 8.1 -0.16 2.49 7.25 - 4.41 1.03
27 J0926-76 NGC2915 0.0265$\pm$0.0038 8.4 0.10 4.26 8.22 -2.32 4.64 2.27
28 J1043-37 ESO376-G016 - 8.0 0.07 4.00 7.26 - - -
29 J1047-38 ESO318-G013 0.0057$\pm$0.0017 7.4 0.60 - 6.85 -3.35 - 0.62
30 J1057-48 ESO215-GQ009 0.0442$\pm$0.0067 8.0 0.89 17.38 6.92 -1.19 - 4.92
31$^l$ J1118-32 NGC3621 0.7127$\pm$0.0775 9.7 0.28 3.44 8.28 -2.60 4.15 3.18
32 J1131-31 new - - - - - - - -
33 J1132-32 new - 7.7 -0.70 - 7.04 - - -
34 J1137-39 ESO320-G014 - 8.1 -0.84 - 7.26 - - 0.47
35 J1154-33 ESO379-G007 - 7.3 0.18 - - - - -
36 J1204-35 ESO379-G024 - 7.5 -0.37 - 7.28 - - -
37 J1214-38 ESO321-G014 0.0010$\pm$0.0001 7.0 0.06 1.79 7.06 -3.12 - 0.97
38 J1219-79 IC3104 - 8.2 -1.19 1.18 7.71 - 4.08 0.37
39 J1244-35 ESO381-G018 - 7.2 0.04 - 7.31 - - 1.51
40 J1246-33 ESO381-G020 0.0097$\pm$0.0015 7.4 0.95 2.37 6.67 -3.06 - 2.33
41 J1247-77 new - 7.1 -0.07 - 6.94 - - 1.39
42 J1305-40 CEN06 0.0016$\pm$0.0005 7.8 -0.20 - 6.91 -3.57 - 0.53
43 J1305-49 NGC4945 $>$0.9566$\pm$0.1455 10.1 -1.01 0.90 8.58 $>$-2.74 2.56 5.39
44 J1310-46 ESO269-G058 0.0018$\pm$0.0004 8.5 -1.21 0.66 7.78 -3.81 1.60 0.62
45 J1321-31 new 0.0005$\pm$0.0001 - - - - -3.97 - -
-------- --------------- ----------------- ---------------------- --------------- ----------------- ----------------------------- --------------------------------- ----------------------------------- ------------------------------- ---------------------------------
$^h$: $H$-band photometry based.\
$^l$: Mark the sub-sample of large galaxies for detailed analysis (see Section \[sec:sample\] and \[sec:analysis2\]).\
\[tab:galprop\]
-------- --------------- ----------------- ---------------------- ----------------- ----------------- --------------------------- --------------------------------- ----------------------------------- ----------------------------- -------------------------------
[ID]{} [HIPASS ID]{} [Galaxy Name]{} SFR [log $M_*^e$]{} [$\log \fHI$]{} [${\rm R_{HI}/R_{25}}$]{} [$\log \Sigma_{{\rm *,eff}}$]{} [$\log \Sigma_{\rm SFR,r<R25}$]{} [$\Sigma_{\rm HI,r<R25}$]{} [$\Sigma_{\rm HI,r<3.2rs}$]{}
( 1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
46 J1321-36 NGC5102 - 9.5 -1.11 1.14 8.68 - 1.77 1.39
47 J1324-30 AM1321-304 0.0004$\pm$0.0006 7.9 -0.95 - 7.15 -3.99 - 0.65
48 J1324-42 NGC5128 0.7782$\pm$0.1038 11.0 - - 9.00 -3.05 - -
49 J1326-30 IC4247 0.0037$\pm$0.0006 7.5 -0.21 - 7.42 -2.97 - 1.12
50 J1327-41 ESO324-G024 0.0118$\pm$0.0019 8.0 0.20 2.36 7.23 -2.85 6.97 3.69
51 J1334-45 ESO270-G017 $>$0.0053$\pm$0.0012 8.7 0.10 1.36 7.41 $>$-4.33 4.02 3.34
52 J1336-29 UGCA365 0.0010$\pm$0.0003 7.2 0.25 - 7.01 -3.39 - -
53$^l$ J1337-29 NGC5236 2.9211$\pm$0.3408 10.5 -0.59 2.62 8.50 -2.11 4.25 3.92
54 J1337-39 new 0.0012$\pm$0.0018 - - - - -3.55 - -
55 J1337-42 NGC5237 0.0031$\pm$0.0007 8.3 -0.81 0.89 8.11 -3.07 1.99 1.03
56 J1337-28 ESO444-G084 0.0031$\pm$0.0005 6.9 1.00 2.89 6.75 -2.89 - 3.12
57 J1339-31 NGC5253 0.3606$\pm$0.0471 8.6 -0.66 0.95 8.38 -1.79 3.74 10.06
58 J1340-28 IC4316 0.0026$\pm$0.0004 7.9 -0.83 - 7.38 -2.26 - 0.33
59 J1341-29 NGC5264 0.0099$\pm$0.0015 8.7 -0.96 1.01 7.71 -3.20 2.65 0.43
60 J1345-41 ESO325-GQ011 - 7.1 0.72 1.79 6.71 - 6.77 2.79
61 J1348-37 new - - - - - - - 0.29
62 J1348-53 ESO174-GQ001 - 7.6 0.57 - 7.65 - - 2.36
63 J1349-36 ESO383-G087 $>$0.0022$\pm$0.0005 8.9 -1.09 0.92 7.85 $>$-4.07 2.47 0.63
64 J1351-47 new - - - - - - - -
65 J1403-41 NGC5408 $>$0.0197$\pm$0.0031 8.2 0.19 2.70 7.73 $>$-2.79 - 3.88
66 J1413-65 Circinus $>$3.0452$\pm$0.4645 10.1 -0.26 5.63 - $>$-1.46 3.72 3.88
67 J1428-46 UKS1424-460 - 7.5 0.18 - 7.04 - - 3.71
68 J1434-49 ESO222-G010 - 7.8$^h$ -0.09 - - - - 0.96
69 J1441-62 new - 7.8 -0.47 - 8.16 - - -
70 J1443-44 ESO272-G025 - 8.0 -0.93 - 7.72 - - -
71 J1501-48 ESO223-G009 - 8.9 0.13 - 7.67 - - 5.32
72 J1514-46 ESO274-G001 $>$0.0101$\pm$0.0016 9.0 -0.52 1.26 7.96 $>$-3.78 3.67 3.94
73 J1526-51 new - 6.9 0.66 - - - - -
74 J1620-60 ESO137-G018 - 8.4 0.19 1.36 7.55 - 8.19 0.84
75 J1747-64 IC4662 0.0485$\pm$0.0058 8.0 0.15 4.68 7.92 -1.59 - 8.48
76 J2003-31 ESO461-G036 - 8.1 -0.12 5.12 7.09 - - 1.73
77$^l$ J2052-69 IC5052 0.0613$\pm$0.0068 9.1 -0.20 1.33 8.02 -3.30 3.92 2.42
78$^l$ J2202-51 IC5152 0.0161$\pm$0.0022 8.2 -0.20 1.85 7.88 -2.59 8.03 4.73
79 J2326-32 UGCA438 0.0013$\pm$0.0002 6.9 -0.25 - 6.86 -2.83 - 0.37
80 J2343-31 UGCA442 0.0080$\pm$0.0012 7.3 1.01 1.81 6.95 -3.33 7.02 3.47
81 J2352-52 ESO149-G003 0.0073$\pm$0.0011 7.4 0.36 - 7.52 -3.14 2.12 -
82$^l$ J2357-32 NGC7793 0.2526$\pm$0.0307 9.5 -0.53 1.30 8.03 -2.64 6.22 4.22
-------- --------------- ----------------- ---------------------- ----------------- ----------------- --------------------------- --------------------------------- ----------------------------------- ----------------------------- -------------------------------
\[tab:galprop\]
Analysis for the large LVHIS galaxies {#sec:analysis2}
=====================================
Resolution-matched multi-wavelength photometry {#sec:phot_conv}
----------------------------------------------
We have selected ten well-resolved LVHIS galaxies for an analysis of radial distributions of multi-wavelength light. We use the pipeline of @Wang10 to perform resolution-matched photometric measurements from the GALEX, WISE and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} images. Instead of using the original GALEX and WISE images, we use the cleaned images (see Section \[sec:uvphot\] and \[sec:irphot\]). We remove the background and set all pixels beyond the ellipse defined by the last measured point in the SB profile to zero. By doing so, we minimise possible contamination from nearby sources in the following step of resolution matching.
Because the PSFs of WISE images are known to have significant side-lobes, we convolve the WISE cleaned images with the kernels from @Aniano11, which convert the WISE PSFs to PSF shapes described by Gaussian functions. After this convolution process, the W1, 2 and 4 images have Gaussian-shaped PSFs with FWHM of 7.5, 8.0 and 15.0 arcsec, respectively.
For each galaxy the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} image, which has the lowest resolution, is used as the reference image. All other images are registered and convolved to have the same geometry and PSF resolution as the reference image. The resolution matched images for the whole sub-sample have a physical spatial resolution between 0.3 to 2.5 kpc, and a median value of 1.5 kpc. We show the stellar mass and SFR images derived based on these images, as well as the relation matched [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} images in Figure \[fig:atlas\_conv\].
We measure radial SB profiles from these images. SB is azimuthally averaged in elliptical rings with a position angle and axis ratio determined from the WISE W1 images (Section \[sec:irphot\]). We apply a projection correction to SB by multiplying them with the axis ratio b$/$a (assuming very thin discs). The major axis of each ring is taken as the profile radius. We limit the outermost radius of each profile to be the same as the profile measured from the original image before convolution, because we have removed all pixels outside that radius before the image resolution matching step. We use the background noise from the original images to estimate photometric errors, as the noise has become correlated in the resolution-matched images.
Finally, $M_*$ and SFR profiles are calculated based on these SB profiles. When deriving $M_*$, we use the W1 SB profile and a fixed mass-to-light ratio determined by the total W1$-$W2 colour, because the W2 SB profiles do not extend as far as the W1 SB profiles in most of the galaxies while our study focuses on the outskirts of galaxies. When deriving SFR profiles, the W4 SB profiles also end at much smaller radius than the FUV SB profiles, so we exponentially extrapolating the SFR$_{\rm W4}$ profiles in order to obtain total SFR profiles. In most galaxies, the extrapolated SFR$_{\rm W4}$ profile in the outer region is lower than the SFR$_{\rm FUV}$ profile, which alleviates the uncertainties caused by extrapolation.
We obtained the surface densities ($\Sigma$) at the radius where $\SHI=3~ \ms {\rm pc^{-2}}$, where $\Sigma_*=3~ \ms {\rm pc^{-2}}$ and at $r$ = 1.25 R$_{25}$. This allows us to study the $\Sigma_{\rm SFR}$-$\SHI$ relation when fixing $\Sigma_*$ at $3~ \ms {\rm pc^{-2}}$, the $\Sigma_{\rm SFR}$-$\Sigma_*$ relation when fixing $\SHI$ at $3~ \ms {\rm pc^{-2}}$, and both relations at 1.25 R$_{25}$. These investigations will be presented in Section \[sec:loc\_SFE\]. When the radial positions are outside the radial range of a profile (in all cases, $\Sigma_*$ profiles), we obtain the surface densities through linearly extrapolating in logarithmic space the exponential outer part of the radial profiles. Most of the extrapolated radii are within 1.2 times the last data point of the profiles. The error of an extrapolated surface density is a combination of the extrapolated measurement error and the f-test based model error.
These resolution-matched $\mHI$, $M_*$ and SFR radial profiles and measurements are presented in the last column of Figure \[fig:atlas\_conv\].
Regarding extrapolating the $\Sigma_*$ profiles, one may worry about the fact that optical and mid-infrared radial profiles frequently bend up or downward in the galactic outskirts [@Pohlen06]. However, there is evidence that $\Sigma_*$ radial profiles are much less bent than the surface brightness profiles [@Bakos08], and we are not extrapolating far away from the last measurable data points. These two facts to some extent alleviate the problem of bending profiles. Nevertheless, optical or MIR images that are deep, of high-resolution (to suppress confusion noise), and of large field-of-view (to homogeneously cover the outskirts of the galaxies) are needed for a future confirmation of our results. Such kind of data will be available with the completion of wide-field optical imaging surveys, like the Skymapper Southern Sky Survey [@Keller07].
HI tilted ring models {#sec:tiltedring}
---------------------
### Building the HI tilted ring models
We make use of Fully Automated TiRiFiC [@Kamphuis15 FAT], which builds on the 3D Tilted Ring Fitting Code [TiRiFiC @Jozsa07 J07 hereafter], to derive tilted ring models and rotation curves for the large LVHIS galaxies. The approaching and receding velocity sides are modelled separately. Part of the sample has already been processed by @Kamphuis15 using FAT, including NGC 253, NGC 1313 and IC5152. We present the model rotation curves and radial profiles of position angles, inclinations and surface densities in Figure \[fig:atlas\_highrs\] in the appendix.
In Figure \[fig:atlas\_highrs\], we can see that at the optical radius R$_{25}$, the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} and optical inclinations usually agree within 20${\ensuremath{^\circ}}$. At R$_{25}$, PA (position angle) from the two bands also agree within 20${\ensuremath{^\circ}}$ for most of the galaxies, except for NGC 5236 (PA from the two bands differ by $\sim60{\ensuremath{^\circ}}$). Tilted ring models mostly provide a fit considering kinematics, while in the optical they are not taken into account. Below an inclination of 40${\ensuremath{^\circ}}$ it becomes very difficult to accurately measure PAs without kinematical information, which might explain the discrepancy between the optical and kinematical PAs for NGC 5236.
If we assume the extent of fluctuations to reflect fitting uncertainties, the 3D de-projected radial [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} profiles (as an output of FAT) generally agree well with the radial [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} profiles which are measured and corrected for projection effects from the 2D moment-0 images (see Section \[sec:sample\] and @Wang16).
We use the model results on both the approaching and receding velocity sides of galaxies to analyse [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp related properties (see next section), and use the average of the two sides to study the relation between SFR and mass radial surface densities (Section \[sec:best\_res\] and \[sec:loc\_SFE\]).
### Tiltograms and identification of warps
As summarised in the introduction of @Jozsa07b, an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp is the inner plane of an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} disc kinematically bending in the outer region, and very often transiting into another plane. @Jozsa07b demonstrated that the “tiltograms” are very useful for identifying [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps.
As outlined in @Jozsa07b, a tiltogram is a pixel map showing the mutual inclination angle ($\theta_{tilt}$) between rings at different radius. The pixel value ($\theta_{tilt}$) at x$=$r1 and y$=$r2 in the map measures the angle between the normal vectors of the tilted rings at radius r1 and r2. The direction of the normal vector for each tilted ring is fully described by the position angle and inclination angle of the tilted ring. A completely flattened disc plane has a tiltogram with zero value in all pixels.
A tiltogram can be divided into four regions at a radius $r$, where the bottom-left square region with $x<r$ and $y<r$ shows the intrinsic mutual inclination of the disc within $r$ ($\theta_{tilt}^{in}$), the top-right square region with $x>r$ and $y>r$ shows the intrinsic mutual inclination of the disc outside $r$ ($\theta_{tilt}^{out}$), and the remaining two rectangles (which are identical to each other) showing the mutual inclination between the discs within and beyond $r$ ($\theta_{tilt}^{mut}$).
If a galaxy has an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp starting from radius $r_{warp}$, $r_{warp}$ will divide the tiltogram into four regions as described above, where $\theta_{tilt}^{in}$ are close to zero and nearly constant, while $\theta_{tilt}^{mut}$ are large. When an outer disc plane exists beyond $r_{warp}$, $\theta_{tilt}^{out}$ are also close to zero and nearly constant. So we calculate the average $\theta_{tilt}$ ($\overline{\theta_{tilt}}$) in the four regions divided by each radius $r$, and set the criteria that an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} disc is classified as warped if a radius $r_{warp}$ can be found where
1. $|\overline{\theta_{tilt}^{in}}-\overline{\theta_{tilt}^{mut}}|$ is a local maximum,
2. $\overline{\theta_{tilt}^{in}} < 5$ deg, and
3. $\overline{\theta_{tilt}^{mut}} > 10$ deg.
$\overline{\theta_{tilt}^{in}}$ indicates the flatness of the inner disc, $\overline{\theta_{tilt}^{mut}}$ indicates the warp angle, and $\overline{\theta_{tilt}^{out}}$ indicates the flatness of the outer disc.
We set the criteria in a relatively arbitrary way, which can be biased toward selecting the strongly warped [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} discs. But it is sufficient for our scientific aim of investigating the influence of warps on SFE, as strong warps should have stronger effects than weak warps.
Because a galaxy can be warped on only one side of the galaxy, we analyse the approaching and receding side of the velocity field seperately. We display the tiltograms for the approaching velocity side of the LVHIS large galaxies in the fifth column of Figure \[fig:atlas\_highrs\]. The tiltogram of NGC 300 is a perfect example showing warp features, where $r_{warp}$ is marked by purple lines. We notice from the tiltograms that sometimes one ring deviates abruptly from adjacent rings (e.g. at the radii 7.5 in IC5052). This kind of feature occurs when the tilted rings warping dramatically from one side to the other side of the inner plane.
We list the $\overline{\theta_{tilt}}$ and $r_{warp}$ measurements in Table \[tab:warp\_app\] for the approaching and receding velocity side respectively. We find [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps on the approaching velocity side in seven out of the ten large LVHIS galaxies except for NGC 253, ESO115-G021 and NGC 1313 for which $\overline{\theta_{tilt}^{mut}}$ are too small. The same result is found on the receding velocity side.
----------------- --------------------------------- ---------------------------------- ---------------------------------- -------------- --------------------- --------------------------------- ---------------------------------- ---------------------------------- -------------- ---------------------
Galaxy $\overline{\theta_{tilt}^{in}}$ $\overline{\theta_{tilt}^{out}}$ $\overline{\theta_{tilt}^{mut}}$ $r_{warp}^1$ $r_{warp}/$R$_{25}$ $\overline{\theta_{tilt}^{in}}$ $\overline{\theta_{tilt}^{out}}$ $\overline{\theta_{tilt}^{mut}}$ $r_{warp}^1$ $r_{warp}/$R$_{25}$
name \[deg\] \[deg\] \[deg\] \[arcmin\] \[deg\] \[deg\] \[deg\] \[arcmin\]
NGC 55 2.6 10.3 21.7 14.7 0.93 3.2 4.9 12.6 20.0 1.36
NGC 300 0.9 9.4 20.5 10.2 1.05 0.6 8.3 17.8 10.2 1.05
NGC 253$^*$ 0.0 1.1 2.3 4.2 0.28 0.0 1.0 2.1 4.2 0.28
ESO115-G021$^*$ 0.2 5.7 8.9 3.7 1.50 0.4 4.7 7.2 3.5 1.42
NGC 1313$^*$ 0.3 4.2 3.8 2.2 0.41 0.2 6.8 7.5 2.7 0.50
NGC 3621 1.4 5.5 15.4 12.4 2.56 3.6 5.5 21.0 14.7 3.02
NGC 5236 3.0 7.3 22.6 10.0 1.30 1.8 20.5 29.5 6.9 0.89
IC5052 2.0 17.2 18.6 3.1 0.88 0.9 7.4 14.5 3.4 0.95
IC5152 0.0 10.0 17.5 1.9 0.76 0.2 8.8 17.3 2.0 0.82
NGC 7793 3.1 5.1 17.5 6.3 1.22 1.6 3.6 11.1 6.2 1.19
----------------- --------------------------------- ---------------------------------- ---------------------------------- -------------- --------------------- --------------------------------- ---------------------------------- ---------------------------------- -------------- ---------------------
$^1$: the radius identified based on the warp identification criteria 1.\
$^*$: galaxies that do not meet the warp identification criteria 2 and 3 in Section \[sec:tiltedring\]. As we have adopted relatively strict criteria that are biased toward strong warps, we provide warp related parameters for these un-warped galaxies to show the full parameter space of the sample.
\[tab:warp\_app\]
### Dark matter and stellar volume densities {#sec:best_res}
In Section \[sec:loc\_SFE\], motivated by the mid-plane pressure models for star formation [@Ostriker10; @Krumholz13], we will study the dependence of SFR surface densities on the volume densities of stars and dark matter added together ($\sigma_{\rm sd}$). In order to obtain a rough estimate of the parameter $\sigma_{\rm sd}$, we make use of the tilted ring models derived in Section \[sec:tiltedring\], which have been de-convolved with their PSF. So when we derive the $\Sigma_*$ and $\Sigma_{\rm SFR}$ radial profiles, we perform a resolution matched photometry as in Section \[sec:phot\_conv\], but at the resolution of the W4 images.
We use the radial [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} profiles from the tilted ring models. We use Equation 3 provided by @Bigiel08 to estimate the molecular hydrogen surface density $\SH2$ from $\Sigma_{\rm SFR}$. The atomic and molecular hydrogen gas surface densities are multiplied by 1.36 to account for the helium gas. We use the sum of gas and stellar surface densities to approximate the baryonic mass densities. We use the rotation curves to estimate the total mass at each radius ($M_{\rm tot}$). We subtracted from $M_{\rm tot}$ the contributions of stars and gas, and obtain the mass of the dark matter halo. We then assume a simple spherical model to calculate the volume density of the dark matter halo, $\sigma_{\rm dm}$. These calculations are crude, while a more detailed decomposition of the different masses is beyond the scope of this paper.
Following the method in @Leroy08, we assume a fixed ratio of 7.3 between the scale-length and the scale-height ($h$) for the stellar discs, and calculate the volume density of stars as $\sigma_*=\Sigma_*/(4h)=\Sigma_*/(0.54 {r_s})$. Finally $\sigma_{\rm sd}=\sigma_*+\sigma_{\rm dm}$.
LVHIS galaxies and the HI radial distributions
==============================================
We have built a multi-wavelength dataset for the LVHIS sample, based on which we have derived $\mHI$, $M_*$ and SFR. In this section, we compare the distribution of these parameters with other surveys to understand the selection effect and the advantage of LVHIS. We also investigate how [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} is radially distributed with respect to the optical discs and how it depends on the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-richness of galaxies. By assessing the diversity of radial [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} distributions, we better understand what new information is gained with interferometry data, and why they are useful for understanding global scaling relations in the next section.
LVHIS compared to other HI samples {#sec:scale_relation}
----------------------------------
This section compares the the $\fHI$ and SFR distributions of the LVHIS sample with other [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} samples.
In Figure \[fig:surveys\], we can see that $M_*$ of LVHIS ranges from 10$^6~\ms$ to 10$^{10.5}~\ms$ and has a median $\sim 10^8~\ms$. The sample covers nearly two orders of magnitude in $\fHI$ ($=\mHI/M_*$) at a fixed $M_*$. For example, at $M_*\sim10^8~\ms$, $\fHI$ ranges from -1.2 to 1 dex.
The left panel of Figure \[fig:surveys\] compares LVHIS with several single-dish [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} datasets. We can see that most of the LVHIS galaxies have lower $M_*$ than the median $M_*$ of ALFALFA, GASS and HRS. Their $\fHI$ are typically lower the median $\fHI$ of ALFALFA at a fixed $M_*$, and comparable to an extrapolation of the $\fHI$-$M_*$ distribution to the lower $M_*$ side for GASS and HRS.
In the right panel of Figure \[fig:surveys\], we compare LVHIS to several [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} interferometry surveys. We can see that both LVHIS and WHISP tend to connect the existing resolved [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} samples of irregular galaxies (like LITTLE THINGS and FIGGS) and massive spiral galaxies (like THINGS) in the $\fHI$-$M_*$ space. WHISP provides a wider $M_*$ distribution than LVHIS, while LVHIS detects galaxies with on average lower $\fHI$ at a fixed $M_*$ than WHISP. On the other hand, the LVHIS large galaxies sample, in comparison to the rest LVHIS galaxies, are biased toward systems with high $\fHI$ for their $M_*$.
These comparisons suggest LVHIS as (or close to) a volume-limited [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} sample covers a wide range of both $M_*$ and $\fHI$. It is not significantly biased to high $\fHI$ galaxies compared to single dish samples, hence has the ability to explore the cause for global scaling relations using spatially resolved [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} distributions; it includes both spiral and irregular galaxies, hence has the ability to explore general physics at work in these two types of galaxies. This paper will make use of these advantages.
In Figure \[fig:SFMS\], we show that most of the LVHIS galaxies lie below the star-forming Main Sequence (MS) from SDSS [@Renzini15] and also below the mean SFR-$M_*$ relations for ALFALFA [@Huang12b], WISE-GAMA [@Jarrett17] and WISE-GALEX [@Grootes13]. This is consistent with the result that most LVHIS galaxies lie below the mean $\fHI$-$M_*$ relation for ALFALFA, which is close to the MS of low redshift galaxies [@Huang12b]. It confirms the previous finding that the SFR-$M_*$ and $\fHI$-$M_*$ relations mimic one another [@Saintonge16; @Brown15]. On the other hand, most of the LVHIS galaxies have log sSFR ($={\log \rm SFR}/M_*$)$>-11.5$, indicating most of them to be star-forming galaxies.
{width="14cm"}
![The distribution of LVHIS galaxies around the star-forming Main Sequence (MS). The open circles and upward arrows show the lower limits of SFR estimate. The green line is the MS from the SDSS sample [@Renzini15], the purple, red and blue dashed lines are the mean relations from ALFALFA [@Huang12b], the WISE-GAMA sample [@Jarrett17] and the GALEX-GAMA sample [@Grootes13]. The black dotted line shows the position for log sSFR ($={\log \rm SFR}/M_*$)$=-11.5$.[]{data-label="fig:SFMS"}](plot/SFsequence.pdf){width="8cm"}
Radial distribution of HI gas {#sec:dist_HI}
-----------------------------
This section investigates the diversity of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} radial distributions in the LVHIS galaxies.
In Figure \[fig:HI\_distr\], we find that galaxies with higher $\fHI$ on average have larger [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} discs with respect to the optical discs (${\rm R_{HI}/R_{25}}$, top panel), and lower fractions of total [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} masses enclosed within the optical discs (${M_{{\rm HI,r<R25}}/M_{{\rm HI}}}$, middle panel), but higher average [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities within the optical discs (${\rm \Sigma_{HI,r<R25}}$, bottom panel). However, the three correlations have considerable scatter. For example, when $\fHI=1$, ${\rm R_{HI}/R_{25}}$ ranges from 1.25 to 6.3, ${M_{{\rm HI,r<R25}}/M_{{\rm HI}}}$ ranges from 0.1 to 0.7, and ${\rm \Sigma_{HI,r<R25}}$ ranges from 4 to 8 $\mspc$, within the LVHIS sample.
So there are a large variety of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} radial distributions at a fixed global properties. Because the inner disc within R$_{25}$ contains the bulk of star formation in galaxies, directly measuring the properties of [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas within R$_{25}$ is important for understanding the intrinsic physics underlying the star formation-[[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} global scaling relations.
{width="14cm"}
\[fig:HI\_distr\]
SFR-HI relations on galactic scales and within R$_{25}$ {#sec:glob_SFE}
=======================================================
In this section, we investigate on galactic scales the relation between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass and the dependence of SFE on other parameters. The reciprocal of SFE is also referred to as the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} depletion time, $t_{dep}$. Because both SFR and SFE are higher within the optical discs (within R$_{25}$) than beyond the optical discs [@Bigiel10], we also analyse SFR, $\mHI$ and SFE measured within R$_{25}$.
The SFR-$\mHI$ relation
-----------------------
In the left-top panel of Figure \[fig:SFL\], we show tight scaling relations between SFR and $\mHI$ and between SFR and $M_{{\rm HI,r<R25}}$ (left panels). This is consistent with the finding of ALFALFA [@Huang12b]. On the other hand, @Bigiel08 found that $\Sigma_{\rm SFR}$ and $\SHI$ are not correlated on kpc-scales within R$_{25}$. A question is whether there is an underlying intrinsic correlation between $\Sigma_{\rm SFR}$ and $\SHI$ on galactic scales that differs from the relation on kpc-scales, because after all, [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} is the reservoir of material for forming molecular gas and subsequently stars.
We investigate the correlations between $\Sigma_{\rm SFR}$ and $\SHI$ averaged within R$_{\rm HI}$ and R$_{25}$ (right panels), and find they are weak, in agreement with findings by @Bigiel08. Hence we conclude that on galactic scales and within R$_{25}$, SFR and $\mHI$ are correlated mostly because both they both are correlated with the size of the galaxies, but not because of a physical link between star formation and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas.
We obtain a partial correlation coefficient of 0.84 for the relation between SFR and $\mHI$ with the effect of $M_*$ removed. So the SFR-$\mHI$ relation is not caused by a correlation of both parameters with M$_*$.
![The relations between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} mass, and between their masses and surface densities within R$_{25}$. The open circles with an upward arrow mark the lower limits of SFE measurements for some galaxies; these data points are excluded when calculating the correlation coefficient $\rho$. When the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} related parameter is defined with a diameter smaller than 2B$_{\rm maj}$, the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data is considered unresolved and the galaxy is excluded. The correlation coefficient is denoted on the top-left corner and the typical error bar is shown on the bottom-right corner of each panel. In the left-top panel, the dashed line is a liner fit to the data points (upper limits excluded), and the dotted line is the relation from @Huang12b. []{data-label="fig:SFL"}](plot/SFL_reason.pdf){width="8.5cm"}
Star Formation Efficiency
-------------------------
We investigate the dependence of SFE on other galactic properties, and compare the trends measured globally and within R$_{25}$. If the global trend reflects a mechanism that directly regulates the efficiency for [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} to fuel star formation, then we should observe a correlation within R$_{25}$ that is stronger than or similar to the one measured globally; but if the global trend is caused by the fact that the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} discs are much more extended than the stellar discs (so the gas cannot directly fuel the bulk of the star formation), then the trend should disappear when the relevant properties are measured within R$_{25}$.
In Figure \[fig:SFeff\_global\], we can see that the SFE is significantly correlated with $\Sigma_{\rm *,eff}$, moderately anti-correlated with $\fHI$, weakly anti-correlated with $M_{{\rm HI,r<r25}}/M_{{\rm HI}}$, hardly correlated with $M_*$, and not correlated with $\Sigma_{\rm HI,r<RHI}$ or sSFR. In Figure \[fig:SFeff\_r25\], we can see that when we use $M_{{\rm HI,r<r25}}$ instead of $\mHI$ in these relations, [SFE$_{\rm r<R25}$ ($=$SFR$/M_{{\rm HI,r<R25}}$) are still significantly correlated with $\Sigma_{\rm *,eff}$ and moderately (weakly) anti-correlated with ${\rm f_{HI,r<R25}=M_{HI,r<R25}/M_*}$, while the weak anti-correlation with $M_{{\rm HI,r<r25}}/M_{{\rm HI}}$ disappears. All other correlations remains weak considering the error bars.]{} The direct implications seem to be that $\Sigma_{\rm *,eff}$ and $\fHI$ are related to mechanisms that directly regulate star formation fuelling in the stellar disc, while the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} radial distributions (indicated by $M_{{\rm HI,r<r25}}/M_{{\rm HI}}$) may only have a rather minor effect on the global SFE values.
Because $\fHI$ and $\Sigma_{\rm *,eff}$ are strongly correlated, we calculate the partial correlation coefficient between ${\rm SFE_{r<R25}}$ and ${\rm f_{HI,r<R25}}$ with the dependence on $\Sigma_{\rm *,eff}$ removed, and the value is $-0.03\pm0.23$. It suggests that SFE depends on ${\rm f_{HI,r<R25}}$ only through their dependence on $\Sigma_{\rm *,eff}$.
The correlation between SFE and $\Sigma_{\rm *,eff}$ is consistent with the best-fit linear relation from @Wong16. The slope of the relation was explained in @Wong16 with a model where the galaxy discs are in a marginally stable status described by a constant two-fluid (gas and stars) Toomre $Q$ parameter and the mid-plane pressure determines the molecular-to-atomic ratio. In such a model, a galaxy with high stellar surface density will have low gas surface density as adjusted by the constant $Q$; the combined effects of high stellar surface density and low gas surface density on the mid-plane pressure model result in very little change in the molecular gas mass (which SFR directly scales with); hence finally SFE as the ratio between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface densities increases.
To summarize, on galactic scales and within R$_{25}$, among the parameters investigated here SFE depends most strongly on the averaged stellar surface density $\Sigma_{\rm *,eff}$.
{width="14cm"}
{width="14cm"}
Star formation beyond R$_{25}$ {#sec:loc_SFE}
==============================
In this section we study the star formation on local scales in the outskirts of galaxies (beyond R$_{25}$), where both gas and stars are of low surface density and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} dominates the cold gas.
The dependence of $\Sigma_{\rm SFR}$ on $\Sigma_*$ and $\SHI$
-------------------------------------------------------------
The investigation is based on the sub-sample of ten large galaxies (Section \[sec:sample\]). First of all, we show in Figure \[fig:SFeff\_inpix\] the relation between $\SHI$ and $\Sigma_{\rm SFR}$ measured in pixels with sizes equivalent to the PSF FWHM beyond R$_{25}$. The distribution of data points and the mean relation are consistent with those from @Bigiel10.
As discussed in Section \[sec:introduction\], in dwarf galaxies the radial variation of $\Sigma_{\rm SFR}$ follows the variation of $\Sigma_*$ better than that of $\SHI$ [@Hunter98]. We ask whether we can observe the same phenomenon for the galactic outskirts of more massive galaxies. Because $\Sigma_*$, $\SHI$ and $\Sigma_{\rm SFR}$ all generally decrease with increasing radius, we fix (control) $\SHI$ when studying the dependence of $\Sigma_{\rm SFR}$ on $\Sigma_*$, and fix $\Sigma_*$ when studying the dependence of $\Sigma_{\rm SFR}$ on $\SHI$. Limited by the data depth for $\Sigma_*$, we rely on the azimuthally averaged measurements in order to get enough signal-to-noise (S$/$N) ratio for $\Sigma_*$ measurements in these low surface brightness galactic outer regions. Radial analysis may have the risk of smoothing out$\slash$integrating correlations, but is the best we can do for now. We note that very deep optical or MIR images (to trace stellar mass) will be needed in the future to confirm our results with a pixel-based analysis in a similar way as in Figure \[fig:SFeff\_inpix\].
We analyse our data in two different ways, basing on resolution-matched radial profiles and profiles at the best resolution possible, respectively.
### Analysis based on the resolution-matched radial measurements
We use the resolution-matched radial profiles derived in Section \[sec:phot\_conv\]. The result is shown in Figure \[fig:SFeff\_conv\]. The top-left panel shows the data points measured at 1.2R$_{25}$ confirming the $\Sigma_{\rm SFR}$-$\SHI$ relation from pixel-based analysis (Figure \[fig:SFeff\_inpix\] and Bigiel et al. 2010). The data points suggest an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} depletion time longer than the Hubble time, but are systematically higher in $\Sigma_{\rm SFR}$ than the median relation from @Bigiel10. This difference comes from two effects. The first effect is the trend of SFE decreasing with galactocentric radius (the @Bigiel10 relation is a median relation for the radial region 1-2 R$_{25}$); the second effect is a difference between the mean and median relations noticed before in @Roychowdhury15. The best-fit relation from the data in this figure has a power law slope of 1.24$\pm$0.36. The slope is close to unity but with a large error bar, suggesting that it is hard to find any dependence of SFE on $\SHI$ in our data.
The top-right panel shows that, as in irregular galaxies, there is also a trend of $\Sigma_{\rm SFR}$ to increase with higher $\Sigma_*$ at 1.2R$_{25}$. Which correlation might be more intrinsic? We find that the partial correlation coefficient between $\Sigma_{\rm SFR}$ and $\SHI$ with the effect of $\Sigma_*$ removed is $-0.22\pm0.50$, while the partial correlation coefficient between $\Sigma_{\rm SFR}$ and $\Sigma_*$ with the effect of $\SHI$ removed is 0.84$\pm0.27$. The partial correlation analysis suggests that the $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation is more intrinsic than the $\Sigma_{\rm SFR}$-$\SHI$ correlation and the latter is caused at least partly by the former.
To confirm this result, in the bottom panels of Figure \[fig:SFeff\_conv\], we further investigate the $\Sigma_{\rm SFR}$ correlations firstly at a fixed $\Sigma_*$ of $3~\mspc$ and then at a fixed $\SHI$ of $3~\mspc$ [^5]. When $\Sigma_*=3~\mspc$, $\SHI$ ranges from 10$^{0.49}$ to 10$^{0.88}~\mspc$, while when $\SHI=3~\mspc$, $\Sigma_*$ ranges from 10$^{-0.20}$ to 10$^{1.13}~\mspc$. In these regions, $\Sigma_*$ values are consistent with the value typical for [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-dominated regions, $\Sigma_* < 81~\mspc$ [@Leroy08]. $\SHI$ is close to the average surface density for transition between atomic and molecular gas (5$~\mspc$, with contribution from helium not included). However as shown in [@Leroy08] the atomic-to-molecular conversion is much more strongly correlated with $\Sigma_*$ than with $\SHI$, hence most of these measurements are obtained from [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-dominated regions. We find that there is a strong correlation of higher $\Sigma_{\rm SFR}$ with increasing $\Sigma_*$ at a fixed $\SHI=3~\mspc$, but very weak correlation between $\Sigma_{\rm SFR}$ and $\SHI$ at a fixed $\Sigma_*=3~\mspc$. This result implies that $\Sigma_*$ is driving changes in $\Sigma_{\rm SFR}$ more dramatically than $\SHI$ does in the galactic outskirts. These results are consistent with the finding from @Hunter98.
In the bottom panels of Figure \[fig:SFeff\_conv\], the dynamic range of $\Sigma_*$ is greater than that of $\SHI$ by nearly one dex for this sample. The wide dynamic range of $\Sigma_*$ as well as the strong $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation have both contributed to producing the significant $\Sigma_{\rm SFR}$-$\SHI$ correlation (without fixing $\Sigma_*$) shown in the top row of Figure \[fig:SFeff\_conv\].
We notice from all panels in Figure \[fig:SFeff\_conv\] the large scatter of $\Sigma_{\rm SFR}$ in the elliptical rings where each radial profile value is measured (the crosses in the figure). Limited by the variation in image resolutions in the sample and the small sample size, we can not evaluate how the scatter varies as a function of $\SHI$ or $\Sigma_*$, but it can be inferred that the scatter in the $\Sigma_{\rm SFR}$-$\Sigma_*$ relation might be as large as that in the $\Sigma_{\rm SFR}$-$\SHI$ relation. The large scatter indicates complex physical factors other than the gravity provided by stars and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas that regulate star formation.
### Analysis based on the radial measurements at the best possible resolution
For the second way of analysing the data, we use the radial profiles derived in Section \[sec:best\_res\]. In addition to repeating the analysis in the bottom row of Figure \[fig:SFeff\_conv\], we further look into the dependence of $\Sigma_{\rm SFR}$ on $\Sigma_*$ and $\sigma_{\rm sd}$. $\sigma_{\rm sd}$ has been predicted by star formation models [@Ostriker10; @Krumholz13] to be a key parameter in setting the SFR in the outskirt of galaxies (see more details in Section \[sec:SFE\_model\]). These results are presented in the top row of Figure \[fig:SFeff\_highrs\]. We warn that we do not have a proper error estimate for $\Sigma_*$ measured from the tilted ring models, and the estimate of $\sigma_{\rm sd}$ is based on simple assumptions.
From the left two panels in the top row of Figure \[fig:SFeff\_highrs\], the trends are largely consistent with what we have found in the bottom panels of Figure \[fig:SFeff\_conv\]. There is a moderate correlation between $\Sigma_{\rm SFR}$ and $\SHI$ here, but it is significantly weaker than the $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation. In the last two panels of the top row, we also find significant correlations of $\Sigma_{\rm SFR}$ increasing as a function of $\sigma_*$ and $\sigma_{\rm sd}$. The slope for the $\Sigma_{\rm SFR}$-$\sigma_{\rm sd}$ relation is 1.07$\pm$0.41.
In the bottom row of Figure \[fig:SFeff\_highrs\], we show the trends of SFE as function of $\SHI$, $\Sigma_*$, $\sigma_*$ and $\sigma_{\rm sd}$ (defined at Section \[sec:best\_res\]). SFE and $\SHI$ are not correlated. The other trends are similar as the trends of $\Sigma_{\rm SFR}$ depending on those parameters.
We note that at $\SHI=3~\mspc$, $\sigma_{\rm sd}$ is on average 3.5 times higher than $\sigma_*$ in our sample, but the two volume densities correlate with each other with $\rho=0.64\pm0.20$. $\Sigma_*$ is correlated with $\sigma_*$ ($\rho=0.85\pm0.28$) because of the way it is calculated, but is not correlated with $\sigma_{\rm sd}$ ($\rho=0.35\pm0.25$). The partial correlation between SFE and $\sigma_*$ with the effect of $\sigma_{\rm sd}$ removed is 0.55$\pm0.45$, while the partial correlation between SFE and $\sigma_{\rm sd}$ with the effect of $\sigma_*$ removed is 0.39$\pm0.48$. Hence the weak statistics does not allow us to conclude whether $\sigma_{\rm sd}$ or $\sigma_*$ ($\Sigma_*$) has independent effects on setting SFE in the galactic outskirts. We will discuss it in more details in Section \[sec:model\_SFE\].
![The dependence of $\Sigma_{\rm SFR}$ on $\SHI$. Measurements are derived from resolution-matched images. $\SHI$ and $\Sigma_{\rm SFR}$ are measured in pixels with sizes equivalent to the PSF FWHM beyond R$_{25}$. We exclude those elements in which SFR$<$0.0003 $M_*~{\rm yr^{-1}}$, to avoid the problem of insufficient IMF sampling [@Lee09; @Lee11]. The blue contours show the distribution of 50% and 90% of the data points, and the blue dot-dashed line show the median relation from @Bigiel10. Correlation coefficient $\rho$ is denoted in the left-top corner. []{data-label="fig:SFeff_inpix"}](plot/SFR_inpix.pdf){width="8cm"}
![The dependence of $\Sigma_{\rm SFR}$on $\SHI$ and $\Sigma_*$. Parameters are derived from resolution matched images. Only measurements with an error less than 0.5 dex are displayed and considered in the analysis. Solid and open circles are measurements derived from radial profiles. Solid circles mark direct measurements and open circles mark the $\Sigma_*$ measurements obtained through extrapolating radial profiles. Vertically from each circle, the green crosses show $\Sigma_{\rm SFR}$ measured in pixels with a size equivalent to the PSF FWHM (see Figure \[fig:SFeff\_inpix\]) along the elliptical ring where the radial profile value is measured. Due to the very different galactic sizes and image resolutions within the small sample, these crosses are only illustrative of the scatters in $\Sigma_{\rm SFR}$ along each elliptical ring and cannot be compared between different galaxies. In the top panels, data points are measured at 1.2R$_{25}$. In the bottom-left panel, data points are measured on the elliptical ring where the averaged $\Sigma_*=3~\mspc$; in the bottom-right panel, data points are measured along the elliptical ring where the average $\SHI=3~\mspc$. These characteristic positions are denoted in the top-right corner of each panel. The black dashed line in the top-left panel is a linear fit to the data points, the blue dash-dotted line is the median relation from @Bigiel10, and the red dotted line marks an SFE=10$^{10}~{\rm yr}^{-1}$ or a gas depletion time of 10 Gyr. In the bottom panels, the black dotted lines mark the density of 3 $\mspc$. Correlation coefficient $\rho$ is denoted in the left-top corner of each panel. []{data-label="fig:SFeff_conv"}](plot/SFeff_mphot.pdf){width="8cm"}
{width="14cm"}
{width="14cm"}
Comparison with existing observations {#sec:SFE_model}
-------------------------------------
As we have mentioned, @Bigiel10 showed the $\Sigma_{\rm SFR}$-$\SHI$ relation for the outskirts of spiral discs. $\Sigma_{\rm SFR}$ and $\SHI$ were measured in kpc-scale pixels, and the median relation has a power law slope of 1.7. Measuring the mean $\Sigma_{\rm SFR}$ in each $\SHI$ bin instead, @Roychowdhury15 found a uniform power law slope between 1.4 and 1.7 for the relations between log $\SHI$ and log $\Sigma_{\rm SFR}$ for irregular galaxies and the outskirts of spiral and massive disc galaxies. The non-unity slope suggests a correlation between log $\SHI$ and log SFE in the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-dominated regime. On the other hand, @Yildiz16 only found a close-to-unity power law slope for the mean relation in the outskirts of early-type galaxies. When we derive the $\Sigma_{\rm SFR}$-$\SHI$ relation at 1.25 R$_{25}$, we average over both star-forming and non-star-forming regions along the elliptical ring, but the power law slope of 1.24$\pm$0.36 is consistent with the literature values within the error bar.
A large scatter in the $\Sigma_{\rm SFR}$-$\SHI$ relation has been shown in all these studies. In @Bigiel10, the distribution of $\Sigma_{\rm SFR}$ has a large scatter at a fixed $\SHI$ ($\sim$1.5 dex at $\SHI=1\mspc$). There is also large offset between the median and mean values of $\Sigma_{\rm SFR}$ at a fixed $\SHI$ ($\sim$0.5 dex at $\SHI=1\mspc$) as revealed in @Roychowdhury15. Based on the FIGGS sample of irregular galaxies, @Roychowdhury09 found that each galaxy has its own slope and intercept for $\Sigma_{\rm SFR}$-$\SHI$ relation when such a correlation exists, while for 20% of the irregular galaxies (most of them disturbed) in their sample, $\Sigma_{\rm SFR}$ does not correlate with $\SHI$. The large scatter indicates complex physics regulating the relationship between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas.
Despite the uncertainty in slopes and the large scatter, the $\Sigma_{\rm SFR}$-$\SHI$ relations observed from these different studies [also see @Yim16] have one common feature: the inferred SFE is low and the gas depletion time is longer than the Hubble time.
As mentioned briefly in Section \[sec:introduction\], the tight $\Sigma_{\rm SFR}$-$\Sigma_*$ relation was found for the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-dominated irregular galaxies by @Hunter98. They tested a variety of physical quantities as possible regulators for star formation in irregular galaxies, including [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas critical densities for gravity instabilities in different disc models. They found that only $\Sigma_*$ correlates well with $\Sigma_{\rm SFR}$ along radial profiles in their investigation. In this paper, we find a strong $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation also exist in the outskirts of spiral galaxies.
Comparison with theoretical models {#sec:model_SFE}
----------------------------------
In this section we try to understand our results for the $\Sigma_{\rm SFR}$-$\SHI$ and $\Sigma_{\rm SFR}$-$\Sigma_*$ relations in the context of modern star formation models.
Simple modern star formation models consider that 1) the atomic inter-stellar medium consists of a warm and a cold phase, and only the cold mode gas phase forms molecular clouds and stars; and that the 2) SFR in the HI-dominated regime is regulated by the hydrostatic equilibrium between thermal and dynamic pressure on gas at the mid-plane (Ostriker et al. 2010, O10 hereafter, Krumholz et al. 2013, K13 hereafter). The exact recipe for determining the pressure equilibrium between gas phases differs in different models. For example, in the O10 model, the atomic gas is in two-phase equilibrium everywhere and the thermal pressure is set by the radiation field and thus, it is a consequence of star formation feedback, while in the K13 model the cold phase density, sensitively depends on an assumed maximum temperature, above which the cold phase cannot exist. Other parameters of the model, such as the gas surface density and the SFR, are regulated under this condition. In these models, the relation between the SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surface density are of complex nature. This is because, at the densities where [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} dominates, other factors besides gravity contribute strongly to the equilibrium condition of the gas, such as temperature, sound speed, magnetic fields, cosmic rays and turbulence. K13 predicted a large scatter in the relation between $\Sigma_{\rm SFR}$ and $\Sigma_{\rm gas}$ (see their Figure 3 and Appendix A1). O10 predicted a complex dependence (their equation 22) of $\Sigma_{\rm SFR}$ as a function of $\SHI$, which contains several terms contributing equally. The large scatter and the complex functional forms make it hard to constrain the models, because measuring the $\SHI$-$\Sigma_{\rm SFR}$ relation from a small data set like LVHIS comes with observational uncertainties, and allow for significant ambiguity in the models.
Because old stars and dark matter contribute to the mid-plane pressure, both O10 and K13 models predict a SFR that increases with increasing $\sigma_{\rm sd}$ at a fixed gas density, which is what we observe in our data. Especially, the O10 model predicts $\Sigma_{\rm SFR}\propto \sqrt{\sigma_{sd}}$ for the low gas surface density region [also see @Kim11]. Our data reveal a power law index of 1.07$\pm$0.41, which does not necessarily disprove the O10 model, considering our crude calculation of the DM densities and the simplifications that go into the modelling. The $\Sigma_{\rm SFR}$-$\sigma_{\rm sd}$ relation also manifests itself as the $\Sigma_{\rm SFR}$-$\Sigma_*$ relation. Although the scatter in the $\Sigma_{\rm SFR}$-$\Sigma_*$ relation might be as large as that of the $\Sigma_{\rm SFR}$-$\SHI$ relation given the same complex physical processes involved, the much wider dynamic range of $\Sigma_*$ compared to $\SHI$ has helped us to reveal a trend between $\Sigma_*$ and $\Sigma_{\rm SFR}$.
The $\Sigma_{\rm SFR}$-$\Sigma_*$ relation might also partly reflect the SFE-metallicity relation (K13). Because metals are expected to provide the necessary shielding for star-forming clouds [@Sternberg14 and K13], and there is a strong correlation between $\Sigma_*$ and local metallicity [@Moran12; @Sanchez13; @LaraLopez13; @Carton15; @Sanchez17]. @Roychowdhury15 found that the SFE in the outskirts of irregular and massive spiral galaxies are the same. Because irregular galaxies on average have much lower metallicity than the massive spiral galaxies, they concluded that the SFE in the galactic outskirts seem to be independent of metallicity. However, metallicities in the galactic outskirts might differ significantly from the global metallicities [@Moran12; @Sanchez12; @LopezSanchez15] which is what Roychowdhury et al. (2015) used for their analysis. Optical spectroscopy for the galactic outskirts is needed before we could make relevant conclusions, and we may need rely on the next generation IFU instruments like HECTOR [@BlandHawthorn15].
In the past it has been suggested that star formation can have a positive feedback effect, effectively triggering more star formation [@Gerola80]. In this picture, winds and the supernovae ejecta can cause gas compression at the shocks front, increasing the density of the gas and triggering new star formation. This is referred to as “self-propagation of star formation", which seemingly could explain the $\SHI$-$\Sigma_*$ correlation. However, modern simulations @Dale11 [@Dale15] show that, although this form of positive feedback sometimes locally increases the gas density, enhancing star formation, the overall effect is negative. Ionising radiation effectively disperses low-to-intermediate mass molecular clouds while outflows disperse the more massive molecular clouds. In observations, there are examples of positive feedback at play and also a few of self-propagating (or subsequently triggered) star formation, such as the W5 HII region in the MW and the stellar cluster NGC 602 in the SMC [@Koenig08; @Carlson11]. The conditions for self-propagating star formation to take place and whether this physical phenomenon could play an important role at low gas surface density regions [as predicted by @Gerola80] remains unclear.
To summarise, the strong $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation at a fixed $\SHI$ and the relatively weak $\Sigma_{\rm SFR}$-$\SHI$ correlation at a fixed $\Sigma_*$ in the galactic outskirts are broadly consistent with modern star formation models. More detailed comparisons rely on more data points from larger samples in the future.
Star Formation Efficiency near HI warps {#sec:warp_SFR}
========================================
We have shown in Section 4.2 that seven out of the ten LVHIS large galaxies display warped [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} discs. Theoretical studies like that of @SanchezBlazquez09 have predicted a drop in SFE caused by the onset of an [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warp. Because SFE generally decreases with radius because of decreasing gas and stellar surface densities [@Bigiel10], we expect the additional effect from a warp would steepen the decrease. In this section, we investigate whether such a trend can be observed with our data.
We mark the warp radius $r_{warp}$ on the SFE radial profiles in Figure \[fig:SFEprofile\]. When calculating the SFE radial profiles, we derive $\Sigma_{\rm SFR}$ profiles at the resolution of the W4 images as in Section \[sec:best\_res\], but separate the approaching and receding velocity sides of a galaxy. We use $\SHI$ profiles both directly derived from the mom-0 images and taken from the tilted-ring models. The former are less noisy and the latter are better resolved, but they resulted in close SFE profiles as shown in Figure \[fig:SFEprofile\] because $\SHI$ vary in a much narrower dynamic range than $\Sigma_{\rm SFR}$.
We find the relations between [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps, R$_{25}$ and SFE radial profile shapes show great diversity within the very small sample. We summarise the most important results shown in Figure \[fig:SFEprofile\]:
1. $r_{warp}$ is located close to R$_{25}$ (with an offset $<$0.3 R$_{25}$) in most of the warped galaxies (6 out of 7). NGC 3621 has a warp that onsets extremely far away from the galactic centre: beyond 2.5 R$_{25}$, and beyond the detection limit of our $\Sigma_{\rm SFR}$ data. In four out of the seven warped galaxies, $r_{warp}$ on the two velocity sides differ by more than 0.4 R$_{25}$ (NGC 55, NGC 3621 and NGC5236).
2. With the limited statistics available, there is no significant difference in the level of SFE beyond R$_{25}$ between warped and un-warped galaxies.
3. Among the six [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warped galaxies where SFE is measurable at $r_{warp}$, the only case for $r_{warp}$ to coincide with a steepened drop in SFE is in NGC 300. On the contrary, in NGC 55 and NGC 7793, $r_{warp}$ coincides with the onset of a shallower SFE profile or even the turning over to a rising SFE profile. In other three warped galaxies, NGC 5236, IC 5052 and IC 5152, there are no clear signs that $r_{warp}$ immediately coincides with a change in the slope of the SFE profile.
4. R$_{25}$ shows a similar complex relation with SFE profile shapes as shown by $r_{warp}$.
In order to better quantify the relation between $r_{warp}$ and change in SFE profile slope, we derive the second order derivative of SFE profiles near $r_{warp}$. For a given radius $r$, we first calculate the first order derivatives of SFE for the two radial segments with distances $-0.25$ to 0 R$_{25}$ and 0 to 0.25 R$_{25}$ from r, ${\rm SFE_-'(r)}$ and ${\rm SFE_+'(r)}$ respectively. And the second order derivative of SFE at r is calculated as ${\rm SFE''(r)=SFE_+'(r)}-{\rm SFE_-'(r)}$. For comparison, we also calculate SFE$''$ at R$_{25}$ and at 100 random radii beyond R$_{25}$ for each galaxy. All these calculations use SFE profiles measured at both the approaching and the receding velocity sides of galaxies. We use the SFE profiles with $\SHI$ derived from [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} column density images (the red and blue curves in Figure \[fig:SFEprofile\]) in this analysis, but we have confirmed that, if we use the SFE profiles with $\SHI$ derived from tilted ring models instead (the pink and cyan curves in Figure \[fig:SFEprofile\]), the conclusion from the relevant results does not change.
Figure \[fig:SFdev\_distr\] shows that SFE$''$ measured at $r_{warp}$ shows no significant difference with that measured at R$_{25}$ or random radii beyond R$_{25}$. It confirms what we have visually summarised from Figure \[fig:SFEprofile\].
Hence what we learn from the limited statistics here is that we do not find strong evidence for [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps to be [*always*]{} linked to suppressed SF activity. A larger sample from future [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} surveys [WALLABY$\slash$Apertif, @Koribalski12] is required to confirm this result or identify the possibly weak effect of warps.
{width="16cm"}
![The distributions of second order derivatives of SFE measured at $r_{warp}$ (violet), R$_{25}$ (green) and random radii ($r_{random}$) beyond R$_{25}$ (black). The black histogram has been normalised to have comparable height as the violet and green histograms. The top right corner shows the Kolmogorov-Smirnov (KS) test probabilities for the null hypothesis that the two distributions are drawn from the same parental distribution. The numbers in orange, violet and green show the KS test probabilities for the comparisons SFE$''(r_{warp})$ vs SFE$''{\rm (R_{25})}$, SFE$''(r_{warp})$ vs SFE$''(r_{random})$ and SFE$''{\rm (R_{25})}$ vs SFE$''(r_{random})$ respectively.[]{data-label="fig:SFdev_distr"}](plot/SFdev_distr.pdf){width="8.5cm"}
Summary and conclusion {#sec:summary}
======================
We have analysed GALEX, WISE and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} images and characterised the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} $M_*$ and SFR properties for galaxies in the LVHIS sample. We use the resolved [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} measurements to understand previously found global relations$\slash$trends (based on the LVHIS galaxies with R$_{25}>2 {\rm B_{maj}}$) and investigate star formation properties in the galactic outskirts (beyond R$_{25}$, based on the ten large galaxies). We try to obtain a coherent picture of how SFR relates to the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas on different scales and in different regions. The major findings are:
1. The [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} radial distribution with respect to R$_{25}$ is correlated with $\fHI$ of galaxies (Figure \[fig:HI\_distr\]). However the large scatter in the relations highlights the importance of using spatially resolved [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} data to understand global relations.
2. The tight SFR-$\mHI$ relation is caused by each being the product of galaxy area multiplying with $\Sigma_{\rm SFR}$ and $\SHI$, respectively, and does not reflect an intrinsic link between star formation and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas. The correlation between the globally averaged $\Sigma_{\rm SFR}$ and $\SHI$ is weak (Figure \[fig:SFL\]). SFE significantly depends on the average stellar surface density $\Sigma_{\rm *,eff}$, which was explained in @Wong16 with a marginally stable disc model.
3. The $\Sigma_{\rm SFR}$-$\SHI$ correlation in the galactic outskirts becomes significantly weaker when $\Sigma_*$ is fixed, while on the other hand there is a strong $\Sigma_{\rm SFR}$-$\Sigma_*$ correlation at a fixed $\SHI$ (Figure \[fig:SFeff\_conv\] and \[fig:SFeff\_highrs\]). These trends are consistent with star formation models where young stars are directly formed from the molecular gas, and are indirectly linked with [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas through complex physical processes. In these models, old stars have a significant influence on the star formation process by providing gravity or through their link with metallicity.
4. [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} warps hardly affect the SFE in the galactic outskirts (Figure \[fig:SFdev\_distr\]).
These results suggest that the direct link between SFR and [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} gas is generally weak in star forming galaxies. The SFR-[[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} relation measured globally and in the galactic outskirts are caused at least partly by side effects (both correlating to a third parameter, galaxy size for the former and $\Sigma_*$ for the latter) and are intrinsically weaker than they appear to be. The gap between them and the (lack of) SFR-[[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} relations measured within galactic inner regions on kpc-scales [@Bigiel08] thus become much smaller. On the other hand, old stars may play a significant role in regulating star formation in both inner and outer regions of galaxies.
One major limitation of the analysis in this paper is the relatively small number of statistics, especially for the sub-sample of large galaxies. The relevant results will be strengthened by larger statistics, deeper images and follow-up optical spectroscopy. We look forward to high quality data and larger samples from future multi-wavelength surveys and instruments, such as WALLABY [@Koribalski12], Skymapper [@Keller07] and HECTOR [@BlandHawthorn15].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous referee for constructive comments. We gratefully thank G. Muerer, M. Krumholz and B. Wang for useful discussions.
The [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{} 21-cm observations were obtained with the Australia Telescope Compact Array which is funded by the Commonwealth of Australia for operations as a National Facility managed by CSIRO.
GALEX (Galaxy Evolution Explorer) is a NASA Small Explorer, launched in April 2003, developed in cooperation with the Centre National d’Études Spatiales of France and the Korean Ministry of Science and Technology.
This publication makes use of data products from the Wide-field Infrared Survey Explorer, which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute of Technology, funded by the National Aeronautics and Space Administration.
This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
ZYL is supported by the National Natural Science Foundation of China under grant no. 11403072, by the China-Chile joint grant from CASSACA, and by Shanghai Yangfan Research Grant (no. 14YF1407700).
The work of LCH was supported by National Key Program for Science and Technology Research and Development grant 2016YFA0400702.
JMvdH acknowledges support from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013)/ERC Grant Agreement no. 291531.
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 679627).
Measurements tables
===================
GALEX and WISE photometric measurements for LVHIS galaxies are presented in Tables \[tab:uvphot\] and \[tab:irphot\] respectively.
Table \[tab:uvphot\] gives the fitted FUV and NUV scale-length of the outer disc (r$_{\rm s}$), the apparent magnitude corrected for foreground extinction (m), the effective radius that enclose half of the light (r$_{\rm eff}$) and the effective surface brightness (the average brightness within r$_{\rm eff}$, $\mu_{\rm eff}$). See Section \[sec:uvphot\] for details.
Table \[tab:irphot\] the four WISE band luminosities (${\rm \log L_{W1-4}}$), the W1-W2 colour and the W1 structural parameters. The superscript $f$ marks the galaxies not detected in the W2-band, for which we have enforced the W1 aperture on the W2 images in order to measure the W2 fluxes. The expected (model) stellar continuum are removed from the luminosities ${\rm \log L_{W3}}$ and ${\rm \log L_{W4}}$. The W1 in-band luminosity (${\rm L_{W1,\odot}}$) is normalised by the equivalent solar value and used for calculating stellar mass-to-light ratio. R$_{75}$ and R$_{25}$ are the 75% and 25% light radius and the ratio between the two is a measure of the central concentration. See Section \[sec:irphot\] for details.
------- ---------------- ---------------- --------------- ------------------- ---------------- ---------------- --------------- -------------------
index r$_{\rm s}$ m r$_{\rm eff}$ $\mu_{\rm eff}$ r$_{\rm s}$ m r$_{\rm eff}$ $\mu_{\rm eff}$
arcsec mag arcsec mag arcsec$^{-2}$ arcsec mag arcsec mag arcsec$^{-2}$
(1) (2) (3) (4) (5) (6) (7) (8) (9)
1 8.1$\pm$0.4 17.16$\pm$0.06 9.1 27.55 8.6$\pm$0.5 16.97$\pm$0.03 10.1 24.03
2 33.7$\pm$7.1 17.67$\pm$0.58 8.0 28.43 20.0$\pm$1.0 17.58$\pm$0.03 6.4 23.68
3 8.5$\pm$1.1 18.33$\pm$0.11 13.5 28.67 13.6$\pm$0.4 17.26$\pm$0.04 8.7 24.01
4 128.6$\pm$0.1 10.31$\pm$0.05 111.9 20.67 121.0$\pm$0.1 9.92$\pm$0.03 107.7 22.08
5 206.6$\pm$9.0 10.31$\pm$0.05 130.5 20.66 177.3$\pm$13.7 10.10$\pm$0.03 120.1 22.49
6 227.4$\pm$0.8 11.24$\pm$0.05 160.5 21.59 256.3$\pm$0.5 10.63$\pm$0.03 167.0 23.74
7 145.9$\pm$5.1 11.47$\pm$0.05 161.3 21.82 208.9$\pm$10.6 11.18$\pm$0.03 132.9 23.79
8 97.2$\pm$7.0 13.79$\pm$0.05 3.0 24.14 108.4$\pm$3.8 13.30$\pm$0.03 3.4 17.95
9 27.8$\pm$0.7 13.74$\pm$0.05 28.2 24.10 29.1$\pm$1.0 13.61$\pm$0.03 27.8 22.83
10 62.1$\pm$20.1 16.21$\pm$0.07 42.5 26.52 72.8$\pm$2.7 15.01$\pm$0.03 32.9 24.66
11 46.4$\pm$0.3 14.68$\pm$0.05 38.0 25.04 47.6$\pm$0.3 14.40$\pm$0.03 37.3 24.26
12 38.1$\pm$0.2 13.89$\pm$0.05 48.5 24.25 39.4$\pm$0.1 13.70$\pm$0.03 44.6 23.94
13 6.1$\pm$0.6 17.69$\pm$0.24 10.0 28.06 6.6$\pm$0.5 17.58$\pm$0.08 10.0 24.60
14 112.3$\pm$3.4 10.52$\pm$0.05 50.8 20.89 79.1$\pm$1.8 10.37$\pm$0.03 44.8 20.63
15 20.9$\pm$0.5 14.81$\pm$0.05 13.1 25.16 30.4$\pm$2.3 14.46$\pm$0.03 12.3 21.91
16 13.4$\pm$3.9 19.23$\pm$1.00 21.2 28.75 9.7$\pm$1.8 18.90$\pm$1.13 18.1 27.20
17 22.0$\pm$0.6 14.48$\pm$0.05 27.2 24.82 50.3$\pm$3.7 14.25$\pm$0.03 18.6 22.59
18 60.6$\pm$1.5 13.40$\pm$0.05 - - 78.6$\pm$4.3 13.45$\pm$0.03 -20.4 21.99
19 - - - - 35.9$\pm$2.9 15.26$\pm$0.05 11.6 22.59
23 9.1$\pm$0.3 16.62$\pm$0.13 8.6 26.93 8.4$\pm$1.9 16.43$\pm$0.08 10.4 23.49
27 24.3$\pm$3.0 13.38$\pm$0.05 12.2 23.71 87.7$\pm$3.5 12.97$\pm$0.03 -1.1 15.28
28 - - - - 7.3$\pm$0.4 15.74$\pm$0.07 6.6 21.89
29 16.4$\pm$1.1 16.17$\pm$0.27 13.2 26.57 17.1$\pm$0.6 15.90$\pm$0.09 13.6 23.59
30 128.4$\pm$6.2 13.49$\pm$0.05 102.3 23.82 164.6$\pm$2.4 12.94$\pm$0.03 99.3 24.92
31 157.4$\pm$0.8 11.66$\pm$0.05 123.8 22.03 377.8$\pm$8.5 11.09$\pm$0.03 57.3 21.87
32 1.6$\pm$0.3 18.81$\pm$0.10 - - 4.5$\pm$0.4 18.08$\pm$1.07 8.6 24.89
33 - - - - 9.4$\pm$0.4 17.89$\pm$0.40 18.8 25.97
34 20.1$\pm$36.5 17.85$\pm$1.83 - - 5.3$\pm$1.2 17.80$\pm$0.47 15.6 25.55
36 - - - - 6.0$\pm$1.1 17.21$\pm$1.77 11.9 24.93
37 11.4$\pm$1.1 16.55$\pm$0.06 12.9 26.90 20.5$\pm$2.3 16.14$\pm$0.03 9.5 23.03
40 27.2$\pm$1.5 15.21$\pm$0.06 25.2 25.51 25.7$\pm$1.2 14.90$\pm$0.03 24.1 23.78
42 11.2$\pm$4.4 17.29$\pm$0.31 15.3 27.57 7.6$\pm$1.3 17.34$\pm$0.09 12.4 24.71
44 7.4$\pm$1.4 16.28$\pm$0.18 13.0 26.47 26.7$\pm$2.3 14.97$\pm$0.09 10.0 21.96
45 10.2$\pm$1.4 18.28$\pm$0.17 14.2 28.65 24.0$\pm$2.5 17.64$\pm$0.04 8.5 24.33
47 5.4$\pm$2.3 18.33$\pm$1.65 12.5 28.62 10.3$\pm$1.9 17.43$\pm$0.91 17.3 25.67
48 149.0$\pm$14.4 11.81$\pm$0.05 66.6 22.12 227.5$\pm$0.5 10.86$\pm$0.03 70.8 22.12
49 7.0$\pm$0.1 16.05$\pm$0.05 7.8 26.39 21.2$\pm$1.9 15.56$\pm$0.03 5.5 21.24
50 18.1$\pm$0.9 14.17$\pm$0.08 51.9 24.43 31.7$\pm$0.4 13.94$\pm$0.03 20.3 22.50
52 8.0$\pm$1.6 17.58$\pm$0.22 18.2 27.89 20.0$\pm$0.9 16.52$\pm$0.12 11.5 23.97
53 185.7$\pm$38.1 10.14$\pm$0.05 74.5 20.49 327.0$\pm$1.6 9.54$\pm$0.03 46.3 19.86
54 4.9$\pm$0.9 17.23$\pm$1.60 13.1 27.56 7.4$\pm$0.3 16.96$\pm$0.12 9.3 23.79
55 8.5$\pm$2.3 15.42$\pm$0.17 10.1 25.82 35.4$\pm$10.2 14.59$\pm$0.05 9.9 21.39
56 8.4$\pm$0.3 16.09$\pm$0.05 27.5 26.37 129.5$\pm$54.7 15.67$\pm$0.03 9.5 22.55
57 224.8$\pm$14.4 12.39$\pm$0.05 - - 139.1$\pm$2.1 11.91$\pm$0.03 -17.4 20.10
58 6.5$\pm$0.5 16.17$\pm$0.07 4.4 26.53 9.2$\pm$0.6 15.88$\pm$0.05 4.5 21.18
59 136.5$\pm$30.1 14.79$\pm$0.06 9.2 25.17 66.2$\pm$7.8 14.34$\pm$0.03 11.7 21.70
74 - - - - 29.9$\pm$2.8 13.30$\pm$0.06 25.7 22.36
75 200.8$\pm$1.6 12.02$\pm$0.05 22.6 22.43 305.4$\pm$12.1 12.10$\pm$0.03 -50.9 22.58
77 134.7$\pm$12.8 13.99$\pm$0.05 35.3 24.34 92.1$\pm$4.6 13.48$\pm$0.03 34.8 23.18
78 140.0$\pm$7.9 12.55$\pm$0.05 21.8 22.91 107.3$\pm$1.5 12.21$\pm$0.03 20.4 20.75
79 57.4$\pm$4.7 15.39$\pm$0.06 11.4 25.75 82.3$\pm$7.9 14.88$\pm$0.04 10.0 22.03
80 23.2$\pm$0.4 14.89$\pm$0.05 24.1 25.25 25.0$\pm$0.3 14.69$\pm$0.03 22.4 23.44
81 21.3$\pm$0.5 15.69$\pm$0.05 11.8 26.06 21.5$\pm$0.3 15.65$\pm$0.03 13.3 23.28
82 131.7$\pm$1.7 11.23$\pm$0.05 67.8 21.58 217.9$\pm$3.3 10.98$\pm$0.03 46.9 21.33
------- ---------------- ---------------- --------------- ------------------- ---------------- ---------------- --------------- -------------------
\[tab:uvphot\]
-------- --------------------- ----------------- ----------------- ----------------- -------------------------- --------- ------------ ------------ ----------------------- -----------------------
index ${\rm \log L_{W1}}$ ${\rm L_{W2}}$ ${\rm L_{W3}}$ ${\rm L_{W4}}$ ${\rm L_{W1,\odot}}$ W1-W2 r$_s$ r$_{eff}$ $\mu_{eff}$ ${\rm R_{75}/R_{25}}$
\[L$_{\odot}$\] \[L$_{\odot}$\] \[L$_{\odot}$\] \[L$_{\odot}$\] \[${\rm L_{\odot,W1}}$\] \[mag\] \[arcsec\] \[arcsec\] \[mag arcsec$^{-2}$\]
1$^f$ 5.295$\pm$0.047 4.850$\pm$0.174 - - 6.655$\pm$0.051 -0.123 65.1 23.0 22.281 2.18
2 5.472$\pm$0.015 5.068$\pm$0.049 - - 6.832$\pm$0.025 -0.020 21.6 20.7 20.722 2.68
3 5.772$\pm$0.020 5.452$\pm$0.045 - - 7.131$\pm$0.028 0.192 28.9 26.0 20.977 2.37
4 8.267$\pm$0.011 7.909$\pm$0.019 7.699$\pm$0.026 7.899$\pm$0.023 9.627$\pm$0.023 0.095 419.9 357.4 18.382 3.64
5 8.255$\pm$0.011 7.860$\pm$0.019 7.655$\pm$0.019 7.588$\pm$0.029 9.615$\pm$0.023 0.003 255.8 230.3 19.213 2.71
6 9.602$\pm$0.011 9.285$\pm$0.019 9.632$\pm$0.012 9.732$\pm$0.011 10.962$\pm$0.023 0.199 117.7 208.5 15.646 3.18
7 8.467$\pm$0.011 8.059$\pm$0.019 7.805$\pm$0.057 7.711$\pm$0.087 9.827$\pm$0.023 -0.031 348.5 282.5 19.388 2.62
8 7.707$\pm$0.012 7.341$\pm$0.020 7.343$\pm$0.033 7.747$\pm$0.024 9.066$\pm$0.023 0.076 51.6 64.6 18.527 3.59
9 6.987$\pm$0.013 6.532$\pm$0.028 5.710$\pm$0.183 - 8.346$\pm$0.024 -0.144 98.9 76.8 21.442 2.43
10 4.668$\pm$0.015 4.219$\pm$0.205 - - 6.028$\pm$0.025 -0.132 139.3 107.9 22.496 2.37
11 7.156$\pm$0.012 6.786$\pm$0.023 5.982$\pm$0.151 5.998$\pm$0.333 8.516$\pm$0.023 0.067 78.3 66.4 20.365 2.84
12 7.440$\pm$0.012 6.998$\pm$0.024 6.246$\pm$0.088 6.662$\pm$0.190 8.800$\pm$0.023 -0.113 252.1 131.1 21.020 3.17
13 5.452$\pm$0.084 5.118$\pm$0.273 - - 6.812$\pm$0.086 0.156 28.6 11.6 22.006 2.36
14 8.482$\pm$0.011 8.113$\pm$0.019 8.225$\pm$0.052 8.359$\pm$0.044 9.842$\pm$0.023 0.068 105.2 147.5 19.004 3.35
15 7.479$\pm$0.012 7.110$\pm$0.022 6.315$\pm$0.079 6.652$\pm$0.106 8.838$\pm$0.023 0.070 43.7 44.0 19.342 3.11
16 5.071$\pm$0.061 4.866$\pm$0.277 - - 6.431$\pm$0.064 0.478 - - - -
17 7.508$\pm$0.012 7.097$\pm$0.022 6.693$\pm$0.067 6.972$\pm$0.067 8.867$\pm$0.023 -0.035 36.2 37.3 19.175 2.47
18 7.298$\pm$0.012 6.941$\pm$0.021 6.513$\pm$0.047 6.726$\pm$0.060 8.658$\pm$0.023 0.097 36.6 18.2 18.005 4.53
19 7.898$\pm$0.022 7.534$\pm$0.053 - - 9.258$\pm$0.030 0.080 31.5 28.4 21.569 3.16
20 6.849$\pm$0.021 6.519$\pm$0.050 - - 8.209$\pm$0.029 0.166 51.3 45.7 21.793 2.30
21 6.944$\pm$0.014 6.580$\pm$0.029 5.894$\pm$0.110 5.928$\pm$0.163 8.303$\pm$0.024 0.081 15.3 15.4 18.984 2.60
22 7.981$\pm$0.012 7.625$\pm$0.021 7.395$\pm$0.033 7.622$\pm$0.037 9.341$\pm$0.023 0.100 80.3 53.6 18.575 2.40
23 5.679$\pm$0.029 5.269$\pm$0.075 - - 7.039$\pm$0.035 -0.033 40.7 19.3 21.808 2.07
24$^f$ 5.811$\pm$0.058 5.514$\pm$0.154 - - 7.171$\pm$0.061 0.248 19.8 11.9 21.791 2.25
25 5.331$\pm$0.029 4.988$\pm$0.069 5.397$\pm$0.115 - 6.690$\pm$0.035 0.134 3.7 13.0 21.241 4.77
26 6.870$\pm$0.014 6.472$\pm$0.025 - - 8.229$\pm$0.024 -0.005 71.5 50.1 20.894 2.71
27 7.322$\pm$0.012 6.942$\pm$0.021 6.159$\pm$0.035 6.404$\pm$0.043 8.681$\pm$0.023 0.043 62.7 32.0 18.168 3.33
28 5.710$\pm$0.044 5.290$\pm$0.115 - - 7.069$\pm$0.048 -0.058 15.6 15.5 21.177 2.81
29 6.414$\pm$0.020 6.070$\pm$0.047 - - 7.774$\pm$0.028 0.131 67.8 34.4 21.039 2.15
30 5.714$\pm$0.041 5.258$\pm$0.107 - - 7.074$\pm$0.046 -0.150 27.9 31.7 22.013 3.56
31 8.999$\pm$0.020 8.696$\pm$0.034 8.996$\pm$0.016 8.826$\pm$0.021 10.359$\pm$0.028 0.233 184.2 64.0 17.032 2.95
32 - - - - - - - - - -
33 5.602$\pm$0.041 5.256$\pm$0.136 - - 6.961$\pm$0.046 0.125 16.5 12.5 21.711 2.51
34 5.900$\pm$0.030 5.443$\pm$0.095 5.599$\pm$0.227 - 7.260$\pm$0.036 -0.152 36.2 16.2 21.178 2.22
35$^f$ 5.261$\pm$0.074 5.036$\pm$0.187 - - 6.620$\pm$0.077 0.429 - - - -
36 5.533$\pm$0.048 5.253$\pm$0.104 4.808$\pm$0.235 - 6.892$\pm$0.052 0.290 16.1 10.7 21.112 4.06
37 6.076$\pm$0.024 5.658$\pm$0.070 - - 7.435$\pm$0.031 -0.052 40.6 36.5 21.664 2.31
38 7.010$\pm$0.012 6.616$\pm$0.021 - - 8.369$\pm$0.023 0.007 82.8 78.0 19.671 2.72
39 5.878$\pm$0.027 5.458$\pm$0.073 - - 7.238$\pm$0.034 -0.059 22.5 14.6 21.050 2.36
40 6.479$\pm$0.018 6.140$\pm$0.038 - - 7.838$\pm$0.027 0.144 66.2 48.3 21.403 2.30
41 5.033$\pm$0.044 4.845$\pm$0.094 - 5.098$\pm$0.243 6.392$\pm$0.048 0.520 30.6 15.5 21.956 3.80
42 5.626$\pm$0.036 5.146$\pm$0.157 - - 6.985$\pm$0.041 -0.208 51.7 22.5 22.045 2.35
43 9.323$\pm$0.011 9.008$\pm$0.019 9.242$\pm$0.011 9.146$\pm$0.011 10.682$\pm$0.023 0.203 82.4 182.2 16.285 2.95
44 6.907$\pm$0.012 6.540$\pm$0.021 5.983$\pm$0.060 5.819$\pm$0.083 8.266$\pm$0.023 0.075 51.8 53.8 19.062 3.30
45 - - - - - - - - - -
46 8.252$\pm$0.011 7.855$\pm$0.019 6.322$\pm$0.073 6.288$\pm$0.106 9.611$\pm$0.023 0.000 122.6 80.4 17.290 5.32
47 5.840$\pm$0.032 5.377$\pm$0.121 - - 7.199$\pm$0.038 -0.166 28.3 19.1 21.438 2.47
48 9.859$\pm$0.011 9.472$\pm$0.019 9.240$\pm$0.011 9.113$\pm$0.012 11.218$\pm$0.023 0.024 138.4 215.1 16.350 5.66
49 6.578$\pm$0.015 6.224$\pm$0.033 - - 7.937$\pm$0.025 0.106 29.7 23.0 19.767 2.56
50 6.712$\pm$0.015 6.248$\pm$0.033 - - 8.071$\pm$0.025 -0.170 59.5 57.1 21.254 2.40
51 8.353$\pm$0.012 7.967$\pm$0.021 7.785$\pm$0.102 7.767$\pm$0.069 9.712$\pm$0.023 0.028 131.6 231.2 20.299 3.36
52 6.368$\pm$0.020 6.030$\pm$0.044 - - 7.727$\pm$0.028 0.147 29.5 21.7 20.530 2.91
53 9.536$\pm$0.011 9.188$\pm$0.019 9.520$\pm$0.011 9.550$\pm$0.012 10.896$\pm$0.023 0.119 160.6 144.6 16.983 3.28
54 4.408$\pm$0.105 4.288$\pm$0.220 4.399$\pm$0.275 - 5.767$\pm$0.107 0.693 - - - -
55 6.632$\pm$0.012 6.243$\pm$0.021 5.002$\pm$0.037 5.078$\pm$0.095 7.992$\pm$0.023 0.019 48.3 28.1 18.595 3.07
56 6.018$\pm$0.027 5.668$\pm$0.073 - - 7.377$\pm$0.034 0.117 39.8 23.6 21.394 2.17
57 7.952$\pm$0.011 7.889$\pm$0.019 8.374$\pm$0.019 8.814$\pm$0.011 9.312$\pm$0.023 0.832 22.1 31.9 16.786 5.67
58 6.690$\pm$0.015 6.288$\pm$0.031 5.640$\pm$0.113 5.649$\pm$0.198 8.049$\pm$0.025 -0.013 49.8 39.7 20.636 3.00
59 7.336$\pm$0.012 6.944$\pm$0.021 6.453$\pm$0.033 6.454$\pm$0.098 8.695$\pm$0.023 0.010 85.9 53.4 19.660 2.70
60 6.474$\pm$0.019 6.123$\pm$0.042 5.433$\pm$0.192 - 7.833$\pm$0.028 0.114 53.6 45.4 21.500 2.20
61 4.794$\pm$0.087 4.459$\pm$0.349 - - 6.153$\pm$0.089 99.000 41.6 13.2 22.924 3.96
62 6.819$\pm$0.023 6.404$\pm$0.039 - - 8.178$\pm$0.031 -0.046 78.0 40.3 20.169 2.50
63 7.003$\pm$0.012 6.591$\pm$0.021 6.044$\pm$0.035 6.002$\pm$0.063 8.362$\pm$0.023 -0.038 112.2 78.6 19.622 2.80
64 4.690$\pm$0.107 4.408$\pm$0.345 - - 6.050$\pm$0.109 99.000 9.6 6.0 22.199 2.10
65 7.287$\pm$0.012 6.927$\pm$0.022 7.002$\pm$0.022 7.603$\pm$0.021 8.647$\pm$0.023 0.090 50.1 35.2 19.096 2.15
66 9.413$\pm$0.011 9.285$\pm$0.019 9.734$\pm$0.011 9.795$\pm$0.012 10.772$\pm$0.023 0.671 60.3 57.8 14.739 12.0
67 5.494$\pm$0.044 5.195$\pm$0.089 - - 6.854$\pm$0.048 0.244 31.7 28.5 21.727 2.57
68 5.674$\pm$0.079 5.420$\pm$0.148 - - 7.033$\pm$0.082 0.357 45.5 11.9 21.046 2.47
69 6.461$\pm$0.028 6.034$\pm$0.051 - - 7.820$\pm$0.034 -0.074 8.1 10.9 18.927 4.75
70 6.890$\pm$0.017 6.496$\pm$0.033 5.831$\pm$0.111 6.119$\pm$0.263 8.249$\pm$0.026 0.007 31.1 23.6 19.661 2.59
71 7.403$\pm$0.014 6.975$\pm$0.026 - - 8.762$\pm$0.024 -0.079 44.5 50.3 20.137 2.52
72 8.118$\pm$0.012 7.726$\pm$0.020 7.006$\pm$0.030 7.582$\pm$0.020 9.478$\pm$0.023 0.011 84.4 213.9 19.032 3.33
73 6.125$\pm$0.033 5.784$\pm$0.067 3.867$\pm$0.220 - 7.484$\pm$0.039 0.140 - - - -
74 7.677$\pm$0.013 7.358$\pm$0.024 6.982$\pm$0.060 7.207$\pm$0.083 9.037$\pm$0.024 0.193 126.8 45.8 18.892 2.38
75 7.032$\pm$0.012 6.691$\pm$0.021 6.648$\pm$0.021 7.225$\pm$0.013 8.391$\pm$0.023 0.140 36.3 40.0 18.307 2.75
-------- --------------------- ----------------- ----------------- ----------------- -------------------------- --------- ------------ ------------ ----------------------- -----------------------
\[tab:irphot\]
-------- --------------------- ----------------- ------------------ ------------------ -------------------------- --------- ------------ ------------- ----------------------- -------------------------
index ${\rm \log L_{W1}}$ ${\rm L_{W2}}$ ${\rm L_{W3}}^1$ ${\rm L_{W4}}^1$ ${\rm L_{W1}}^2$ W1-W2 r$_s^3$ r$_{eff}^4$ $\mu_{eff}^4$ ${\rm R_{75}/R_{25}}^5$
\[L$_{\odot}$\] \[L$_{\odot}$\] \[L$_{\odot}$\] \[L$_{\odot}$\] \[${\rm L_{\odot,W1}}$\] \[mag\] \[arcsec\] \[arcsec\] \[mag arcsec$^{-2}$\]
76$^f$ 6.110$\pm$0.048 5.608$\pm$0.203 - - 7.470$\pm$0.052 -0.265 32.6 17.2 21.587 2.11
77 8.122$\pm$0.011 7.748$\pm$0.019 7.506$\pm$0.026 7.769$\pm$0.029 9.481$\pm$0.023 0.059 99.0 93.8 18.584 2.77
78 7.266$\pm$0.040 6.911$\pm$0.059 6.555$\pm$0.049 6.493$\pm$0.079 8.625$\pm$0.045 0.103 74.3 60.1 18.653 3.02
79 5.823$\pm$0.018 5.463$\pm$0.043 - - 7.183$\pm$0.027 0.091 55.5 42.9 21.272 2.45
80 6.475$\pm$0.015 6.130$\pm$0.034 - - 7.835$\pm$0.025 0.129 66.8 52.5 20.794 2.52
81 6.261$\pm$0.026 5.827$\pm$0.075 - 5.492$\pm$0.329 7.620$\pm$0.033 -0.092 17.4 15.6 20.520 2.91
82 8.514$\pm$0.011 8.153$\pm$0.019 8.235$\pm$0.014 8.063$\pm$0.018 9.873$\pm$0.023 0.090 131.5 119.2 18.357 2.61
-------- --------------------- ----------------- ------------------ ------------------ -------------------------- --------- ------------ ------------- ----------------------- -------------------------
\[tab:irphot\]
Atlas of the large galaxies {#sec:appendix_figure}
===========================
In Figure \[fig:atlas\_conv\] we present the resolution-matched distributions of $M_*$, SFR and $\mHI$ (left side) and radial SB profiles (right side) for the ten large LVHIS galaxies (see Section \[sec:phot\_conv\] and \[sec:loc\_SFE\]). Ellipses in the images and vertical lines in the SB profile figures mark three radii, at 1.25 R$_{25}$ (dotted), $\Sigma_*=3~\mspc$ (dashed), and $\SHI = 3~\mspc$ (dash-dotted), which are also annotated with star symbols in the $\Sigma_*$ and $\SHI$ profiles.
In Figure \[fig:atlas\_highrs\] we present images for the same sample of 10 large LVHIS galaxies showing the distribution of $M_*$ (at the resolution of the WISE W1-band), SFR (at the resolution of the WISE W4-band), and the mean ATCA HI velocity field at the original resolution (left side) as well as our tilted ring models and derived tiltograms (right side). Ellipses in the images highlight the onset of the warp ($r_{warp}$, solid line), if present, and the R$_{25}$ isophote (dashed line). For each tilted ring model we show Ð from top to bottom Ð the rotational velocity ($v_{\rm rot}$), position angle (PA), inclination angle (i) and HI surface brightness ($\SHI$) as a function of radius. The approaching, receding and mean values are coloured in blue, red and black, respectively. The tiltograms are for the approaching velocity side of each galaxy; the magenta dashed lines mark $r_{warp}$, if a warp is present.
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="12cm"} {width="5cm"}
{width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"}
{width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"} {width="10cm"} {width="3cm"} {width="4cm"}
[^1]: Email: hyacinthwj@gmail.com
[^2]: Email:Baerbel.Koribalski@csiro.au
[^3]: http://www.atnf.csiro.au/people/Baerbel.Koribalski/LVHIS/LVHIS-galaxies.html
[^4]: Masking foreground stars through visual inspection is subject to uncertainties, but generally speaking pixels associated with foreground stars identified from the NUV$/$FUV values are a small fraction ($<0.02\%$) of all the pixels. Errors introduced in the NUV fluxes through this step do not affect the science analysis (which uses the FUV but not the NUV fluxes) in this paper.
[^5]: Fixing surface densities at 3 $\mspc$ rather than lower values comes from a compromise between limiting our analysis in the [[H]{}[<span style="font-variant:small-caps;">i</span>]{}]{}-dominated regions and the limited depth of the $\Sigma_*$ data. If we use 2 $\mspc$ instead of 3 $\mspc$, the general trends do not change much but we will be at the risk of relying on extrapolated $\Sigma_*$ values for more than half of the sample. If we use 4 $\mspc$ instead of 3 $\mspc$, we will be at the risk of reaching the molecular dominated high gas surface density regions.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper presents an efficient technique to prune deep and/or wide convolutional neural network models by eliminating redundant features (or filters). Previous studies have shown that over-sized deep neural network models tend to produce a lot of redundant features that are either shifted version of one another or are very similar and show little or no variations; thus resulting in filtering redundancy. We propose to prune these redundant features along with their connecting feature maps according to their differentiation and based on their relative cosine distances in the feature space, thus yielding smaller network size with reduced inference costs and competitive performance. We empirically show on select models and CIFAR-10 dataset that inference costs can be reduced by 40% for VGG-16, 27% for ResNet-56, and 39% for ResNet-110.'
author:
- |
Babajide O. Ayinde & Jacek M. Zurada\
Department of Electrical and Computer Engineering\
University of Louisville\
Louisville, KY 40229, USA\
`{babajide.ayinde,jacek.zurada}@louisville.edu`
bibliography:
- 'ref.bib'
title: Building Efficient ConvNets using Redundant Feature Pruning
---
Introduction {#headings1}
============
Recent studies indicate that over-sized deep learning models typically result in largely over-determined (or over-complete) systems [@denil2013predicting; @rodriguez2016regularizing; @bengio2009slow; @changpinyo2017power; @ayinde2017nonredundant; @han2015deep; @han2016dsd]. The resulting architectures may therefore be less computationally efficient due to their size, over-parameterization, and largely due to their high inference cost. To account for the scale, diversity and the difficulty of data these models learn from, the architectural complexity and the excessive number of weights and units are often deliberately built in into the deep neural network models by design [@bengio2007scaling; @changpinyo2017power]. These over-sized models have expensive inference costs especially for applications with constrained computational and power resources such as web services, mobile and embedded devices. In addition to good accuracy, such resource-limited applications benefit greatly from lower inference cost [@li2016pruning; @szegedy2016rethinking].\
\
In this paper, we focus on controlled network size reduction of well-trained deep learning models based on feature agglomeration followed by *one-shot* elimination of redundant features and retraining heuristic. By leveraging on the observations that large capacity CNNs usually have significant redundancy among different features, we propose a simple, intuitive, and efficient way to remove such redundancy without undermining the efficiency or introducing sparsity that would require specialized library and/or hardware. We find that our pruning technique improves inference cost over a recently proposed technique [@li2016pruning] across benchmark models and dataset considered without modifying existing hyperparameters.
Related Work {#headings2}
============
Storage and computational cost reduction via network pruning techniques has a long history [@NIPS1989_250; @hassibi1993second; @mariet2015diversity; @ioannou2015training; @polyak2015channel; @molchanov2016pruning]. For instance, Optimal Brain Damage [@NIPS1989_250] and Optimal Brain Surgeon [@hassibi1993second] use second-order derivative information of the loss function to prune redundant network parameters. Other related work include but is not limited to @anwar2017structured which prunes based on particle filtering, @mathieu2013fast uses FFT to avoid overhead due to convolution operation, and @howard2017mobilenets uses depth multiplier method to scale down the number of filters in each convolutional layer. Closely related to our work, @li2016pruning sorts and prunes filters based on the sum of their absolute weights and @han2015learning prunes weights with magnitude below a set threshold.
Feature Clustering and Pruning {#headings3}
==============================
The objective here is to discover $n_f$ clusters in the set of $n'$ original filters that are identical or very similar in weight space according to a well-defined similarity measure, where $n_f\leq n'$.
{width="0.85\linewidth"}
Achieving this involves choosing suitable similarity measures to express the inter-feature distances between features $\mathbf{\phi}_i$ that connect the feature map $Z_{l-1}$ of layer $l-1$ to neurons of layer $l$. In other words, $\mathbf{\phi}^{(l)}_i$ , i=1,...$n'_l$, are feature vectors in layer $l$, each $\mathbf{\phi}^{(l)}_i \in \mathbb{R}^{p}$ corresponds to the $i$-th column of the kernel matrix $\mathbf{W}^{(l)} = [\mathbf{\phi}^{(l)}_1, \;\;...\mathbf{\phi}^{(l)}_{n'_l}] \in \mathbb{R}^{p\times n'_l}$ where $p= k^2 n'_{l-1}$ and $k$ is the size of square 2D kernel $\Psi \in \mathbb{R}^{k\times k}$. A number of suitable agglomerative similarity testing/clustering algorithms can be applied for localizing redundant features. Based on a comparative review, a clustering approach from @walter2008fast [@ding2002cluster] has been adapted and reformulated for this purpose. By starting with each weight vector $\mathbf{\phi}_i$ as a potential cluster, agglomerative clustering is performed by merging the two most similar clusters $C_a$ and $C_b$ as long as the average similarity between their constituent feature vectors is above a chosen cluster similarity threshold denoted as $\tau$ [@leibe2004combined; @manickam2000intelligent]. The pair of clusters $C_a$ and $C_b$ exhibits average mutual similarities as follows: $$\label{MyEq11}
\begin{split}
\overline{SIM_C}(C_a, C_b) &= \frac{\sum_{\mathbf{\phi}_i\in C_a, \mathbf{\phi}_j\in C_b}SIM_C(\mathbf{\phi}_i,\mathbf{\phi}_j)}{|C_a|\times |C_b|} > \tau \\
& a,b = 1,...n'_l;\,\, a\neq b ;\,\,\, i = 1,...|C_a|; \\
& j=1,...|C_b|; \;\;\;\ and \;\;\;\; i\neq j
\end{split}$$ where $SIM_C(\mathbf{\phi}_1,\mathbf{\phi}_2) = \frac{<\mathbf{\phi}_1,\mathbf{\phi}_2>}{\parallel\mathbf{\phi}_1\parallel\parallel\mathbf{\phi}_2\parallel}$ is the cosine similarity between two features and $<\mathbf{\phi}_1,\mathbf{\phi}_2>$ is the inner product of arbitrary feature vectors $\mathbf{\phi}_1$ and $\mathbf{\phi}_2$, and $\tau$ is a set threshold.\
\
The redundant-feature-based pruning procedure for $l^{th}$ convolutional layer is summarized as follows:
1. Group all the filters $\phi_i$ (columns of the kernel matrix) into $n_f$ clusters whose average similarities are above a set threshold $\tau$.
2. Two heuristics are considered: (A) Randomly sample one representative filter from each of the $n_f$ clusters and prune the remaining filters and their corresponding feature maps. (B) Randomly prune $n'-n_f$ filters and their corresponding feature maps. The weights of the pruned feature maps in $l^{th}$ layer are also removed in layer $(l+1)^{th}$ as shown in Figure \[pruning\].
3. A new kernel matrix is defined for both $l^{th}$ and $(l+1)^{th}$ layer of a new smaller model.
Experiments {#headings4}
===========
The network pruning was implemented in Pytorch deep learning library [@paszke2017automatic]. We evaluated the proposed redundant-feature-based pruning on three deep networks, namely: VGG-16 [@simonyan2015very] and two residual networks (ResNet-56 and 110) [@he2016deep] trained on CIFAR-10. The baseline model and accuracy for residual networks were obtained by training the model following the procedures highlighted in @he2016deep. See Appendix for implementation details and supplemental results for both VGG-16 and residual networks.
VGG-16 on CIFAR-10
------------------
As seen in Table \[eval\_table1\], for $\tau=0.54$ our approaches (both A and B) outperform that in @li2016pruning and are able to prune more than $78\%$ of the parameters resulting in $40\%$ FLOP reduction and a competitive classification accuracy. We suspect that our pruning approach outshines that of @li2016pruning, which ranks importance of filters based on the sum of absolute value of their weights, because it localizes and prunes similar or shifted versions of filters that do not add extra information to the feature hierarchy. This notion is reinforced from information theory standpoint that the activation of one unit should not be predictable based on the activations of other units of the same layer [@rodriguez2017regularizing]. Another crucial observation is that heuristic A achieves a better accuracy than heuristic B because random pruning might prune filters that are dissimilar.
RESNET-56/110 on CIFAR-10
-------------------------
For RestNet-56/110, we only prune the first layer of the residual block to avoid dimensions mismatch due to unavailability of projection mapping for selecting the identity mapping (see @he2016deep for details). We found that redundant-feature-based pruning are competitive to that in @li2016pruning in terms of $\%$ FLOP reduction. However, for RestNet-56, our approach prunes $10\%$ more parameters than @li2016pruning and upon retraining, it achieves better classification accuracy.
Conclusion {#headings5}
==========
Motivated by the observations of recent studies that modern CNNs often have large number of overlapping filters amounting to unnecessary filtering redundancy and large inference cost. By using hierarchical agglomerative clustering to group all filters at each layer in the weight space according to a predefined measure, redundant filters are pruned and inference cost (FLOPS) reduced by $40\%$ for VGG-16, $28\%$/$39\%$ for ResNet-56/110 trained on CIFAR-10. To recover the accuracy after pruning, we retrained the model for a few iterations without the need to modify hyper-parameters.
### Acknowledgments {#acknowledgments .unnumbered}
This work was supported by the NSF under grant 1641042.
Appendix {#appendix .unnumbered}
========
Implementation Details and Results
==================================
All experiments were performed on Intel(r) Core(TM) i7-6700 CPU @ 3.40Ghz and a 64GB of RAM running a 64-bit Ubuntu 14.04 edition. The software implementation has been in Pytorch library [^1] on two Titan X 12GB GPUs and the filter clustering was implemented in SciPy ecosystem @scipy. The agglomeration of filters using hierarchical clustering is practical for very wide and deep networks even though its complexity is $\emph{O}((n'_l)^2\text{log}(n'_l))$. For instance, clustering VGG-16 feature vectors empirically takes on the average on our machine $14.1$ milliseconds and this is executed only once during training. This amounts to a negligible computational overhead for most deep architectures.\
\
The implementation of our filter pruning strategy is similar to that in @li2016pruning in the sense that when a particular filter of a convolutional layer is pruned, its corresponding feature map is also pruned and the weights of the pruned feature map in the filter of the next convolutional layer are equally pruned. It must be emphasized that after pruning the feature maps of last convolutional layer, the input to the linear layer has changed and its weight matrix has to be pruned accordingly. CIFAR-10 dataset was used in all the experiments to train and validate the models. The dataset contains a labeled set of 60,000 32x32 color images belonging to 10 classes: airplanes, automobiles, birds, cats, deer, dogs, frogs, horses, ships, and trucks. The dataset is split into $50000$ and $10000$ training and testing sets, respectively. We used FLOP to compare the computational efficiency of the models considered because its evaluation is independent of the underlying software and hardware. In order to fairly compare our method with that in @li2016pruning, we also calculated the FLOP only for the convolution and fully connected layers.
VGG-16 on CIFAR-10
------------------
For the first set of experiments, we used a modified version of the popular convolutional neural network known as the VGG-16 (@simonyan2015very), which has 13 convolutional layers and 2 fully connected layer. In the modified version of VGG-16, each layer of convolution is followed by a Batch Normalization layer [@ioffe2015batch]. Our base model was trained for 350 epochs, with a batch-size of 128 and a learning rate 0.1 as highlighted in the repository[^2]. The learning rate was reduced by a factor of 10 at 150 and 250 epochs. We have shared our pruning implementation and trained model for reproducibility of results [^3]. After pruning we retrain the network with learning rate of 0.001 for 80 epochs to fine tune the weights of the remaining connections to regain the accuracy.\
Figure \[filter\_vgg16\] shows the number of nonredundant filters per layer for different $\tau$ values. As can be seen that some convolutional layers in VGG are prone to extracting features with very high correlation such as Conv layer 1, 11, 12, and 13. Another very important observation is that later layers of VGG are more susceptible to extracting redundant filters than earlier layers and can be pruned heavily. Figure \[acc\_vgg16\](a) shows the sensitivity of the convolutional layer of VGG-16 to pruning and it can be observed that layers such as Conv 1, 3, 4, 9, 11, and 12 are very sensitive. However, as can be observed in Figure \[acc\_vgg16\](c), accuracy can be restored after pruning filters in later layers (Conv 9, 11, and 12) compared to early ones (Conv 1, 3, and 4).
[[(a) Prune redundant filters]{}]{}
[[(b) Prune $n'-n_f$ random filters]{}]{}
[[(c) Prune and retrain]{}]{}
For our final test score, we fine tuned on the entire training set. For pruning, we performed a grid search over $\tau$ values within 0.1 and 1.0, and found 0.54 gave the least test error. Table \[eval\_table2\] reports the pruning performance for $\tau=0.54$ and it can be easily seen that more than 90% of most of the latter layers have been pruned and most of the sensitive earlier layers are minimally pruned. Figure \[acc\_vgg16\](b) depicts the sensitivity of trained VGG-16 model to pruning using heuristic B that calculates the number of redundant filters ($n'-n_f$) and randomly prunes them.
layer $v_l\times h_l$ \#Maps FLOP \#Params \#Maps FLOP%
---------- ----------------- -------- --------- ---------- -------- --------
Conv\_1 $32\times 32$ 64 1.8E+06 1.7E+03 32 50.0%
Conv\_2 $32\times 32$ 64 3.8E+07 3.7E+04 58 54.7%
Conv\_3 $16\times 16$ 128 1.9E+07 7.4E+04 125 11.5%
Conv\_4 $16\times 16$ 128 3.8E+07 1.5E+05 128 2.3%
Conv\_5 $8\times 8$ 256 1.9E+07 2.9E+05 256 0%
Conv\_6 $8\times 8$ 256 3.8E+07 5.9E+05 254 0.8%
Conv\_7 $8\times 8$ 256 3.8E+07 5.9E+05 252 2.3%
Conv\_8 $4\times 4$ 512 1.9E+07 1.2E+06 299 42.5%
Conv\_9 $4\times 4$ 512 3.8E+07 2.4E+06 164 81.3%
Conv\_10 $4\times 4$ 512 3.8E+07 2.4E+06 121 92.4%
Conv\_11 $2\times 2$ 512 9.4E+06 2.4E+06 59 97.3%
Conv\_12 $2\times 2$ 512 9.4E+06 2.4E+06 104 97.7%
Conv\_13 $2\times 2$ 512 9.4E+06 2.4E+06 129 94.9 %
: Pruning performance on CIFAR dataset using VGG-16 model at $\tau=0.54$[]{data-label="eval_table2"}
RESNET-56/110 on CIFAR-10
-------------------------
The architecture of residual networks is more complex than VGG and also the number of parameters in the fully connected layer is relatively smaller and this makes it a bit challenging to prune a large proportion of the parameters. Both ResNet-56 and ResNet-110 have three stages of residual blocks for feature maps with of differing sizes. The sizes ($v_l\times h_l$) of feature maps in stages 1,2, and 3 are $32\times32$, $16\times16$, and $8\times8$, respectively. Each stage has 9 and 18 residual blocks for ResNet-56 and ResNet-110, respectively. A residual block consists of two convolutional layer each followed by a Batch Normalization layer. Preceding the first stage is a convolutional layer followed by a Batch Normalization layer[^4]. Only the redundant filters in first convolution layer of each block are pruned due to unavailability of mapping for selecting the identity feature maps.\
\
As can be observed in Figures \[filter\_resnet56\] and \[filter\_resnet110\] that convolutional layers in first stage are prone to extracting more redundant features than those of second stage, and the convolutional layers in the second stage are susceptible to extracting redundant filters than those of third block, which is contrary to the observations with VGG-16. In effect, more filters could be pruned from layers in first stage than the latter ones without losing much to accuracy. More specifically, many layers in the first stage of ResNet-56, such as Conv 2,8,10, and 26, have filters that are correlated more $80\%$ and could be heavily pruned. Similarly, convolutional layers in the first stage of ResNet-110 exhibit similar tendency to produce more filters that are redundant. As a result of these differing tendencies at each stage, $\tau$ for all the stages is set to different values. In pruning ResNet-56, we set $\tau$ to $0.253$, $0.223$, $0.20$ as thresholds for stages 1,2, and 3, respectively. Similarly for ResNet-110 we used $0.18$, $0.12$, and $0.17$.\
\
Figure \[acc\_resnet56\] shows the sensitivity of the convolutional layer of ResNet-56 to pruning and it can be observed that layers such as Conv 10, 14, 16, 18, 20, 34, 36, 38, 52 and 54 are more sensitive to filter pruning than other convolutional layers. Likewise for ResNet-110, the sensitivity of the convolutional layer to pruning is depicted in Figure \[acc\_resnet110\] and it can be observed that Conv 1, 2, 38, 78, and 108 are sensitive to pruning. In order to regain the accuracy by retraining the pruned model, we skip these sensitive layers while pruning.\
\
As seen on Table \[eval\_table1\] for ResNet-56, redundant-feature-based pruning (both A and B) have competitive performance in terms of FLOP reduction but outperform that in @li2016pruning in reducing the number of effective parameters by $10\%$ with relatively better classification accuracy after retraining. However, we were able to marginally increase the effective number of parameters pruned in ResNet-110 from $38.6\%$ to $39.1\%$, which gives rise to approximately $2\%$ increase. Also, the accuracy after fine tuning the pruned ResNet-110 model is not as good as that in @li2016pruning.\
\
Model FLOP Pruned % Time(s) Saved %
--------------------------------- --------------------- ---------- --------- --------- --
VGG-16 $ 3.13 \times 10^8$ 1.47
VGG-16-pruned-A (this work) 1.86 $\times 10^8$ 40.5% 0.94 34.01%
ResNet-56 1.25 $\times 10^8$ 1.16
ResNet-56-pruned-A (this work) 9.07 $\times 10^7$ 27.9% 0.96 17.2%
ResNet-110 2.53 $\times 10^8$ 2.22
ResNet-110-pruned-A (this work) 1.54 $\times 10^8$ 39.1% 1.80 18.9%
: FLOP and wall-clock time reduction for inference. Operations in convolutional and fully connected layer are considered for computing FLOP[]{data-label="eval_table4"}
The inference time of both original and pruned models was recorded and reported in Table \[eval\_table4\]. 10000 test images of CIFAR-10 dataset were used for the timing evaluation conducted in Pytorch version 0.2.0\_3 with Titan X (Pascal) GPU and cuDNN v8.0.44, using a mini-batch of size 100. It can be observed that %FLOP reduction also translates almost directly into inference clock time savings.
Prune and Train from Scratch
----------------------------
In order to see the effect of copying weights from the original (larger) model to a pruned (smaller) model, we pruned two models (VGG-16 and ResNet-56) as described above and re-initialized their weights and trained them from scratch. As shown in Table \[eval\_table3\] that fine tuning a pruned model is almost always better than re-initializing and training a pruned model from scratch. We believe that already-trained filters may serve as good initialization for a smaller network which might on its own be difficult to train. Other observation from Table \[eval\_table3\] is that redundant-feature-based pruning results in an architecture that when trained attains a better performance than its counterpart in @li2016pruning. This may indicate that redundant-feature-based pruning might be a potential approach to determining the architectural width of modern deep neural network models.
[^1]: <http://pytorch.org/>
[^2]: Implementation of modified version of VGG-16 can be found in <https://github.com/kuangliu/pytorch-cifar>
[^3]: <https://github.com/babajide07/Redundant-Feature-Pruning-Pytorch-Implementation>
[^4]: We used the Pytorch implementation of ResNet56/110 in <https://github.com/D-X-Y/ResNeXt-DenseNet> as baseline models
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The Erdős-Ginzburg-Ziv constant of an abelian group $G$, denoted $\mathfrak{s}(G)$, is the smallest $k\in\mathbb{N}$ such that any sequence of elements of $G$ of length $k$ contains a zero-sum subsequence of length $\exp(G)$. In this paper, we use the partition rank from [@Naslund2017Partition], which generalizes the slice rank, to prove that for any odd prime $p$, $$\mathfrak{s}\left(\mathbb{F}_{p}^{n}\right)\leq(p-1)2^{p}\left(J(p)\cdot p\right)^{n}$$ where $0.8414<J(p)<0.91837$ is the constant appearing in Ellenberg and Gijswijt’s bound on arithmetic progression-free subsets of $\mathbb{F}_{p}^{n}$ [@EllenbergGijswijtCapsets]. For large $n$, and $p>3$, this is the first exponential improvement to the trivial bound. We also provide a near optimal result conditional on the conjecture that $\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}$ satisfies *property $D$,* as defined in [@GaoGeroldinger2006ZeroSumSurvey], showing that in this case $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\leq(k-1)4^{n}+k.$$'
author:
- Eric Naslund
title: 'Exponential Bounds for the Erdős-Ginzburg-Ziv Constant'
---
Introduction
============
For an abelian group $G$, let $\exp(G)$ denote the exponent of $G$, which is the maximal order of any element in $G$. The Erdős-Ginzburg-Ziv constant of $G$, denoted $\mathfrak{s}(G)$, is defined to be the smallest $k\in\mathbb{N}$ such that any sequence of elements from $G$ of length $k$ contains a zero-sum subsequence of length $\exp(G)$. In 1961, Erdős, Ginzburg, and Ziv [@ErdosGinzburgZiv1961Original] proved that $$\mathfrak{s}\left(\mathbb{Z}/k\mathbb{Z}\right)=2k-1.$$ That is, among any sequence of $2k-1$ integers there is a subsequence of length $k$ which sums to $0$, and furthermore this is not true if $2k-1$ is replaced by $2k-2$. In 2007, Reiher [@Reiher2007KemnitzConj] resolved a longstanding conjecture of Kemnitz [@Kemnitz1983Conj], and proved that $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{2}\right)=4k-3.$$ When $G$ is a power of a cyclic group, $\mathfrak{s}(G)$ has a geometric interpretation: $\mathfrak{s}((\mathbb{Z}/k\mathbb{Z})^{n})$ is the smallest integer such that any set of $\mathfrak{s}((\mathbb{Z}/k\mathbb{Z})^{n})$ points in $\mathbb{Z}^{n}$ contains $k$ points whose average is again a lattice point. When $G=\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}$ with $n$ large, very little is known. Harborth [@Harborth1973zerosumsets] gave the elementary bounds $$(k-1)2^{n}+1\leq\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\leq(k-1)k^{n}+1,$$ and Elsholtz [@Elsholtz2003EGZLowerBounds] improved the lower bound to $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\geq(k-1)2^{n}\cdot\left(\frac{9}{8}\right)^{\left[\frac{n}{3}\right]}+1\approx(k-1)(2.08)^{n}.$$ In a different direction, Alon and Dubiner [@AlonDubiner1993Zerosumsets] proved that $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\leq\left(cn\log n\right)^{n}k$$ for some $c>0$, which implies that for fixed $n$, $\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)$ grows linearly in $k$. In this paper, we give a conditional improvement to Alon and Dubiner’s bound, that is within a factor of $2^{n}$ of the lower bound, and we give an unconditional exponential improvement to the upper bounds for $\mathfrak{s}\left(\left(\mathbb{Z}/p\mathbb{Z}\right)^{n}\right)$ for large $n$. Our main result is:
\[thm:Main\_Theorem\]Suppose that $A\subset\mathbb{F}_{p}^{n}$ does not contain $p$ distinct elements $x_{1},\dots,x_{p}$ such that $$x_{1}+\cdots+x_{p}=0.$$ Then $$|A|\leq(2^{p}-p-2)(pJ(p))^{n}$$ where $0.8414\leq J(p)\leq0.918$ is the constant appearing in the Ellenberg-Gijswijt result on three term progressions in $\mathbb{F}_{p}^{n}$. Consequently, $$\mathfrak{s}\left(\mathbb{F}_{p}^{n}\right)\leq(p-1)2^{p}(pJ(p))^{n}.$$
Note that the case $p=3$ is a consequence of Ellenberg and Gijswijt’s theorem [@EllenbergGijswijtCapsets]. The bound for $\mathfrak{s}(\mathbb{F}_{p}^{n})$ has since been improved to $2p\cdot(pJ(p))^{n}$ by Fox and Sauermann [@FoxSauermann2017EGZ]. In section \[sec:Reduction-to-Prime\], we will see that bounding $\mathfrak{s}\left((\mathbb{Z}/k\mathbb{Z})^{n}\right)$ for general $k$ can be reduced to bounding $\mathfrak{s}(\mathbb{F}_{p}^{n})$ for $p|k$, due to the results of Fox and Sauermann [@FoxSauermann2017EGZ]. Using this reduction, we have the following corollary:
\[cor: cp\_asymp\]Let $p$ be the largest prime power dividing $k$. Then $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\leq\left(J(p)p+o(1)\right)^{n},$$ and the $o(1)$ term tends to $0$ as $n\rightarrow\infty$.
The function $J(p)$ is given explicitely by the minimization $$J(p)=\frac{1}{p}\min_{0<x<1}\frac{1-x^{p}}{1-x}x^{-\frac{p-1}{3}}.$$ In [@BlasiakChurchCohnGrochowNaslundSawinUmans2016MatrixMultiplication prop. 4.12] it was shown that $J(p)$ is a decreasing function of $p$ that satisfies $$J(3)=\frac{1}{8}\sqrt[3]{207+33\sqrt{33}}=0.9183\dots,$$ and $$\lim_{p\rightarrow\infty}J(p)=\inf_{z>1}\frac{z-z^{-2}}{3\log z}=0.8414\dots.$$ The proof of theorem \[thm:Main\_Theorem\] uses a variant of *slice rank method* [@TaosBlogCapsets] that we call the *partition rank method* [@Naslund2017Partition]. The slice rank method was introduced by Tao [@TaosBlogCapsets] following the work of Ellenberg and Gijswijt [@EllenbergGijswijtCapsets], and the breakthrough result of Croot, Lev, and Pach [@CrootLevPachZ4]. Our bound for the Erdős-Ginzburg-Ziv constant relies on the more general notion of the partition rank, as defined in [@Naslund2017Partition], which allows us to handle the indicator tensor that appears when we force the variables to be distinct. The slice rank method has seen numerous applications, such as Ellenberg and Gijswijt’s upper bound for progression-free sets in $\mathbb{F}_{p}^{n}$ [@TaosBlogCapsets; @EllenbergGijswijtCapsets], new bounds for the Erdős-Szemerédi sunflower problem [@NaslundSawinSunflower], disproving certain conjectures concerning fast matrix multiplication [@BlasiakChurchCohnGrochowNaslundSawinUmans2016MatrixMultiplication], and right angles in $\mathbb{F}_{q}^{n}$ [@GeShangguan2017NoRightAngles]. We refer the reader to [@BlasiakChurchCohnGrochowNaslundSawinUmans2016MatrixMultiplication] for a more detailed discussion of the properties of the slice rank, and it’s relationship to geometric invariant theory.
We say that a group $G$ satisfies *property $D$* (see [@GaoGeroldinger2006ZeroSumSurvey Sec. 7]) if every sequence $S$ of length $\mathfrak{s}(G)-1$ that does not contain $\exp(G)$ elements summing to $0$, when viewed as a multi-set, takes the form $$S=\cup_{i=1}^{\exp(G)-1}T$$ for some set $T$. That is, if $S$ is a maximal sequence that does not contain $\exp(G)$ elements summing to zero, then every element in $S$ appears exactly $\exp(G)-1$ times. Gao and Geroldinger conjecture [@GaoGeroldinger2006ZeroSumSurvey Conj. 7.2] that $\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}$ satisfies property $D$ for every $k$ and $n$, and under this conjecture we can improve upon theorem \[thm:Main\_Theorem\], corollary \[cor: cp\_asymp\], and Alon and Dubiner’s bound for the Erdős-Ginzburg-Ziv constant. This also improves upon a recent result of Hegedűs [@Hegedus2017EGZ] that used the polynomial method to show that if $\mathbb{F}_{p}^{n}$ satisfies property $D$, then $$\mathfrak{s}\left(\mathbb{F}_{p}^{n}\right)\leq(p-1)p^{\left(1-\frac{(p-2)^{2}}{2p^{2}\log p}\right)n+1}+1.$$
\[thm:property D theorem\] Let $q$ be a prime power, and assume that $\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ satisfies property $D$. Then $$\mathfrak{s}\left(\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}\right)\leq(q-1)4^{n}+1.$$
Using the results of section \[sec:Reduction-to-Prime\], theorem \[thm:property D theorem\] implies the following result for general $k$:
\[thm:prop\_d\_for\_general\_k\] Suppose that $k=p_{1}^{r_{1}}\cdots p_{m}^{r_{m}}$, and $$\left(\mathbb{Z}/p_{i}^{r_{i}}\mathbb{Z}\right)^{n}$$ satisfies property $D$ for each $1\leq i\leq m$. Then $$\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)\leq(k-1)4^{n}+k.$$
The proof of theorem \[thm:q\_el\_not\_distinct\] uses the slice rank method, and does not require the added flexibility of the partition rank. To handle prime powers, we use an idea appearing in [@BlasiakChurchCohnGrochowNaslundSawinUmans2016MatrixMultiplication], where they extend Ellenberg and Gijswijt’s bounds for progression-free sets in $\mathbb{F}_{p}^{n}$ to progression-free sets in $(\mathbb{Z}/k\mathbb{Z})^{n}$ using a binomial coefficient indicator function.
\[sec:Reduction-to-Prime\]Reduction to Prime Powers
===================================================
In this section, we state several lemmas that provide upper bounds for $\mathfrak{s}\left(\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}\right)$ based on $\mathfrak{s}\left(\left(\mathbb{Z}/p\mathbb{Z}\right)^{n}\right)$ for the primes $p$ dividing $k$. These lemmas allow us to deduce corollary \[cor: cp\_asymp\] and theorem \[thm:prop\_d\_for\_general\_k\] from theorems \[thm:Main\_Theorem\] and \[thm:property D theorem\], respectively.
\[lem:EGZ\_Lemma\] ([@ChiDingGaoGeroldingerSchmid2005ZeroSumSubsequences proposition 3.1], see also [@FoxSauermann2017EGZ lemma 3.1]) Let $G$ be a finite abelian group. Let $H\subset G$ be a subgroup such that $\exp(G)=\exp(H)\exp\left(G/H\right)$. Then $$\mathfrak{s}(G)\leq\exp\left(G/H\right)\left(\mathfrak{s}(H)-1\right)+\mathfrak{s}\left(G/H\right).$$
By repeated application of this lemma, Fox and Sauermann gave the following two lemmas:
\[lem:g\_reduced\]([@FoxSauermann2017EGZ Lemma 3.4]) Let $G$ be a non-trivial finite abelian group, and suppose that $G\cong G_{1}\times\cdots\times G_{m}$ where the exponents $\exp\left(G_{i}\right)$ are pairwise relatively prime. Then $$\mathfrak{s}\left(G\right)\leq\sum_{i=1}^{m}\exp\left(G_{1}\right)\dotsm\exp\left(G_{i-1}\right)\mathfrak{s}\left(G_{i}\right).$$
\[lem:k\_reduced\_to\_p\]([@FoxSauermann2017EGZ Lemma 3.5]) Let $k=p_{1}^{\alpha_{1}}\cdots p_{r}^{\alpha_{r}}$. Then we have that $$\mathfrak{s}\left((\mathbb{Z}/k\mathbb{Z})^{n}\right)<k\cdot\left(\frac{\mathfrak{s}\left(\mathbb{F}_{p_{1}}^{n}\right)}{p_{1}-1}+\cdots+\frac{\mathfrak{s}\left(\mathbb{F}_{p_{r}}^{n}\right)}{p_{r}-1}\right).$$
Lemma \[lem:k\_reduced\_to\_p\] combined with theorem \[thm:Main\_Theorem\] implies that $$\begin{aligned}
\mathfrak{s}\left((\mathbb{Z}/k\mathbb{Z})^{n}\right) & <k\cdot\left(2^{p_{1}}(J(p_{1})p_{1})^{n}+\cdots+2^{p_{r}}(J(p_{r})p_{r})^{n}\right)\\
& \leq kr2^{p}(J(p)p)^{n}\\
& =(J(p)p+o(1))^{n}\end{aligned}$$ where $p=\max_{i}p_{i}$, and so corollary \[cor: cp\_asymp\] follows. To deduce theorem \[thm:prop\_d\_for\_general\_k\] from theorem \[thm:property D theorem\], write $$\left(\mathbb{Z}/k\mathbb{Z}\right)^{n}=\left(\mathbb{Z}/q_{1}\mathbb{Z}\right)^{n}\times\cdots\times\left(\mathbb{Z}/q_{m}\mathbb{Z}\right)^{n}$$ where $q_{1}\leq\cdots\leq q_{m}$ are pairwise relatively prime prime powers dividing $k$. By lemma \[lem:g\_reduced\] and theorem \[thm:property D theorem\], we have that $$\mathfrak{s}\left((\mathbb{Z}/k\mathbb{Z})^{n}\right)\leq k\cdot\frac{(q_{m}-1)4^{n}+1}{q_{m}}+k\cdot\frac{(q_{m-1}-1)4^{n}+1}{q_{m}q_{m-1}}+\cdots+k\cdot\frac{(q_{1}-1)4^{n}+1}{q_{m}\cdots q_{1}}.$$ The series telescopes, and we obtain $$\mathfrak{s}\left((\mathbb{Z}/k\mathbb{Z})^{n}\right)\leq(k-1)4^{n}+k\left(\frac{1}{q_{m}}+\frac{1}{q_{m}q_{m-1}}+\cdots+\frac{1}{q_{m}\cdots q_{1}}\right).$$ Since $q_{i}\geq2$, the sum $\frac{1}{q_{m}}+\frac{1}{q_{m}q_{m-1}}+\cdots+\frac{1}{q_{m}\cdots q_{1}}$ will be at most $1$, and theorem \[thm:prop\_d\_for\_general\_k\] follows.
The Partition Rank
==================
To motivate the definition of the slice rank and the partition rank we begin by recalling the definition of the tensor rank. For finite sets $X_{1},\dots,X_{n}$, a non-zero function $$h:X_{1}\times\cdots\times X_{k}\rightarrow\mathbb{F},$$ is called a *rank 1 function* there exists $f_{1},\dots,f_{k}$ such that $$h(x_{1},\dots,x_{k})=f_{1}(x_{1})f_{2}(x_{2})\cdots f_{k}(x_{k}).$$ The tensor rank of $$F:X_{1}\times\cdots\times X_{k}\rightarrow\mathbb{F}$$ is defined to be the minimal $r$ such that $$F=\sum_{i=1}^{r}g_{i}$$ where the $g_{i}$ are rank 1 functions. This function $F$ can be thought of as a $|X_{1}|\times\cdots\times|X_{k}|$ array of elements of $\mathbb{F}$, and the rank $1$ functions correspond to the outer product of $k$ vectors. To define the partition rank, we first need some notation. Given variables $x_{1},\dots,x_{k}$, and a set $S\subset\{1,\dots,k\}$, $S=\left\{ s_{1},\dots,s_{m}\right\} $, let $\vec{x}_{S}$ denote the $m$-tuple $$x_{s_{1}},\dots,x_{s_{m}},$$ so that for a function $g$ of $m$ variables, we have $$g(\vec{x}_{S})=g(x_{s_{1}},\dots,x_{s_{m}}).$$ For example, if $k=5$, and $S=\{1,2,4\}$, then $g\left(\vec{x}_{S}\right)=g(x_{1},x_{2},x_{4})$, and $f\left(\vec{x}_{\{1,\dots,k\}\backslash S}\right)=f(x_{3},x_{5})$. A partition of $\{1,2,\dots,k\}$ is a collection $P$ of non-empty pairwise disjoint subsets of $\{1,\dots,k\}$ such that $$\bigcup_{A\in P}A=\{1,\dots,k\}.$$ We say that $P$ is the *trivial* partition if it consists only of a single set, $\{1,\dots,k\}$.
Let $X_{1},\dots,X_{k}$ be finite sets, and let $$h:X_{1}\times\cdots\times X_{k}\rightarrow\mathbb{F}.$$ We say that $h$ has *partition rank* $1$ if there exists some non-trivial partition $P$ of the variables $\{1,\dots,k\}$ such that $$h(x_{1},\dots,x_{k})=\prod_{A\in P}f_{A}\left(\vec{x}_{A}\right)$$ for some functions $f_{A}$. We say that $h$ has *slice rank* $1$, if in addition one of the sets $A\in P$ is a singleton.
The tensor $h:X_{1}\times\cdots\times X_{k}\rightarrow\mathbb{F}$ will have partition rank $1$ if and only if it can be written in the form $$h(x_{1},\dots,x_{k})=f(\vec{x}_{S})g(\vec{x}_{T})$$ for some $f,g$ and some $S,T\neq\emptyset$ with $S\cup T=\{1,\dots,k\}$. Additionally, $h$ will have slice rank $1$ if it can be written in the above form with either $|S|=1$ or $|T|=1$. In other words, a function $h$ has partition rank $1$ if the tensor can be written as a non-trivial outer product, and it has slice rank $1$ if it can be written as the outer product between a vector and a $k-1$ dimensional tensor.
Let $X_{1},\dots,X_{k}$ be finite sets. The *partition rank* of $$F:X_{1}\times\cdots\times X_{k}\rightarrow\mathbb{F},$$ is defined to be the minimal $r$ such that $$F=\sum_{i=1}^{r}g_{i}$$ where the $g_{i}$ have partition rank $1$. The *slice rank* of $F$ is the minimal $r$ such that $$F=\sum_{i=1}^{r}g_{i}$$ where the $g_{i}$ have slice rank $1$.
The partition rank is the minimal rank among all possible ranks obtained from partitioning the variables. The slice rank can be viewed as the rank which results from the partitions of $\{1,\dots,k\}$ into a set of size $1$ and a set of size $k-1$, and so we have that $$\operatorname{partition-rank}\leq\operatorname{slice-rank}.$$ For two variables, the slice rank, partition rank, and tensor rank are equivalent since there is only one non-trivial partition of a set of size $2$. For three variables, the partition rank and the slice rank are equivalent, and for $4$ or more variables, all three ranks are different. A key property of the partition rank is the following lemma, given in [@Naslund2017Partition Lemma 11], which generalizes [@TaosBlogCapsets Lemma 1].
\[lem:Critical\_Lemma\]Let $X$ be a finite set, and let $X^{k}$ denote the $k$-fold Cartesian product of $X$ with itself. Suppose that $$F:X^{k}\rightarrow\mathbb{F}$$ is the diagonal identity tensor, that is $$F(x_{1},\dots,x_{k})=\delta(x_{1},\dots,x_{k})=\begin{cases}
1 & x_{1}=\cdots=x_{k}\\
0 & \text{otherwise}
\end{cases}.$$ Then $$\operatorname{partition-rank}(F)=|X|.$$
See [@Naslund2017Partition Lemma 11].
Unconditional Bounds for $\mathfrak{s}\left(\mathbb{F}_{p}^{n}\right)$
======================================================================
We begin with a lemma that bounds from above the number of monomials of degree at most $d$ over $\mathbb{F}_{q}^{n}$. This is a standard application of Markov’s inequality.
\[lem:number\_of\_monomials\]The number of monomials of degree at most $d$ over $\mathbb{F}_{q}^{n}$ equals $$\#\left\{ v\in\{0,\dots,q-1\}^{n}:\ \sum_{i=1}^{n}v_{i}\leq d\right\} ,\label{eq:number_monomials}$$ and this is at most $$\min_{0<x<1}\left(\frac{1-x^{q}}{1-x}x^{-\frac{d}{n}}\right)^{n}.$$
Let $X_{1},\dots,X_{n}$ denote independent uniform random variables on $\{0,\dots q-1\}$, and let $X=\sum_{i=1}^{n}X_{i}$. Then (\[eq:number\_monomials\]) is at most $$q^{n}\mathbb{P}\left(X\leq d\right),$$ and by Markov’s inequality, for any $0<x<1$ $$\begin{aligned}
q^{n}\mathbb{P}\left(X\leq d\right) & =q^{n}\mathbb{P}\left(x^{d}\leq x^{X}\right)\\
& \leq q^{n}\mathbb{E}\left(x^{X}\right)x^{-d}\\
& =q^{n}\left(\mathbb{E}\left(x^{X_{1}}\right)\right)^{n}x^{-d}\\
& =\left(\frac{1-x^{q}}{1-x}x^{-\frac{d}{n}}\right)^{n}.\end{aligned}$$
Let $p>2$ be a prime. For $x_{1},\dots,x_{p}\in\mathbb{F}_{p}^{n}$, define $$F_{p}:\left(\mathbb{F}_{p}^{n}\right)^{p}\rightarrow\mathbb{F}_{p}$$ by $$F_{p}(x_{1},\dots,x_{p})=\prod_{i=1}^{n}\left(1-\left(x_{1i}+x_{2i}+\cdots+x_{pi}\right)^{p-1}\right),\label{eq:F_p definition}$$ where $x_{ji}$ is understood to be the $i^{th}$ coordinate of the vector $x_{j}\in\mathbb{F}_{p}^{n}$, and the notation $\left(\mathbb{F}_{p}^{n}\right)^{p}$ is used to denote the $p$-fold Cartesian product of $\mathbb{F}_{p}^{n}$ with itself. Then $$F_{p}\left(x_{1},\dots,x_{p}\right)=\begin{cases}
1 & \text{if }x_{1}+x_{2}+\cdots+x_{p}=0\\
0 & \text{otherwise}
\end{cases}.$$ This tensor does not take into account whether or not the variables are distinct, and so in particular for any $x\in\mathbb{F}_{p}^{n}$ $$F_{p}(x,\dots,x)=1.$$ In order to modify this tensor so that it picks up only distinct $k$-tuples of elements summing to zero, we use the technique in [@Naslund2017Partition], and introduce an indicator which is a sum over the permutations in the symmetric group $S_{p}$. For every $\sigma\in S_{p}$, define $$f_{\sigma}:X^{p}\rightarrow\mathbb{F}\label{eq:f_sigma def}$$ to be the function that is $1$ if $(x_{1},\dots,x_{p})$ is a fixed point of $\sigma$, and $0$ otherwise. The following is [@Naslund2017Partition Lemma 14]:
\[lem:Group-action-lemma\]We have the identity $$\sum_{\sigma\in S_{k}}\operatorname{sgn}(\sigma)f_{\sigma}(x_{1},\dots,x_{k})=\begin{cases}
1 & \text{if }x_{1},\dots,x_{k}\text{ are distinct}\\
0 & \text{otherwise}
\end{cases},$$ where $\operatorname{sgn}(\sigma)$ is the sign of the permutation.
By definition, $$\sum_{\sigma\in S_{k}}\operatorname{sgn}(\sigma)f_{\sigma}(x_{1},\dots,x_{k})=\sum_{\sigma\in\text{Stab}(\vec{x})}\operatorname{sgn}(\sigma)$$ where $\text{Stab}(\vec{x})\subset S_{k}$ is the stabilizer of $\vec{x}$. Since the stabilizer is a product of symmetric groups, this will be non-zero precisely when $\text{Stab}(\vec{x})$ is trivial, and hence $x_{1},\dots,x_{k}$ must be distinct. This vector is then fixed only by the identity element, and so the sum equals $1$.
Using this lemma, we give the following modification of [@Naslund2017Partition Lemma 15]:
\[lem:critical-indicator-function\]Let $\mathcal{C}_{i}\subset S_{k}$ denote the set all elements in $S_{k}$ which are the product of exactly $i$ disjoint cycles. Define $$R_{k}(x_{1},\dots,x_{k})=\sum_{\begin{array}{c}
\sigma\in S_{k}\\
\sigma\notin C_{1},C_{2}
\end{array}}\operatorname{sgn}(\sigma)f_{\sigma}(x_{1},\dots,x_{k}).$$ Then for $k\geq3$, $$R_{k}(x_{1},\dots,x_{k})=\begin{cases}
1 & \text{if }x_{1},\dots,x_{k}\text{ are distinct}\\
(-1)^{k-1}(k-1)!\sum_{j=2}^{k-1}\frac{1}{j} & \text{if }x_{1}=\cdots=x_{k}\\
\alpha(x_{1},\dots,x_{k}) & \text{if }x_{1},\dots,x_{k}\text{ take on }2\text{ distinct values}\\
0 & \text{otherwise}
\end{cases},$$ where $$\alpha(x_{1},\dots,x_{k})=(-1)^{k-1}\#\left\{ \sigma\in\mathcal{C}_{2}\text{ that fix }(x_{1},\dots,x_{k})\right\} .$$
If there are $3$ or more distinct elements among $x_{1},\dots,x_{k}$ then $f_{\sigma}(x_{1},\dots,x_{k})=0$ for any $\sigma\in\mathcal{C}_{2},\mathcal{C}_{1}$, and so the identity holds by lemma \[lem:Group-action-lemma\]. When there are exactly two distinct elements among $x_{1},\dots,x_{k}$, $f_{\sigma}(x_{1},\dots,x_{k})=0$ for any $\sigma\in\mathcal{C}_{1}$, and so $$R_{k}(x_{1},\dots,x_{k})=(-1)^{k-1}\#\left\{ \sigma\in\mathcal{C}_{2}\text{ that fix }(x_{1},\dots,x_{k})\right\}$$ since $\operatorname{sgn}(\sigma)=(-1)^{k}$ for any $\sigma\in\mathcal{C}_{2}$. Lastly, when $x_{1}=\cdots=x_{k}$, we have that $$R_{k}(x_{1},\dots,x_{k})+(-1)^{k}|\mathcal{C}_{2}|+(-1)^{k-1}|\mathcal{C}_{1}|=0$$ by lemma \[lem:Group-action-lemma\] since $\operatorname{sgn}\sigma=(-1)^{k-1}$ for $\sigma\in\mathcal{C}_{1}$. Since $$|\mathcal{C}_{1}|=(k-1)!\ \ \ \text{and}\ \ \ |\mathcal{C}_{2}|=(k-1)!\sum_{j=1}^{k-1}\frac{1}{j},$$ it follows that $$R_{k}(x_{1},\dots,x_{k})=(-1)^{k-1}\sum_{j=2}^{k-1}\frac{1}{j}$$ when $x_{1}=\cdots=x_{k}$.
\[rem:split\_R\_k\]Let $$\delta(x_{1},\dots,x_{k})=\begin{cases}
1 & x_{1}=\cdots=x_{k}\\
0 & \text{otherwise}
\end{cases},\label{eq:delta_def}$$ and for a partition $P$ of $\{1,\dots,k\}$ define $$\delta_{P}(x_{1},\dots,x_{k})=\prod_{A\in P}\delta(\vec{x}_{A})$$ where if $A=\{j\}$ is a singleton, $\delta(\vec{x}_{A})$ is defined to be the constant function $1$. Let $\sigma\in S_{k}$, and consider it’s disjoint cycle decomposition. Let $S_{1},\dots,S_{m}$ denote the sets of indices corresponding to the disjoint cycles. Then $P=\{S_{1},\dots,S_{m}\}$ is a partition of $\{1,\dots,k\}$ and $$f_{\sigma}(x_{1},\dots,x_{k})=\delta_{P}(x_{1},\dots,x_{k})=\prod_{i=1}^{m}\delta(\vec{x}_{S_{i}}).$$ It follows that we can write $R_{k}(x_{1},\dots,x_{k})$ exactly as a linear combination of products of disjoint delta functions, including the constant function. When $k=3$ $$R_{3}(x_{1},x_{2},x_{3})=1,$$ when $k=4$ $$\begin{aligned}
R_{4}(x_{1},x_{2},x_{3},x_{4})= & 1-\delta(x_{1},x_{2})-\delta_{2}(x_{2},x_{3})-\delta(x_{3},x_{4})\\
& -\delta(x_{4},x_{1})-\delta(x_{1},x_{3})-\delta(x_{2},x_{4}),\end{aligned}$$ and when $k=5$ $$\begin{aligned}
R_{5}(x_{1},x_{2},x_{3},x_{4},x_{5})= & 1-\sum_{i<j\le5}\delta(x_{i},x_{j})+2\sum_{i<j<l\leq5}\delta(x_{i},x_{j},x_{l})\\
& \sum_{i<j\leq5}\delta(x_{i},x_{j})\sum_{\begin{array}{c}
l<m\leq5\\
l,m\neq i,j
\end{array}}\delta(x_{l},x_{m}).\end{aligned}$$ Starting with $k=5$, there will be terms that are the product of multiple delta functions in multiple variables, and to handle these we need to use the partition rank. For example, as a function on $X^{4}$, $$\delta(x_{1},x_{2})\delta(x_{3},x_{4})$$ has slice rank equal to $|X|$ but partition rank equal to $1$ [@TaoSawinBlog].
We now use the function $R_{p}(x_{1},\dots,x_{p})$ to modify $F_{p}$, and arrive at our desired indicator tensor.
\[lem:I\_p def\]For $x_{1},\dots,x_{p}\in\mathbb{F}_{p}^{n}$ define $$I_{p}(x_{1},\dots,x_{p})=R_{p}(x_{1},\dots,x_{p})F_{p}(x_{1},\dots,x_{p})$$ where $F_{p}$ is defined in (\[eq:F\_p definition\]) and $R_{p}$ is defined in lemma \[lem:critical-indicator-function\]. Then $$I_{p}(x_{1},\dots,x_{p})=\begin{cases}
1 & \text{if }x_{1},\dots,x_{p}\text{ are distinct and sum to zero}\\
1 & \text{if }x_{1}=\cdots=x_{p}\\
0 & \text{otherwise}
\end{cases}.$$
By pairing every element in $\mathbb{F}_{p}^{\times}$ with it’s inverse we have that $$\sum_{j=2}^{p-1}\frac{1}{j}=-1+\sum_{j=1}^{p-1}\frac{1}{j}=-1,$$ and so by Wilson’s theorem $$(-1)^{p-1}(p-1)!\sum_{j=2}^{p-1}\frac{1}{j}=(-1)^{p-1}=1$$ and hence if $x_{1}=\cdots=x_{p}$ then $I_{p}(x_{1},\dots,x_{p})=1$. If there are exactly two distinct elements among $x_{1},\dots,x_{p}$, say $x,y\in\mathbb{F}_{p}^{n}$, then $$x_{1}+\cdots+x_{p}=cx+(p-c)y=c(x-y)$$ where $1\leq c\leq p-1$. Since $x\neq y$, this cannot equal $0$, and so $F_{p}(x_{1},\dots,x_{p})=0$ in this case. When $x_{1},\dots,x_{p}$ are distinct and sum to zero, then $F_{p}(x_{1},\dots,x_{p})=1$ and $R_{p}(x_{1},\dots,x_{p})=1$, and so the result follows.
We are now ready to prove theorem \[thm:Main\_Theorem\].
Suppose that $A\subset\mathbb{F}_{p}^{n}$ does not contain distinct $x_{1},\dots,x_{p}$ such that $$x_{1}+\cdots+x_{p}=0.$$ Let $I_{p}$ be defined as in lemma \[lem:I\_p def\]. When we restrict $I_{p}$ to $A^{p}$, it will be the diagonal tensor with $1$’s on the diagonal, and so by lemma \[lem:Critical\_Lemma\] $$|A|\leq\operatorname{partition-rank}(I_{p}).$$ To bound the partition rank of $I_{p}(x_{1},\dots,x_{p})$, we split $R_{k}$ into a sum of $f_{\sigma}$ terms, and examine $$f_{\sigma}(x_{1},\dots,x_{p})F_{p}(x_{1},\dots,x_{p}).\label{eq:Breaking_up_I_p_equ}$$ By remark \[rem:split\_R\_k\], we may write $$f_{\sigma}(x_{1},\dots,x_{p})=\prod_{i=1}^{m}\delta\left(\vec{x}_{S_{i}}\right)$$ where the $S_{i}$ are disjoint non-empty sets, and where we have the convention $\delta(\vec{x}_{S})=1$ when $S$ is a singleton set. For each $i$, let $s_{i}$ be some element in $S_{i}$. Then we have the equality of functions $$f_{\sigma}(x_{1},\dots,x_{p})F_{p}(x_{1},\dots,x_{p})=\prod_{i=1}^{m}\delta\left(\vec{x}_{S_{i}}\right)\prod_{j=1}^{n}\left(1-\left(\sum_{i=1}^{m}r_{i}x_{s_{i}j}\right)^{p-1}\right)$$ where $\sum_{i=1}^{,m}r_{i}=p$. Note that by the disjointness of the sets $S_{1},\dots,S_{m}$, the variable $x_{s_{i}}$ will not appear in any delta function $\delta\left(\vec{x}_{S_{j}}\right)$ for $j\neq i$. We may expand the polynomial above as a linear combination of monomials of the form $$\left[\delta\left(\vec{x}_{S_{1}}\right)x_{s_{1}1}^{e_{11}}\cdots x_{s_{1}n}^{e_{1n}}\right]\left[\delta\left(\vec{x}_{S_{2}}\right)x_{s_{2}1}^{e_{21}}\cdots x_{s_{2}n}^{e_{2n}}\right]\cdots\left[\delta\left(\vec{x}_{S_{m}}\right)x_{s_{1}1}^{e_{m1}}\cdots x_{s_{m}n}^{e_{mn}}\right],$$ where $e_{ij}\leq p-1$ and $\sum_{i=1}^{m}\sum_{j=1}^{n}e_{ij}\leq(p-1)n$. For each term, there will be some coordinate $i$ such that the monomial $$x_{s_{i}1}^{e_{i1}}\cdots x_{s_{i}n}^{e_{in}}$$ has degree at most $\frac{(p-1)n}{m}$. Futhermore, we must have $m\geq3$ since the permutations $\sigma$ with $1$ or $2$ cycles in their disjoint cycle decomposition do not appear in $R_{k}$. Now, consider the simultaneous expansion of all $f_{\sigma}(x_{1},\dots,x_{p})F_{p}(x_{1},\dots,x_{p})$ for each permutation $\sigma$, that is the complete expansion of $I_{p}(x_{1},\dots,x_{p})$. By the above analysis, for each term $v(x_{1},\dots,x_{p})$ appearing in the expansion, there exists a set $S\subset\{1,2,\dots,p\}$, of size $|S|\leq p-2$, such that for some $s\in S$, $$v(x_{1},\dots,x_{p})=\left[\delta\left(\vec{x}_{S}\right)x_{s1}^{e_{s1}}\cdots x_{sn}^{e_{sn}}\right]h\left(\vec{x}_{\{1,\dots,p\}\backslash S}\right)$$ and $$\sum_{i=1}^{n}e_{si}\leq\frac{(p-1)n}{3},$$ where $h$ is some function. By grouping terms by delta function with lowest degree monomial, and counting the number of monomials of degree at most $\frac{n(p-1)}{3}$, we find that the partition rank of $I_{p}$ is at most $$\sum_{\begin{array}{c}
S\subset\{1,\dots,p\}\\
1\leq|S|\leq p-2
\end{array}}\#\left\{ v\in\{0,1,\dots,p-1\}^{n}:\ \sum_{i=1}^{n}v_{i}\le\frac{n(p-1)}{3}\right\} ,\label{eq:true_bound}$$ and so by lemma \[lem:number\_of\_monomials\] we have the upper bound $$\begin{aligned}
\operatorname{partition-rank}(I_{p}) & \leq\left(2^{p}-p-2\right)\left(\inf_{0<x<1}\frac{1-x^{p}}{1-x}x^{-\frac{p-1}{3}}\right)^{n}\\
& =(2^{p}-p-2)(pJ(p))^{n}.\end{aligned}$$ The bound for $\mathfrak{s}(\mathbb{F}_{p}^{n})$ follows from the fact that any sequence can have at most $p-1$ copies of the same element without trivially introducing a solution to $$x_{1}+\cdots+x_{p}=0.$$
Conditional Bounds for $\mathfrak{s}\left(\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}\right)$
=========================================================================================
To prove theorem \[thm:property D theorem\], we first prove a theorem for subsets of $\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$.
\[thm:q\_el\_not\_distinct\]Suppose that $A\subset\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ satisfies $$|A|>\gamma_{q}^{n}$$ where $$\gamma_{q}=\inf_{0<x<1}\frac{1-x^{q}}{1-x}x^{-\frac{q-1}{q}}.$$ Then $A$ contains $q$ not necessarily distinct, but not all equal, elements $x_{1},\dots,x_{q}$ which sum to $0$.
Note that for any $q$, $$\gamma_{q}<\inf_{0<x<1}\frac{1}{1-x}x^{-1}=4,\label{eq:gamma_q_bound}$$ and so the statement above holds with $\gamma_{q}^{n}$ replaced with $4^{n}$. Theorem \[thm:property D theorem\] follows as a corollary of theorem \[thm:q\_el\_not\_distinct\] and (\[eq:gamma\_q\_bound\]).
(Theorem \[thm:q\_el\_not\_distinct\] implies theorem \[thm:property D theorem\]) Let $S$ be a sequence in $\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ of length $s\left(\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}\right)-1$ which does not contain $q$ elements that sum to $0$. By property $D$, it follows that as a multi-set $$S=\cup_{i=1}^{q-1}T$$ for some $T\subset\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$. This implies that the only solution to $$x_{1}+\cdots+x_{q}=0$$ for $x_{i}\in T$, with no distinctness requirement, occurs when $x_{1}=x_{2}=\cdots=x_{q}$. From theorem \[thm:q\_el\_not\_distinct\] we obtain $|T|\leq\gamma_{q}^{n},$ and hence $\mathfrak{s}\left(\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}\right)\leq(q-1)4^{n}+1$.
To prove theorem \[thm:q\_el\_not\_distinct\] we use the slice rank method. The following lemma appears in proposition 4.14 and theorem 4.15 in [@BlasiakChurchCohnGrochowNaslundSawinUmans2016MatrixMultiplication]:
Let $q$ be a prime power. For any $x_{1},\dots,x_{k}\in\mathbb{Z}/q\mathbb{Z}$ we have that $$\sum_{m_{1}+\cdots+m_{k}\leq q-1}(-1)^{m_{1}+\cdots+m_{k}}\binom{x_{1}}{m_{1}}\cdots\binom{x_{k}}{m_{k}}=\begin{cases}
1 & \text{if }x_{1}+\cdots+x_{k}=0\\
0 & \text{otherwise}
\end{cases}.$$
This follows since $$\sum_{m\leq q-1}(-1)^{m}\binom{x_{1}+\cdots+x_{k}}{m}=\begin{cases}
1 & \text{if }x_{1}+\cdots+x_{k}=0\\
0 & \text{otherwise}
\end{cases},$$ and $$\binom{x_{1}+\cdots+x_{k}}{m}=\sum_{m_{1}+\cdots+m_{k}=m}\binom{x_{1}}{m_{1}}\cdots\binom{x_{k}}{m_{k}}.$$
Note that the binomial coefficient modulo $q$ is well defined due to Lucas’ theorem. For $x_{1},\dots,x_{k}\in\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ let $$E_{q}:\left(\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}\right)^{q}\rightarrow\mathbb{F}_{q}$$ be defined by $$E_{q}(x_{1},\dots,x_{q})=\prod_{i=1}^{n}\left(\sum_{m_{1}+\cdots+m_{q}\leq q-1}(-1)^{m_{1}+\cdots+m_{q}}\binom{x_{1i}}{m_{1}}\cdots\binom{x_{qi}}{m_{q}}\right)\label{eq:E_q_def}$$ where $x_{ji}$ denotes the $i^{th}$ coordinate of the vector $x_{j}$. For any $q$ we have that $$E_{q}\left(x_{1},\dots,x_{q}\right)=\begin{cases}
1 & \text{if }x_{1}+x_{2}+\cdots+x_{q}=0\\
0 & \text{otherwise}
\end{cases},$$ and $E_{q}(x,\dots,x)=1$ for any $x\in\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$.
\[prop: slice\_rank\_E\_q\]The slice-rank of $E_{q}$ on $\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ is at most $q\cdot\gamma_{q}^{n}$.
Expanding the product form for $E_{q}\left(x_{1},\dots,x_{q}\right)$ appearing in (\[eq:E\_q\_def\]) as a polynomial, we may write $E_{q}\left(x_{1},\dots,x_{q}\right)$ as a linear combination of terms of the form $$\left(\binom{x_{11}}{m_{11}}\cdots\binom{x_{1n}}{m_{1n}}\right)\cdots\left(\binom{x_{q1}}{m_{q1}}\cdots\binom{x_{qn}}{m_{qn}}\right)$$ where $m_{ij}\leq q-1$ for every $i,j$ and $$\sum_{i=1}^{q}\sum_{j=1}^{n}m_{ij}\leq(q-1)n.$$ Thus, for each term there is a coordinate $i$ such that $$\sum_{j=1}^{n}m_{ij}\leq\frac{(q-1)n}{q},$$ and by always slicing away the lowest degree piece, it follows that $$\operatorname{slice-rank}(E_{q})\leq q\cdot\#\left\{ v\in\{0,1,\dots,q-1\}^{n}:\ \sum_{i=1}^{n}v_{i}\le\frac{n(q-1)}{q}\right\} .\label{eq:intermediate_eq_1}$$ By lemma \[lem:number\_of\_monomials\] this will be at most $$q\cdot\left(\frac{1-x^{q}}{1-x}x^{-\frac{q-1}{q}}\right)^{n}$$ and so we conclude that $$\operatorname{slice-rank}(E_{q})\leq q\cdot\gamma_{q}^{n}.$$
We now prove theorem \[thm:q\_el\_not\_distinct\].
Suppose that $A\subset\left(\mathbb{Z}/q\mathbb{Z}\right)^{n}$ does not contain a non-trivial $q$-tuple that sums to zero. Then when restricted to $A^{q}$, $E_{q}$ will be a diagonal tensor taking the value $1$ on the diagonal. Hence lemma \[lem:Critical\_Lemma\] implies that $$|A|\leq\operatorname{slice-rank}(E_{q}),$$ and so $$|A|\leq q\cdot\gamma_{q}^{n}.$$ The factor of $q$ can be removed by an amplification argument. The $m$-fold Cartesian product $A^{m}=A\times\cdots\times A$ viewed as a subset of $\left(\mathbb{Z}/q\mathbb{Z}\right)^{nm}$ will not contain a non-trivial $k$-tuple summing to $0$, and so by the same slice rank argument $$|A|^{m}\leq q\cdot\gamma_{q}^{nm},$$ and hence $$|A|\leq q^{\frac{1}{m}}\gamma_{q}^{n}.$$ Letting $m\rightarrow\infty$, it follows that $|A|\leq\gamma_{q}^{n}$.
I would like to thank Will Sawin for his helpful comments and for his simple proofs of lemmas \[lem:Group-action-lemma\] and \[lem:critical-indicator-function\]. I would also like to thank Lisa Sauermann for her many helpful comments, corrections, and suggestions, and for pointing out lemmas \[lem:EGZ\_Lemma\], \[lem:g\_reduced\], and \[lem:k\_reduced\_to\_p\]. This work was partially supported by the NSERC PGS-D scholarship, and by Ben Green’s ERC Starting Grant 279438, Approximate Algebraic Structure and Applications.
[10]{}
N. Alon and M. Dubiner, *Zero-sum sets of prescribed size*, Combinatorics, [P]{}aul [E]{}rdős is eighty, [V]{}ol. 1, Bolyai Soc. Math. Stud., János Bolyai Math. Soc., Budapest, 1993, pp. 33–50. [MR ]{}[1249703]{}
Jonah Blasiak, Thomas Church, Henry Cohn, Joshua A. Grochow, Eric Naslund, William F. Sawin, and Chris Umans, *On cap sets and the group-theoretic approach to matrix multiplication*, Discrete Anal. (2017), Paper No. 3, 27. [MR ]{}[3631613]{}
R. Chi, S. Ding, W. Gao, A. Geroldinger, and W. A. Schmid, *On zero-sum subsequences of restricted size. [IV]{}*, Acta Math. Hungar. **107** (2005), no. 4, 337–344. [MR ]{}[2150796]{}
Ernie Croot, Vsevolod F. Lev, and Péter Pál Pach, *Progression-free sets in [$\Bbb Z^n_4$]{} are exponentially small*, Ann. of Math. (2) **185** (2017), no. 1, 331–337. [MR ]{}[3583357]{}
Jordan S. Ellenberg and Dion Gijswijt, *On large subsets of [$\Bbb F^n_q$]{} with no three-term arithmetic progression*, Ann. of Math. (2) **185** (2017), no. 1, 339–343. [MR ]{}[3583358]{}
Christian Elsholtz, *Lower bounds for multidimensional zero sums*, Combinatorica **24** (2004), no. 3, 351–358. [MR ]{}[2085361]{}
Paul Erdos, Abraham Ginzburg, and Abraham Ziv, *A theorem in additive number theory*, Bull. Res. Council Israel **10F** (1961), 41–43.
Jacob Fox and Lisa Sauermann, *[Erdős-Ginzburg-Ziv constants by avoiding three-term arithmetic progressions]{}*, ArXiv e-prints (2017), [arXiv:[1708.09100]{}]{}
Weidong Gao and Alfred Geroldinger, *Zero-sum problems in finite abelian groups: a survey*, Expo. Math. **24** (2006), no. 4, 337–369. [MR ]{}[2313123]{}
Gennian Ge and Chong Shuangguan, *Rank counting and maximum subsets of $\mathbb{F}_q^n$ containing no right angles*, ArXiv e-prints (2016), [arXiv:[1612.08255]{}]{}
Heiko Harborth, *Ein [E]{}xtremalproblem für [G]{}itterpunkte*, J. Reine Angew. Math. **262/263** (1973), 356–360, Collection of articles dedicated to Helmut Hasse on his seventy-fifth birthday. [MR ]{}[0327666]{}
Gábor Hegedüs, *The erdös-ginzburg-ziv constant and progression-free subsets*, ArXiv e-prints (2017), [arXiv:[1701.01038]{}]{}
Arnfried Kemnitz, *On a lattice point problem*, Ars Combin. **16** (1983), no. B, 151–160. [MR ]{}[737118]{}
Eric Naslund, *The partition rank of a tensor and $k$-right corners in $\mathbb{F}_p^n$*, ArXiv e-prints (2017), [arXiv:[1701.04475v2]{}]{}
Eric Naslund and Will Sawin, *Upper bounds for sunflower-free sets*, Forum Math. Sigma **5** (2017), e15, 10. [MR ]{}[3668469]{}
Christian Reiher, *On [K]{}emnitz’ conjecture concerning lattice-points in the plane*, Ramanujan J. **13** (2007), no. 1-3, 333–337. [MR ]{}[2281170]{}
Will Sawin and Terence Tao, *Notes on the slice rank of tensors*, 2016, [](https://terrytao.wordpress.com/2016/08/24/notes-on-the-slice-rank-of-tensors/).
Terence Tao, *A symmetric formulation of the croot-lev-pach-ellenberg-gijswijt capset bound*, 2016, [](https://terrytao.wordpress.com/2016/05/18/a-symmetric-formulation-of-the-croot-lev-pach-ellenberg-gijswijt-capset-bound/)
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper the new type of thermal corrections for the helium and helium-like atomic systems are introduced. These are thermal one-photon exchange between the bound electrons and nucleus as well as between the bound electrons induced by the blackbody radiation (BBR). All the derivations are given within the rigorous QED theory. It is shown that these thermal corrections are the same order in powers of $\alpha$ (fine structure constant) as the well-known BBR-induced Stark shift but the different behaviour in temperature. The numerical results presented in this paper make possible to expect their significance for modern experiments and testing the fundamental interactions in helium.'
author:
- 'D. Solovyev'
- 'T. Zalialiutdinov'
- 'A. Anikin'
bibliography:
- 'mybibfile.bib'
title: Thermal corrections of lowest order for a helium atom
---
Helium is a few-electron atomic system which is the subject of present theoretical and experimental studies. The great success in experimental measurements of the transition frequencies in the helium atom [@Roo; @Zheng] makes it necassery to carry out very precise theoretical calculations [@Drake_1999; @Korobov_2000] and vice versa. At the present days, experimental accuracy reaches the level of several parts in $10^{-12}$ and requires a detailed analysis of various relativistic, quantum electrodynamics (QED) and etc corrections, see for example [@PYP-2016; @PYP-2017], which are usually evaluated in the framework of nonrelativistic theory. Progress in theoretical calculations and spectroscopic measurements of the energies of bound states in a helium atom makes it possible to determine accurately the charge radii of the nuclei of the ${}^3$He and ${}^4$He isotopes [@MWDrake; @PK-Ch].
The very precise theoretical calculations of the binding energies, the fine structure, and the isotope shift of the low-lying states of helium are inclined to serve as an independent tool for testing of fundamental interactions. Similar to well-studied one-electron atomic systems the measured transition frequencies should be compared with the theoretical calculations pursuing the search of possible discrepancy [@PPY-2017]. The quantum electrodynamics theory developed during last decades for the helium and helium-like ions allows the calculations of different corrections arising within expansion in powers of the fine structure constant $\alpha$ and the electron-to-nucleus mass ratio $m/M$ upto the effects of the order of $\alpha^6 m^2/M$ and $\alpha^7m$ [@PPY-2017].
Such extraordinary calculations however pay attention to the effects of the other type: corrections induced by the external blackbody radiation (BBR) field. The influence of the BBR field is well-known in atomic physics, it is manifested in the existence of a Stark shift of bound states. The theory and corresponding calculations for one- and few-electron atoms were presented in [@farley] in the framework of the Quantum Mechanical (QM) approach. These calculations were continued for the case of atomic clocks (many-particle systems) in [@Saf; @SKC; @Porsev] and are the subject of theoretical investigations in present days. Not long ago, the QED derivation of the Stark shift induced by the BBR field was performed in [@SLP-QED] and, subsequently, applied to calculations in the helium atom [@jphysb2017].
Recently, the QED theory of thermal energy shifts induced by the BBR field was presented in [@S-TQED] for the one-electron atomic systems. In particular it was found that the thermal interaction of a bound electron with a nucleus can lead to an energy shift exceeding the corresponding Stark shift. Although the theory developed in [@S-TQED] is a continuation of thermal quantum electrodynamics (TQED), see, for example, [@Dol; @Don; @DHR], the effect of thermal interaction has not been previously considered. An experimental verification of this effect for a hydrogen atom was also proposed in [@S-TQED]. In this paper, the same corrections are investigated for two-electron atomic system. Increasing experimental accuracy [@Roo; @Zheng] makes it possible to observe this effect in helium. It can be expected that thermal corrections of this type can serve for further testing of fundamental interactions on helium atom [@PPY-2017].
We start by considering the interelectron interaction given as the one thermal photon exchange between two bound electrons. Such interaction can be presented schematically by the Feynman graph in Fig. \[Fig1\].
![The Feynman diagram depicting the one thermal photon exchange between the bound electrons. The double solid lines denote the bound electron. The wavy line with the index $\gamma_T$ means the thermal photon. Indices $i(i')$ and $f(f')$ characterize the initial and final states of bound electrons, respectively.[]{data-label="Fig1"}](Fig-1.png)
The $S$-matrix element is $$\begin{aligned}
\label{1}
S_{fi} = (-ie)^2\int d^4x_1 d^4x_2\left(\bar{\psi}_f(x_1)\gamma^\mu\psi_i(x_1)\right)\times
\\
\nonumber
D_{\mu\nu}^\beta(x_1,x_2) \left(\bar{\psi}_{f'}(x_2)\gamma^\nu\psi_{i'}(x_2)\right),\end{aligned}$$ where $\psi_i$ and $\bar{\psi}_i$ are the one-electron and its Dirac conjugated wave functions, respectively. The initial and final states of bound electrons are denoted by the indeces $i$ and $f$, $\gamma^\mu$ are the Dirac matrices and $x_i=(t_i,\vec{r}_i)$ represents the four-space coordinate vector for each electron. The function $D_{\mu\nu}^\beta(x_1,x_2)$ is the thermal part of photon propagator, see [@Dol; @Don].
In [@S-TQED] is was established that the thermal part of photon propagator $D_{\mu\nu}^\beta(x_1,x_2)$ is given by the Hadamard propagation function [@Akhiezer; @Greiner] and, therefore, admits a different (equivalent) form: $$\begin{aligned}
\label{2}
i D^{\mu \nu}_{\beta}(x, x') =
- 4\pi i g^{\mu\nu}\int\limits_{C_1}\frac{d^4k}{(2\pi)^4} \frac{e^{ik(x-x')}}{k^2}n_\beta(|\vec{k}|),\end{aligned}$$ where $g^{\mu\, \nu}$ is the metric tensor, $k^2=k_0^2-\vec{k}^2$ and $n_{\beta}$ is the Planck’s distribution function. The contour of integration in $k_0$-plane for Eq. (\[2\]) is given in Fig. \[Fig-2\].
![Integration contour $C_1$ in $k_0$ plane of Eq. (\[2\]).Arrows on the contour define the pole-bypass rule. The poles $\pm\omega_k$ are denoted with $\times$ marks.[]{data-label="Fig-2"}](Fig-2.png)
Thermal photon propagator in the form Eq. (\[2\]) has the advantage of allowing the introduction of gauges in complete analogy with the ’ordinary’ QED theory, see [@S-TQED]. Then in the Coulomb gauge the function $D_{\mu\nu}^\beta(x_1,x_2)$ recasts into $$\begin{aligned}
\label{3}
D^{00}_{\beta}(x, x') &=& 4\pi i \int\limits_{C_1}\frac{d^4k}{(2\pi)^4}\frac{e^{i k (x-x')}}{\vec{k}^2}n_\beta(\omega),\qquad
\\
\nonumber
%\label{4}
D^{ij}_{\beta}(x, x') &=& 4\pi i \int\limits_{C_1}\frac{d^4k}{(2\pi)^4}\frac{e^{i k (x-x')}}{k^2}n_\beta(\omega)\left(\delta^{ij}-\frac{k^i k^j}{\vec{k}^2}\right).\end{aligned}$$ In following calculations we restrict ourselves by the consideration of $D^{00}_{\beta}(x, x')$ only, reproducing the thermal Coulomb interaction of two charges in the way similar to the ordinary Coulomb interaction [@LabKlim].
Substitution of the Coulomb part $D^{00}_{\beta}(x, x')$ into Eq. (\[1\]), after the integration in $k_0$-plane and time variables $t_{1,2}$, yields $$\begin{aligned}
\label{4}
S_{fi} = -2\pi i\delta\left(E_f-E_i+E_{f'}-E_{i'}\right)\frac{4e^2}{\pi}\times\qquad
\\
\nonumber
\int d^3r_1 d^3r_2\left(\bar{\psi}_f\psi_i\right)(\vec{r}_1)\int\limits_0^\infty d\kappa\,n_\beta(\kappa)\frac{\sin\kappa r_{12}}{\kappa r_{12}}\left(\bar{\psi}_{f'}\psi_{i'}\right)(\vec{r}_2).\end{aligned}$$ Here $r_{12}\equiv |\vec{r}_1-\vec{r}_2|$ as usual. The energy shift for an arbitrary bound state arises via the relations: $\langle a'|\hat{S}|a\rangle = - 2\pi i \langle a'| U | a \rangle\delta(E_{a'}-E_a)$ and $\Delta E_a = \langle a | U | a \rangle$. Thus, one can find that the integral $\int\limits_0^\infty d\kappa\,n_\beta(\kappa)\frac{\sin\kappa r_{12}}{\kappa r_{12}}$ represents the desired thermal interaction potential averaged on the one-electron wave functions. Evaluation of the diagram in Fig. \[Fig1\] in case of zero temperature does not contain the Planck’s distribution function and is twice less than the expression (\[4\]). Then the integration over $\kappa$ would lead to $\pi/2r_{12}$, i.e. the interelectron Coulomb interaction.
In particular, Eq. (\[4\]) implies a divergence of the type $\int\limits _ 0^\infty d\kappa\,n_\beta(\kappa)$ arising from a series expansion in powers of $\kappa r_{12}$. This divergence however does not depend on the state and can simply be excluded from consideration, since it vanishes in the energy difference (transition frequency). For the conciseness we omit the discussion of possible ways to regularize this divergence and employ the procedure proposed in [@S-TQED], where the appropriate analytical calculations of the integral over $\kappa$ can be found. The final result was obtained as $$\begin{aligned}
\label{5}
V^\beta(r)=-\frac{4e^2}{\pi}
\left(-\frac{\gamma}{\beta}+\frac{i }{2 r}\ln \left[\frac{\Gamma \left(1+\frac{i r}{\beta}\right)}{\Gamma \left(1-\frac{i r}{\beta}\right)}\right]\right),\end{aligned}$$ where $\beta\equiv 1/(k_B T)$ ($k_B$ is the Boltzmann constant and $T$ is the temperature in kelvin) and $r$ represents the modulus (length) of corresponding radius vector for the electron-nucleus or electron-electron interactions. $\Gamma$ is the Gamma function and $\gamma$ is the Euler-Mascheroni constant, $\gamma\simeq 0.577216$.
In the nonrelativistic limit at low temperatures, the potential (\[5\]) can be expanded in powers of $r\rightarrow 0$ or $\beta\rightarrow\infty$ that in the lowest order gives $$\begin{aligned}
\label{6}
V^\beta(r) \approx -\frac{4e^2}{\pi} \left[-\frac{r^2 \zeta(3)}{3 \beta^3}+\frac{r^4 \zeta(5)}{5 \beta^5}+\dots\right],\end{aligned}$$ where $\zeta(n)$ is the Riemann zeta function. It should be emphasized here that the expression (\[6\]) matches exactly to the series expansion in the formula (\[4\]) as if the divergent terms were excluded. Then, averaging the expression (\[6\]) on two-electron wave-functions, the lower-order thermal correction can be written as $$\begin{aligned}
\label{7}
\Delta E^\beta_A = \frac{4\zeta(3)}{3\pi\beta^3}\langle A|e^2 r_{12}^2 -Ze^2 r_1^2 - Ze^2 r_2^2|A\rangle.\end{aligned}$$ Here $A$ represents a two-electron bound state, $r_i$ is the length of the radius vector for each bound electron, respectively, and $r_{12}$ denotes the interelectron distance. The derivation of the thermal interaction of electrons with nucleus repeats the above calculations and we omit it for the short, see [@S-TQED] for more details.
The parametric estimate of $\Delta E^\beta_A$ can be obtained in the ordinary manner, it is $\Delta E_a^{\beta}\sim Z^5\alpha^3/\beta^3$ for the thermal electron-nucleus interaction and $\Delta E_a^{\beta}\sim Z^4\alpha^3/\beta^3$ for the thermal electron-electron interaction in atomic units. The temperature factor is $\beta^{-1}=k_B T\sim 10^{-3}$ at room temperature and, therefore, this correction can be compared with QED corrections of the order of $\alpha^7m$ [@PPY-2017]. In addition, the correction (\[7\]) should be compared with the BBR-induced Stark shift. According to the results of [@SLP-QED] the parametric estimate in $\alpha$ for the Stark shift is the same, but should have an extra factor of $\beta$ at low temperatures. Thus, one can expect that the values of Eq. (\[3\]) will exceed the corresponding BBR-induced Stark shift.
For the numerical calculations in the two-electron atom, we use trial wave functions with quasirandom nonlinear parameters developed in [@Korobov_1999; @Korobov_2000]. As a first step in testing the methods of calculation, the nonrelativistic energies of helium states were evaluated, see Table \[tab:1\], which are in good agreement with [@Drakebook; @Frolov-1998].
State Value obtained in this work Drake [@Drakebook]
---------- ----------------------------- --------------------------
$ 1^1S $ $ -2.9037243770 $ $ -2.9037243770341195 $
$ 2^1S $ $ -2.1459740460 $ $ -2.145974046054419 $
$ 2^3S $ $ -2.1752293782 $ $ -2.17522937823679130 $
$ 2^1P $ $ -2.1238430864 $ $ -2.123843086498093 $
$ 2^3P $ $ -2.1331641908 $ $ -2.133164190779273 $
$ 3^1S $ $ -2.0612719897 $ $ -2.061271989740911 $
$ 3^3S $ $ -2.0686890674 $ $ -2.06868906747245719 $
$ 3^1P $ $ -2.0551463620 $ $ -2.05514636209194 $
$ 3^3P $ $ -2.0580810842 $ $ -2.05808108427428 $
$ 3^1D $ $ -2.0556207328 $ $ -2.055620732852246 $
$ 3^3D $ $ -2.0556363094 $ $ -2.055636309453261 $
$ 4^1S $ $ -2.0335867167 $ $ -2.03358671703072$
$ 5^1S $ $ -2.0211767759 $ $ -2.021176851574363$
: Nonrelativistic energies of helium states obtained in the present work in a.u. by variational method [@Korobov_2000].[]{data-label="tab:1"}
The calculated expectation values of the $r_{12}^2$, $r_1^2$ operators and the corresponding energy shift are collected in Table \[tab:2\].
State $r_{12}^2$ in a.u. $r_1^2$ in a.u. $\Delta E^\beta_A$ in a.u. $\Delta E^\beta_A$ in Hz
---------- -------------------- ----------------- ---------------------------- --------------------------
$ 1^1S $ $ 2.516439313 $ $ 1.193482995 $ $ -3.83766\cdot 10^{-16} $ $ -2.52505 $
$ 2^1S $ $ 32.30238038 $ $ 16.08923325 $ $ -5.44916\cdot 10^{-15} $ $ -35.8537 $
$ 2^3S $ $ 23.04619748 $ $ 11.46432162 $ $ -3.8778\cdot 10^{-15} $ $ -25.5147 $
$ 2^1P $ $ 31.59851603 $ $ 15.76565497 $ $ -5.34878\cdot 10^{-15} $ $ -35.1933 $
$ 2^3P $ $ 26.64279322 $ $ 13.21174046 $ $ -4.45461\cdot 10^{-15} $ $ -29.3099 $
$ 3^1S $ $ 171.8387553 $ $ 85.89015822 $ $ -2.91921\cdot 10^{-14} $ $ -192.075 $
$ 3^3S $ $ 137.4750832 $ $ 68.70840131 $ $ -2.33505\cdot 10^{-14} $ $ -153.638 $
$ 3^1P $ $ 183.7866266 $ $ 91.87290715 $ $ -3.12292\cdot 10^{-14} $ $ -205.478 $
$ 3^3P $ $ 164.3028806 $ $ 82.10989293 $ $ -2.79026\cdot 10^{-14} $ $ -183.59 $
$ 3^1D $ $ 126.4161413 $ $ 63.17681865 $ $ -2.1469\cdot 10^{-14} $ $ -141.259 $
$ 3^3D $ $ 126.2834766 $ $ 63.11075331 $ $ -2.14467\cdot 10^{-14} $ $ -141.112 $
$ 4^1S $ $ 562.8623853 $ $ 281.4144002 $ $ -9.56731\cdot 10^{-14} $ $ -629.499 $
$ 5^1S $ $ 1406.742766 $ $ 703.3605371 $ $ -2.39134\cdot 10^{-13} $ $ -1573.42 $
: Expectation values of $r_1^2$, $r_{12}^2$ operators (in a.u.) and the corresponding energy shift Eq. (\[7\]) at room temperature ($T=300$K) in Hz for the He($M = \infty$) atom. []{data-label="tab:2"}
In particular, from Table \[tab:2\] it follows that the thermal correction $\Delta E^{\beta}_A$, Eq. (\[7\]), is one order of magnitude larger than the corresponding Stark shift for lowest states. However, the correction Eq. (\[7\]) becomes the same order as the Stark shift, see [@farley], for the state $ 3^1S $ and higher but it has the opposite sign.
To demonstrate the correct reasoning on the temperature corrections, the numerical data for the second term in Eq. (\[6\]) are given in Table \[tab:3\]. The order of magnitude was estimated as $Z^7\alpha^5$ for the thermal interaction of electrons with nucleus interaction and, consequently, as $Z^6\alpha^5$ for the thermal electron-electron interaction in atomic units.
State $r_{12}^4$ in a.u. $r_1^4$ in a.u. $\Delta E^{\beta(r^4)}_A$ in Hz
---------- ---------------------------- ---------------------------- ---------------------------------
$ 1^1S $ $ 12.98127136 $ $ 3.973564932 $ $ -8.10541\cdot 10^{-11} $
$ 2^1S $ $ 0.1737415294\cdot 10^4 $ $ 0.8257531786\cdot 10^3 $ $ -4.35629\cdot 10^{-8} $
$ 2^3S $ $ 0.9163894338\cdot 10^3 $ $ 0.4284022731\cdot 10^3 $ $ -2.21827\cdot 10^{-8} $
$ 2^1P $ $ 0.1839207883\cdot 10^4 $ $ 0.8786029923\cdot 10^3 $ $ -4.66127\cdot 10^{-8} $
$ 2^3P $ $ 0.1349634573\cdot 10^4 $ $ 0.6388893042\cdot 10^3 $ $ -6.58923\cdot 10^{-9} $
$ 3^1S $ $ 0.4271527957\cdot 10^5 $ $ 0.2113668547\cdot 10^5 $ $ -1.16396\cdot 10^{-6} $
$ 3^3S $ $ 0.2772173839\cdot 10^5 $ $ 0.1368384508\cdot 10^5 $ $ -7.51657\cdot 10^{-7} $
$ 3^1P $ $ 0.5018603163\cdot 10^5 $ $ 0.2485719672\cdot 10^5 $ $ -1.37018\cdot 10^{-6} $
$ 3^3P $ $ 0.4055841235\cdot 10^5 $ $ 0.2006781742\cdot 10^5 $ $ -1.10501\cdot 10^{-6} $
$ 3^1D $ $ 0.2570029451\cdot 10^5 $ $ 0.1268705703\cdot 10^5 $ $ -6.96961\cdot 10^{-7} $
$ 3^3D $ $ 0.2565480602\cdot 10^5 $ $ 0.1266449525\cdot 10^5 $ $ -6.95715\cdot 10^{-7} $
$ 4^1S $ $ 0.4323006495\cdot 10^6 $ $ 0.2154350194\cdot 10^6 $ $ -1.19492\cdot 10^{-5} $
$ 5^1S $ $ 0.2624327529\cdot 10^7 $ $ 0.1310386371\cdot 10^7 $ $ -7.28243\cdot 10^{-5} $
: Expectation values of second term in Eq. (\[6\]) and corresponding energy shifts at room temperature ($T=300$K) for helium. []{data-label="tab:3"}
Thus, it can be found that the thermal correction corresponding to the second term in Eq. (\[6\]) is negligible.
As the next step in the analysis of the diagram in Fig. \[Fig1\], the thermal correction on finite size of nucleus (NS) can be obtained. This correction occurs by the introducing the charge distribution of the nucleus $\rho\approx 1 - \frac{\vec{k}2}{6}r_N^2$ in momentum space, where $r_N^2$ denotes the mean square value of the nuclear charge radius. Derivation of thermal correction on the finite size of nucleus can be found in [@S-TQED], where the lowest order effect was found in the form: $$\begin{aligned}
\label{8}
\Delta E^{\beta(NS)}_A = \frac{32 Ze^2}{15}\frac{\zeta(5)}{\beta^5}r_N^7\left|\psi_A(0)\right|^2.\end{aligned}$$
To evaluate this correction, one can separate out the ordinary (non-thermal) NS correction of lowest order: $$\begin{aligned}
\label{9}
\Delta E^{\beta(NS)}_A =C_\beta \frac{2\pi}{3}Z e^2\left|\psi_A(0)\right|^2r_N^2,
\\
\nonumber
C_\beta= \frac{16}{5\pi}\zeta(5)\frac{r_N^5}{\beta^5}.\end{aligned}$$ The dimensionless coefficent $C_\beta$ can be estimated via the Compton wavelength $\lambdabar = \hbar/m_e c$ as follows $$\begin{aligned}
\label{10}
C_\beta \equiv \frac{16}{5\pi}\zeta(5)\frac{r_N^5}{\lambdabar^5\beta^5}.\end{aligned}$$ where $r_N$ is given in fm and $\beta = m_e c^2/k_B T$ (in SI units) can be translated in appropriate units. The thermal correction on the finite size of nucleus, Eq. (\[8\]), is insignificant at relevant temperatures, since the coefficient $C_\beta$ is equal to $4.0765\cdot 10^{-50} r_N^5 $ $ \rm{fm}^{-5} $ at room temperature.
The result Eq. (\[7\]) can be extended to a negative hydrogen atom. The corresponding expectation values of the operators are $r_1^2=11.91369967805$ and $r_{12}=25.2020252912$ in atomic units for the ground state, respectively, see [@Drakebook]. Then, the thermal correction (\[7\]) gives the energy shift $\Delta E^\beta_A = 1.53755$ Hz at the temperature $300$ K. In contrast to the helium atom this correction has the opposite sign. Since the hydrogen anion is of particular interest in studying the atmosphere of stars, we also give the value $\Delta E^\beta_A = 12300.4$ Hz at an effective temperature of photosphere $6000$ K, and the same correction is equal to $1.92194\cdot 10^{11}$ Hz (or $0.00079485$ eV) at the corona temperature ($1.5\cdot 10^6$ K) of the Sun.
The results of this paper were obtained within the $S$-matrix formalism. It is found that the replacement of the ordinary photon line by the thermal one gives rise to the thermal potential. Since the thermal potential Eq. (\[5\]) was never recognized before, the new type of thermal corrections is introduced here. The theory presented in this paper for the thermal corrections of lowest order reduces to evaluation of one-particle operators to thermal electron-electron and electron-nuclear interactions in atomic systems with few electrons. Thus, the corresponding mathematical derivations repeat the case of one-electron atom and, as a consequence, the result Eq. (\[6\]) and thermal correction of lowest order Eq. (\[7\]) were obtained within the nonrelativistic limit and the point-nucleus assumption. Going beyond the approximation of a point nucleus the NS thermal correction can be considered, see Eq. (\[8\]), which turns out to be negligibly small.
The parametric estimation of thermal correction Eq. (\[6\]) is the same as for the well-known Stark shift induced by the BBR field. However, this correction differs by the factor of temperature. The numerical results are given in Table \[tab:2\] for the different bound states of helium atom. The values in Tables \[tab:2\] and \[tab:3\] establish that the thermal corrections arising from the thermal photon exchange of electron with the nucleus and other electrons reach a level of few kHz for low-lying states in ${}^4$He atom at the room temperature. The contribution grows with an increase of the principal quantum number of the bound state repeating the behaviour in the one-electron atomic system, see [@S-TQED]. We limited our calculations upto the $5^1S$ state, since the descipancy between the numerical values of the bound energies (comparing with the results of [@Drakebook]) becomes visible, see Table \[tab:1\]. Although such accuracy is not required (ten digits after the decimal point), the calculations, in principle, can be continued to higher excited states. We leave this problem for future works. It should noted here that the evaluation of diagrams with two, three and etc thermal photon exchange between electrons or an electron and a nucleus would lead to a thermal correction of the same order in powers of $\alpha$ according to the QED theory. However, each thermal photon line would produce an additional factor of temperature (cubic in our case) and, therefore, such diagrams should lead to thermal corrections of the next orders of smallness. Partially, a rapid decrease in temperature is confirmed by an estimate of the thermal correction corresponding to the second term in the formula (\[6\]). The numerical results for this correction are compiled in Table \[tab:3\] which shows the insignificance of the effect.
Finally, the data listed in Tables \[tab:2\], \[tab:3\] were obtained for a temperature $300$ K. They can be easily extended to other temperatures by dint of the factors $(\frac{T}{300})^3$ and $(\frac{T}{300})^5$, respectively. For example, at temperature $77$ K (the boiling point of nitrogen) the coefficient $(\frac{T}{300})^3$ is $0.0169086$. Thus, one can expect that the presence of corrections resulting from the thermal photon exchange between bound electrons as well as bound electron and nucleus can be verified experimentally in a helium atom by varying the temperature. In conclusion, this new type of thermal corrections can serve to test fundamental interactions on helium and helium-like ions.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by Russian Foundation for Basic Research (grant 20-02-00111). T. Z. acknowledges foundation for the advancement of theoretical physics and mathematics “BASIS”. The authors are indebted to V. I. Korobov for permission to use the [*Fortran*]{} code for the construction of the He variational wave functions.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The known approaches of number-phase problem (for a quantum oscillator) are mutually contradictory. All of them are subsequent in respect with the Robertson-Schrödinger uncertainty relation (RSUR). In opposition here it is proposed a new aproacch aimed to be ab origine as regard RSRUR. From the new perspective the Dirac’s operators for vibrational number and phase appear as correct mathematical tools while the alluded problem receives a natural solution.'
author:
- 'S. Dumitru'
bibliography:
- 'articol-prl.bib'
title: 'On the number-phase problem'
---
Introduction
============
A recent outstanding work [@1] reviews the publications referring to the problem of theoretical description for the vibrational number $N$ and phase $\Phi$ of a quantum oscillator (QO). So one discloses the fact that the respective problem known various approaches differing among them both quantitatively and qualitatively. Moreover it is pointed out that until now in scientific literature an agreement regarding the mentioned problem does not exist. As we shall show below all the alluded approaches are subsequent in respect with the Robertson-Schrödinger uncertainty relation (RSUR). In such a context we think that another approach, ab origine regarding RSRUR, as the one which we present in the next sections, can be of nontrivial interest.
Briefly on known facts
======================
The story of $N-\Phi$ problem [@1] started with the Dirac’s idea to transcribe the ladder (annihilation and creation) operators $\hat{a}$ and $\hat{a}^+$ in the forms $$\label{eq:1}
\hat{a} = e^{i\hat\Phi} \sqrt{\hat N} \qquad
\hat{a}^+ = \sqrt{\hat N} e^{-i\hat\Phi}$$ For an oscilator $\hat N$ and $\hat\Phi$ were identified with the operators of vibrational number respectively of phase. Due to the fact that $[ \hat a, \hat a^+]= \hat a \hat a^+ -\hat a^+ \hat a = 1$ from it follows $$\label{eq:2}
[\hat N , \hat\Phi] = i$$
Then this relation was regarded in connection with RSUR $$\label{eq:3}
\Delta A \cdot \Delta B \geq \frac{1}{2} \vert \langle [ \hat A,\hat B ]
\rangle \vert$$ In the standard deviations $\Delta A$ and $\Delta B$ respectively the operators $A$ and $B$ refer to the two arbitrary observables $A$ and $B$. By a direct application of to the $N-\Phi$ case it was introduced the relation $$\label{eq:4}
\Delta N \cdot \Delta\Phi \geq \frac{1}{2}$$ But, lately it was found that relation is false - at least in some well-specified situations. Such a situation appears in the case of QO eigenstate corresponsding to the energy eigenvalue $E_n = \hbar\omega ( n+ \frac{1}{2})$. The respective state is described by the wave function $\Psi_n$ for which one obtains $\hat N\Psi_n = \hat a^+\hat a \Psi_n = n\Psi_n$, $\Delta N =0$ respectively $\Delta\Phi \leq 2\pi$ (the noted restriction for the value of $\Delta\Phi$ results [@1] from the fact that the range of definition for $\Phi$ is the interval $[0, 2\pi)$). With the mentioned falsity of the $N-\Phi$ problem reached a deadlock. For avoiding the mentioned deadlock in literature various alternative approaches were promoted (see [@1] and references). But it is easy to emark that all the alluded approaches are subsequent (and dependent) in respect with the RSUR in the following sense. The respective approaches consider as an absolutely valid formula and try to adjust accordingly the description of the pair $N-\Phi$ for QO. So the operators $\hat N$ and $\hat\Phi$ defined in were replaced by some alternative operators $\hat N_a$ and $\hat\Phi_a$ whose standard deviations $\Delta N_a$ and $\Delta\Phi_a$ satisfy relations resembling more or less with . But it is very doubtful that the variables $N_a$ and $\Phi_a$ have a natural (or even useful) physical significance. Probably that this is one of the reasons why until now, in scientific community, it does not exist an agreement regarding the $N-\Phi$ problem.
An ab origine approach
======================
In contrast with the known approaches of $N-\Phi$ problem alluded above we think that the same problem can be approached on a new way which is ab origine (i.e. non-subsequent) in respect with the RSUR . Such a new approach can be done by investigating the true origin of the relation as well as the conditions of validity for the respective relation. For puting in practice our thinking we act as follows. Firstly let us remind some elements/notations from quantum mechanics. We consider a system (particularly an oscilator) of quantum nature. The state of the system and its observables $A_j (j=1,2,\ldots,r)$ are described by the wave function $\Psi$ respectively by the operators $\hat A_j$. The scalar product of two functions $\Psi_\alpha$ and $\Psi_\beta$ will be denoted by $(\Psi_\alpha,\Psi_\beta)$. In the state described by $\Psi$ the expected (mean) value of the quantity $A_j$ is given by $\langle A_j \rangle= (\Psi,\hat A_j\Psi)$ and the operator $\delta\hat A_j = \hat A_j -\langle A_j\rangle$ can be defined. Then for two osevables $A_1=A$ and $A_2=B$ one can write the following Schwartz relation $$\label{eq:5}
(\delta \hat A\Psi, \delta \hat A\Psi) \, (\delta\hat B\Psi, \delta\hat B\Psi)
\geq \vert (\delta\hat A \Psi,\delta \hat B \Psi) \vert^2$$
But $(\delta\hat A\Psi,\delta\hat A\Psi)=(\Delta A)^2$ where $\Delta A$ denotes the standard deviation of $A$. So from one obtains $$\label{eq:6}
\Delta A \cdot \Delta B \geq \vert ( \delta\hat A \Psi, \delta \hat B\Psi)\vert$$ Note that the relation is generally valid for any wave function $\Psi$ and any observables $A$ and $B$. The respective relation imply the less general formula which is RSUR only in particular circumstances. The alluded circumstances can be specified as follows. If in respect with the wave function $\Psi$ the operators $\hat A=\hat A_1$ and $\hat B=\hat A_2$ satisfy the conditions $$\label{eq:7}
(\hat A_j \Psi, \hat A_k \Psi)=(\Psi, \hat A_j \hat A_k \Psi) \qquad
(j=1,2;\, k=1,2)$$ one can write $$\label{eq:8}
(\delta \hat A \Psi, \delta \hat B \Psi)=\frac{1}{2}
(\Psi, \{ \delta \hat A, \delta\hat B\}\Psi) - \frac{i}{2}
(\Psi, i[\hat A , \hat B]\Psi)$$ where both $(\Psi, \{ \delta\hat A,\delta\hat B\}\Psi) = (\Psi,(\delta\hat A
\, \delta\hat B +\delta\hat B\, \delta\hat A)\Psi)$ and $(\Psi,i[\hat A, \hat B]\Psi)$ are real quantities. This means that in the circumstances strictly delimited by the conditions the relation imply directly the RSUR . In all other circumstances RSUR is false but the relation remains always valid. The above-presented considerations, regarded in connection with the here investigated $N-\Phi$ problem, justify the following observations (**Ob**):
- For a state described by an eigenfunction $\Psi_n$ mentioned in Sec.2, by using the formula $\hat N\Psi_n = n\Psi_n$ together with the relation , it results $$\label{eq:9}
(\hat N\Psi_n,\hat\Phi\Psi_n) = (\Psi_n, \hat N\hat\Phi\Psi_n)+i$$ Such a result clearly shows that in the considered situation $N$ and $\Phi$ do not satisfy the conditions . This means that, in the respective situation, for $N$ and $\Phi$ the RSUR and (consequently) the formula are not valid. However in the same situation the relation remain true. But then $\delta\hat N \Psi_n=0$, $\Delta N = 0$ and $\Delta\Phi \leq 2\pi$ (more exactly $\Delta\Phi = \pi/\sqrt{3}$ - see below in Appendix). So in the alluded state degenerates into trivial equality $0 = 0$.
- : The cases when the state of the oscilator is described by a wave function of the form $\Psi=\sum_n C_n\Psi_n$ ( with $\sum_n \vert C_n\vert^2 =1$) must be discussed distinctly. Depending on the values of coefficients $C_n$ in respect with such a function $\hat N$ and $\hat\Phi$ defined by can satisfy the conditions . Then relation is valid with $$\label{eq:10}
\Delta N= \Big[\sum_n \vert C_n\vert^2 n^2 - (\sum_n \vert
C_n\vert^2 n)^2\Big]^{1/2} \, , \quad \Delta\Phi \leq 2\pi$$
- By means of the operators $\hat N$ and $\hat\Phi$ from can be composed a simple procedure able to reveal the important characteristics of a QO (see below the Appendix).
Conclusions {#conclusions .unnumbered}
===========
In scientific literarure the $N-\Phi$ problem persists as a controversial question. In the main the controversies originate from the fact that the operators $\hat N$ and $\hat\Phi$ defined by are incompatible with the RSUR /. The known approaches of the problem are subsequent in the respect with RSUR . In oposition we propose a new approach which is ab origine in relation with RSUR . Within the framework of the proposed approach we point out the delimitative conditions in which the RSUR is valid. Also we noted that in fact RSUR originates from a more general formula namely from the Schwartz relation / which is ab origine and always valid. Then as it is shown in in the cases of energy eigenstates of a QO the operators $\hat N$ and $\hat\Phi$ do not satisfy the conditions of type required by RSUR . However in the respective cases the pair $N-\Phi$ satisfies the Schwartz’s relation of type which reduces to the trivial equality $0 = 0$. As we have shown in **Ob.1** the mentioned facts give a natural elucidation of the known incompatibility (and resulting troubles) existing between the Dirac’s operators $\hat N$ and $\hat\Phi$ from and the RSUR /. In direct connection with the respective elucidation in **Ob.2** we indicate a situation when the Dirac’s operators $N$ and $\Phi$ satisfy RSUR /. Additionally in **Ob.3** we noted that the Dirac’s operators $\hat N$ and $\hat\Phi$ prove themselves to be useful tools for composing a simple mathematical procedure able to reveal the important characteristics of a QO The resulting conclusion of the above discussions is that that the ab origine approach presented here gives a complete and natural solution to the controversial $N-\Phi$ problem. Accordingly the Dirac’s operators $\hat N$ and $\hat\Phi$ \[defined in \] remains as correct thoretical concepts with clear significance and utility for physics. From the perspective of the mentioned approach, in the case of a QO, the appeals to other “alternative operators for number and phase” seem to be pure mathematical exercices without major significance for physics.
The oscillator equation in $\Phi$-representation
================================================
In obtaining the characteristics of a QO the usual procedures operate with the wave function $\Psi$ taken in x-representation - i.e. $\Psi=\Psi(x)$ ($x$ = coordinate). Mathematically the respective procedures imply relative lablaborious work (for a direct solution of QO Schrödinger equation or, equivalently, for the iterative handling of the ladder operators $\hat a$ and $\hat a^+$ in the same equation). A more simple procedure can be done if the wave function $\Psi$ taken in $\Phi$-representation - i.e. $\Psi=\Psi(\Phi)$ \[with $\Phi$ from \]. In such a representation from it results directly that $\hat N$ has the form $\hat N=(i\partial/\partial\Phi)$. Then for the QO Hamiltonian $\hat H =(\hat p^2/2m)+(m\omega^2\hat x^2/2)=(\hbar\omega/2)
(\hat a\hat a^+ + \hat a^+\hat a)$ one obtains $$\label{eq:A.1}
\hat H = \hbar\omega \left(\hat N +\frac{1}{2}\right)
= \hbar\omega \left( i\frac{\partial}{\partial\Phi}+ \frac{1}{2}\right)$$ Consequently the Schrödinger equation for QO becomes $$\label{eq:A.2}
\hbar\omega \left( i\frac{\partial}{\partial\Phi} + \frac{1}{2}\right)
\Psi = E\Psi$$ From this equation it is very easy to infer the results $$\label{eq:A.3}
\Psi(\Phi)=\Psi_n(\Phi) = C e^{-in\Phi} \, , \qquad
E=E_n=\hbar\omega \left(n+\frac{1}{2}\right)$$ The integration constants $C$ and $n$ can be precised as follows. The condition $\Psi_n(0) = \lim_{\Phi\to 2\pi} \Psi_n(\Phi)$ requires for $n$ to be an integer number. From normalisation condition $(\Psi_n,\Psi_n) = 1$ it result $C=1/\sqrt{2\pi}$. Then requiring that, similarly with the case of classical oscillator, to have $E > 0$ one finds that $n \geq 0$. With the wave function $\Psi_n(\Phi)$ determined as above can be evaluated the characteristics of the QO. So for the states described by $\psi_n (\Phi)$ given by (which regard just the situations debated in literature [@1]) one obtains $$\label{eq:A.4}
\Delta N =0 \, , \quad \Delta\Phi = \pi/\sqrt{3} \, , \quad
(\delta\hat{N}\psi, \delta\hat{\Phi}\psi)=0$$
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Xavier Garcia[^1]'
- 'Jennifer Kunze[^2]'
- 'Thomas Rudelius[^3]'
- 'Anthony Sanchez[^4]'
- 'Sijing Shao[^5]'
- 'Emily Speranza[^6]'
- 'Chad Vidden[^7]'
title: Invariant Measures for Hybrid Stochastic Systems
---
In this paper, we seek to understand the behavior of dynamical systems that are perturbed by a parameter that changes discretely in time. If we impose certain conditions, we can study certain embedded systems within a hybrid system as time-homogeneous Markov processes. In particular, we prove the existence of invariant measures for each embedded system and relate the invariant measures for the various systems through the flow. We calculate these invariant measures explicitly in several illustrative examples.
***Keywords:** Dynamical Systems; Markov processes; Markov chains; stochastic modeling*
Introduction
============
An understanding of dynamical systems allows one to analyze the way processes evolve through time. Usually, such systems are given by differential equations that model real world phenomena. Unfortunately, these models are limited in that they cannot account for random events that may occur in application. These stochastic developments, however, may sometimes be modeled with Markov processes, and in particular with Markov chains. We can unite the two models in order to see how these dynamical systems behave with the perturbation induced by the Markov processes, creating a hybrid system consisting of the two components. Complicating matters, these hybrid systems can be described in either continuous or discrete time.
The focus of this paper is studying the way these hybrid systems behave as they evolve. We begin by defining limit sets for a dynamical system and stochastic processes. We next examine the limit sets of these hybrid systems and what happens as they approach the limit sets. Concurrently, we define invariant measures and prove their existence for hybrid systems while relating these measures to the flow. In addition, we supply examples with visuals that provide insight to the behavior of hybrid systems.
The Stochastic Hybrid System
============================
In this section, we define a hybrid system.
\[timehomogeneous\] A Markov process $X_t$ is called time-homogeneous on $T$ if, for all $t_1, t_2, k \in T$ and for any sets $A_1,A_2\in S$, $$P(X_{t_1+k} \in A_1 | X_{t_1} \in A_2) = P(X_{t_2+k} \in A_1 | X_{t_2} \in A_2) .$$ Otherwise, it is called time-inhomogeneous.
\[markovchain2\] A Markov chain $X_n$ is a Markov process for which perturbations occur on a discrete time set $T$ and finite state space $S$.
For a Markov chain on the finite state space $S$ with cardinality $|S|$, it is useful to describe the probabilities of transitioning from one state to another with a transition matrix $$Q \equiv \left ( \begin{array}{ccccc}
P_{1 \rightarrow 1} & .&.&.&P_{1 \rightarrow |S|} \\
.&&&&. \\
.&&&&. \\
.&&&&. \\
P_{|S| \rightarrow 1} & . & . & . & P_{|S| \rightarrow |S|}\\
\end{array} \right )$$ where $P_{i \rightarrow j}$ is the probability of transitioning from state $s_i\in S$ to state $s_j\in S$.
Also, for the purposes of this paper, we suppose that our Markov chain transitions occur regularly at times $t=nh$ for some length of time $h\in\mathbb{R}^+$ and for all $n\in\mathbb{N}$.
\[markovchain\] Let $\{X_n\}$, for $X_n \in S$ and $ n \in \mathbb{N}$, be a sequence of states determined by a Markov chain.
For $t \in \mathbb{R}^+$, define the Markov chain perturbation $Z_t=X_{\left \lfloor\frac{t}{h} \right \rfloor}$, where $\left \lfloor \frac{t}{h} \right \rfloor$ is the greatest integer less than or equal to $\frac{t}{h}$.
Note that $Z_t$, instead of being defined only on discrete time values like a Markov chain, is instead a stepwise function defined on continuous time.
\[varphi\] Given a metric space $M$ and state space $S$ as above, define a dynamical system $\varphi$ with random perturbation function $Z_t$, as given in Definition \[markovchain\], by $$\varphi: \mathbb{R}^+ \times M \times S \rightarrow M$$ with $$\varphi(t,x_0,Z_0) = \varphi_{Z_t}(t-nh, \varphi_{Z_{nh}}(h, ... \varphi_{Z_{2h}}(h, \varphi_{Z_h}(h,\varphi_{Z_0}(h,x_0)))))$$ where $\varphi_{Z_k}$ represents the deterministic dynamical system $\varphi$ evaluated in state $Z_k$ and $nh$ is the largest multiple of $h$ less than $t$.
For ease of notation, let $$x_t=\varphi(t, x_0, Z_0)\in M$$ represent the position of the system at time $t$.
\[hybridsys\] Let $$Y_t = \left ( \begin{array} {c}
x_t \\
Z_t
\end{array} \right )$$ define the hybrid system at time $t$. In other words, the hybrid system consists of both a position $x_t=\varphi(t, x_0, Z_0) \in M$ and a state $Z_t \in S$.
The $\omega$-limit set has the following generalization in a hybrid system.
\[C\] The stochastic limit set $C(x)$ for an element of our state space $x\in M$ and the hybrid system given above is the subset of $M$ with the following three properties:
1. Given $y\in M$ and $t_k\rightarrow\infty$ such that $x_{t_k}\rightarrow y$, $P(y\in C(x))=1$.
2. $C(x)$ is closed.
3. $C(x)$ is minimal: if some set $C'(x)$ has properties 1 and 2, then $C \subseteq C'$.
The Hybrid System as a Markov Process
=====================================
Each of the following is a Markov process:\
(i) any deterministic dynamical system $\varphi (t,x_0)$, as in Definition \[dynsys\].\
(ii) any Markov chain perturbation $Z_t$, as in Definition \[markovchain\].\
(iii) the corresponding hybrid system $Y_t$, as in Definition \[hybridsys\].
\(i) Any deterministic system is trivially a Markov process, since $\varphi(t,x_0)$ is uniquely determined by $\varphi(\tau,x_0)$ at any single past time $\tau\in\mathbb{R}^+$.
\(ii) By definition, a Markov chain is a Markov process. However, the Markov chain perturbation $Z_t$ is not exactly a Markov chain. A Markov chain exists on a discrete time set, in our case given by $T=\{t\in\mathbb{R}^+ | t=nh$ for some $n\in\mathbb{N}\}$; conversely, the time set of $Z_t$ is $\mathbb{R}^+$, with transitions between states ocurring on the previous time set (that is, at $t \equiv 0$ mod $h$). Despite this difference, $Z_t$ maintains the Markov property: we can compute $P(Z_t \in A)$ for any set $A$ based solely on $Z_{\tau_1}$ and the values of the times $t$ and $\tau_1$. Explicitly, the probability that $Z_t$ will be in state $s_i$ at time $t$ is given by $$P(Z_t = s_i) = \left( (Q^T)^n \right)_{ij}$$ where $n$ is the number of integer multiples of $h$ (i.e. the number of transitions that occur) between $t$ and $\tau_1$. Clearly, this is independent of the states $Z_{\tau_i}$ for $i > 1$, so that the random perturbation is indeed a Markov process.
\(iii) Now, keeping in mind that the hybrid system $Y_t$ consists of both a location $x_t\in M$ in the state space and a value $Z_t\in S$ of the random component, we can combine (i) and (ii) to see that the entire system is also a Markov process. We see from (ii) that $Z_t$ follows a Markov process. Furthermore, $P(x_t \in A_x)$ at time $t$ depends solely on the location $x_{\tau_1}$ at any time $\tau_1<t$ and the states of the random perturbation sequence $Z$ between $t$ and $\tau_1$, regardless of any past behavior of the system. Hence, for any collection of sets $A_{\alpha}$, $\alpha\in\mathbb{N}$, $$P(Z_t \in A_z | Z_{\tau_1} \in A_{z_1}, Z_{\tau_2} \in A_{z_2},...,Z_{\tau_n} \in A_{z_n}) = P(Z_t \in A_z | Z_{\tau_1} \in A_{z_1})$$ and $$P(x_t \in A_x | x_{\tau_1} \in A_{x_1}, x_{\tau_2} \in A_{x_2},...,x_{\tau_n} \in A_{x_n}) = P(x_t \in A_x | x_{\tau_1} \in A_{x_1}).$$ So, $$P(Y_t \in A_y | Y_{\tau_1} \in A_{y_1}, Y_{\tau_2} \in A_{y_2},...,Y_{\tau_n} \in A_{y_n}) = P(Y_t \in A_y | Y_{\tau_1} \in A_{y_1}).$$ Thus, the hybrid system is a Markov process.
Unfortunately, the hybrid system is not time-homogeneous. Recall that state transitions of $Z_t$ occur at times $t=nh$ for $n\in\mathbb{N}$. So, the state of the system at time $\frac{h}{4}$ uniquely determines the system at $\frac{3h}{4}$, since there is no transition in this interval. However, the system at time $\frac{5h}{4}$ is not determined uniquely by the system at $\frac{3h}{4}$, since a stochastic transition occurs at $t = h \in [\frac{3}{4},\frac{5}{4}]$. Therefore, with $t_1 = \frac{h}{4}$, $t_2 = \frac{3h}{4}$, and $k = \frac{1}{2}$, $$P(Y_{\frac{h}{4}+ \frac{1}{2}}\in A | Y_\frac{h}{4} \in A_0) \neq P(Y_{\frac{3h}{4}+\frac{1}{2}} \in A | Y_\frac{3h}{4} \in A_0) ,$$ violating Definition \[timehomogeneous\]. However, in order to satisfy the hypotheses of the Krylov-Bogolyubov theorem [@Hairer; @Ergodic] found in Theorem \[Krylov-Bogolyubov\], the hybrid system must be time-homogeneous.
To create a time-homogeneous system, we restrict the time set on which our Markov process is defined. Instead of allowing our time set $\{t, \tau_1, \tau_2, \tau_3, ... , \tau_n\} \subset \mathbb{R}^+$ to be any decreasing sequence of real numbers, we create time sets $t_0+nh$ for each $t_0\in [0,h)$ and $n\in\mathbb{N}$. In other words, we define a different time set for each value $t_0<h$ with
$ \{t\in\mathbb{R}^+ | t = t_0+nh$ for some $n\in\mathbb{N}\} $ .
We call the hybrid system on these multiple, restricted time sets the discrete system.
\[discretetomarkov\] The discrete hybrid system above is a time-homogeneous Markov process.
First, we must show that the discrete hybrid system is a Markov process at all. This follows immediately from the proof that our original hybrid system is a Markov process. Since the Markov property (Definition \[markovprocess\]) holds for all $t, \tau_1, \tau_2, ..., \tau_n \in \mathbb{R}^+$, it must necessarily hold for the specific time sets
$\{t\in\mathbb{R}^+ | \exists n\in\mathbb{N}$ such that $t = t_0+nh\}$
for each $t_0<h$.
Now, it remains to show that this system is time-homogeneous. Recall that the time-continuous hybrid system failed to be time-homogeneous because its $Z_t$ component was not time-homogeneous. Although transitions occurred only at regular, discrete time values, a test interval could be of any length; an interval of size $\frac{h}{2}$, for example, might contain either $0$ or $1$ transitions. However, because our discrete system creates separate time sets, any time interval - starting and ending within the same time set - must be of length $nh$ for some $n\in\mathbb{N}$, and thus will contain precisely $n$ potential transitions. So, taking $t_1,t_2\in\mathbb{R}^+$, we know $$P(Y_{t_1+nh} \in A | Y_{t_1} \in A_0) = P(Y_{t_2+nh} \in A | Y_{t_2} \in A_0) .$$ Note that the first component of the hybrid system, $x_t$, is also time-homogeneous under the discrete time system. Given $Z_t$, it can be treated as a deterministic system, and therefore time-homogeneous. Thus, the discrete hybrid system is time-homogeneous.
Invariant Measures for the Hybrid System
========================================
We now introduce several definitions that will lead to the main results of this paper.
\[markovinvm\] Consider a hybrid system $Y_t$ and a $\sigma$-algebra $\Sigma$ on the space $M$. A measure $\mu$ on $M$ is invariant if, for all sets $A\in\Sigma$ and all times $t\in\mathbb{R}^+$, $$\mu (A) = \int_{x_0 \in M} P(x_t\in A) \mu (dx) .$$
\[tight\] Let $(M, \mathcal{T})$ be a topological space, and let $\Sigma$ be a $\sigma$-algebra on $M$ that contains the topology $\mathcal{T}$. Let $\mathcal{M}$ be a collection of probability measures defined on $\Sigma$. The collection $\mathcal{M}$ is called tight if, for any $\epsilon > 0$, there is a compact subset $K_{\epsilon}$ of $M$ such that, for any measure $\mu$ in $\mathcal{M}$, $$\mu (M \backslash K_{\epsilon}) < \epsilon .$$
Note that, since $\mu$ is a probability measure, it is equivalent to say $\mu(K_{\epsilon}) > 1 - \epsilon$.
The following definitions are from [@Hairer].
\[Prokhorov\] Let $(M,\rho)$ be a separable metric space. Let $\{\mathcal{P}(M)\}$ denote the collection of all probability measures defined on $M$ (with its Borel $\sigma$-algebra). A collection $K \subset \{\mathcal{P}(M)\}$ of probability measures is tight if and only if $K$ is sequentially compact in the space equipped with the topology of weak convergence.
\[markovoperator\] Consider $M$ with $\sigma$-algebra $\Sigma$. Let $C^0(M,\mathbb{R})$ denote the set of continuous functions from $M$ to $\mathbb{R}$. The probability measure $\mathcal{P}(t,x,\cdot)$ on $\Sigma$ induces a map $$\mathcal{P}_t(x):C^0(M,\mathbb{R}) \rightarrow \mathbb{R}$$ with $$\mathcal{P}_t(x)(f) = \int_{y \in M}{f(y)\mathcal{P}(t,x,dy)} .$$ $\mathcal{P}_t$ is called a Markov operator.
\[Feller\] A Markov operator $\mathcal{P}$ is Feller if $\mathcal{P} \varphi$ is continuous for every continuous bounded function $\varphi: X \rightarrow R$. In other words, it is Feller if and only if the map $x \mapsto \mathcal{P}(x, \cdot)$ is continuous in the topology of weak convergence.
We state the Krylov-Bogolyubov Theorem without proof.
\[Krylov-Bogolyubov\] [**(Krylov-Bogolyubov)**]{} Let $\mathcal{P}$ be a Feller Markov operator over a complete and separable space X. Assume that there exists $x_0 \in X$ such that the sequence $\mathcal{P}^{n}(x_0, \cdot)$ is tight. Then, there exists at least one invariant probability measure for $\mathcal{P}$.
We now show that the conditions of the theorem are satisfied by the discrete hybrid system, yielding the existence of invariant measures as a corollary.
\[pFeller\] Given $t_0 \in [0,h)$, the discrete hybrid system Markov operators $\mathcal{P}_n$ for $n \in \mathbb{N}$ given by $$\mathcal{P}_nf(Y) \equiv \int_{M \times S}{f(Y_1)\mathcal{P}(nh,Y,dY_1)}$$ are Feller.
We begin by showing that $\mathcal{P}_1$ is Feller. By induction, it follows that $\mathcal{P}_n$ is Feller for all $n\in\mathbb{N}$. It is clear that there are only finitely many possible outcomes of running the hybrid system for time $h$. Namely, there are at most $|S|$ possible outcomes, where $|S|$ denotes the cardinality of $S$. Given $$Y_0 =
\left ( \begin{array} {c}
x_0\\
Z_0=s_i
\end{array} \right )\in M \times S ,$$ the only possible outcomes at time $t=1$ are $$Y_1^j = \left ( \begin{array} {c}
\varphi_j(t_0,\varphi_i(h-t_0, x)) \\
s_j
\end{array} \right )$$ for $j \in \{1,...,|S|\}$ where $\varphi_i, \varphi_j$ are the flows of the dynamical systems corresponding to states $s_i$ and $s_j$, respectively. The probability of the $j^{th}$ outcome is given by $ P_{i \rightarrow j}$, the probability of transitioning from state $s_i$ to state $s_j$. Therefore, $$\mathcal{P}_1f(Y) = \int_{M \times S}{f(Y_1)\mathcal{P}(h,Y,dY_1)} = \sum_{j = 1}^{|S|}{P_{i \rightarrow j}f(Y_1^j)}$$
Each $\varphi_i$ is continuous under the assumption that each flow is continuous with respect to its initial conditions. The map from $s_i$ to $s_j$ is continuous since $S$ is finite, so every set is open and hence the inverse image of any open set is open. The function $f$ is continuous by hypothesis, and any finite sum of continuous functions is also continuous. Therefore $\mathcal{P}_1f$ is also continuous, and hence $\mathcal{P}_1$ is Feller.
We see now that the conditions of the Krylov-Bogolyubov Theorem (\[Krylov-Bogolyubov\]) hold. Namely, because $M$ and $S$ are compact (the former by assumption, the latter since it is finite), $M \times S$ is compact. Thus, any collection of measures is automatically tight, since we can take $K_{\epsilon} = X$. It is well-known that any compact metric space is also complete and separable. Applying Theorem \[Krylov-Bogolyubov\], then, gives the following corollary, which is one of the primary results of the paper.
\[discretemeasure\] The discrete hybrid system has an invariant measure for each $t_0 \in [0, h)$.
So, rather than speaking of an invariant measure for the time-continuous hybrid system, we can instead imagine a periodic invariant measure cycling continuously through $h$. That is, for each time $t_0 \in [0,h)$, there exists a measure $\mu_{t_0}$ such that for $t \equiv 0$ (mod $h$), $$\mu_{t_0} (A) = \int_{Y \in M \times S} \mathcal{P}(t,Y,A) d\mu_{t_0} .$$ The measure $\mu_{t_0}$ above is a measure on the product space $M \times S$, since this is where the hybrid system lives. However, what we are really after is an invariant measure on just $M$, the space where the dynamical system part of the hybrid system lives. Fortunately, we can define a measure on $M$ by the following construction.
\[inducedmeasure\] Given $\mu_t$, an invariant probability measure on $M \times S$, the function $$\tilde{\mu}_t(A) \equiv \mu_t(A,S)$$ where $A \subseteq M$ is an invariant probability measure on $M$.
The fact that $\tilde{\mu}_t$ is a probability measure follows almost immediately from the fact that $\mu_t$ is a probability measure. The probability that $x_t \in \emptyset$ is 0, so $\tilde{\mu}_t(\emptyset) = 0$. The probability that $x_t \in M$ is $1$, so $\tilde{\mu}_t(M) = 1$. Countable additivity of $\tilde{\mu}_t$ follows from countable additivity of $\mu_t$. Therefore, $\tilde{\mu}_t$ is a probability measure on $M$.
Thus far, we have proven the existence of a measure $\mu_{t_0}$ for $t_0 \in [0,h)$ such that for $t \equiv 0$ (mod $h$), $$\mu_{t_0} (A) = \int_{x_0 \in M, s \in S} P(\varphi(t,x_0,s)\in A) d\mu_{t_0} .$$
The following theorem relates the collection of invariant measures $\{ \tilde{\mu}_{t_0} \}$ using the flow $\varphi$. This is the main result of the paper.
\[relatedmeasures\] Given invariant measure $\mu_0$, the measure $\mu_{t}$ defined by $$\mu_t(A) = \displaystyle \sum_{s \in S}{\int_{x_0 \in M}{P(\varphi(t,x_0,s)\in A)d\mu_0}}$$ is also invariant in the sense that $\mu_{t} = \mu_{t+nh}$ for $n \in \mathbb{N}$.
We will show that $\mu_t = \mu_{t+h}$. By induction, this implies that $\mu_t = \mu_{t+nh}$ for all $n \in \mathbb{N}$. We have $$\mu_{t+h}(A) = \sum_{s \in S}{\int_{x_0 \in M}{P(\varphi(t+h,x_0,s)\in A)d\mu_0}} .$$ Applying the definition of conditional probability, $$\sum_{s \in S}{\int_{x_0 \in M}{P(\varphi(t+h,x_0,s) \in A)d\mu_0}} =$$ $$\sum_{r \in S}{\int_{y \in M}{\left [P(\varphi(t,y,r) \in A) \sum_{s \in S}{\int_{x_0 \in M}{P(\varphi(h,x_0,s)\in dy \times \{r\})d\mu_0}}\right ]}} .$$ Loosely speaking, the probability that a trajectory beginning at $(x,s)$ will end in a set $A$ after a time $t+h$ is the product of the probability that a trajectory beginning at $(y,r)$ will end in $A$ after a time $t$ multiplied by the probability that a trajectory beginning at $(x,s)$ will end at $(y,r)$ after a time $h$, integrating over all possible pairs $(y,r)$. Here, we have implicitly used the fact that the hybrid system is a Markov process to ensure that the state of the system at time $t+h$ given the state at time $h$ is independent of the initial state, and we have avoided the problem of time-inhomogeneity by considering trajectories that only begin at times congruent to $0$ (mod $h$).
Furthermore, we have that $$\mu_h(dy \times \{r\}) = \sum_{s \in S}{\int_{x_0 \in M}{P(\varphi(h,x_0,s)\in dy \times \{r\})d\mu_0}}$$ and $$\mu_h(dy \times \{r\}) = d\mu_h(y,r);$$ so, $$\mu_{t+h} (A) = \sum_{r \in S}{\int_{y \in M}{P(\varphi(t,y,r)\in A)d\mu_h}} .$$ Since $\mu_0$ is invariant by assumption, $\mu_0 = \mu_h$. Therefore, $$\mu_{t+h} (A) = \sum_{r \in S}{\int_{y \in M}{P(\varphi(t,y,r)\in A)d\mu_0}} = \mu_{t} (A) .$$
Examples
========
Some examples of hybrid systems can be found in [@fitzhughnagurno; @climatemodel]. Here, we will examine two simple cases to illustrate the theory developed above.
A 1-D Hybrid System
-------------------
We begin with a 1-dimensional linear dynamical system with a stochastic perturbation: $$\dot{x} = -x + Z_t$$ where $Z_t \in \{-1,1\}$. Both components of this system have a single, attractive equilibrium point: for $Z_t=1$, this is $x=1$, and for $Z_t=-1$, $x=-1$. At timesteps of length $h=1$, $Z_t$ is perturbed by a Markov chain given by the transition matrix $Q$. $Q$ is therefore a $2 \times 2$ matrix of nonnegative entries, $$Q = \left ( \begin{array}{cc}
P_{1 \rightarrow 1} & P_{1 \rightarrow -1} \\
P_{-1 \rightarrow 1} & P_{-1 \rightarrow -1} \\
\end{array} \right ) ,$$ where $P_{i \rightarrow j}$ gives the probability of the equilibrium point transitioning from $i$ to $j$ at each integer time step. Since the total probability measure must equal $1$, $$\sum_{j}{P_{i \rightarrow j}} = 1, \, \, \, i, j \in \{1,-1\} .$$ Furthermore, to avoid the deterministic case, we take $P_{i \rightarrow j} \neq 0$ for all $i,j$.
\[C2\] The stochastic limit set $C(x_0) = [-1,1]$ for all $x_0\in\mathbb{R}$.
We begin by showing that $C(x)\subset[-1,1]$: that is, that every possible trajectory in our system will eventually enter and never leave $[-1,1]$, meaning that no it is only possible to have $t^*\rightarrow\infty$ such that $x^*=y$ for $y\in[-1,1]$. First, consider $x_0\in[-1,1]$. If we are in state $Z_t=1$, then the trajectory is attracted upwards and bounded above by $x=1$; in state $Z_t=-1$, the trajectory is attracted downwards and bounded below by $x=-1$. In both cases, the trajectory cannot move above $1$ or below $-1$, and so will remain in $[-1,1]$ for all time.
Now, consider $x_0\notin[-1,1]$. If the trajectory ever enters $[-1,1]$, by similar argument as above, it will remain in that region for all time. So, it remains to show that $\varphi(t,x_0,Z_0)\in[-1,1]$ for some $t\in\mathbb{R}$. First, take $x_0>1$. In either state, the trajectory will be attracted downward, and will eventually enter $[1,2]$ at time $t_2$. Once there, at the first timestep in which $Z_t=-1$ it will cross $x=1$ and enter $[-1,1]$. And since we have taken all entries of the transition probability matrix $Q$ to be nonzero, there almost surely exists a time $t_3>t_2$ for which the state is $Z_t=-1$; then, the trajectory will enter $[-1,1]$ and never leave. By similar argument, any trajectory starting at $x_0<-1$ will enter and never leave $[-1,1]$. Thus, $C(x)\subset[-1,1]$.
Now, we must show that $[-1,1]\in C(x)$: that is, that for every trajectory $\varphi(t,x_0,Z_0)$ and every point $y\in[-1,1]$ there is $t^*\rightarrow\infty$ such that $\varphi(t^*,x_0,Z_0)\rightarrow y$. To do this, we really only need to show that given any point $x_0 \in [-1,1]$ and any transition matrix $Q$, there almost surely exists some time $t^*$ with $\varphi(t^*,x_0,Z_0) = x^*$. If one such time $t^*$ is guaranteed to exist, then we can iterate the process for a solution beginning at $(t^*, x^*)$ to produce an infinite sequence of times. To show that $t^*$ exists, we calculate a lower bound on the probability that $\varphi(t_n,x_0,Z_0) = x^*$.
Without loss of generality, suppose that $x_0 > x^*$. We have already shown that any solution will enter $[-1,1]$, so take sup$(x_0) = 1$. From here, we can calculate the minimum number of necessary consecutive periods, $k$, for which $Z_n = -1$ in order for a solution with $x_0 = 1$ to decay to $x^*$. The probability of this sequence of $k$ consecutive periods occurring is given by $$P_{1}(k) = (P_{1 \rightarrow -1})(P_{-1 \rightarrow -1})^{k-1}$$ if $Z_0 = 1$ and $$P_{-1}(k) = (P_{-1 \rightarrow -1})^k$$ if $Z_0 = -1$. Thus, for some $t^* \in [0, k]$, $$P(\varphi(t^*,x_0,Z_0)=x^*) \geq \mbox{min}(P_{1}(k), P_{-1}(k)) > 0$$ since $P_{i \rightarrow j} > 0$. So, $$P(t^* \notin [0,k]) \leq 1 - P(x^*) < 1$$ and $$P(t^* \notin [0,mk]) \leq (1 - P(x^*))^m .$$ As $m \rightarrow \infty$, $(1 - P(x^*))^m \rightarrow 0$. So, with probability $1$, there exists $t^*$ with $\varphi(t^*,x_0,Z_0) = x^*$.
By similar argument, for $x_0<x^*$ and all $x^*\in(-1,1)$, we can find a time sequence $\{t_n\}$ such that $\varphi(t_n,x_0,Z_0) = x^*$. So, we know that for all $x^*\in (-1,1)$, $x^*\in C(x)$.
So, we have proven that $[-1,1]\subseteq C(x)$ and $(-1, 1) \subseteq C(x)$. Since $C(x)$ must by definition be closed, $C(x) = [-1,1]$.
We can study the behavior of this system numerically. Figure 1 (left) depicts a solution calculated for the transition matrix $$Q_1 = \left ( \begin{array}{cc}
.4 & .6 \\
.5 & .5 \\
\end{array} \right )$$ with initial values $x_0 = 2$, $Z_0 = 1$.
As expected, the trajectory enters the interval $(-1,1)$ and stays there for all time, oscillating between $x = -1$ and $x = 1$. Intuitively, it seems that the trajectory will cross any $x^*$ in this interval repeatedly, so that indeed $C(x) = [-1,1]$. This is not quite so clear for the transition matrix $$Q_2 = \left ( \begin{array}{cc}
.1 & .9 \\
.1 & .9 \\
\end{array} \right )$$ which yields the trajectory shown in Figure 1 (right) for $x_0 = 2$, $Z_0 = 1$.
Figure 1. A sample trajectory for a hybrid system with transition matrix $Q_1$ (left) and $Q_2$ (right).
It may appear that some set of points near $x = 1$ might be crossed by our path only a finite number of times. But, as proven above, any point in $(-1,1)$ will almost surely be reached infinitely many times as $t \rightarrow \infty$, so $C(x) = [-1,1]$.
Now, we consider the eigenvalues and eigenvectors of the transition matrices. The eigenvector of $Q_1^T$ with eigenvalue $1$ is $$\vec{v} = \left ( \begin{array}{cc}
\frac{5}{11} \\
\frac{6}{11} \\
\end{array}\right ) ,$$ and the eigenvector of $Q_2^T$ with eigenvalue $1$ is $$\vec{v}\,' = \left ( \begin{array}{cc}
\frac{9}{10} \\
\frac{1}{10} \\
\end{array}\right ) .$$
These eigenvectors give the invariant measures on the state space $S$. We know from Proposition \[inducedmeasure\] that there also exists an invariant measure on $M$. Here, since any trajectory in $M$ will almost surely enter $C(x) = [-1,1]$, the support of the invariant measure must be contained in $C(x)$. It is not difficult to see that this invariant measure cannot be constant for all $t \in \mathbb{R}^+$. Given any point $x_0 \in [-1,1]$, we know that at $t=1$, one of two things will have happened to the trajectory:
\(i) it will have decayed exponentially toward $x=1$, if $Z_1 = 1$, or
\(ii) it will have decayed exponentially toward $x = -1$, if $Z_1 = -1$.
In case (i), if a solution begins at $x_0=-1$ for $t=0$, then the solution will have decayed to a value of $1-(2 e^{-1})\approx 0.264$ by $t=1$ . In case (ii) a solution beginning at $x_0 = 1$ for $t=0$ will decay to a value of $-1+2 e^{-1}\approx -0.264$. Thus, if we are in case (i), all trajectories in $[-1,1]$ at $t = n$ will be located in $[0.264,1]$ at $t = n+1$. If we are in case (ii), all will be in $[-1, -0.264]$. It is not possible for any trajectory to be located in $[-0.264,0.264]$ at an integer time value. But, clearly, some solutions will cross into this region, as depicted in Figure 2. Therefore, no probability distribution will remain constant for all $t$ in the timeset $\mathbb{R}^+$.
{width="100mm"}
Figure 2. A spider plot showing all possible trajectories starting at $x_0=0$.
However, as Figure 2 suggests, there is some distribution that is invariant under $t \rightarrow t+n$ for $n \in \mathbb{N}$. Approximations of the invariant measures at $t \in [0,1]$ for transition matrix $Q_1$ are shown in Figure 3.
{width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"}
Figure 3. The invariant measure $\tilde{\mu}_{0}$ for a hybrid system with transition matrix $Q_1$.
A 2-D Hybrid System
-------------------
Our second example is a two-dimensional system used to model the kinetics of chemical reactors. The general system $f(x_1,x_2)$ is given by
$$\dot{x_1} = -\lambda x_1 - \beta (x_1 - x_c ) + BDa f(x_1,x_2)$$ $$\dot{x_2} = -\lambda x_2+Daf(x_1,x_2) .$$ [@chemflow]
Here, we use the following simplified application of the system:
$$\dot{x_1} = -x_1 - .15(x_1-1)+.35(1-x_2)e^{x_1}+Z_t(1-x_1)$$ $$\dot{x_2} = -x_2 + .05(1-x_2)e^{x_1} .$$ This system is used to describe a Continuous Stirred Tank Reactor (CSTR). This type of reactor is used to control chemical reactions that require a continuous flow of reactants and products and are easy to control the temperature with. They are also useful for reactions that require working with two phases of chemicals.
To understand the behavior of this system mathematically, we set our stochastic variable $Z_t=0$ and treat it as a deterministic system. This system has three fixed points, approximately at $(.67,.09)$, $(2.64,.41)$, and $(5.90,.95)$; the former and latter are attractor points, while the middle is a saddle point, as shown in Figure 4. The saddle point $(2.64,.41)$ creates a separatrix, a repelling equilibrium line between the two attracting fixed points. These points, $(.67,.09)$ and $(5.90,.95)$, comprise the $\omega$-limit set of our state space.
\
Figure 4. Phase plane of the deterministic system, $Z_n=0$.
With this information, we proceed to analyze the stochastic system. As discussed above, the random variable here is $Z_t$, which in applications can take values between $-.15$ and $.15$. To understand the full variability of this system, we take $$Z_t \in \{-.15, 0, .15\}$$ with the transition matrix $$\left ( \begin{array} {ccc}
.3&.3&.4 \\
.3&.3&.4 \\
.3&.3&.4
\end{array}\right ) ,$$ yielding the phase plane in Figure 5. We use red to indicate state 1 ($Z_t=-.15$), blue to indicate state 2 ($Z_t=0$), and green to indicate state 3 ($Z_t=.15$) for the corresponding fixed points, separatrices, and portions of trajectories.
{width="100mm"}\
Figure 5. Phase plane with randomness.
We see that, for $x_0$ away from the separatrices, $\varphi(t,x_0,Z_0)$ behaves similarly to $\varphi(t,x_0)$. Although state changes create some variability in a given trajectory, these paths move toward the groups of associated attracting fixed points, which define the stochastic limit sets for this system. However, $\varphi(t,x_0,Z_0)$ for $x_0$ between the red and green separatrices is unpredictable; depending on the sequence of state changes for a given trajectory, it might move either to the right or the left of the region defined by the separatrices. This area is the *bistable region*, because a trajectory beginning within it has two separate stochastic limit sets.
For example, we have in Figure 6 a spider plot beginning in the bistable region at $(3.5,0.75)$. A spider plot shows all possible trajectories starting from a single point in a hybrid system by, at each time step, taking every possible state. Our previous coloring scheme still applies.
{width="100mm"}\
Figure 6. Spider plot.
Thus, we see that the introduction of a stochastic element to a deterministic system can grossly affect the outcome of the system, as a trajectory can now cross any of the separatrices by being in a different state.
The stochastic element also affects the behavior of the hybrid system around the invariant region. In Figure 7, we show the path of a single trajectory in the invariant region defined by the fixed points near $(.67,0.9)$. Plotting this trajectory for a long period of time approximates the invariant region that would appear if we ran a spider plot from the same point, but much more clearly.
{width="100mm"}\
Figure 7. Random trajectory.
As we saw in the $1$-dimensional system, considering the counts taken at specific times in the interval between two state changes, $h=1$ (since our state transitions occur on $\mathbb{N}$), yields a periodic set of invariant measures. Similarly to Figure 3, Figure 8 shows the positions of our random trajectory in the invariant region at time $t$, mod $h$.
{width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"} {width="50mm"}\
Figure 8. Count of trajectory paths within one timestep.
A denser series of count images would show more clearly that the invariant measure at $t$ mod $h$ cycles continuously.
Conclusion
==========
We have studied hybrid systems consisting of a finite set $S$ of dynamical systems over a compact space $M$ with a Markov chain on $S$ acting at discrete time intervals. Such a hybrid system is a Markov process, which can be made time-homogeneous by discretizing the system. Then, there exists a family of invariant measures on the product space $M \times S$, which can be projected onto a family of measures on $M$. We have demonstrated a relation between the members of this family.
We have studied both a one-dimensional and a two-dimensional example of a hybrid system. These examples provide insight into the stochastic equivalent of $\omega$-limit sets and yield graphical representations of the invariant measures on these sets.
Acknowledgements
================
We wish to recognize Kimberly Ayers for her helpful discussions and Professor Wolfgang Kliemann for his instruction and guidance. We would like to thank the Department of Mathematics at Iowa State University for their hospitality during the completion of this work. In addition, we would like to thank Iowa State University, Alliance, and the National Science Foundation for their support of this research. Figure 4 was drawn using the ‘pplane8.m’ MATLAB program.
[9]{}
Ackerman, J., K. Ayers, E. J. Beltran, J. Bonet, and D. Lu. *A Behavioral Characterization of Discrete Time Dynamical Systems using Directed Graphs*. Iowa State University, Ames, 2011.
Ayala-Hoffmann, J., P. Corbin, K. McConville, F. Colonius, W. Kliemann, J. Peters. *Morse Decompositions, Attractors and Chain Recurrence*. Iowa State University, Ames, 2007.
Ayers, K.D. *Stochastic Perturbations of the Fitzhugh-Nagurno Equations*. Bowdoin College, Maine, 2010.
Baldwin, M.C. *Stochastic Analysis of Marotzke and Stone climate model*. Iowa State University, Ames, 2007.
Hairer, M. *Convergence of Markov Processes*. University of Warwick, Coventry, UK, 2010.
Hairer, M. *Ergodic Properties of Markov Processes*. University of Warwick, Coventry, UK, 2006.
Poore, A. B. *A Model Equation Arising from Chemical Reactor Theory*. Arch. Rational Mech. Anal. 52 (1974), pp 358-388.
Van Zanten, H. *An Introduction to Stochastic Processes in Continuous Time*. Eindhoven University of Technology, Centrum, NL, 2010.
[^1]: Department of Mathematics, University of Minnesota - Twin Cities, Minneapolis, MN 55455, USA (garci363@umn.edu). Research supported by DMS 0750986 and DMS 0502354
[^2]: Department of Mathematics, Saint Mary’s College of Maryland, St Marys City, MD 20686, USA ($\mbox{jckunze@smcm.edu}$). Research supported by DMS 0750986
[^3]: Department of Mathematics, Cornell University, Ithaca, NY 14850, USA (twr27@cornell.edu). Research supported by DMS 0750986
[^4]: Department of Mathematics, Arizona State University, Tempe, AZ 85281, USA ($\mbox{Anthony.Sanchez.1@asu.edu}$). Research supported by DMS 0750986 and DMS 0502354
[^5]: Department of Mathematics, Iowa State University, Ames, IA 50014, USA (sshao@iastate.edu). Research supported by Iowa State University
[^6]: Department of Mathematics, Carroll College, Helena, MT 59625, USA ($\mbox{esperanza@carroll.edu}$). Research supported by DMS 0750986
[^7]: Department of Mathematics, Iowa State University, Ames, IA 50014, USA ($\mbox{cvidden@iastate.edu}$). Research supported by Iowa State University.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
Identification of discrete outcome models is often established by using special covariates that have full support. This paper shows how these identification results can be extended to a large class of commonly used semiparametric discrete outcome models when all covariates are bounded. I apply the proposed methodology to multinomial choice models, bundles models, and finite games of complete information.
Keywords: Discrete outcome, multinomial choice, random coefficients, games of complete information
author:
- |
Nail Kashaev [^1]\
nkashaev@uwo.ca
bibliography:
- 'references.bib'
date: 'This version: November 2018'
title: 'Identification of semiparametric discrete outcome models with bounded covariates[^2]'
---
Introduction
============
Covariates that have full support (i.e., supported on the whole Euclidean space) provide an elegant and powerful approach for establishing identification in many discrete outcome models.[^3] Existence of such covariates is often necessary for nonparametric identification of distributions of latent variables.[^4] “Identification-at-infinity” arguments are based on existence of covariates with full support.[^5] However, finding such covariates in applied work is often problematic.
In this paper, I show that common parametric assumptions about the distribution of *some* unobservables (e.g. one normally distributed random coefficient in a model with multiple random coefficients) can fully restore the identification power of covariates with full support, even if covariates are in fact bounded. I provide two results connecting semiparametric discrete outcome models with covariates that have bounded support and (non)parametric discrete outcome models with special covariates that have full support. The first result is general and can be applied to a large class of semiparametric models. However, it requires a preliminary identification of a finite-dimensional parameter by using some auxiliary arguments. The second result does not require any extra identification steps, and uses one of the most popular parametrizations in applied work – the Gaussian distribution. I apply the proposed approaches to three well-known models: multinomial choice models with random coefficients, bundles models, and finite games of complete information.
The results of this paper rest on two commonly used assumptions. First, I assume existence of excluded (special) covariates that affect the distribution over outcomes via a latent index. Second, I impose commonly used parametric restrictions on the distribution of this index (e.g., the Gumbel or the Gaussian distribution). If the distribution of the index is sufficiently “rich”, then I show how to identify the distribution over outcomes conditional on the realization of the observed covariates and the latent index. Since the index distribution is usually assumed to have full support, I can treat the latent index as observed covariate with full support and apply *any* identification technique that requires existence of such covariates.
The index has different interpretation in different settings. For instance, in random coefficients model one of the random coefficients can be treated as the latent index. In games, the role of the index is played by a component of random utilities corresponding to different outcomes. “Richness” of the latent index distribution is formalized by a notion of bounded completeness.[^6]
This paper also contributes to the literature on partially identified models. I show that in many partially identified models the identified sets are “thin” in the following sense. The model parameters (including infinite-dimensional ones) are identified up to a finite-dimensional parameter of a lower dimension. This finding may lead to substantial computational gains in constructing confidence sets for partially identified parameters (e.g. @CTT) and sheds some light on the properties of identified sets in these models.
The paper is organized as follows. Section \[sec:binary example\] provides a motivating example. In Section \[sec:model\] I describe the setting and derive general identification results. Section \[sec: gaussian\] specializes the results from Section \[sec:model\] for widely used normally distributed latent variables. In Sections \[sec: applications\] I apply the result from Section \[sec:model\] to three different discrete outcome models. Section \[sec:conclusion\] concludes. All proofs can be found in Appendix \[app:proofs\].
Motivating Example: Binary Choice {#sec:binary example}
=================================
Consider a simple single agent binary choice problem.[^7] A utility maximizing agent has to choose ${y}\in{Y}=\{0,1\}$. The utility of alternative ${y}=0$ is normalized to $0$. The utility of option ${y}=1$ is $$\begin{aligned}
\label{eq:motivex 2}
{\mathbf{z}}_{2,1}{\mathbf{{v}}}_1+{\mathbf{g}}_1.\end{aligned}$$ Random variables ${\mathbf{v}}_1$ and ${\mathbf{g}}_1$ represent the random slope coefficient corresponding to covariate ${\mathbf{z}}_{2,1}\in{Z}_{2,1}\subseteq{{\mathds{R}}}$ and the random intercept, respectively.
The objective of the econometrician is to recover the c.d.f. of ${\mathbf{g}}_1$, $F_{{\mathbf{g}}_1}$, from observed distribution of choices ${\mathbf{{y}}}$ and covariates ${\mathbf{{z}}}$. Typically it is assumed that ${\mathbf{v}}_1=1\:{\ensuremath{\mathrm{a.s.}}}$, and ${\mathbf{z}}_{2,1}$ and ${\mathbf{g}}_1$ are independent. In this case the main identifying conditions is $$\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_{2,1}={z}_{2,1})=1-F_{{\mathbf{g}}_1}(-{z}_{2,1}),$$ for all ${z}_{2,1}\in{Z}_{2,1}$. Hence, if ${\mathbf{g}}_1$ is supported on ${{\mathds{R}}}$, then nonparametric identification of its distribution can be achieved if and only if ${\mathbf{z}}_{2,1}$ has full support. Next I will show how another commonly used assumption about ${\mathbf{v}}_1$ can lead to identification of $F_{{\mathbf{g}}_1}$ even if all observed covariates have bounded support.
Assume that there exists a covariate ${\mathbf{{z}}}_1\in{Z}_1\subseteq{{\mathds{R}}}$ such that for ${\mathbf{{z}}}=({\mathbf{{z}}}_1,{\mathbf{z}}_{2,1}){^{{\sf T}}}\in{Z}={Z}_1\times{Z}_{2,1}$ $${\mathbf{v}}_1|({\mathbf{z}}={z})\sim N(\beta_0+\beta_1{z}_1,1),\quad \forall {z}\in{Z},$$ where $\beta_1\neq 0$ and $N(\mu,\sigma^2)$ denotes the normal distribution with mean $\mu$ and variance $\sigma^2$. For this section only I assume that $\beta=(\beta_0,\beta_1){^{{\sf T}}}$ is known or can be identified. Later on I will show how one can pointidentify $\beta$. Assuming that ${\mathbf{g}}_1$ is independent of ${\mathbf{{z}}}$ and ${\mathbf{v}}_1$, we can rewrite (\[eq:motivex 2\]) as $$\begin{aligned}
{\mathbf{z}}_{2,1}(\beta_0+\beta_1{\mathbf{z}}_1+{\mathbf{{e}}}_1)+{\mathbf{g}}_1,\end{aligned}$$ where ${\mathbf{{e}}}_1$ is a standard normal random variable that is independent of ${\mathbf{{z}}}$.[^8]
The key step is to show that we can identify $$\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}={z},{\mathbf{{v}}}_1={v}_1)$$ for all $v_1\in{{\mathds{R}}}$ and ${z}\in{Z}$. If we can identify $\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}=\cdot,{\mathbf{{v}}}_1=\cdot)$, then we can analyze the model as if ${\mathbf{{v}}}_1$ is observed. In this case ${\mathbf{z}}_{2,1}{\mathbf{{v}}}_1$ becomes an observed special covariate with *full* support since ${\mathbf{v}}_1$ is supported on ${{\mathds{R}}}$. By using variation in ${\mathbf{{v}}}_1$ we can recover $F_{{\mathbf{g}}_1}$ even if the support of ${\mathbf{{z}}}$ is bounded. In other words, the above binary choice model with just one normally distributed random slope coefficient with bounded covariates, in terms of identification features, is equivalent to the binary choice model where the utility from choosing ${y}=1$ is equal to $$\label{eq:motiv22}
{\mathbf{r}}_{1}+{\mathbf{g}}_1,$$ where ${\mathbf{r}}_1$ and ${\mathbf{g}}_1$ are independent, and ${\mathbf{r}}_1$ is observed covariate supported on $$\{r_1\in{{\mathds{R}}}\::\:r_1=v_1{z}_{2,1}, v_1\in{{\mathds{R}}}, {z}_{2,1}\in{Z}_{2,1}\}={{\mathds{R}}}.$$
It is left to show that $\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}={z},{\mathbf{{v}}}_1=\cdot)$ can be identified. First, note that under independence condition $$\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_1={z}_1,{\mathbf{{z}}}_{2,1}={z}_{2,1},{\mathbf{{v}}}_1={v}_1)=\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_{2,1}={z}_{2,1},{\mathbf{{v}}}_1={v}_1).$$ Second, note that for any fixed ${z}_{2,1}$ we have the following integral equation $$\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_1={z}_1,{\mathbf{{z}}}_{2,1}={z}_{2,1})=\int_{{{\mathds{R}}}}{\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_{2,1}={z}_{2,1},{\mathbf{{v}}}_1={v})\phi(v-(\beta_0+\beta_1{z}_1))dv},$$ for all ${z}_1\in{Z}_1$. Since variation in $z_1$ does not affect $\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_{2,1}={z}_{2,1},{\mathbf{{v}}}_1=\cdot)$, we can use this variation to identify it. In other words, if the family of normal distributions $\{\phi(\cdot-\beta_0-\beta_1{z}_1)\::\:z_1\in{Z}_1\}$ is sufficiently “rich”, then the integral equation has a unique solution. For discrete distributions the “richness” condition is usually characterized by the rank condition. For continuous distributions the “richness” condition is associated with a notion of “completeness”. In our example, the “completeness” condition is satisfied since the family of $\{\phi(\cdot-\beta_0-\beta_1{z}_1)\::\:z_1\in{Z}_1\}$ is complete if $\operatorname*{int}(Z_1)\neq\emptyset$.[^9] Since the choice of ${z}_{2,1}$ was arbitrary, we recover $\Pr({\mathbf{{y}}}=1|{\mathbf{{z}}}_{2,1}=\cdot,{\mathbf{{v}}}_1=\cdot)$ and thus can work with well-known model (\[eq:motiv22\]).
Two important assumptions needed for the result to hold are: (i) existence of a parametric index ${\mathbf{v}}_1$ such that the choices are affected by some excluded covariates only through the index; and (ii) the distribution of the index conditional on excluded covariates is sufficiently rich (complete). The rest of the paper generalizes these key assumptions to environments with multiple agents and outcomes, and establishes identification of $\beta$.
General Model {#sec:model}
=============
Each instance of the environment is characterized by an endogenous outcome ${\mathbf{{y}}}$ from a known finite set ${Y}$, a vector of observed exogenous characteristics ${\mathbf{{x}}}\in{X}\subseteq{{\mathds{R}}}^{d_{x}}$, $d_{x}<\infty$, that can be partitioned into ${x}=({z}{^{{\sf T}}},{w}{^{{\sf T}}}){^{{\sf T}}}$, and a vector of unobserved structural variables ${\mathbf{{v}}}\in{V}\subseteq{{\mathds{R}}}^{d_v}$.[^10] It is assumed that the econometrician observes the joint distribution of $({\mathbf{{y}}},{\mathbf{{x}}}{^{{\sf T}}}){^{{\sf T}}}$.
\[ass:exclusion restriction general\] There exist ${Y}^*\subseteq{Y}$ and ${h}_0:{Y}^*\times{W}\times{V}\to[0,1]$, such that $$\begin{aligned}
\Pr({\mathbf{{y}}}={y}|{\mathbf{{z}}}={z},{\mathbf{{w}}}={w},{\mathbf{{v}}}={v})={h}_0({y},{w},{v}),
\end{aligned}$$ for all ${y}\in{Y}^*$, ${x}=({z}{^{{\sf T}}},{w}{^{{\sf T}}}){^{{\sf T}}}\in{X}$, and ${v}\in{V}$.
Assumption \[ass:exclusion restriction general\] is an exclusion restriction that requires covariates ${\mathbf{{z}}}$ to affect distribution over some outcomes only via the distribution of the latent ${\mathbf{{v}}}$. Note that the exclusion restriction does not need to be imposed on all outcomes. For instance, in single agent decision models one can identify the payoff parameters by observing only the probability of choosing the outside option (e.g., @thompson1989identification and @lewbel2000semiparametric). Assumption \[ass:exclusion restriction general\] does not rule out existence of other latent variables (different from ${\mathbf{v}}$) since exclusion restrictions are imposed on the distribution over outcomes conditional on ${\mathbf{{x}}}={x}$ and ${\mathbf{{v}}}={v}$.[^11]
The next assumption is a parametric restriction on the latent variable whose distribution is affected by the excluded covariates ${\mathbf{{z}}}$.
\[ass:bounded completeness general\] For every ${w}\in{W}$, there exists ${Z}'\subseteq{Z}_{w}$ such that the family of distributions $\left\{F_{{\mathbf{{v}}}|{\mathbf{{z}}},{\mathbf{{w}}}}(\cdot|{z},{w}),{z}\in{Z}'\right\}$ is boundedly complete. That is, $$\forall{z}\in{Z}',\:\int_{V}g(t)dF_{{\mathbf{{v}}}|{\mathbf{{z}}},{\mathbf{{w}}}}(t|{z},{w})=0\implies g({\mathbf{{v}}})=0 \:{\ensuremath{\mathrm{a.s.}}}.$$
Completeness assumptions have been widely used in econometric analysis. Completeness is typically imposed on the distribution of observables (e.g. @newey03). However, many commonly used parametric restrictions on the distribution of unobservables imply Assumption \[ass:bounded completeness general\]. For instance, it is satisfied for the Gaussian distribution and the Gumbel distribution. [^12] $$\mu({y}|{x})=\Pr({\mathbf{{y}}}={y}|{\mathbf{{x}}}={x})$$ is known for every ${y}\in{Y}^*$ and ${x}\in{X}$.
\[prop:identification of h up to v\] Under Assumptions \[ass:exclusion restriction general\] and \[ass:bounded completeness general\], ${h}_0$ is identified from $\mu$ up to $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$.
Proposition \[prop:identification of h up to v\] implies that under exclusion restrictions if one assumes that the latent variable has a known distribution belonging to a boundedly complete family, then one can work with the model as if the realizations of latent variables are observed in the data since we can identify $h_0({y},{w},\cdot)=\Pr({\mathbf{{y}}}={y}|{\mathbf{{w}}}={w},{\mathbf{{v}}}=\cdot)$. Thus, if I know or can identify $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ (see Section \[sec: gaussian\]), for identification I can interpret *latent* variables (${\mathbf{{v}}}$) as *observed* covariates. If these latent variables have full support (e.g. normal errors), then all identification techniques that require existence of covariates with full support can be applied (e.g., @fox2016nonparametric in the context of random coefficients model and @BHR in the context of games). In other words, I can transform a model with covariates that have bounded support into the model with covariates that have full support, and then use existing methods to identify different objects of interest.
The following example demonstrates how knowing ${h}_0$ can help to identify some underlying aspects of the model.
Suppose that ${Y}={Y}^*=\{0,1\}$ and ${h}_0$ is identified.[^13] Assume that $\gamma({\mathbf{{w}}})+{\mathbf{v}}$ represents the utility agent gets from choosing ${y}=1$, where $\gamma$ is some unknown function of ${w}$. Assume, moreover, that the utility from choosing ${y}=0$ is normalized to zero. The identified ${h}_0$ is consistent with the utility maximizing behavior if and only if $${h}_0(1,{w},{v})={\mathds{1}\left(\,\gamma({w})+v\geq 0\,\right)}$$ for almost all ${v}\in{V}$ and ${w}\in{W}$. Thus, if ${V}={{\mathds{R}}}$, then one can test for utility maximizing behavior, and can identify $\gamma(\cdot)$ if the agent maximizes utility.
Proposition \[prop:identification of h up to v\] establishes identification of $h_0$ only up to $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$. Typically $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ is known up to a finite-dimensional parameter. In Section \[sec: gaussian\] I show how one can identify this parameter and thus $h_0$ if $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ is assumed to be the Gaussian distribution. In Section \[subsec:binary games of comp info\] I show how Proposition \[prop:identification of h up to v\] can be used to characterize the identified set of a partially identified game of complete information.
Gaussian Distribution {#sec: gaussian}
=====================
Proposition \[prop:identification of h up to v\] can be applied with any known parametric distribution $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ as long as the family of the distributions generated by the variation in excluded covariates is complete. The most prominent example of such families is the exponential family of distributions. In this section I specialize the results from the previous section to probably one of the most common parametrization in applied work – Gaussian errors. I show how to identify $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ for such models. The following assumption is sufficient for Assumption \[ass:bounded completeness general\] to hold.
\[ass:normal errors\]
1. The latent ${\mathbf{{v}}}$ satisfies $${\mathbf{{v}}}_i={\mathbf{{z}}}_{2,i}[\beta_{0,i}({\mathbf{{w}}})+\beta_{1,i}({\mathbf{{w}}}){\mathbf{z}}_{1,i}+{\mathbf{{e}}}_i]\quad{\ensuremath{\mathrm{a.s.}}}$$ where $\beta_{0,i}(\cdot)$ and $\beta_{1,i}(\cdot)$, $i=1,\dots,d_{{v}}$, are some unknown measurable functions such that $\beta_{1,i}({w})\neq 0$ for all ${w}\in{W}$;
2. $\{{\mathbf{{e}}}_i\}_{i=1,\dots,d_{v}}$ are independent identically distributed standard normal random variables that are independent of ${\mathbf{{x}}}$;
3. For every ${w}\in{W}$ the support of ${\mathbf{{z}}}|({\mathbf{{w}}}={w})$, ${Z}_{{w}}$, contains an open ball;
4. The sign of either $\beta_{0,i}({w})$ or $\beta_{1,i}({w})$ is known for every ${w}\in{W}$ and $i$.
Assumption \[ass:normal errors\](i) is motivated by random coefficient models. The covariate ${\mathbf{{z}}}_{2,i}$ can be interpreted as choice (“product”) specific characteristic. The random coefficients $[\beta_{0,i}({\mathbf{{w}}})+\beta_{1,i}({\mathbf{{w}}}){\mathbf{z}}_{1,i}+{\mathbf{{e}}}_i]$ captures agent specific heterogeneity in tastes. The only support restriction is imposed on ${\mathbf{z}}$ (Assumption \[ass:normal errors\](iii)). Non of the covariates are assumed to have full or unbounded support. Assumptions \[ass:normal errors\](i)-(iii) are sufficient for Assumption \[ass:bounded completeness general\] since the family of normal distributions indexed by the mean parameter is complete as long as the parameter space for the mean parameter contains an open ball.[^14]
Assumptions \[ass:normal errors\](iv) is a normalization. It requires that either the sign of the marginal effect of ${z}_{1,i}$ or the sign of the intercept (as long at it is not equal to zero) are known (or can be identified). In discrete outcome models with almost surely unique equilibrium (e.g. multinomial choice) the sign of $\beta_{1,i}(\cdot)$ can often be identified because of monotonicity of ${h}$ in utility indexes $v$. For instance, in multinomial choice models the probability of choosing an outside option is decreasing in mean utilities of other choices. Without additional restrictions the sign can not be identified as the following example demonstrates.
Suppose that Assumptions \[ass:exclusion restriction general\] and \[ass:normal errors\](i)-(iii) are satisfied with $d_v=1$ and $x=(z_{11},z_{21}){^{{\sf T}}}$. Take ${y}\in{Y}^*$ and note that $$\begin{aligned}
\mu({y}|{x})&=\int_{-\infty}^{+\infty}{h}_0(y,v_1)\phi(v_1/z_{2,1}-\beta_{0,1}-\beta_{1,1}z_{1,1})dv_1=\\
&=\int_{-\infty}^{+\infty}{h}_0(y,v_1)\phi(-v_1/z_{2,1}+\beta_{0,1}+\beta_{1,1}z_{1,1})dv_1=\\
&=\int_{-\infty}^{+\infty}{h}_0(y,-v_1)\phi(v_1/z_{2,1}+\beta_{0,1}+\beta_{1,1}z_{1,1})dv_1=\\
&=\int_{-\infty}^{+\infty}{h}_0(y,-v_1)\phi(v_1/z_{2,1}-(-\beta_{0,1})-(-\beta_{1,1})z_{1,1})dv_1,\end{aligned}$$ where the second equality follows from the symmetry of $\phi(\cdot)$ and the third equality follows from the change of variables. Hence, if ${h}_0$ and $(\beta_{0,1}, \beta_{1,1}){^{{\sf T}}}$ can generate the data, then $\tilde{h}_0$ and $(-\beta_{0,1}, -\beta_{1,1}){^{{\sf T}}}$ such that $\tilde{h}_0({y},{v})={h}_0({y},-{v})$ for all $y$ and ${v}$ can generate the data too.
Proposition \[prop:identification of h up to v\] only identifies ${h}_0$ up to the distribution of unobservables. In particular, under Assumption \[ass:normal errors\], ${h}_0$ is identified up to $\{\beta_{0,i}(\cdot),\beta_{1,i}(\cdot)\}_{i=1}^{d_{v}}$. The following assumption allows us to identify the distribution of unobservables and thus ${h}_0$.
Let $z_{1,-i}=({z}_{1,k})_{k\neq i}$. For a fixed ${y}^*\in{Y}^*$, ${z}_{1,-i}$ and ${z}_{2}$, let $\eta:{Z}_{1,i|{w},{z}_{1,-i},{z}_2}\to[0,1]$ be such that for ${x}=(({z}_{1,i},{z}_{1,-i}){^{{\sf T}}},{z}_2{^{{\sf T}}},{w}{^{{\sf T}}}){^{{\sf T}}}$ $$\eta({z}_{1,i})=\mu({y}^*|{x}).$$
\[ass:support restriction\] For every ${w}\in{W}$ and $i=1,2,\dots,d_{{y}}$, there exists ${y}^*\in{Y}^*$ and $z_{2,i}\in{Z}_{2,i|{w}}\setminus\{0\}$ such that $\eta(\cdot)$ is neither an exponential nor an affine function of $z_{1,i}$.
Assumption \[ass:support restriction\] means that if we fix all covariates but one, then the probability of observing one excluded outcome conditional on covariates is neither affine nor exponential function of the non-fixed covariate. Assumption \[ass:support restriction\] is not very restrictive since it rules out only some exponential and linear probability models. Moreover, it is testable.
\[prop:identification of beta\] Suppose that Assumptions \[ass:exclusion restriction general\], \[ass:normal errors\], and \[ass:support restriction\] hold. Then ${h}_0$ and $\{\beta_{0,i}(\cdot),\beta_{1,i}(\cdot)\}_{i=1}^{d_{v}}$ are identified.
Proposition \[prop:identification of beta\] establishes identification of ${h}_0$ and $F_{{\mathbf{{v}}}|{\mathbf{{x}}}}$ for normally distributed latent variables. Next I will show how one can use ${h}_0$ in different discrete outcome models.
Applications {#sec: applications}
============
Multinomial Choice {#subsec: multinomial}
------------------
Consider the following random coefficients model. The agent has to choose between $J$ inside goods and an outside option of no purchase. That is, $y\in{Y}=\{0,1,\dots,J\}$. I normalize the utility from alternative ${y}=0$ to $0$. The random utility from choosing an alternative ${y}\neq 0$ is of the form $$\begin{aligned}
\label{eq:random coeff parametrization}
&{\mathbf{u}}_{y}={\mathbf{{z}}}_{2,{y}}\big[\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\big]+{\mathbf{g}}_{y}.\end{aligned}$$ The random coefficient $\big[\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\big]$ represents individual specific heterogeneous tastes associated with product characteristic ${\mathbf{z}}_{2,{y}}$. The latent random vector ${\mathbf{g}}=({\mathbf{g}}_{{y}})_{{y}\in{Y}\setminus\{0\}}$ captures other sources of unobserved heterogeneity which are different from $\big[\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\big]$. The observed covariates are ${\mathbf{{x}}}=({\mathbf{{z}}}_1, {\mathbf{z}}_{2}{^{{\sf T}}},{\mathbf{{w}}}{^{{\sf T}}}){^{{\sf T}}}$, where ${\mathbf{{z}}}_2=({\mathbf{{z}}}_{2,{y}})_{{y}\in{Y}\setminus\{0\}}$. Assume that the agents are utility maximizers.[^15]
\[ass:normal error random coef\]
1. ${\mathbf{{e}}}_1$ is an independent of $({\mathbf{g}}{^{{\sf T}}},{\mathbf{{x}}}{^{{\sf T}}}){^{{\sf T}}}$ standard normal random variable;
2. $\beta_1({\mathbf{{w}}})\neq 0\quad {\ensuremath{\mathrm{a.s.}}}$;
3. For every ${w}\in{W}$ the support of ${\mathbf{{z}}}|{\mathbf{{w}}}={w}$, ${Z}_{{w}}$, contains an open ball.
Similarly to the existing treatment of random coefficients model, I assume that the random coefficients in front of ${\mathbf{{z}}}_{2,{y}}$ are the same for each alternative ${y}$.[^16] However, I do not impose sign restrictions on $\big[\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\big]$. Note that $$\Pr(\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\geq 0|{\mathbf{{x}}}={x})=\Phi(\beta_0({w})+\beta_{1}({w}){z}_1),$$ where $\Phi(\cdot)$ is the standard normal c.d.f. Thus, since there are no restrictions on $\beta_0(\cdot)$, the random coefficient $\big[\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\big]$ can be positive (negative) with probability that is arbitrarily close to $1$.
\[ass:random coefficients support and exclusion\]
1. Random shocks ${\mathbf{g}}$ are conditionally independent of ${\mathbf{{z}}}$ conditional on ${\mathbf{{w}}}={w}$. That is, for all ${x}=({z}{^{{\sf T}}},{w}{^{{\sf T}}}){^{{\sf T}}}\in{X}$ $$F_{{\mathbf{g}}|{\mathbf{{z}}},{\mathbf{{w}}}}(\cdot|{z},{w})=F_{{\mathbf{g}}|{\mathbf{{w}}}}(\cdot|{w}),$$ where $F_{{\mathbf{g}}|{\mathbf{{z}}},{\mathbf{{w}}}}(\cdot|{z},{w})$ and $F_{{\mathbf{g}}|{\mathbf{{w}}}}(\cdot|{w})$ are conditional c.d.fs of ${\mathbf{g}}|{\mathbf{{z}}}={z},{\mathbf{{w}}}={w}$ and ${\mathbf{g}}|{\mathbf{{w}}}={w}$, respectively;
2. For every ${w}\in{W}$, there exists $z_2\neq 0$ such that it is contained in ${Z}_{2|{w}}$ with some open neighborhood and ${z}_{2,{y}}={z}_{2,{y}'}$ for any ${y},{y}'\in{Y}\setminus\{0\}$;
3. For every ${w}\in{W}$ there exists ${z}_2\in{Z}_{2|{w}}$ such that $$\Pr({\mathbf{{y}}}=0|{\mathbf{z}}_1=\cdot,{\mathbf{z}}_2={z}_2,{\mathbf{{w}}}={w})$$ is neither exponential nor affine function.
Assumption \[ass:random coefficients support and exclusion\](i) together with utility maximizing behavior guarantees that Assumption \[ass:exclusion restriction general\] is satisfied. Assumption \[ass:random coefficients support and exclusion\](i) is the only restriction on ${\mathbf{g}}$. I allow ${\mathbf{g}}$ to be discrete or continuous random variable with unknown support. Assumptions \[ass:random coefficients support and exclusion\](ii)-(iii) are testable restrictions. Assumptions \[ass:random coefficients support and exclusion\](ii) imply that one can find ${z}_{2,1}={z}_{2,2}=\dots={z}_{2,J}\neq 0$ such that Assumption \[ass:bounded completeness general\] is satisfied for $v={z}_{2,1}(\beta_0({w})+\beta_{1}({w}){z}_1+{e}_1)$. Assumption \[ass:random coefficients support and exclusion\](iii) is sufficient for Assumption \[ass:support restriction\] to hold.
\[prop:random coefficients\] Suppose Assumptions \[ass:normal error random coef\] and \[ass:random coefficients support and exclusion\] hold. Then
1. $\beta_0(\cdot)$ and $\beta_1(\cdot)$ are identified;
2. The above model inherits all identifying properties of the following random coefficients model: $$\begin{aligned}
&{\mathbf{u}}_{{y}}={\mathbf{r}}_{{y}}+{\mathbf{g}}_{{y}}, \quad {y}\neq 0,\\
&{\mathbf{u}}_{y}=0,\quad y=0,\end{aligned}$$ where ${\mathbf{r}}=({\mathbf{r}}_{{y}})_{{y}\in{Y}\setminus\{0\}}$ is an observed covariate independent of ${\mathbf{g}}=({\mathbf{g}}_{y})_{{y}\in{Y}\setminus\{0\}}$ conditional on ${\mathbf{{w}}}$ with the conditional support $$R_w=\left\{r\in{{\mathds{R}}}^J\::\:r=\lambda {z}_2,\: \lambda\in{{\mathds{R}}},\:{z}_2\in{Z}_{2|{w}}\right\}.$$
Proposition \[prop:random coefficients\] implies that the original random coefficient model can be represented in the “special-covariate-with-full-support” framework without assuming existence of such covariates. Moreover, if the set of directions that ${z}_2/{\left\lVertz_2\right\rVert}$ can cover is sufficiently rich, then $R_{{w}}={{\mathds{R}}}^{J}$ and all the identification results that require existence of special covariates with full support (e.g., @lewbel2000semiparametric, @berry2009nonparametric, and @fox2016nonparametric) can be applied. For instance, if $\{z_2/{\left\lVertz_2\right\rVert}\::\:z_2\in Z_{2|w}\}$ is equal to a unit sphere in ${{\mathds{R}}}^{J}$ for every ${w}$, then $R_{{w}}={{\mathds{R}}}^{J}$ and I can nonparametrically identify $F_{{\mathbf{g}}|{\mathbf{{w}}}}$.[^17]
Bundles
-------
Consider the following bundles model motivated by @gentzkow2007valuing, @dunker2017nonparametric, and @fox2017note. There are $J$ goods and the agent can purchase any bundle consisting of these goods. The vector ${y}$ describes the purchasing decision of the agent. That is, $y\in{Y}=\{0,1\}^{J}$. For instance, $y=(0,1,0,1,0,\dots,0){^{{\sf T}}}$ corresponds to the case when the agent purchased a bundle of goods $2$ and $4$. I normalize the utility from “not buying”, ${y}=0$, to $0$. The random utility from choosing an alternative ${y}\neq 0$ is of the form $$\begin{aligned}
\label{eq:bundle parametrization}
{\mathbf{u}}_{y}=(\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1)\sum_{j=1}^J{y}_j{\mathbf{{z}}}_{2,j}+{\mathbf{g}}_{y}.\end{aligned}$$
Although the model (\[eq:bundle parametrization\]) looks similar to the model (\[eq:random coeff parametrization\]), there is one important difference: there is no bundle specific covariate since ${z}_{2,j}$ affects not only the utility from buying good $j$ alone, but also every bundle that includes it.
\[prop:bundle\] Suppose Assumptions \[ass:normal error random coef\] and \[ass:random coefficients support and exclusion\] hold. Then
1. $\beta_0(\cdot)$ and $\beta_1(\cdot)$ are identified;
2. The above model inherits all identifying properties of the following bundles model: $$\begin{aligned}
&{\mathbf{u}}_{y}=\sum_{j=1}^J {y}_j {\mathbf{r}}_{j}+{\mathbf{g}}_{y}\quad,\\
&{\mathbf{u}}_0=0.\end{aligned}$$ where ${\mathbf{r}}=({\mathbf{r}}_{j})_{j=1,\dots,J}$ is an observed covariate independent of ${\mathbf{g}}=({\mathbf{g}}_{y})_{{y}\in{Y}\setminus\{0\}}$ conditional on ${\mathbf{{w}}}$ with the conditional support $$R_w=\left\{r\in{{\mathds{R}}}^J:r=\lambda {z}_2, \lambda\in{{\mathds{R}}},\:{z}_2\in{Z}_{2|{w}}\right\}.$$
Note that for $J=2$ if one assumes that $$\begin{aligned}
{\mathbf{g}}_{(1,0)}&=f_1({\mathbf{{w}}})+{\boldsymbol{\epsilon}}_1,\\
{\mathbf{g}}_{(0,1)}&=f_2({\mathbf{{w}}})+{\boldsymbol{\epsilon}}_2,\\
{\mathbf{g}}_{(1,1)}&={\mathbf{g}}_{(1,0)}+{\mathbf{g}}_{(0,1)}+{\mathbf{\xi}}f_3({\mathbf{{w}}}),\end{aligned}$$ where $f_i(\cdot)$, $i=1,2,3$, are some unknown functions, and $({\boldsymbol{\epsilon}}_1,{\boldsymbol{\epsilon}}_2,{\boldsymbol{\xi}}){^{{\sf T}}}\in{{\mathds{R}}}^2\times{{\mathds{R}}}_{+}$, then model (\[eq:random coeff parametrization\]) is equivalent to the one in @fox2017note, and one can apply their Theorem 1 to identify $f_i(\cdot)$, $i=1,2,3$, and the distributions of ${\boldsymbol{\epsilon}}_i|{\mathbf{{w}}}$, $i=1,2$, and ${\boldsymbol{\xi}}|{\mathbf{{w}}}$.
Binary games of complete information {#subsec:binary games of comp info}
------------------------------------
In the multinomial choice and the bundles models I am able to establish identification of the objects of interest without requiring covariates with full support. The example that is considered in this section is different: the model is not pointidentified. However, the Lebesgue measure of the identified set is zero. In particular, all parameters of the model are identified up to a finite-dimensional parameter of lower-dimension.
There are ${\left\lVert{{I}}\right\rVert}<\infty$ players indexed by $i\in{{I}}$. Every player must choose ${y}_i\in\{0,1\}$. Thus, the outcome space is ${Y}=\{0,1\}^{{\left\lVert{{I}}\right\rVert}}$.[^18] Players $i$’s payoff from choosing action ${y}_i$ when the other agents are choosing ${y}_{-i}$ is given by $$\begin{aligned}
{\mathbf{\pi}}_{0i}({y}) = \left[\alpha_{0,i}({\mathbf{{w}}})+[\beta_{0,i}({\mathbf{{w}}}){\mathbf{{z}}}_i+{\mathbf{{e}}}_i] + \sum_{j\in{{I}}\setminus\{i\}}\delta_{0,i,j}({\mathbf{{w}}}){y}_{j}\right]{y}_i,\end{aligned}$$ where ${\mathbf{{e}}}_{i}$, ${{i}}\in{{I}}$, are observed by players but unobserved by the econometrician shocks; $\alpha_{0,i}(\cdot)$, $\beta_{0,i}(\cdot)$ and $\delta_{0,i,j}(\cdot)$ are unknown functions. The econometrician observes a joint distribution of $({\mathbf{{y}}},{\mathbf{{x}}}{^{{\sf T}}}){^{{\sf T}}}$, where ${\mathbf{{x}}}=({\mathbf{{z}}}{^{{\sf T}}},{\mathbf{{w}}}{^{{\sf T}}}){^{{\sf T}}}\in{X}$ with ${\mathbf{{z}}}=({\mathbf{{z}}}_i)_{i\in{{I}}}$, is a vector of observed covariates. Let $\beta_0(\cdot)=(\beta_{0,i}(\cdot))_{i\in{{I}}}$, $\alpha_0(\cdot)=(\alpha_{0,i}(\cdot))_{i\in{{I}}}$, and $\delta_{0}(\cdot)=(\delta_{0,i,j}(\cdot))_{i\neq j\in{{I}}}$.
The following two assumptions are sufficient for Assumptions \[ass:exclusion restriction general\] and \[ass:bounded completeness general\].
\[ass:binary game exclusion and support\]
1. Assumption \[ass:exclusion restriction general\] is satisfied with ${\mathbf{{v}}}=(\beta_{0,i}({\mathbf{{w}}}){\mathbf{{z}}}_i+{\mathbf{{e}}}_i)_{i\in{{I}}}$;
2. For every $i,j\in{{I}}$, $i\neq j$, the cardinality of $$\Big\{\big(0\big)_{k\in{{I}}},\:\big({\mathds{1}\left(\,k\in\{i,j\}\,\right)}\big)_{k\in{{I}}},\:\big({\mathds{1}\left(\,k=i\,\right)}\big)_{k\in{{I}}},\:\big({\mathds{1}\left(\,k=j\,\right)}\big)_{k\in{{I}}}\Big\}\bigcap{Y}^*$$ is at least $2$.
Assumption \[ass:binary game exclusion and support\](i) implies that excluded covariates affect the distribution of some outcomes via payoffs only. Assumption \[ass:binary game exclusion and support\](ii) imposes restrictions on the set of those outcomes. If one thinks of an entry game where ${y}_i=1$ corresponds to the entry decision, the outcomes in Assumption \[ass:binary game exclusion and support\](ii) have the following interpretation. The outcome $\big(0\big)_{k\in{{I}}}$ corresponds to the market where nobody enters. The outcome $\big({\mathds{1}\left(\,k\in\{i,j\}\,\right)}\big)_{k\in{{I}}}$ corresponds to the market where only players $i$ and $j$ enter. Similarly, $\big({\mathds{1}\left(\,k=i\,\right)}\big)_{k\in{{I}}}$ means that only players $i$ enters. Note that although the cardinality of ${Y}$ is $2^{{\left\lVertI\right\rVert}}$ , the cardinality of ${Y}^*$ can be as low as ${\left\lVertI\right\rVert}+1$.
\[ass:binary game normality\]
1. The shocks $\{{\mathbf{{e}}}_i\}_{i\in{{I}}}$ are i.i.d. standard normal random variables and independent of ${\mathbf{{x}}}$;
2. $\beta_{0,i}({\mathbf{{w}}})\neq 0\:{\ensuremath{\mathrm{a.s.}}}$ for all $i\in{{I}}$;
3. For every ${w}\in{W}$, the support of ${\mathbf{{z}}}|{\mathbf{{w}}}={w}$, ${Z}_{{w}}$, contains an open ball.
Assumption \[ass:binary game normality\] is a parametric restriction on the distribution of payoffs. It is satisfied by the parametrization used in @BHR and @ciliberto09.
I consider three complete information equilibrium/solution concepts: (i) minimax play, (ii) “collusive” behavior, and (iii) rationalizability. Under minimax solution concept every player picks the action that maximizes her minimum possible payoff. Under “collusive” behavior players play the outcome that maximizes the sum of individual profits. Under rationalizability players play any action that survives iterated strict dominance elimination. The most commonly used solution concept of Nash equilibrium is nested within rationalizability. Hence, all identification results that are derived under rationalizability are valid for Nash equilibrium (both in pure and mixed strategies).
These three solution concepts have different implications for identification. For instance, in the two player case, if players display “collusive” behavior, then there is no hope of separate identification of “competition” effects $\delta_{0,i,j}$. At best, one can identify $\delta_{0,1,2}+\delta_{0,2,1}$. Fortunately, it is possible to distinguish between these three models up to $\beta_0(\cdot)$ since under Assumptions \[ass:binary game exclusion and support\] and \[ass:binary game normality\] I am able to nonparametrically recover ${h}_0$ up $\beta_0(\cdot)$.
(-4,-8)(10,6) (3,-7)[(a) minimax]{} (3,1)(8,5) (-2,-3)(3,1) (-2,1)(8,1) (3,-3)(3,5) [ ]{} (-4,1) (3,-4) (0.5,-1) (6,-1) (0,3) (5.5,3) [ ]{}
(-5,-8)(13,6) (3,-7)[(b) “collusion”]{} (6,2)(11,5) (-3,-3)(2,0) (-3,0)(2,0)(6,2)(6,5) (2,-3)(2,0)(6,2)(11,2) [ ]{} (-3,2)(6,2) (6,-3)(6,2) (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (8.5,3.5) (1,4) (7.5,-2) (-0.5,-1.5) [ ]{}
(-5,-8)(11,6) (3,-7)[(c) rationalizability]{} (2,0)(6,2) (6,2)(11,5) (-3,-3)(2,0) (-3,0)(6,0)(6,5) (2,-3)(2,2)(11,2) [ ]{} (-3,2)(6,2) (6,-3)(6,2) (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (8.5,3.5) (1,4) (7.5,-2) (-0.5,-1.5) (4,1)[[mixture]{}]{} [ ]{}
Figure \[fig:2player,all,main\] illustrates predictions in a two player binary game for different realizations of $v$ when $\delta_{0,1,2}<0$ and $\delta_{0,2,1}<0$. The thresholds $a_i,b_i,c_i$, $i=1,2$, are determined by $\alpha_{0,i}$ and $\delta_{0,i,j}$, $i,j=1,2$. If one can identify ${h}_0$ for two outcomes, say $(0,0)$ and $(1,1)$, then one can distinguish between all three equilibrium concepts as long as the following three conditions are satisfied. $$\begin{aligned}
(a_1,a_2){^{{\sf T}}}&\neq (b_1,b_2){^{{\sf T}}}\quad(\text{minimax vs. collusion})\label{sep:mmvscol},\\
(\tilde a_1,\tilde a_2){^{{\sf T}}}&\neq (\tilde b_1,\tilde b_2){^{{\sf T}}}\quad(\text{minimax vs. rationalizability})\label{sep:mmvsrat},\\
\dfrac{\tilde b_2-\tilde a_2}{\tilde b_1-\tilde a_1}&\neq \dfrac{b_2-a_2}{b_1-a_1}\quad(\text{rationalizability vs. collusion})\label{sep:colvsrat}.\end{aligned}$$ Conditions (\[sep:mmvscol\])-(\[sep:colvsrat\]) hold if $\delta_{0,1,2}^2({w})\neq \delta_{0,2,1}^2({w})$ for some ${w}\in{W}$. Thus, if I can identify the correct solution concept, then I can recover the threshold values. For instance, if I know that agents play rationalizable strategies, then I can identify $\tilde a_i, \tilde b_i$, $i=1,2$. Since $\tilde a_i=-\alpha_{0,i}$, and $\tilde b_i=-\alpha_{0,i}-\delta_{0,i,j}$, $i,j=1,2$, I also can identify $\alpha_0$ and $\delta_{0}$.
The above intuition generalizes to games with more than two players, more than two actions, and without any sign restrictions on $\delta_{0,i,j}$. The following result establishes identification in binary games.
\[prop:binary game\] Under Assumptions \[ass:binary game exclusion and support\] and \[ass:binary game normality\]
1. If there exist $i,j\in{{I}}$, $i\neq j$, such that $\Pr(\delta_{0,i,j}^2({\mathbf{{w}}})\neq\delta_{0,j,i}^2({\mathbf{{w}}}))>0$, then one can determine whether players behave according to minimax play, “collusive” behavior, or rationalizability up to $\beta_{0}(\cdot)$;
2. If players behave according to rationalizability, then $\alpha_0(\cdot)$ and $\delta_{0}(\cdot)$ are identified up to $\beta_{0}(\cdot)$;
3. If players behave according to “collusive” behavior, then $\alpha_0(\cdot)$ and $\left\{\delta_{0,i,j}(\cdot)+\delta_{0,j,i}(\cdot)\right\}_{i\neq j\in{{I}}}$ are identified up to $\beta_{0}(\cdot)$;
4. If players behave according to minimax play, then $\left\{\alpha_{0,i}(\cdot)+\min\{\delta_{0,i,j}(\cdot),0\}\right\}_{i\neq j\in{{I}}}$ are identified up to $\beta_{0}(\cdot)$.
Proposition \[prop:binary game\] states that if marginal effects of some excluded covariates on payoffs are known, then one can identify the solution concept, together with some (or sometimes all) payoff parameters. Assumption \[ass:binary game normality\](i) implies that the errors in payoffs are not correlated. I can allow for unknown variance covariance matrix $V$. In this case all objects in Proposition \[prop:binary game\] are identified up to $\beta_0(\cdot)$ and $V$.
Full identification in this binary game can be achieved if one can identify $\beta_0(\cdot)$. The standard identification-at-infinity argument requires existence of player-action-specific covariates that have full support (unbounded from above *and* below). However, the full support assumption is only needed to separately identify the intercepts of the mean utilities ($\alpha_{0}$ and $\delta_{0}$). In contrast, in order to identify $\beta_{0}$ one only needs to have player-action-specific covariates with unbounded support (e.g., unbounded from above only).[^19] In other words, full identification can be achieved in a substantially bigger set of applications (e.g., prices or income can potentially take arbitrary large positive values, but cannot be negative).[^20]
Conclusion {#sec:conclusion}
==========
This paper shows that commonly used exclusion restrictions and parametric assumptions about the distribution of some unobservables may lead to identification in discrete outcome models. The proposed identification framework allows one to extend the results from a large literature that uses special covariates with full support to environments where such full-support covariates are not available.
The partial identification result can substantially decrease computational complexity of constructing confidence sets for partially identified parameters. For instance, the likelihood ratio statistic of @CTT is asymptotically $\chi^2$ distributed after profiling $\beta_0$ under the null hypothesis, since the model in this case is identified. Thus, there is no need to use bootstrap and one can take critical values from $\chi^2$ distribution.
Proofs {#app:proofs}
======
Proof of Proposition \[prop:identification of h up to v\]
---------------------------------------------------------
Fix some ${y}\in{Y}^*$ and ${w}\in{W}$ (for brevity I will drop ${w}$ in the notation below). Under Assumption \[ass:exclusion restriction general\] I have the following integral equation $$\forall {z}\in{Z}\::\:\mu({y}|{z})=\int_{V}{h}({y}^*,v)dF_{v|z}(v|z).$$ Suppose that there exists ${h}$ with ${h}({y}^*,v)\neq{h}_0({y}^*,v)$ for all $v$ in some nonzero measure set $V'$ such that $$\forall {z}\in{Z}\::\:\mu({y}|{z})=\int_{V}{h}({y}^*,v)dF_{v|z}(v|z)=\int_{V}{h}_0({y}^*,v)dF_{v|{z}}(v|{z}).$$ Which implies that the nonzero function ${h}({y},\cdot)-{h}_0({y},\cdot)$ integrates to $0$ for all $z\in{Z}'$. The latter contradicts to Assumption \[ass:bounded completeness general\]. The fact that the choice of ${y}$ and ${w}$ was arbitrary completes the proof.
Proof of Proposition \[prop:identification of beta\]
----------------------------------------------------
Note that Proposition \[prop:identification of h up to v\] implies that ${h}_0$ is identified up to $\{\beta_{0,i}(\cdot),\beta_{1,i}(\cdot)\}_{i=1}^{d_{v}}$. Hence, I only need to show that $\{\beta_{0,i}(\cdot),\beta_{1,i}(\cdot)\}_{i=1}^{d_{v}}$ is identified.
Fix some ${w}$, $i\in\{1,2,\dots,d_{v}\}$, $z_{2,-i}$, $z_{1,-i}$, and take ${y}^*$ from Assumption \[ass:support restriction\]. To simplify notation let $F_0:{{\mathds{R}}}\to{{\mathds{R}}}$ and $\eta:{{\mathds{R}}}^2\to{{\mathds{R}}}$ such that $$F_0(v_i)=\int_{{{\mathds{R}}}^{d_v-1}}{h}_0(y^*,{w},v)\prod_{k\neq i}\dfrac{\phi\left(v_k/z_{2,k}-\beta_{0,k}({w})-\beta_{1,k}({w})z_{1,k}\right)}{z_{2,k}}dv_{k},$$ where $\phi(\cdot)$ is the standard normal p.d.f., and $\eta(z_{1,i},z_{2,i})=\mu({y}^*|{z},{w})$.
Assumptions \[ass:exclusion restriction general\] and \[ass:normal errors\] imply that $$\eta(z_{1,i},z_{2,i})=\int_{{{\mathds{R}}}} F_0(v_i)\dfrac{\phi(v_i/z_{2,i}-\beta_{0,i}-\beta_{1,i}z_{1,i})}{z_{2,i}}dv_i,$$ After dropping index $i$ from the notation and some rearrangements I get $$\label{eq:rc 1}
\tilde\eta(z_{1},z_{2})=\int_{{{\mathds{R}}}} F_0(t)\phi(t/z_{2}-\beta_{0}-\beta_{1}z_{1})dt,$$ where $\tilde\eta(z_{1},z_{2})=z_2\eta(z_{1},z_{2})$.
Next, note that since $\phi''(x)=-\phi(x)-x\phi'(x)$ the following system of equations holds $$\begin{aligned}
\partial_{z_1}\tilde\eta(z_{1},z_{2})&=-\beta_1\int F_0(t)\phi'(t/z_2-\beta_0-\beta_1z_1)dt,\\
\partial^2_{z_1^2}\tilde\eta(z_{1},z_{2})&=\beta^2_1\int F_0(t)\phi''(t/z_2-\beta_0-\beta_1z_1)dt
=-\beta^2_1\tilde\eta(z_{1},z_{2})-\beta_1(\beta_0+\beta_1z_1)\partial_{z_1}\tilde\eta(z_1,z_2)-\\
&-\beta_1^2\int tF_0(t)\phi'(t/z_2-\beta_0-\beta_1z_1)dt/z_2;\end{aligned}$$ Moreover, $$\begin{aligned}
\partial_{z_2}\tilde\eta(z_{1},z_{2})&=-\int F_0(t)t\phi'(t/z_2-\beta_0-\beta_1z_1)dt/z_2^2.\end{aligned}$$ Hence, $$\begin{aligned}
\partial^2_{z_1^2}\tilde\eta(z_{1},z_{2})&=-\beta^2_1\tilde\eta(z_{1},z_{2})-\beta_1(\beta_0+\beta_1z_1)\partial_{z_1}\tilde\eta(z_1,z_2)+\beta_1^2z_2\partial_{z_2}\tilde\eta(z_{1},z_{2});\end{aligned}$$ Equivalently $$\begin{aligned}
\dfrac{\beta_0}{\beta_1}=\dfrac{z_2\partial_{z_2}\tilde\eta(z_{1},z_{2})-\tilde\eta(z_{1},z_{2})}{\partial_{z_1}\tilde\eta(z_{1},z_{2})}-z_1 -\dfrac{\partial^2_{z^2}\tilde\eta(z_{1},z_{2})}{\partial_{z_1}\tilde\eta(z_{1},z_{2})}\dfrac{1}{\beta_1^2}. \end{aligned}$$ Thus, $\beta_0/\beta_1$ is identified up to $\beta^2_1$. Moreover, the last equality implies that $\beta_1^2$ is identified if for some $z_2$ and $z_1$ $$\partial_{z_1}\left(\dfrac{\partial_{z_1^2}\tilde\eta(z_{1},z_{2})}{\partial_{z_1}\tilde\eta(z_{1},z_{2})}\right)\neq0.$$ Suppose this is not the case. That is, for all $z_2$ and $z_1$ $$\partial^2_{z_1^2}\left(\log(\partial_{z_1}\tilde\eta(z_{1},z_{2}))\right)=0.$$ The latter would imply that either $$\tilde\eta(z_{1},z_{2})=K_1(z_2)e^{K_2(z_2)z_1}+K_3(z_2)$$ or $$\tilde\eta(c,z)=K_4(z_2)z_1+K_3(z_2)$$ for some functions $K_i(\cdot)$, $i=1,2,3,4$. Since it is assumed that $\tilde\eta(\cdot,z_2)=z_2\eta(\cdot,{z}_2)$ is neither exponential nor affine function, I can conclude that $\beta^2_1$ is identified (hence, ${\left\lvert\beta_1\right\rvert}$ is also identified). Hence, I identify $\beta_0/\beta_1$. If $\beta_0/\beta_1=0$, then the sign of $\beta_1$ is identified from Assumption \[ass:normal errors\](iv). If $\beta_0/\beta_1\neq 0$, then the sign of either $\beta_1$ or $\beta_0$ is identified from Assumption \[ass:normal errors\](iv). Knowing the sign of, say, $\beta_0$ and $\beta_0/\beta_1$ identifies $\beta_1$ and $\beta_0$.
Proof of Propositions \[prop:random coefficients\] and \[prop:bundle\]
----------------------------------------------------------------------
(i). Note that under Assumption \[ass:random coefficients support and exclusion\].(ii) there exists $z_2$ with some open neighbourhood such that $z_{2,y}=z_{2,{y}'}$ for all ${y},{y}'\in{Y}$. Let $${\mathbf{v}}_1=-{\mathbf{z}}_{2,1}(\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1)\:{\ensuremath{\mathrm{a.s.}}}.$$ Assumption \[ass:normal error random coef\] implies that Assumptions \[ass:normal errors\](i)-(iii) are satisfied for the above ${\mathbf{v}}_1$. Moreover, $$\Pr({\mathbf{y}}=0|{\mathbf{x}}={x})=\int_{{{\mathds{R}}}}F_{{\mathbf{g}}|{\mathbf{{w}}}}(-z_{2,1}(\beta_0({w})+\beta_{1}({w}){z}_1+{e}_1),\dots,-z_{2,1}(\beta_0({w})+\beta_{1}({w}){z}_1+{e}_1)|{w})\phi(e_1)de_1$$ identifies the sign of $\beta_{1}({w})$ since $F_{{\mathbf{g}}|{\mathbf{{w}}}}(\cdot|{w})$ is weakly monotone. Thus, Assumption \[ass:normal errors\](iv) is also satisfied.
Assumption \[ass:exclusion restriction general\] is satisfied for ${Y}^*=\{0\}$ and for $h(0,w,v)=F_{{\mathbf{g}}|{\mathbf{{w}}}}(v,v,\cdot,v|{w})$. Assumption \[ass:support restriction\] is implied by Assumption \[ass:random coefficients support and exclusion\]. Hence, by Proposition \[prop:identification of h up to v\] $\beta_0(\cdot)$ and $\beta_{1}(\cdot)$ are identified.
(ii). Since $\beta_0(\cdot)$ and $\beta_{1}(\cdot)$ are identified I can redefine the index $v$. Let $${\mathbf{v}}=\beta_0({\mathbf{{w}}})+\beta_{1}({\mathbf{{w}}}){\mathbf{{z}}}_1+{\mathbf{{e}}}_1\:{\ensuremath{\mathrm{a.s.}}}.$$ Note that $$\begin{aligned}
\Pr({\mathbf{y}}=0|{\mathbf{x}}={x})&=\int_{{{\mathds{R}}}}F_{{\mathbf{g}}|{\mathbf{{w}}}}(-z_{2,1}v,\dots,-z_{2,J}v|{w})\phi(v-\beta_0({w})+\beta_{1}({w}){z}_1)dv=\\
&=\int_{{{\mathds{R}}}}h(z_2,w,v)\phi(v-\beta_0({w})+\beta_{1}({w}){z}_2)dv\end{aligned}$$ Since assumption of Proposition \[prop:identification of h up to v\] are satisfied I pointidentify $$h(z_2,w,v)=F_{{\mathbf{g}}|{\mathbf{{w}}}}(-z_{2,1}v,\dots,-z_{2,Y}v|{w})$$ for all $z_2,w,v$. Note that since $v$ can take any value in ${{\mathds{R}}}$ for any direction $-z_2/{\left\lVertz_2\right\rVert}$ in the support of $z_2$ I can recover $F_{{\mathbf{g}}|{\mathbf{{w}}}}(g|w)$ for any $g$ such that $g=-z_2v/{\left\lVertz_2\right\rVert}$ for some $v\in{{\mathds{R}}}$. That is, I identify $F_{{\mathbf{g}}|{\mathbf{{w}}}}(\cdot|{w})$ over the set $$R_w=\left\{r\in{{\mathds{R}}}^J\::\:r=\lambda {z}_2,\: \lambda\in{{\mathds{R}}},\:{z}_2\in{Z}_{2|{w}}\right\}.$$
Proof of Proposition \[prop:binary game\]
-----------------------------------------
First, I fix some ${w}\in{W}$ and for notation simplicity I drop dependence on ${w}$. Since assumptions of Proposition \[prop:identification of h up to v\] are satisfied, I identify $$\Pr({\mathbf{{y}}}={y}|{\mathbf{v}}=\cdot)={h}_0({y},\cdot)$$ for all ${y}\in{Y}^*$. Since $v\in{{\mathds{R}}}^{{\left\lVertI\right\rVert}}$ and $v_i$ enters only payoffs of player $i$, I can make a payoff of any player arbitrary small (“close” to $-\infty$). Hence, under all solution concepts under consideration I can force any player to choose ${y}_i=0$. Take any two players $i\neq j$ and consider $${h}_{0,i,j}({y},v_i,v_j)=\lim_{v_k\to-\infty,k\in{{I}}\setminus\{i,j\}}{h}_0({y},v)$$ for all ${y}\in{Y}^*$ and $v_{i},v_j$. Note that ${h}_{0,i,j}$ corresponds to a two player binary game with payoffs $$\left[\alpha_{0,i} + \delta_{0,i,j}{y}_{j} + v_i\right]{y}_i$$ and $$\left[\alpha_{0,j} + \delta_{0,j,i}{y}_{i} + v_j\right]{y}_j.$$ Moreover, by Assumption \[ass:binary game exclusion and support\](ii), in this two player game at least two outcomes from $$\{(0,0),(1,0),(0,1),(1,1)\}$$ satisfy exclusion restrictions. Assume that $(0,0)$ and $(1,1)$ satisfy the exclusion restriction (the proof for any other case, e.g., $(0,0)$ and $(1,0)$, is almost the same).
To discriminate between solution concepts I analyze their predictions about outcomes $(0,0)$ and $(1,1)$. Without loss of generality let $i=1$ and $j=2$.
Case 1. Suppose that agents are behaving according to minimax play. Then for any $\alpha_{0,1}$, $\alpha_{0,2}$, $\delta_{0,1,2}$, and $\delta_{0,2,1}$, the predictions about outcomes $(0,0)$ and $(1,1)$ depending on the value of $(v_1,v_2){^{{\sf T}}}\in{{\mathds{R}}}^2$ can be depicted as in Figure \[fig:2player,mm\], where $c_1=-\alpha_{0,1}-\min\{\delta_{0,1,2},0\}$ and $c_2=-\alpha_{0,2}-\min\{\delta_{0,2,1},0\}$
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (3,1)(8,5) (-2,-3)(3,1) (-2,1)(8,1) (3,-3)(3,5) [ ]{} (-4,1) (3,-4) (0.5,-1) (6,-1) (0,3) (5.5,3) [ ]{}
Case 2. Suppose that agents are behaving according to “collusive” solution concept. Then for any $\alpha_{0,1}$, $\alpha_{0,2}$, $\delta_{0,1,2}$, and $\delta_{0,2,1}$, the predictions about outcomes $(0,0)$ and $(1,1)$ depending on the value of $v\in{{\mathds{R}}}^2$ can be depicted as in figures \[fig:2player,col1\]-\[fig:2player,col2\], where $a_1=-\alpha_{0,1}$, $a_2=-\alpha_{0,2}$, $b_1=-\alpha_{0,1}-(\delta_{0,1,2}+\delta_{0,2,1})$, and $b_2=-\alpha_{0,2}-(\delta_{0,1,2}+\delta_{0,2,1})$.
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (6,2)(11,5) (-3,-3)(2,0) (-3,0)(2,0)(6,2)(6,5) (2,-3)(2,0)(6,2)(11,2) [ ]{} (-3,2)(6,2) (6,-3)(6,2) (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (8.5,3.5) (1,4) (7.5,-2) (-0.5,-1.5) [ ]{}
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (2,5)(2,2)(6,0)(11,0)(11,5) (-3,2)(2,2)(6,0)(6,-3)(-3,-3) (-3,2)(2,2)(6,0)(6,-3) (2,5)(2,2)(6,0)(11,0) [ ]{} (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (6,2) (-1,4) (9,-2) (1,0) [ ]{}
Case 3. Suppose that agents are playing rationalizable strategies. Then for any $\alpha_{0,1}$, $\alpha_{0,2}$, $\delta_{0,1,2}$, and $\delta_{0,2,1}$, the predictions about outcomes $(0,0)$ and $(1,1)$ depending on the value of $v\in{{\mathds{R}}}^2$ can be depicted as in figures \[fig:2player,ltr1\]-\[fig:2player,ltr4\], where $a_1=-\alpha_{0,1}$, $a_2=-\alpha_{0,2}$, $b_1=-\alpha_{0,1}-\delta_{0,1,2}$, and $b_2=-\alpha_{0,2}-\delta_{0,2,1}$.
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (2,0)(6,2) (6,2)(11,5) (-3,-3)(2,0) (-3,0)(6,0)(6,5) (2,-3)(2,2)(11,2) [ ]{} (-3,2)(6,2) (6,-3)(6,2) (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (8.5,3.5) (1,4) (7.5,-2) (-0.5,-1.5) (4,1)[[multiple]{}]{} [ ]{}
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (2,0)(6,2) (2,5)(2,2)(6,2)(6,0)(11,0)(11,5) (-3,2)(2,2)(2,0)(6,0)(6,-3)(-3,-3) (-3,2)(2,2)(2,0)(6,0)(6,-3) (2,5)(2,2)(6,2)(6,0)(11,0) [ ]{} (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (7,3) (-1,4) (9,-2) (0,-1) [ ]{}
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (2,0)(6,2) (2,5)(2,2)(6,2)(11,2)(11,5) (-3,0)(2,0)(6,0)(6,-3)(-3,-3) (-3,0)(6,0)(6,-3) (2,5)(2,2)(11,2) [ ]{} (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (7,3) (-1,4) (9,-2) (0,-1) [ ]{}
(-9,-6.5)(13,6) (-8.5,3)(-8.5,5) (-8.5,5.4)[${v}_2$]{} (-9,3.5)(-7,3.5) (-6.8,3.5)[${v}_1$]{} [ ]{} (2,0)(6,2) (6,5)(6,0)(11,0)(11,5) (-3,2)(2,2)(2,-3)(-3,-3) (-3,2)(2,2)(2,-3) (6,5)(6,0)(11,0) [ ]{} (-3.5,2) (-3.5,0) (6,-3.5) (1.5,-3.5) (8,3) (-1,4) (9,-2) (0,-1) [ ]{}
As a result, minimax is consistent with the data if and only if ${h}_{0,i,j}((0,0),\cdot)$ and ${h}_{0,i,j}((1,1),\cdot)$ match Figure \[fig:2player,mm\]. If ${h}_{0,i,j}((0,0),\cdot)$ and ${h}_{0,i,j}((1,1),\cdot)$ match Figure \[fig:2player,ltr3\] or Figure \[fig:2player,ltr4\], then only rationalizability can explain the data. There is still a possibility that ${h}_{0,i,j}((0,0),\cdot)$ and ${h}_{0,i,j}((1,1),\cdot)$ match Figure \[fig:2player,col1\] and Figure \[fig:2player,ltr1\], or Figure \[fig:2player,col2\] and Figure \[fig:2player,ltr2\]. Hence, one might think that rationalizability and “collusive” behavior both can explain the data. However, since in Proposition \[prop:binary game\](i) I assume that there exist two players such that $\delta_{0,i,j}^2\neq\delta_{0,j,i}^2$, the “multiplicity” region under rationalizability is never a square (the lengths of the region are $\|\delta_{0,1,2}\|$ and $\|\delta_{0,2,1}\|$). In contrast, under “collusive” behavior the “multiplicity” region is always a square (the lengths of the region are $\|\delta_{0,1,2}+\delta_{0,2,1}\|$ and $\|\delta_{0,1,2}+\delta_{0,2,1}\|$). Thus by analyzing these two “asymmetric” players I can determine the correct solution concept.
If I determined that the correct solution concept is minimax, then at most I can identify $\alpha_{0,1}+\min\{\delta_{0,1},0\}$ and $\alpha_{0,2}+\min\{\delta_{0,2},0\}$. If the correct solution concept is “collusive” behavior, then I can pointidentify $\alpha_{0,1}$, $\alpha_{0,2}$, and $\delta_{0,1,2}+\delta_{0,2,2}$. If agents play rationalizable strategies, then I can pointidentify all payoff parameters.
The result then follows from the fact that the choice of players $i,j$ and ${w}\in{W}$ was arbitrary.
[^1]: Department of Economics, University of Western Ontario.
[^2]: I would like to thank Roy Allen, Victor H. Aguiar, Tim Conley, Salvador Navarro, David Rivers, and Bruno Salcedo for useful comments and discussions.
[^3]: See, for example, @manski1985semiparametric, @matzkin1992nonparametric, @ichimura1998maximum, @lewbel1998semiparametric, @lewbel2000semiparametric, @tamer03, @matzkin2007heterogeneous, @berry2009nonparametric, @BHR, @gautier2013nonparametric, @fox2016nonparametric, @dunker2017nonparametric, @fox2017note, and @fox2018unobserved.
[^4]: Full support assumption is not necessary if it is assumed that latent variables have bounded support. In this case the support of the special covariates have to be large enough to cover the support of unobservables.
[^5]: For example, @manski1988identification, @heckman1990varieties, and @tamer03.
[^6]: Completeness of a family distribution is a well-known concept both in the Statistics and Econometrics literature. See for example @mattner1993some, @newey03, @chernozhukov2005iv, @blundell07, @chernozhukov2007instrumental, @hu2008instrumental, @andrews11, @darolles11, and @d2011completeness.
[^7]: Throughout the paper, deterministic vectors and functions are denoted by lower-case regular font Latin letters (e.g., ${x}$), random objects by bold letters (e.g., ${\mathbf{x}}$). Capital letters are used to denote supports of random variables (e.g., ${\mathbf{{x}}}\in{X}$). I denote the support of a conditional distribution of ${\mathbf{x}}$ conditional on ${\mathbf{z}}=z$ by $X_z$. Also, given a family $x = (x_k)_{k\in K}$ and a particular index value $k\in K$, I use the notation $x_{-k}$ for $(x_j)_{j\in K\setminus\{k\}}$. $F_{{\mathbf{{x}}}}(\cdot)$ ($f_{{\mathbf{{x}}}}(\cdot)$) and $F_{{\mathbf{{x}}}|{\mathbf{{z}}}}(\cdot|{z})$ ($f_{{\mathbf{{x}}}|{\mathbf{{z}}}}(\cdot|{z})$) denote the c.d.f. (p.d.f.) of ${\mathbf{{x}}}$ and ${\mathbf{{x}}}|{\mathbf{{z}}}=z$, respectively.
[^8]: Assumption that the variance of ${\mathbf{{e}}}_1$ is $1$ is a scale normalization.
[^9]: $\operatorname*{int}(Z_1)$ denotes the interior of $Z_1$.
[^10]: My analysis allows for countable sets of outcomes, but for the exposition purposes I focus on finite sets.
[^11]: I consider a model with unobserved heterogeneity that is not fully captured by ${\mathbf{v}}$ in Section \[subsec: multinomial\].
[^12]: For testability of the completeness assumptions see @canay13.
[^13]: For binary outcome models assuming that both outcomes satisfy the exclusion restriction is equivalent to assuming that just one of them does.
[^14]: Any distribution from the exponential family of distributions (e.g., the Gumbel distribution) would be sufficient for Assumption \[ass:bounded completeness general\] to hold.
[^15]: The results below hold even if utility maximizing behavior is not assumed as long as Assumption \[ass:exclusion restriction general\] is satisfied for at least one outcome.
[^16]: The model extends to the case when the random slope coeffcient is choice specific as in Assumption \[ass:normal errors\]. But in this case one would need to find more excluded covariates.
[^17]: In general, I can identify $F_{{\mathbf{g}}|{\mathbf{{w}}}}(\cdot|{w})$ over $R_{{w}}$ only.
[^18]: I work with binary action spaces for ease of exposition. The result can be extended to multiple players and actions games.
[^19]: See, for instance, @tamer03, @BHR, and @kashaev17.
[^20]: One may argue that it is always possible to transform any covariate into a covariate with full support. For instance, one can always treat logarithm of income or price as a covariate. Thus, when price goes to $0$ the logarithm of price goes to $-\infty$. However, in order to interpret linear parametrization of a payoff function in this case, one would need to explain why prices that are close to zero would lead to extremely negative (or positive) profits.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We provide upper bounds for the size of subsets of finite fields lacking the polynomial progression $$x, x+y, ..., x+(m-1)y, x+y^m, ..., x+y^{m+k-1}.$$ These are the first known upper bounds in the polynomial Szemerédi theorem for the case when polynomials are neither linearly independent nor homogeneous of the same degree. We moreover improve known bounds for subsets of finite fields lacking arithmetic progressions with a difference coming from the set of $k$-th power residues, i.e. configurations of the form $$x, x+y^k, ..., x+(m-1)y^k.$$ Both results follow from an estimate of the number of such progressions in an arbitrary subset of a finite field.'
author:
- Borys Kuca
title: Further bounds in the polynomial Szemerédi theorem over finite fields
---
Introduction
============
Generalizing Szemerédi’s theorem on arithmetic progressions in subsets of integers [@szemeredi_1975], Bergelson and Leibman proved that each dense subset of ${\mathbb{Z}}$ contains a configuration of the form $x, x+P_1(y), ..., x+P_m(y)$, where $y\in{\mathbb{Z}}\setminus{\{0\}}$ and $P_1, ..., P_m$ are polynomials with integer coefficients and zero constant term [@bergelson_leibman_1996]. Their proof, based on ergodic theory, does not give explicit quantitative bounds. Although no general bounds are known so far, they exist in certain special cases, for instance for $x, x+y^k, ..., x+(m-1)y^k$ with $m\geqslant 2$ and $k>1$ [@prendiville_2017] or for $x, x+y, x+y^2$[@peluse_prendiville_2019]. In the finite field analogue of the question, when we are looking for bounds on the size of $A\subset{\mathbb{F}}_q$ lacking $x, x+P_1(y), ..., x+P_m(y)$, bounds are known in the case of $P_1, ..., P_m$ being linearly independent [@peluse_2019].
In this paper, we give the first explicit upper bounds for the sizes of subsets of finite fields lacking certain polynomial progressions. Our main result is the following.
\[main theorem, special case\] Let $m,k\in{\mathbb{N}}_+$, and $p$ be a prime. Suppose that $A\subset{\mathbb{F}}_p$ lacks the progression $$\begin{aligned}
\label{union of AP and GP}
x,\; x+y,\; ...,\; x+(m-1)y,\; x+y^m,\; ...,\; x+y^{m+k-1}\end{aligned}$$ with $y\neq 0$. Then $$|A|\ll\begin{cases}
p^{1-c},\; &m = 1,2,\\
p\dfrac{(\log \log p)^4}{\log p},\; &m = 3,\\
p{(\log p)^{-c}},\; &m = 4,\\
{p}{(\log\log p)^{-c}}, \; &m>4
\end{cases}$$ where all constants are positive, and the implied constant depends on $k$ and $m$ while $c$ depends only on $m$. For $m>4$, one can take the exponent $c$ to equal $c = 2^{-2^{m+9}}$.
It is worth noting that the exponent $c$ appearing in Theorem \[main theorem, special case\] for $c>4$ is the same as the exponent that appeared in Gowers’ bounds in Szemerédi theorem [@gowers_2001].
One can think of (\[union of AP and GP\]) as the union of an arithmetic progression and a shifted geometric progression. The cases $m=1$ and $m=2$ are in fact identical, and the bound in this case comes from the work of Peluse [@peluse_2019]. Our contribution is the $m>2$ case, for which there are no previous bounds in the literature. This is the first polynomial progression for which quantitative bounds are known where polynomials in $y$ are neither linearly independent nor homogeneous of the same degree. Theorem \[main theorem, special case\] is a special case of a more general result, which generalizes [@peluse_2019] and uses it as a base case for induction.
\[main theorem, general case\] Let $m,k\in{\mathbb{N}}_+$, $m\geqslant 3$, and $P_m$, ..., $P_{m+k-1}$ be polynomials in ${\mathbb{Z}}[y]$ such that $$a_m P_m + ... + a_{m+k-1}P_{m+k-1}$$ has degree at least $m$ unless $a_m = ... = a_{m+k-1}=0$ (in particular, $P_m$, ..., $P_{m+k-1}$ are linearly independent and each of them has degree at least $m$). Let $r_m(p)$ be the size of the largest subset of ${\mathbb{F}}_p$ lacking $m$-term arithmetic progressions and ${s_m: [p_0,\infty)\to (0,1]}$ be a decreasing function satisfying ${r_m(p)\leqslant p\cdot s_m(p)}$ for all primes $p\geqslant p_0>0$, with $s_m(n)\to 0$ as $n\to\infty$. If $A\subset {\mathbb{F}}_p$ lacks $$\begin{aligned}
\label{generalized union of AP and GP}
x, x+y, ..., x+(m-1)y, x+P_m(y), ..., x+P_{m+k-1}(y)\end{aligned}$$ with $y\neq 0$, then $$\begin{aligned}
|A|\ll p\cdot s_m(c p^{c})\end{aligned}$$ where the constants $C$, $c$, and the implied constant depend on $m, k$, and $P_m$, ..., $P_{m+k-1}$ but not on the choice of $s_m$.
The best bounds for $r_m$ currently in the literature are of the form $$r_m(p)\ll \begin{cases}
p\dfrac{(\log \log p)^4}{\log p},\; &m = 3 \;\text{\cite{bloom_2016}},\\
p(\log p)^{-c},\; &m = 4\; \text{\cite{green_tao_2017}},\\
p(\log\log p)^{-c},\; &m > 4\; \text{\cite{gowers_2001}}
\end{cases}$$ yielding the bounds given in Theorem \[main theorem, special case\]. The content of Theorem \[main theorem, general case\] is that up to the values of constants, our bounds are of the same shape as the bounds in Szemerédi theorem. One cannot hope to do better, as each set containing (\[generalized union of AP and GP\]) necessarily contains an $m$-term arithmetic progression. The function $s_m$ plays only an auxiliary role, allowing us to conveniently express known bounds in Szemerédi’s theorem as functions defined over positive real numbers.
We prove Theorem \[main theorem, general case\] by first proving an estimate for how many polynomial progressions a set $A\subset{\mathbb{F}}_p$ has. This counting result is the heart of this paper; once it is proved, deducing Theorem \[main theorem, general case\] is straightforward.
\[counting theorem for the more difficult configuration\] Let $m\in{\mathbb{N}}_+$ and $P_m$, ..., $P_{m+k-1}$ be polynomials in ${\mathbb{Z}}[y]$ such that $$a_m P_m + ... + a_{m+k-1}P_{m+k-1}$$ has degree at least $m$ unless $a_m = ... = a_{m+k-1}=0$ (in particular, $P_m$, ..., $P_{m+k-1}$ are linearly independent and each of them has degree at least $m$). Suppose that $f_0, ..., f_{m+k-1}:{\mathbb{F}}_p\to{\mathbb{C}}$ satisfy $|f_j(x)|\leqslant 1$ for each $0\leqslant j\leqslant m+k-1$ and $x\in{\mathbb{F}}_p$. Then $$\begin{aligned}
\label{equality in main counting theorem}
&{\mathbb{E}}_{x,y\in{\mathbb{F}}_p}\prod_{j=0}^{m-1}f_j(x+jy)\prod_{j=m}^{m+k-1}f_j(x+P_j(y))\\
\nonumber
=\; &{\mathbb{E}}_{x,y\in{\mathbb{F}}_p}\prod_{j=0}^{m-1}f_j(x+jy)\left(\prod_{j=m}^{m+k-1}{\mathbb{E}}f_j\right) + O(p^{-c})\end{aligned}$$ where all the constants are positive and depend on $m, k$ and polynomials $P_m$, ..., $P_{m+k-1}$ but not on $f_0, ..., f_{m+k-1}$.
Using the language of probability theory, we can interpret this result as “discorrelation": up to an error $O(p^{-c})$, the polynomials $P_m, ..., P_{m+k-1}$ occur independently from $m$-term arithmetic progressions.
The condition imposed on the polynomials $P_m, ..., P_{m+k-1}$ may seem artificial, but Theorem \[counting theorem for the more difficult configuration\] fails if this condition is not satisfied. As an example of failure, consider the configuration $x, x+y, x+2y, x+y^2$. Because $y^2$ has degree 2, which is less than the length of the arithmetic progression, $y^2$ is contained in the span of $x^2, (x+y)^2, (x+2y)^2$. Thus, there exist quadratic polynomials $Q_0, Q_1, Q_2$ and a nonzero linear polynomial $Q_3$ satisfying $$\begin{aligned}
Q_0(x)+Q_1(x+y)+Q_2(x+2y)+Q_3(x+y^2) = 0.
\end{aligned}$$ As a consequence, if we take $f_j(t) = e_p(a Q_i(t))$ for $a\neq 0$, then $$\begin{aligned}
{\mathbb{E}}_{x,y}f_0(x)f_1(x+y)f_2(x+2y)f_3(x+y^2) = 1
\end{aligned}$$ while the right-hand side of (\[equality in main counting theorem\]) in this case is $O(p^{-c})$, as ${\mathbb{E}}f_3 = 0$. More generally, if a linear combination of $P_m, P_{m+1}, ..., P_{k+m-1}$ has degree $d<m$, then there is a nontrivial polynomial relation connecting $x, x+y, ..., x+(m-1)y$ with some of $P_m, ..., P_{k+m-1}$, and this relation prevents discorrelation from happening. A natural question that one could ask at this point is whether Theorems \[main theorem, special case\], \[main theorem, general case\] and \[counting theorem for the more difficult configuration\] generalise to ${\mathbb{F}}_q$ when $q$ is a prime power and not just a prime number. Indeed, Theorem \[counting theorem for the more difficult configuration\] remains true if we replace ${\mathbb{F}}_p$ by ${\mathbb{F}}_q$, with the error $O(q^{-c})$ instead of $O(p^{-c})$. However, Theorems \[main theorem, special case\] and \[main theorem, general case\] no longer need to hold. In the process of going from Theorem \[counting theorem for the more difficult configuration\] to Theorems \[main theorem, special case\] and \[main theorem, general case\], one needs to apply known upper bounds for the largest subset of ${\mathbb{F}}_q$ lacking $m$-term arithmetic progressions. These bounds differ in two extreme cases, one being ${\mathbb{F}}_p$ and another being ${\mathbb{F}}_q$ with $q=p^n$ and $p$ fixed. In the former case, the upper bounds for the largest subset lacking $m$-term arithmetic progressions vary from $O(\frac{p}{\log p^{1-o(1)}})$ to $O(\frac{p}{(\log \log p)^c})$ depending on the length $m$ of the arithmetic progression, as indicated earlier. For the latter, Ellenberg and Gijswijt proved a bound of the form $O(q^{1-c})$ for 3-term arithmetic progressions [@ellenberg_gijswijt_2016]. The fact that polynomial bounds like that cannot be attained in ${\mathbb{F}}_p$ comes from the celebrated construction of Behrend [@behrend_1946]. Therefore, the bounds that we gave in Theorems \[main theorem, special case\] and \[main theorem, general case\] are given for ${\mathbb{F}}_p$ and not for all ${\mathbb{F}}_q$. If we wanted to work in the fixed characteristic case, then the largest subset of ${\mathbb{F}}_q$ lacking the progression (\[generalized union of AP and GP\]) would have size $O(q^{1-c})$ for $m=3$, with constants depending on $p, k, P_3, ..., P_{2+k}$. Polynomial bounds would also hold for $m>3$ contingent on generalising polynomial bounds in Ellenberg and Gijswijt’s result to longer arithmetic progressions. In the fixed characteristic case, one can thus do strictly better than in ${\mathbb{F}}_p$. However, between these two extremes there is a grey area of finite fields ${\mathbb{F}}_q$ which have both large characteristic $p$ and large dimension $n$ as a ${\mathbb{F}}_p$-vector space, for which our methods seem difficult to adapt.
To complement these results, we prove an upper bound for the size of subsets of ${\mathbb{F}}_p$ lacking progressions of the form $$\begin{aligned}
\label{APs with restricted common differences}
x, x+y^k, .., x+(m-1)y^k \end{aligned}$$ i.e. arithmetic progressions with $k-$th power common difference. An upper bound on subsets of ${\mathbb{Z}}$ lacking this configuration of the form $C\frac{N}{(\log \log N)^{c}}$, with constants depending on $m$ and $k$, was proved by Prendiville [@prendiville_2017] using density increment, and it naturally carries over to subsets of finite fields. Our bound works only for finite fields, where it is of the same shape as Prendiville’s for $m>4$, albeit with a better exponent, and strictly improves on it for $m=3,4$.
\[Sets lacking arithmetic progressions with $k$-th power differences\] Suppose $A\subset {\mathbb{F}}_p$ contains no arithmetic progression of length $m$ and common difference coming from the set of $k$-th power residues. Then $$|A|\ll\begin{cases}
p\dfrac{(\log \log p)^4}{\log p},\; &m = 3,\\
{p}{(\log p)^{-c}},\; &m = 4,\\
{p}{(\log\log p)^{-c}}, \; &m>4.
\end{cases}$$ The constant $c$ depends only on $m$, and in fact for $m>4$, we can take $c = 2^{-2^{m+9}}$. More generally, $$\begin{aligned}
|A|&\ll p\cdot s_m(c'\cdot p^{c'})\end{aligned}$$ where $s_m$ is defined as in Theorem \[main theorem, general case\]. The constants $C, c'$ and the implied constants are positive and depend on $k$ and $m$.
Again, up to the values of constants involved, our bounds are optimal in the sense that they are of the same shape as the bounds in Szemerédi theorem.
We derive the bounds in Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\] using a simple argument that heavily exploits the density and equidistribution of $k$-th power residues in the finite fields. With this argument, we prove the following more general counting theorem which implies Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\].
\[counting theorem for linear forms with restricted variables\] Let $L_1, ..., L_m$ be pairwise linearly independent linear forms in $x_1, ..., x_d$. Let $k_1,...,k_d$ be positive integers. Moreover, if $k_j>1$, assume that no linear form $L_i$ is of the form $L_i(x_1,...,x_d)=cx_j$. If $f_1,...,f_m$ satisfy $|f_i(x)|\leqslant 1$ for each $1\leqslant i\leqslant m$ and each $x\in{\mathbb{F}}_p$, then $$\begin{aligned}
\label{main equation}
{\mathbb{E}}_{x_1,...,x_d\in{\mathbb{F}}_p}\prod\limits_{j=1}^m f_j(L_j(x_1^{k_1},...,x_d^{k_d})) ={\mathbb{E}}_{x_1,...,x_d\in{\mathbb{F}}_p}\prod\limits_{j=1}^m f_j(L_j(x_1,...,x_d))+O(p^{-c}).\end{aligned}$$ In particular, $$\begin{aligned}
&|\{(x_1,...,x_d)\in {\mathbb{F}}_p^d: L_i(x_1^{k_1},...,x_d^{k_d})\in A\;{\rm{for}}\;1\leqslant i\leqslant m\}|\\
&=|\{(x_1,...,x_d)\in {\mathbb{F}}_p^d: L_i(x_1,...,x_d)\in A\;{\rm{for}}\;1\leqslant i\leqslant m\}|+O(p^{-c}).
$$
Similarly to the discussion following Theorem \[counting theorem for the more difficult configuration\], Theorem \[counting theorem for linear forms with restricted variables\] remains true for ${\mathbb{F}}_q$, but going from Theorem \[counting theorem for linear forms with restricted variables\] to Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\] forces us to work in ${\mathbb{F}}_p$ instead of ${\mathbb{F}}_q$. An analogue of Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\] for ${\mathbb{F}}_{q}$, $q=p^n$ with $p$ fixed and $n$ being the asymptotic parameter is that sets lacking $x, x+y^k, x+2y^k$ have size $O(q^{1-c})$, with the implied constant dependent on $k$. A similar result would hold for longer progressions if Ellenberg and Gijswijt’s bound [@ellenberg_gijswijt_2016] could be generalised.
Known results {#Known results}
-------------
In this section we enumerate known bounds for subsets of ${\mathbb{F}}_q$ or $[N]$ lacking polynomial progressions $$\begin{aligned}
x,\; x+P_1(y),\; ...,\; x+P_m(y)\end{aligned}$$ where not all of $P_1, ..., P_m$ are linear. There are some differences between the integral and finite field settings. Most importantly, finite fields contain significantly more polynomial progressions of a given form if at least one polynomial is nonlinear. That is because a nonlinear polynomial $P$ of degree $d>1$ has only $\Theta(N^{\frac{1}{d}})$ images in $[N]$, but it is a dense subset of ${\mathbb{F}}_q$, in the sense that there are at least $\frac{q}{d}$ images of $P$ in ${\mathbb{F}}_q$.
The case $m=1$ in natural numbers is often referred to as Furstenberg-Sárközy’s theorem, and it is equivalent to finding the largest subset $A$ of natural numbers whose difference set does not intersect the values of $P$ evaluated at integers. This problem has been studied, among others, by Sárközy [@sarkozy_1978a; @sarkozy_1978b], Balog, Pelikán, Pintz, and Szemerédi [@balog_pelikan_pintz_szemeredi_1994], Slijepcević [@slijepcevic_2003], Lucier [@lucier_2006], and Rice [@rice_2019]. They showed that $A$ is sparse if and only if for each natural number $n$ there exists $m\in{\mathbb{N}}$ for which $n$ divides $P(m)$, getting explicit bounds on the way; such polynomials have been called *intersective*. When $P(y)=y^k$ for $k>1$, a lower bound of the form $\Omega(N^c)$ for $0<c<1$ depending on $k$ can be obtained by trivial greedy algorithm, and the value of $c$ has been improved nontrivially by Ruzsa [@ruzsa_1984]. For finite fields ${\mathbb{F}}_q$, an elementary Fourier analytic argument gives upper bounds of the form $O(p^\frac{1}{2})$ with the implied constant depending on $k$, while the best known lower bounds are of the form $\Omega(\log p\log\log\log p)$ for infinitely many primes $p$ [@graham_ringrose_1990].
In the case $m>1$, bounds have only been known in two extremes. If $P_1$, ..., $P_m$ are all homogeneous of the same degree, i.e. we have a configuration of the form $$\begin{aligned}
x,\; x+c_1 y^k,\; ...,\; x+c_m y^k\end{aligned}$$ then Prendiville [@prendiville_2017] proved that all subsets of $[N]$ lacking this configuration have size $O\left(\frac{N}{(\log\log N)^{c}}\right)$ for some $c>0$ depending on $m$ and $k$. Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\] improves this result over finite fields for configurations of length 3 and 4.
The other extreme is when $P_1$, ..., $P_m$ are all linearly independent. This case has recently been tackled over finite fields by Peluse [@peluse_2019] who has showed that subsets of ${\mathbb{F}}_q$ lacking such progressions have size $O(q^{1-c})$ for $c>0$ depending on $P_1, ..., P_m$. In the case $m=2$, a specific exponent is known due to works of Bourgain and Chang [@bourgain_chang_2017], Peluse [@peluse_2018], and Dong, Li and Sawin [@dong_li_sawin_2017]. Recently, the results on the case $x, x+y, x+y^2$ have been extended to the integers by Peluse and Prendiville [@peluse_prendiville_2019].
Notation, terminology, and assumptions {#Notation}
--------------------------------------
Throughout the paper, $p$ always denotes the characteristic and cardinality of the finite field ${\mathbb{F}}_p$ in which we are currently working.
A function $f$ is *1-bounded* if $||f||_\infty\leqslant 1$. We always assume that $f$ is a $1$-bounded function from ${\mathbb{F}}_p$ to ${\mathbb{C}}$ unless explicitly stated otherwise. Sometimes, we use an expression $\textbf{b}(t_1, ..., t_n)$ to denote a 1-bounded function depending only on the variables $t_1, ..., t_n$ whose exact form is irrelevant and may differ from line to line.
We denote constants by $0<c<1<C$. The exact values of these constants are generally unimportant, only their relative size, therefore we shall often use the same symbol $c, C$ to denote constants whose value changes from line to line or even in the same expression. If there are good reasons to distinguish between two constants in the same expression, we shall denote them as $c, c'$ or $C, C'$ respectively. If we need to fix a constant for the duration of an argument, we give it a numerical subscript, e.g. $c_0$. We also use asymptotic notation $f=O(g), g=\Omega(f), f\ll g$, or $g\gg f$ to denote that $|f(p)|\leqslant C |g(p)|$ for sufficiently large $p$. The constant may depend on parameters such as the length of the polynomial progression or the degrees and leading coefficients of polynomials $P_1$, ..., $P_m$ involved. However, if the asymptotic notation is used in an expression involving arbitrary functions $f_0, ..., f_m$, the constant never depends on the choice of $f_0, ..., f_m$. While it is quite common in additive combinatorics to denote the dependence of the constant on these parameters by e.g. writing $C_m$ when it depends on $m$, we refrain from doing so in order not to clutter the notation. Therefore the reader should always assume that constants depend on the shape and length of the polynomial progression, but never on the functions $f_0, ..., f_m$ weighting the progression. We shall reiterate this in the statements of our lemmas and theorems.
We often use expected values, which we denote by ${\mathbb{E}}_{x\in X}f(x)=\frac{1}{|X|}\sum_{x\in X}f(x)$. If the set $X$ is omitted from the notation, it is assumed that $x$ is taken from ${\mathbb{F}}_p$ or from another specified set.
We denote the indicator function of the set $A$ by $1_A$. The map $\mathcal{C}:x\mapsto\overline{x}$ denotes the conjugation operator. Finally, we set $e_p(x):=e(x/p)=e^{2\pi i x/p}$.
Acknowledgements
----------------
The author is indebted to Sean Prendiville for his unrelenting support, useful suggestions, inspiring discussions, and help with editing the paper.
Basic concepts from additive combinatorics {#Gowers norms}
==========================================
The purpose of this section is to describe a few basic and standard concepts that are used extensively throughout this paper. We only introduce here ideas that are essential for all the arguments. There are tools which shall only be applied in specific proofs, and these will be discussed in relevant sections.
Fourier transform
-----------------
Given a function $f:{\mathbb{F}}_p\to{\mathbb{C}}$ and $\alpha\in{\mathbb{F}}_p$, we define its Fourier transform by the formula $$\begin{aligned}
\hat{f}(\alpha):={\mathbb{E}}_x f(x) e_p(\alpha x).\end{aligned}$$ We also call $\hat{f}(\alpha)$ the *Fourier coefficient of $f$ at $\alpha$*. We define the inner product on ${\mathbb{F}}_p$ as well as $L^s$ and $\ell^s$ norms for functions from ${\mathbb{F}}_p$ to ${\mathbb{C}}$ to be $$\begin{aligned}
\langle f, q\rangle :={\mathbb{E}}_x f(x)\overline{g(x)},\quad
||f||_{L^s} =\left({\mathbb{E}}_x |f(x)|^s\right)^\frac{1}{s},\quad {\rm{and}}\quad
||f||_{\ell^s} =\left(\sum_x |f(x)|^s\right)^\frac{1}{s}.\end{aligned}$$ for $1\leqslant s<\infty$, and we set $||f||_\infty:=||f||_{L^\infty}=||f||_{\ell^\infty}=\max\{|f(x)|: x\in{\mathbb{F}}_p\}$.
Gowers norms {#section on Gowers norms}
------------
Let $\Delta_h f(x)=f(x+h)\overline{f(x)}$ denote the *multiplicative derivative* of $f$ and $\nabla_h f(x)=f(x+h)-f(x)$ be its *additive derivative*. The $U^s$ norm of $f$ is defined as $$\begin{aligned}
\label{Gowers norm}
||f||_{U^s}:=\left({\mathbb{E}}_{x,h_1,...,h_s}\prod_{{{\underline{w}}}\in\{0,1\}^s}\mathcal{C}^{|w|}f(x+{{\underline{w}}}\cdot {{\underline{h}}}) \right)^{\frac{1}{2^s}}\end{aligned}$$ where $|w|=w_1+...+w_s$. If $f=1_A$, then $||1_A||_{U^s}^{2^s}$ is the normalized count of $s$-dimensional parallelepipeds in $A$, i.e. conifgurations of the form $$(x+w_1 h_1 + ... + w_s h_s)_{{{\underline{w}}}\in \{0,1\}^s}.$$ It turns out that $||f||_{U^s}$ is a well-defined norm for $s>1$ and a seminorm for $s=1$ (for the proofs of these and other facts on Gowers norms described in this section, including Lemma \[von Neumann\], consult [@green_2007] or [@tao_2012]). In fact, $||f||_{U^1}=|{\mathbb{E}}_x f(x)|=|\hat{f}(0)|$. Gowers norms enjoy several important properties that are used extensively in this paper. First, they are monotone: $$\begin{aligned}
||f||_{U^1}\leqslant ||f||_{U^2}\leqslant ||f||_{U^3}\leqslant ...\end{aligned}$$ Second, one can express a $U^s$ norm of $f$ in terms of a lower-degree Gowers norm of its multiplicative derivatives: $$\begin{aligned}
||f||_{U^s}^{2^s}={\mathbb{E}}_{h_1, ..., h_{s-k}}||\Delta_{h_1,..., h_{s-k}}f||_{U^k}^{2^k}.\end{aligned}$$ In particular, taking $k=2$ gives: $$\begin{aligned}
||f||_{U^s}^{2^s}={\mathbb{E}}_{h_1, ..., h_{s-2}}||\Delta_{h_1,..., h_{s-2}}f||_{U^2}^4.\end{aligned}$$ The utility of this formula for us is that $U^2$ norm is much easier to understand than the $U^s$ norms for $s>2$. In particular, $||f||_{U^2}=||\hat{f}||_{\ell^4}$, and from the fact that $\max_{\phi\in{\mathbb{F}}_p}|\hat{f}(\phi)|\leqslant||\hat{f}||_{\ell^4} \leqslant\max_{\phi\in{\mathbb{F}}_p}|\hat{f}(\phi)|^{\frac{1}{2}}$ it follows that having a large $U^2$ norm is equivalent to having a large Fourier coefficient, which is the statement of *$U^2$ inverse theorem*. For $s>2$, corresponding inverse theorems exist as well, but they are significantly more involved and we fortunately do not need them.
Gowers norms, introduced by Gowers in his celebrated proof of Szemerédi theorem, occur frequently in additive combinatorics because $||1_A||_{U^s}$ controls the number of $(s+1)$-term arithmetic progressions in $A$ in the following way.
\[von Neumann\] Let $f_0, ..., f_s$ be $1$-bounded. Then $$\begin{aligned}
|{\mathbb{E}}_{x,y}f_0(x)f_1(x+y)...f_s(x+sy)|\leqslant\min_{0\leqslant i\leqslant s}||f_j||_{U^s}.\end{aligned}$$
Counting arithmetic progressions in subsets of finite fields {#Counting arithmetic progressions in subsets of finite fields}
------------------------------------------------------------
In Theorems \[counting theorem for the more difficult configuration\] and \[counting theorem for linear forms with restricted variables\], we show that a certain counting operator can be expressed in terms of $$\begin{aligned}
\Lambda_m(f_0, ..., f_{m-1}):={\mathbb{E}}_{x,y}f_0(x)f_1(x+y)...f_{m-1}(x+(m-1)y)\end{aligned}$$ which counts $m$-term arithmetic progressions weighted by $f_0$, ..., $f_{m-1}$. In particular, $\Lambda_m(1_A)=\Lambda_m(1_A, ..., 1_A)$ is a normalized count of $m$-term arithmetic progressions in $A$. Instead of giving the exact estimates for what this counting operator is, we want to bound it from below by an expression involving $N_m(\alpha)$, which is the smallest natural number such that $p>N_m(\alpha)$ implies that each subset of ${\mathbb{F}}_p$ of size at least $\alpha p$ contains an $m-$term arithmetic progression. The reason why we want to have the estimate for $\Lambda_m$ in terms of $N_m$ is because the functions $N_m$ and $r'_m(n):=r_m(n)/n$ are essentially inverses, where $r_m(p)$ is the size of the largest subset of ${\mathbb{F}}_p$ not containing an $m$-term arithmetic progression. What we mean by this is that if $r'_m$ is bounded from above by a decreasing function $s_m$, then - subject to certain conditions - $N_m$ is bounded from above by $s^{-1}_m$. The following lemma makes this precise.
\[inverse functions\] Let $r_m(p)$ be the size of the largest subset of ${\mathbb{F}}_p$ lacking m-term arithmetic progressions. Let $N_m(\alpha)$ be the smallest natural number such that $p > N_m(\alpha)$ implies that each subset of ${\mathbb{F}}_p$ of size at least $\alpha p$ has an $m$-term arithmetic progression. Suppose that $s_m: [p_0,\infty) \to (0,1]$ is a decreasing function with $\lim\limits_{n\to\infty} s_m(n)=0$. Let $M_m$ be its inverse defined on $(0,\alpha_0]$, where $\alpha_0:=s_m(p_0)$. Then $r_m(p) \leqslant p s_m(p)$ for $p\geqslant p_0$ if and only if $N_m(\alpha) \leqslant M_m(\alpha)$ for $0<\alpha\leqslant\alpha_0$.
Combining Lemma \[inverse functions\] with an averaging argument of Varnavides, we obtain the following lemma, the precise version of which has been borrowed from [@rimanic_wolf_2019].
\[Averaging over progressions\] Suppose $0<\alpha_0\leqslant 1$, and let $M_m:(0,\alpha_0]\to{\mathbb{R}}_+$ be a decreasing function satisfying $N_m\leqslant M_m$. Suppose that $A\subset{\mathbb{F}}_{p}$ has size $|A|=\alpha p$ for some $0<\alpha\leqslant\alpha_0$. Then $|\Lambda_m(1_A)|\gg 1/M_m(\alpha/2)^2$, where the implied constant depends on $m$.
We conclude this section with the proof of Lemma \[inverse functions\].
Assume that $s_m$ is defined as in the statement of the lemma and that $r_m(p)\leqslant s_m(p)p$ for all prime $p\geqslant p_0$. Fix a prime number $p\geqslant p_0$ and $\alpha\in (0,\alpha_0]$. Suppose that $A\subset{\mathbb{F}}_p$ of size $|A|=\alpha p$ lacks an $m$-term arithmetic progression. The assumption of $p\geqslant p_0$ implies that $|A|\leqslant r_m(p)\leqslant s_m(p)p$, or $\alpha\leqslant s_m(p)$. From the monotonicity of $s_m$ it follows that $p\leqslant M_m(\alpha)$.
Thus, if a subset $A\subset {\mathbb{F}}_p$ of size $|A|=\alpha p$ for $0<\alpha\leqslant\alpha_0$ lacks $m$-term arithmetic progression, it must be that either $p\leqslant p_0$ or $p\leqslant M_m(\alpha)$, implying $N_m(\alpha)\leqslant\max\{p_0, M_m(\alpha)\}$. The definition of $p_0$ and monotonicity of $M_m$ imply that $p_0 = M_m(\alpha_0)\leqslant M_m(\alpha)$, and so $N_m(\alpha)\leqslant M_m(\alpha)$.
Conversely, suppose $N_m(\alpha)\leqslant M_m(\alpha)$ for $0<\alpha\leqslant\alpha_0$. Suppose that a set $A\subset{\mathbb{F}}_p$ of size $|A|=\alpha p$ lacks an $m$-term arithmetic progression, and assume $0<\alpha\leqslant\alpha_0$, $p\geqslant p_0$. Then $p\leqslant N_m(\alpha)\leqslant M_m(\alpha)$, and so $\alpha \leqslant s_m(p)$.
It thus follows that if a subset $A\subset {\mathbb{F}}_p$ of size $|A|=\alpha p$ for $p\geqslant p_0$ lacks an $m$-term arithmetic progression, then either $\alpha\leqslant s_m(p)$ or $\alpha >\alpha_0$. If the latter holds, then $\alpha>\alpha_0$ implies $M_m(\alpha)<M_m(\alpha_0)=p_0$, and so this case is impossible whenever $p\geqslant p_0$. Thus we must have that $\alpha\leqslant s_m(p)$ whenever $p\geqslant p_0$.
Deriving upper bounds in Theorem \[main theorem, general case\] {#deriving upper bounds}
===============================================================
This section is devoted to the proof of Theorem \[main theorem, general case\] using Theorem \[counting theorem for the more difficult configuration\] coupled with the notation from Section \[Counting arithmetic progressions in subsets of finite fields\].
Throughout this proof, all the constants are allowed to depend on $m,k$ and $P_m$, ..., $P_{m+k-1}$. From Theorem \[counting theorem for the more difficult configuration\] it follows that $$\begin{aligned}
{\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}{1}_A(x+jy)\prod_{j=m}^{m+k-1}{1}_A(x+P_j(y))
=\left({\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}{1}_A(x+jy)\right)\alpha^k + O(p^{-c})\end{aligned}$$ If $A\subset{\mathbb{F}}_p$ for $p\geqslant p_0$ has size $|A|=\alpha p$ and lacks progressions (\[generalized union of AP and GP\]), then the expression on the left-hand side is $O(p^{-1})$, and so $$\begin{aligned}
\label{inequality with inverse functions}
\left({\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}{1}_A(x+jy)\right)\alpha^k\ll p^{-c}.\end{aligned}$$ Let $M_m$ be the inverse function for $s_m$ on $(0,\alpha_0]$, where $\alpha_0=s_m(p_0)$, and set $M=M_m(\alpha/2)$. The assumption $p\geqslant p_0$ and the fact that $s_m$ is decreasing imply that $0<\alpha\leqslant\alpha_0$. Applying Lemma \[Averaging over progressions\] to (\[inequality with inverse functions\]) gives $\alpha^{k}M^{-2}\ll p^{-c}$. Behrend’s construction implies that $M$ grows faster than polynomially in $\alpha$: that is, for each $C>1$ there exists $c>0$ such that $M\geqslant c \alpha^{-C}$ [@behrend_1946]. Consequently, we have $M^{-3}\ll p^{-c}$ which implies that $M\gg p^c$ for a different constant $0<c<1$. From monotonicity of $s_m$ it follows that $\alpha\leqslant 2s_m(c p^c)$.
To illustrate the last bit of the above proof, we take Gowers’s [@gowers_2001] estimate $$N_m(\alpha)\leqslant 2^{2^{\alpha^{-C}}}$$ for $m > 4$. Combined with $N_m(\alpha/2)\gg p^c$, it gives the inequality $2^{2^{C\alpha^{-C}}} \gg p^c$. After rearranging, it yields $$\begin{aligned}
\alpha\ll\frac{1}{(\log\log p)^c}.\end{aligned}$$ Note that the function $s_m(p)=(\log_2\log_2 p)^{-c}$ is precisely the inverse function of $M_m(\alpha)=2^{2^{\alpha^{-C}}}$ for an appropriate choice of constants.
Proof of Theorem \[counting theorem for the more difficult configuration\] {#main section}
==========================================================================
Finally, we come to the main part of the paper, which is the proof of the counting theorem for the progression (\[generalized union of AP and GP\]). Like before, all the constants here are allowed to depend on $m, k$ and $P_m, ..., P_{m+k-1}$. First, we lexicographically order the set ${\mathbb{N}}^2_+$, i.e. $$\begin{aligned}
(m,k)<(m',k')\iff m<m' \vee (m=m' \wedge k<k').\end{aligned}$$ We induct on $(m,k)$ by following the lexicographic order on ${\mathbb{N}}_+^2$. Let ${\mathcal{S}}(m,k)$ denote the statement of Theorem \[counting theorem for the more difficult configuration\] for $(m,k)$; that is, ${\mathcal{S}}(m,k)$ holds iff for all linearly independent polynomials $P_m, ..., P_{m+k-1}$ of degree at least $m$ that do not span a polynomial of degree less than $m$ there exists a constant $c>0$ such that for all 1-bounded functions $f_0, ..., f_{m+k-1}$, we have $$\begin{aligned}
&{\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy)\prod_{j=m}^{m+k-1}f_j(x+P_j(y))
\\=&\left({\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy)\right)\prod_{j=m}^{m+k-1}{\mathbb{E}}f_j + O(p^{-c}).\end{aligned}$$
${\mathcal{S}}(1,k)$ and ${\mathcal{S}}(2,k)$ follow from the work of Peluse [@peluse_2019], and they shall serve as our base cases. In the inductive step, we have to prove two cases:
1. ${\mathcal{S}}(m,1)$, assuming the statement holds for all $(m',k')<(m,1)$ (although we shall only need to invoke ${\mathcal{S}}(m-1,2)$).
2. ${\mathcal{S}}(m,k)$ for $k>1$, assuming it holds for ${\mathcal{S}}(m,k')$ with $1\leqslant k' < k$.
The first case turns out to be the simpler of the two, and we shall carry it out promptly. The second case is much more involved, and it is where most of the difficulties lie.
Throughout this section, we denote the counting operator appearing in the statement of the Theorem \[counting theorem for the more difficult configuration\] by $\Lambda$ with appropriate subscripts. $$\begin{aligned}
\Lambda_{m, P_m, ..., P_{m+k-1}}(f_0, ..., f_{m+k-1}) := {\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy)\prod_{j=m}^{m+k-1}f_j(x+P_j(y)).\end{aligned}$$ In particular, $\Lambda_m$ denotes the counting operator for $m$-term arithmetic progressions: $$\begin{aligned}
\Lambda_m(f_0, ...,f_{m-1}):= {\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy).\end{aligned}$$ When $m,k$, and $P_m, ..., P_{m+k-1}$ are clear out of the context, we shall suppress the subscripts and denote the operator just by $\Lambda$.
Proof of ${\mathcal{S}}(m,1)$ assuming ${\mathcal{S}}(m-1,2)$
-------------------------------------------------------------
As advertised earlier, we first prove the inductive step for ${\mathcal{S}}(m,1)$. Let $P$ be a polynomial of degree at least $m$. Our goal is to show that the counting operator $$\begin{aligned}
\label{Lambda_{m,P}}
\Lambda_{m, P}(f_0,...,f_m)={\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m-1}f_j(x+jy)\right)f_m(x+P(y))\end{aligned}$$ is in fact controlled by an operator involving an arithmetic progression of length $m-1$ of difference functions of $f_1, ..., f_{m-1}$. To accomplish this, we first rewrite (\[Lambda\_[m,P]{}\]) as $$\begin{aligned}
{\mathbb{E}}_{x}f_0(x){\mathbb{E}}_y \left(\prod_{j=1}^{m-1}f_j(x+jy)\right)f_m(x+P(y)).\end{aligned}$$ Applying the Cauchy-Schwarz inequality in $x$ together with 1-boundedness of $f_0$, changing variables, translating $x\mapsto x-y$, and finally using the triangle inequality, we obtain that $$\begin{aligned}
|\Lambda_{m,P}(f_0,...,f_m)|^2 &\leqslant{\mathbb{E}}_x \left|{\mathbb{E}}_y \left(\prod_{j=1}^{m-1} f_j(x+jy)\right)f_m(x+P(y))\right|^2\\
&\leqslant{\mathbb{E}}_{x,y,h}\left(\prod_{j=1}^{m-1}\Delta_{jh} f_j(x+jy)\right)\overline{f_m(x+P(y))}f_m(x+P(y+h))\\
&\leqslant{\mathbb{E}}_h\left|{\mathbb{E}}_{x,y}\left(\prod_{j=1}^{m-1}\Delta_{jh} f_j(x+(j-1)y)\right)\overline{f_m(x+P(y)-y)}f_m(x+P(y+h)-y)\right|.\end{aligned}$$ By the pigeonhole principle, there exists $h\neq 0$ such that $$\begin{aligned}
|\Lambda_{m,P}(f_0,...,f_m)|^2 &\leqslant \left|{\mathbb{E}}_{x,y}\left(\prod_{j=1}^{m-1}\Delta_{jh} f_j(x+(j-1)y)\right)\overline{f_m(x+P(y)-y)}f_m(x+P(y+h)-y)\right|+O(p^{-1})\\
&=|\Lambda_{m-1, P_m, P_{m+1}}(g_0, ..., g_{m-2}, \overline{f_m}, f_m)|+O(p^{-1})\end{aligned}$$ where we set $$\begin{aligned}
P_m(y) = P(y)-y, \quad P_{m+1}(y) = P(y+h)-y \quad {\rm{and}} \quad
g_j(t) = \Delta_{(j+1)h}f_{j+1}(t).\end{aligned}$$ From $h\neq 0$ it follows that $P_m$, $P_{m+1}$ are linearly independent. Moreover, for any $(a,b)\neq (0,0)$, the polynomial $a P_m + b P_{m+1}$ has degree at least $m-1$, attaining this degree precisely when $a+b=0$. We have thus reduced the study of $\Lambda_{m,P}$ to the analysis of $\Lambda_{m-1,P_m,P_{m+1}}$, and so we are in the ${\mathcal{S}}(m-1,2)$ case. Applying Theorem \[counting theorem for the more difficult configuration\] for this case, we see that $$\begin{aligned}
|\Lambda_{m,P}(f_0,...,f_m)|^2 &\leqslant|\Lambda_{m-1, P_m, P_{m+1}}(g_0, ..., g_{m-2}, \overline{f_m},f_m)|+O(p^{-1})\\
&=|\Lambda_{m-1}(g_0,...,g_{m-2})|\cdot|{\mathbb{E}}f_m|^2 + O(p^{-c})\\
&\leqslant|{\mathbb{E}}f_m|^2 + O(p^{-c})\end{aligned}$$ and hence $$\begin{aligned}
|\Lambda_{m,P}(f_0,...,f_m)|\leqslant |{\mathbb{E}}f_m|+O(p^{-c}).\end{aligned}$$
We have established so far that the $U^1$ norm of $f_m$ controls $\Lambda_{m,P}(f_0,...,f_m)$ up to a power-saving error, i.e. $||f_m||_{U^1}=0$ implies $|\Lambda_{m,P}(f_0,...,f_m)|=O(p^{-c})$. To utilise this fact, we decompose $f_m = {\mathbb{E}}f_m + (f_m - {\mathbb{E}}f_m)$ and split $\Lambda_{m,P}$ accordingly. The term involving $f_m - {\mathbb{E}}f_m$ has size at most $O(p^{-c})$ because ${\mathbb{E}}(f_m-{\mathbb{E}}f_m) = 0$, and so $$\begin{aligned}
\Lambda_{m, P}(f_0, ..., f_m) = \Lambda_m(f_0, ...,f_{m-1}){\mathbb{E}}f_m + O(p^{-c}),\end{aligned}$$ as required.
Proof of ${\mathcal{S}}(m,k)$, $k>1$
------------------------------------
Our next goal is to prove ${\mathcal{S}}(m,k)$ whenever $k>1$. The natural thing to try would be to prove this case in a similar manner we proved ${\mathcal{S}}(m,1)$; that is, to apply the Cauchy-Schwarz inequality to the counting operator $$\begin{aligned}
\Lambda_{m, P_m, ..., P_{m+k-1}}(f_0, ..., f_{m+k-1})\end{aligned}$$ and bound it by the counting operator of $$\begin{aligned}
\Lambda_{m-1,Q_m, R_m, ..., Q_{m+k-1}, R_{m+k-1}}(g_0, ..., g_{m-2}, \overline{f_m}, f_m, ..., \overline{f_{m+k-1}}, f_{m+k-1})\end{aligned}$$ where $$\begin{aligned}
Q_j(y) = P_j(y)-y, \quad
R_j(y) = P_j(y+h)-y \quad {\rm{and}} \quad
g_j(t) = \Delta_{(j+1)h}f_{j+1}(t).\end{aligned}$$ However, this simple extension of the method used to prove ${\mathcal{S}}(m,1)$ does not work because there is no guarantee that $Q_m, R_m, ..., Q_{m+k-1}, R_{m+k-1}$ are linearly independent (and in general, they may not be), nor that any nonzero linear combination of them has degree at least $m-1$. To illustrate this problem, we look at $$\begin{aligned}
x,\; x+y,\; x+2y,\; x+y^3,\; x+y^4.\end{aligned}$$ Applying the Cauchy-Schwarz inequality and translating by $x\mapsto x-y$, we control this configuration by the counting operator of the configuration $$\begin{aligned}
x,\; x+y,\; x+y^3-y,\; x+(y+h)^3-y,\; x+y^4-y,\; x+(y+h)^4-y.\end{aligned}$$ Note that the polynomials $y,\; y^3-y,\; (y+h)^3-y,\; y^4-y,\; (y+h)^4-y$ have degree at most 4, and there are 5 of them, hence there exist $a_1, ..., a_5, b$ not all zero such that $$\begin{aligned}
a_1 y + a_2 (y^3-y) +a_3 ((y+h)^3-y) + a_4 (y^4-y) + a_5 ((y+h)^4-y) = b.\end{aligned}$$ Consequently, one cannot apply induction hypothesis to this configuration. One therefore needs to come up with a different method.
In the remainder of this section, we outline the proof of ${\mathcal{S}}(m,k)$, $k>1$. We formulate consecutive steps of the proof as lemmas to be proved separately in the next section. Let $$\begin{aligned}
\Lambda:=\Lambda_{m, P_m, ..., P_{m+k-1}}.\end{aligned}$$ It shall become clear shortly that proving the general case of ${\mathcal{S}}(m,k)$ for $k>1$ can be reduced to the case of $f_{m+k-1}$ being a character. In this case, the following holds.
\[Vanishing of expression II\] Let $a\in{\mathbb{F}}_p^\times$ and $k\in{\mathbb{N}}_+$. Assume ${\mathcal{S}}(m,k-1)$. Then $$\begin{aligned}
|\Lambda(f_0, ..., f_{m+k-2}, e_p(a\cdot))|\leqslant O(p^{-c})\end{aligned}$$ for a constant $c>0$ depending on $m$, $k$,and the polynomials $P_{m}, ..., P_{m+k-1}$ but not on $a$ or 1-bounded functions $f_0, ..., f_{m+k-2}:{\mathbb{F}}_p\to{\mathbb{C}}$.
Our first task before using this lemma is to show that $\Lambda$ is controlled by some Gowers norm of $f_{m+k-1}$. This follows from the so-called PET induction scheme, which originally appeared in Bergelson and Leibman’s ergodic-theoretic proof of polynomial Szemerédi theorem [@bergelson_leibman_1996] and was subsequently applied in the works of Prendiville and Peluse [@prendiville_2017; @peluse_2019; @peluse_prendiville_2019].
\[PET induction\] Let $P_1, ..., P_l$ be nonconstant polynomials in ${\mathbb{Z}}[y]$ such that $P_i - P_j$ is nonconstant whenever $i\neq j$. Then for any $1\leqslant j\leqslant l$ there exist $s\in{\mathbb{N}}$ and $0<\beta\leqslant 1$, depending only on the degrees and leading coefficients of $P_1, ..., P_l$, such that $$\begin{aligned}
|\Lambda_{x, x+P_1(y), ..., x+P_l(y)}(f_0, ..., f_l)|\leqslant ||f_j||_{U^s}^\beta+O(p^{-\beta}).\end{aligned}$$ for all 1-bounded functions $f_0, ..., f_l:{\mathbb{F}}_p\to{\mathbb{C}}$.
Our statement differs slightly from the statement of Proposition 2.2 in [@peluse_2019] in that Peluse did not mention explicitly our condition that the difference between any two polynomials $P_i$, $P_j$ cannot be constant. However, she assumed throughout her paper that $P_1, ..., P_l$ were distinct polynomials with zero constant terms, which implies our condition. In our paper, the polynomials may have nonzero constant terms, in which case we replace $P_i(y)$ by $P'_i(y):=P_i(y)-P_i(0)$ and $f_i(t)$ by $f_i'(t):=f_i(t+P_i(0))$, so that $f_i(x+P_i(y))=f'_i(x+P_i'(y))$. The facts that $f_i$ and $f'_i$ have the same Gowers norms and that $P_1'$, ..., $P_l'$ are all distinct polynomials with zero constant terms allows us to reduce to the case covered in Proposition 2.2 of [@peluse_2019].
Our next step is to decompose $f_{m+k-1}$ into three terms using a decomposition based on the Hahn-Banach theorem.
\[Hahn-Banach decomposition\] Let $f:{\mathbb{F}}_p\to{\mathbb{C}}$ and $||\cdot||$ be a norm on the space of ${\mathbb{C}}$-valued functions from ${\mathbb{F}}_p$. Suppose $||f||_{L^2}\leqslant 1$. Then there exists a decomposition $$\begin{aligned}
f = f_a + f_b + f_c\end{aligned}$$ with $||f_a||^*\leqslant p^{\delta_1}$, $||f_b||_{L^1}\leqslant p^{-\delta_2}$, $||f_c||_{L^\infty}\leqslant p^{\delta_3}$, $||f_c||\leqslant p^{-\delta_4}$ provided $$\begin{aligned}
\label{delta inequality}
p^{\delta_4-\delta_1}+p^{\delta_2-\delta_3}\leqslant\frac{1}{2}.\end{aligned}$$
This decomposition was pioneered by Gowers and Wolf in their work on true complexity of linear forms [@gowers_wolf_2011a; @gowers_wolf_2011b; @gowers_wolf_2011c; @gowers_2010]. The variant that we are using is due to Peluse and appeared in [@peluse_2019; @peluse_prendiville_2019]. The dual norm in the statement of Lemma \[Hahn-Banach decomposition\] is defined by $||f||^*=\sup\{|\langle f,g\rangle|: ||g||_\infty\leqslant 1\}$.
Using this decomposition, we can write $f_{m+k-1}$ as a sum of three functions: the first has not too big $U^s$-dual norm, the second has small $L^1$ norm, and the third has a small $U^s$ norm and not too big $L^\infty$ norm. By taking appropriate values of $\delta_1, \delta_2, \delta_3, \delta_4$, we get rid of two error terms and only work with $f_a$. This gives us control over $\Lambda_{m, P_m, ..., P_{m+k-1}}$ by the $U^s$ norm of a dual function $$\begin{aligned}
\label{dual function}
F(x):={\mathbb{E}}_y\prod_{j=0}^{m-1} f_j(x+jy-P_{m+k-1}(y))\prod_{j=m}^{m+k-2}f_j(x+P_j(y)-P_{m+k-1}(y))\end{aligned}$$ and allows us to essentially replace $f_{m+k-1}$ in the $\Lambda_{m, P_m, ..., P_{m+k-1}}$ operator by a character. We call $F$ a “dual function" because $\Lambda_{m, P_m, ..., P_{m+k-1}}(f_0,...,f_{m+k-1})=\langle F,f_{m+k-1}\rangle$.
In general, higher degree Gowers norms control lower degree norms but the converse is not true. For the special case of the dual function $F$, we however show that $||F||_{U^s}$ is indeed controlled by $||F||_{U^2}$ for any $s\in{\mathbb{N}}$. We achieve this in the following lemma.
\[degree lowering\] Let $F$ be defined as in (\[dual function\]). For each $s\geqslant 2$, $$\begin{aligned}
||F||_{U^{s-1}}= \Omega(||F||_{U^s}^{2^{2s-1}})-O(p^{-c})\end{aligned}$$ for $c>0$ depending on $m, k,$ and $P_m, ..., P_{m+k-1}$ but not on $f_0, ..., f_{m+k-2}$. As a consequence, $$\begin{aligned}
||F||_{U^2} = \Omega(||F||_{U^s}^{2^{(s-2)(s+2)}}) - O(p^{-c}).\end{aligned}$$
Having a control by the $U^2$ norm of the dual function $F$ is important because this norm is in turn controlled by the $U^{1}$ norms of the component functions $f_m, ..., f_{m+k-1}$, which follows from Lemma \[Vanishing of expression II\] coupled with ${\mathcal{S}}(m,k-1)$.
\[$U^1$ control of the dual\] Let $F$ be defined as in (\[dual function\]). Then $$\begin{aligned}
||F||_{U^2}\leqslant\min_{m\leqslant j\leqslant m+k-2}||f_j||_{U^1}^\frac{1}{2}+O(p^{-c})\end{aligned}$$ for some $c>0$ depending on $m, k,$ and $P_m, ..., P_{m+k-1}$ but not on $f_0, ..., f_{m+k-2}$.
Combining the estimates of two previous lemmas with Hahn-Banach decomposition, we get a control of the $\Lambda$ operator by $U^1$ norms of $f_m$, ..., $f_{m+k-1}$.
\[$U^1$ control of the operator\] There exist constants $c,c'>0$ and $s\in{\mathbb{N}}$ depending only on $m,k, P_m, ..., P_{m+k-1}$ but not on $f_0, ..., f_{m+k-1}$ such that $$\begin{aligned}
|\Lambda(f_0, ..., f_{m+k-1})| &\ll p^{c'}\min\limits_{m\leqslant j\leqslant m+k-2}||f_j||_{U^{1}}^{2^{-s}}+p^{-c}.\end{aligned}$$
Having established Lemma [\[$U^1$ control of the operator\]]{}, it is straightforward to prove ${\mathcal{S}}(m,k)$. We split each of $f_m$, ..., $f_{m+k-2}$ into $f_j = {\mathbb{E}}f_j + (f_j-{\mathbb{E}}f_j)$, and decompose $\Lambda$ accordingly. Then $\Lambda(f_0, ..., f_{m+k-1})$ splits into the main term $$\begin{aligned}
\Lambda(f_0, ..., f_{m-1}, {\mathbb{E}}f_m, ..., {\mathbb{E}}f_{m+k-2}, f_{m+k-1})\end{aligned}$$ and $2^k-1$ error terms, each of which involves at least one $f_j - {\mathbb{E}}f_j$ for $m\leqslant j\leqslant m+k-2$. Using Lemma \[$U^1$ control of the operator\], each of the error terms has size $O(p^{-c})$; thus $$\begin{aligned}
\Lambda(f_0, ..., f_{m+k-1}) &= \Lambda(f_0, ..., f_{m-1}, {\mathbb{E}}f_m, ..., {\mathbb{E}}f_{m+k-2}, f_{m+k-1}) +O(p^{-c})\\
&=\Lambda_{m, P_{m+k-1}}(f_0, ..., f_{m-1}, f_{m+k-1})\prod_{j=m}^{m+k-2}{\mathbb{E}}f_j +O(p^{-c}).\end{aligned}$$ Applying the ${\mathcal{S}}(m,1)$ case, we can split $\Lambda_{m, P_{m+k-1}}(f_0, ..., f_{m-1}, f_{m+k-1})$ into $$\begin{aligned}
\Lambda_{m, P_{m+k-1}}(f_0, ..., f_{m-1}, f_{m+k-1})=\Lambda_m(f_0,...,f_{m-1}){\mathbb{E}}f_{m+k-1}+O(p^{-c})\end{aligned}$$ and hence $$\begin{aligned}
\Lambda(f_0, ..., f_{m+k-1}) = \Lambda_m(f_0, ..., f_{m-1})\prod_{j=m}^{m+k-1}{\mathbb{E}}f_j + O(p^{-c}).\end{aligned}$$ This proves ${\mathcal{S}}(m,k)$ for $k>1$.
Proofs of Lemmas \[degree lowering\], \[$U^1$ control of the dual\] and \[$U^1$ control of the operator\]
---------------------------------------------------------------------------------------------------------
While in the previous section we outlined the proof of ${\mathcal{S}}(m,k)$ for $k>1$, here we derive the technical lemmas which are used in this proof.
This proof follows the path of Proposition 6.6 in [@peluse_prendiville_2019]. The main idea is to write the $U^s$ norm of the dual function $F$ as an average of the $U^2$ norms of derivatives of $F$, extract the maximum Fourier coefficients of $\Delta_{h_1, ..., h_{s-2}}F$, and show that for a dense proportion of $(h_1, ..., h_{s-2})$ these coefficients satisfy certain linear relations provided $||F||_{U^s}\gg p^{-c}$. If $s=3$ and $\phi(h)$ is the phase of the maximum Fourier coefficient of $\Delta_h F$, then we show that $\phi$ is constant on a dense proportion of $h$. For $s>3$, analogous relations are somewhat more complicated. These linear relations turn out to be sufficient to get a control of the $U^s$ norm of $F$ by its $U^{s-1}$ norm with polynomial bounds.
Using the definition of Gowers norms, we have $$\begin{aligned}
\eta :=||F||_{U^s}^{2^s}={\mathbb{E}}_{h_1,...,h_{s-2}}||\Delta_{h_1,...,h_{s-2}}F||_{U^2}^4.\end{aligned}$$ Let $H_1=\{(h_1, ..., h_{s-2})\in{\mathbb{F}}_p^{s-2}:||\Delta_{h_1,...,h_{s-2}}F||_{U^2}^4\geqslant\frac{1}{2}\eta\}$. To simplify the notation, let ${{\underline{h}}}=(h_1, ...,h_{s-2})$ and ${\mathbb{E}}_{{\underline{h}}}:={\mathbb{E}}_{{{\underline{h}}}\in{\mathbb{F}}_p^{s-2}}$. From the popularity principle (see e.g. Exercise 1.1.4 in [@tao_vu_2006]) it follows that $|H_1|\geqslant\frac{1}{2}\eta p^{s-2}$, and so $$\begin{aligned}
\label{Long expression 1}
\frac{1}{4}\eta^2 &\leqslant{\mathbb{E}}_{{{\underline{h}}}}||\Delta_{{{\underline{h}}}}F||_{U^2}^4\cdot 1_{H_1}({{\underline{h}}}).\end{aligned}$$ The $U^2$ inverse theorem, stated in Section \[section on Gowers norms\], implies that the square of the $U^2$ norm of a function is bounded by its maximum Fourier coefficient. Given $\Delta_{{{\underline{h}}}} F$, let $\widehat{\Delta_{{{\underline{h}}}} F}(\phi({{{\underline{h}}}}))$ denote its maximum Fourier coefficient. Then the right hand side of (\[Long expression 1\]) is bounded by $$\begin{aligned}
\label{Long expression 2}
\nonumber
{\mathbb{E}}_{{{\underline{h}}}}|\widehat{\Delta_{{{\underline{h}}}}F}(\phi({{\underline{h}}}))|^2 1_{H_1}({{\underline{h}}})&={\mathbb{E}}_{{\underline{h}}}|{\mathbb{E}}_x\Delta_{{{\underline{h}}}} F(x) e_p(\phi({{\underline{h}}}) x)|^21_{H_1}({{\underline{h}}})\\
&={\mathbb{E}}_{x,x',{{\underline{h}}}}\Delta_{{{\underline{h}}}} F(x)\overline{\Delta_{{{\underline{h}}}}F(x')}e_p(\phi({{\underline{h}}})(x-x'))1_{H_1}({{\underline{h}}}).\end{aligned}$$ To simplify notation, we denote $Q_j=P_j-P_{m+k-1}$ for $0\leqslant j\leqslant m+k-2$, where we extend the definition of $P_j$ to $0\leqslant j\leqslant m-1$ by setting $P_j(y) = jy$ for these values of $j$. Unpacking the definition of the dual function $F$, the expression (\[Long expression 2\]) equals $$\begin{aligned}
\label{Long expression 3}
&{\mathbb{E}}_{x,x',{{\underline{h}}}}\Delta_{{{\underline{h}}}}\left({\mathbb{E}}_y\prod_{j=0}^{m+k-2} f_j(x+Q_j(y)) \right)\\
\nonumber
&\overline{\Delta_{{{\underline{h}}}}\left({\mathbb{E}}_y\prod_{j=0}^{m+k-2} f_j(x'+Q_j(y))\right)}
e_p(\phi({{\underline{h}}})(x-x'))1_{H_1}({{\underline{h}}}).\end{aligned}$$ After writing out the multiplicative derivatives, (\[Long expression 3\]) is equal to $$\begin{aligned}
\label{Long expression 4}
&{\mathbb{E}}_{x,x',{{\underline{h}}}}{\mathbb{E}}_{{{\underline{y}}},{{\underline{y}}}'\in{\mathbb{F}}_p^{\{0,1\}^{s-2}}}\prod_{j=0}^{m+k-2}\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}} f_j(x+{{\underline{w}}}\cdot{{\underline{h}}}+Q_j({{\underline{y}}}_{{\underline{w}}}))\\
\nonumber
&\overline{f_j(x'+{{\underline{w}}}\cdot{{\underline{h}}}+Q_j({{\underline{y}}}'_{{\underline{w}}}))}
e_p(\phi({{\underline{h}}})(x-x'))1_{H_1}({{\underline{h}}}).\end{aligned}$$ The product in (\[Long expression 4\]) contains $2^{s-2}$ copies of $f_j$ for each $j$ and each of $x$ and $x'$. In each of these copies the $y$-variable is different. We would like all the copies of $f_j$ to be expressed in terms of the same $y$-variable. To achieve this, we modify (\[Long expression 4\]) by applying the Cauchy-Schwarz inequality $s-2$ times. First, (\[Long expression 4\]) can be rewritten as $$\begin{aligned}
\label{Long expression 5}
&{\mathbb{E}}_{x,x',h_1,...,h_{s-3}}{\mathbb{E}}_{{{\underline{y}}},{{\underline{y}}}'\in{\mathbb{F}}_p^{\{0,1\}^{s-2}}}\textbf{b}(x,x',h_1,...,h_{s-3},{{\underline{y}}},{{\underline{y}}}'){\mathbb{E}}_{h_{s-2}}\prod_{j=0}^{m+k-2}\prod_{\substack{{{\underline{w}}}\in\{0,1\}^{s-2}\\ w_s=1}}\\
\nonumber
& f_j(x+{{\underline{w}}}\cdot{{\underline{h}}}+Q_j({{\underline{y}}}_{{\underline{w}}}))\overline{f_j(x'+{{\underline{w}}}\cdot{{\underline{h}}}+Q_j({{\underline{y}}}'_{{\underline{w}}}))}
e_p(\phi({{\underline{h}}})(x-x'))1_{H_1}({{\underline{h}}}).\end{aligned}$$ By the Cauchy-Schwarz inequality and change of variables, (\[Long expression 5\]) is bounded by $$\begin{aligned}
\label{Long expression 6}
(&{\mathbb{E}}_{x,x',h_1,...,h_{s-3}, h_{s-2},h'_{s-2}}{\mathbb{E}}_{{{\underline{y}}},{{\underline{y}}}'\in{\mathbb{F}}_p^{\{0,1\}^{s-2}}}\prod_{j=0}^{m+k-2}\prod_{\substack{{{\underline{w}}}\in\{0,1\}^{s-2}\\ w_s=1}}\\
\nonumber
& (f_j(x+\sum_{i=1}^{s-3}w_i h_i+w_{s-2} h_{s-2}+Q_j({{\underline{y}}}_{{\underline{w}}}))\overline{f_j(x+\sum_{i=1}^{s-3}w_i h_i+w_{s-2} h'_{s-2}+Q_j({{\underline{y}}}_{{\underline{w}}}))}
\\\nonumber
&\overline{f_j(x'+\sum_{i=1}^{s-3}w_i h_i+w_{s-2} h_{s-2} +Q_j({{\underline{y}}}'_{{\underline{w}}}))}f_j(x'+\sum_{i=1}^{s-3}w_i h_i+w_{s-2} h'_{s-2} +Q_j({{\underline{y}}}'_{{\underline{w}}})))\\\nonumber
&e_p((\phi(h_1,...,h_{s-3},h_{s-2})-\phi(h_1,...,h_{s-3},h'_{s-2}))(x-x'))\\\nonumber
&1_{H_1}(h_1,...,h_{s-3},h_{s-2})
1_{H_1}(h_1,...,h_{s-3},h'_{s-2}))^\frac{1}{2}.\end{aligned}$$ The presence of so many terms in (\[Long expression 6\]) comes from the fact that in the process of applying the Cauchy-Schwarz inequality and changing variables, each expression $E(h_{s-2})$ (depending possibly on other variables as well) is replaced by ${E(h_{s-2})\overline{E(h'_{s-2})}}$. Therefore the number of expressions in the product doubles, making (\[Long expression 6\]) rather lengthy. Applying Cauchy-Schwarz another $s-3$ times to $h_{s-3}, ..., h_1$ respectively, we bound (\[Long expression 6\]) by $$\begin{aligned}
\label{intermediate long expression 1}
&({\mathbb{E}}_{x,x',y,y',{{\underline{h}}},{{\underline{h}}}'}\prod_{j=0}^{m+k-2}\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}}(1_{H^1}({{\underline{h}}}^{({{\underline{w}}})}) {\mathcal{C}}^{|w|}f_j(x+{{\underline{w}}}\cdot{{\underline{h}}}^{({{\underline{w}}})}+Q_j(y))\\
\nonumber
&\overline{{\mathcal{C}}^{|w|}f_j(x'+{{\underline{w}}}\cdot{{\underline{h}}}^{({{\underline{w}}})}+Q_j(y'))})e_p((\sum_{{{\underline{w}}}\in\{0,1\}^{s-2}}(-1)^{|w|}\phi({{\underline{h}}}^{({{\underline{w}}})})(x-x')))^{\frac{1}{2^{s-2}}}.\end{aligned}$$ where $${{\underline{h}}}^{({{\underline{w}}})}_i =
\begin{cases}
h_i, w_i = 0\\
h'_i, w_i = 1
\end{cases}$$ The expression (\[intermediate long expression 1\]) can be simplified to $$\begin{aligned}
\left({\mathbb{E}}_{{{\underline{h}}},{{\underline{h}}}'}\left|{\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m+k-2} g_j(x+P_j(y))\right)e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')(x+P_{m+k-1}(y))\right)
\right|^2{1}_{\square(H_1)}({{\underline{h}}},{{\underline{h}}}')\right)^\frac{1}{2^{s-2}}\end{aligned}$$ where $$\begin{aligned}
g_j(t):=\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}}{\mathcal{C}}^{|w|}f_j(t+{{\underline{w}}}\cdot{{\underline{h}}}^{({{\underline{w}}})}),\end{aligned}$$ $$\begin{aligned}
\square(A):=\{({{\underline{h}}},{{\underline{h}}}')\in{\mathbb{F}}_p^{2(s-2)}:\forall {{\underline{w}}}\in\{0,1\}^{s-2}{{\underline{h}}}^{({{\underline{w}}})}\in A\}\end{aligned}$$ and $$\begin{aligned}
\psi({{\underline{h}}},{{\underline{h}}}'):=\sum_{{{\underline{w}}}\in\{0,1\}^{s-2}}(-1)^{|w|}\phi({{\underline{h}}}^{({{\underline{w}}})}).\end{aligned}$$ Combining all of this, we obtain the estimate $$\begin{aligned}
\label{important intermediate expression}
{\mathbb{E}}_{{{\underline{h}}},{{\underline{h}}}'}&\left|{\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m+k-2} g_j(x+P_j(y))\right)e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')(x+P_{m+k-1}(y)))\right)
\right|^2\\\nonumber
&{1}_{\square(H_1)}({{\underline{h}}},{{\underline{h}}}')\geqslant\left(\frac{\eta}{2}\right)^{2^{s-1}}.\end{aligned}$$
We are now precisely in the situation of Lemma \[Vanishing of expression II\]. By this lemma, the expression inside the absolute values equals $O(p^{-c})$ unless $\psi({{\underline{h}}},{{\underline{h}}}') = 0$. Therefore, the set $$\begin{aligned}
H_2:=\left\{({{\underline{h}}},{{\underline{h}}}')\in \square(H_1): \psi({{\underline{h}}},{{\underline{h}}}') = 0\right\}\end{aligned}$$ has size at least $$\begin{aligned}
\left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O(p^{-c})\right)p^{2(s-2)}.\end{aligned}$$ In particular, there exists ${{\underline{h}}}\in H_1$ such that the fiber $$\begin{aligned}
H_3:=\{{{\underline{h}}}': ({{\underline{h}}},{{\underline{h}}}')\in H_2\}\end{aligned}$$ has size at least $$\begin{aligned}
\left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O(p^{-c})\right)p^{s-2}.\end{aligned}$$ Fix this ${{\underline{h}}}$. We now show that the phases $\phi$ possess some amount of low-rank structure which we subsequently use to complete the proof of the lemma. By the definitions of $H_2$ and $H_3$, for each ${{\underline{h}}}'\in H_3$ we have $\psi({{\underline{h}}},{{\underline{h}}}')=0$. Define $$\begin{aligned}
\psi_i({{\underline{h}}},{{\underline{h}}}'):=(-1)^s\sum_{\substack{{{\underline{w}}}\in\{0,1\}^{s-2},\\ w_1 = ... = w_{i-1}=1,\\ w_i = 0}}(-1)^{|w|}\phi({{\underline{h}}}^{({{\underline{w}}})}).\end{aligned}$$ Note that, $\psi({{\underline{h}}},{{\underline{h}}}')=\phi(h'_1,...,h'_{s-2})-\psi_1({{\underline{h}}},{{\underline{h}}}')-...-\psi_{s-2}({{\underline{h}}},{{\underline{h}}}')$. Crucially, $\psi_i$ does not depend on $h'_1, ..., h'_i$. Thus, $\psi({{\underline{h}}},{{\underline{h}}}') = 0$ implies that $$\begin{aligned}
\phi(h'_1,...,h'_{s-2})=\sum_{i=1}^{s-2}\psi_i({{\underline{h}}},{{\underline{h}}}').\end{aligned}$$ That is to say, $\phi(h'_1, ...,h'_{s-2})$ can be decomposed into a sum of $s-2$ functions, each of which does not depend on $h_i'$ for a different $i$.
To make the notation a bit more palatable, we illustrate the aforementioned for $s=3$ and 4. For $s=3$, $$\psi(h,h')=\phi(h)-\phi(h')=\psi_1(h)-\phi(h').$$ Hence $\psi(h,h')=0$ implies that $\phi(h')=\phi(h)$. For $s=4$, $$\begin{aligned}
\psi({{\underline{h}}},{{\underline{h}}}') &=\phi(h_1,h_2)-\phi(h_1',h_2)-\phi(h_1,h_2')+\phi(h_1',h_2')\\
&=\psi_1({{\underline{h}}},{{\underline{h}}}')-\psi_2({{\underline{h}}},{{\underline{h}}}')+\phi(h_1',h_2')\end{aligned}$$ and so $\psi({{\underline{h}}},{{\underline{h}}}')=0$ implies that $$\begin{aligned}
\phi(h_1',h_2') = \phi(h_1,h_2')+\phi(h_1',h_2)-\phi(h_1,h_2) = \psi_2({{\underline{h}}},{{\underline{h}}}')-\psi_1({{\underline{h}}},{{\underline{h}}}').\end{aligned}$$
We now estimate the expression $$\begin{aligned}
\label{Long expression 7}
{\mathbb{E}}_{{{\underline{h}}}'}||\Delta_{{{\underline{h}}}'}F||_{U^2}^4{1}_{H_3}({{\underline{h}}}')\end{aligned}$$ from above and below. From below, it is bounded by $$\frac{\eta}{2}\cdot \left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O(p^{-c})\right)\geqslant \left(\frac{\eta}{2}\right)^{2^{s}}-O(p^{-c}).$$ The upper bound is more complicated, and it relies on the fact that we can decompose $\phi({{\underline{h}}}')$ into a sum of $\psi_i$’s such that $\psi_i$ does not depend on $h'_i$. Using $U^2$-inverse theorem, (\[Long expression 7\]) is bounded from above by: $$\begin{aligned}
\label{Long expression 8}
{\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}(\phi({{\underline{h}}}'))\right|^2{1}_{H_3}({{\underline{h}}}') &= {\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2{1}_{H_3}({{\underline{h}}}').\end{aligned}$$ By positivity, we can extend (\[Long expression 8\]) to the entire ${\mathbb{F}}_p^{s-2}$; that is, we have $$\begin{aligned}
{\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2{1}_{H_3}({{\underline{h}}}')
\leqslant {\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2.\end{aligned}$$ Rewritting, we obtain that $$\begin{aligned}
\label{Long expression 9}
\nonumber
{\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2 &= {\mathbb{E}}_{{{\underline{h}}}'}\left|{\mathbb{E}}_x\Delta_{{{\underline{h}}}'}F(x)e_p\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')x\right)\right|^2\\
&={\mathbb{E}}_{x,{{\underline{h}}}',h_{s-1}}\Delta_{{{\underline{h}}}',h_{s-1}}F(x)e_p\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')h_{s-1}\right).\end{aligned}$$ We apply Cauchy-Schwarz $s-2$ times to (\[Long expression 9\]) to get rid of the phases $\psi_i({{\underline{h}}}')$. In the first application, we start by rewriting (\[Long expression 9\]) as $$\begin{aligned}
\label{Long expression 10}
{\mathbb{E}}_{\substack{x,h_2',...,\\ h'_{s-2},h_{s-1}}}\textbf{b}(x, h_2', ..., h'_{s-2}, h_{s-1}){\mathbb{E}}_{h'_1}\Delta_{\substack{h'_2,...,h'_{s-2},h_{s-1}}}F(x+h'_1)e_p\left(\sum_{i=2}^{s-2}\psi_i({{\underline{h}}}')h_{s-1}\right)\end{aligned}$$ and then we bound it by $$\begin{aligned}
\label{Long expression 11}
&({\mathbb{E}}_{x,h'_1, h''_1, h_2',..., h'_{s-2},h_{s-1}}\Delta_{h'_2,...,h'_{s-2},h_{s-1}}\left(F(x+h'_1)\overline{F(x+h''_1)}\right)\\
\nonumber
&e_p\left(\sum_{i=2}^{s-2}(\psi_i(h'_1,h'_2,...,h'_{s-2})-\psi_i(h''_1,h'_2...,h'_{s-2}))h_{s-1}\right))^{\frac{1}{2}}.\end{aligned}$$ After repeatedly applying Cauchy-Schwarz in this manner, we get rid of all the phases and bound (\[Long expression 11\]) by $||F||^2_{U^{s-1}}$. This proves the lemma.
The second proof is simpler.
By $U^2$-inverse theorem, $||F||^2_{U^2}\leqslant\max\limits_{\alpha\in{\mathbb{F}}_p}|\hat{F}(\alpha)|$. By Lemma \[Vanishing of expression II\], this is $O(p^{-c})$ unless $\alpha = 0$, in which case $\hat{F}(\alpha)=\Lambda_m(f_0,f_1,..., f_{m+k-2})$. Thus, $$\begin{aligned}
||F||^2_{U^2} &\leqslant 1_{\alpha=0}|\Lambda_{m, P_m, ..., P_{m+k-2}}(f_0, ..., f_{m+k-1})| + O(p^{-c})\\
&\leqslant|\Lambda_m(f_0, ..., f_{m-1})|\prod_{j=m}^{m+k-2}|{\mathbb{E}}f_j| + O(p^{-c})\leqslant\min_{m\leqslant i\leqslant m+k-2}||f_j||_{U^1} + O(p^{-c})\end{aligned}$$ where the intermediate inequality follows from applying ${\mathcal{S}}(m,k-1)$. Taking square roots on both sides and applying Hölder’s inequality proves the lemma.
Next we prove Lemma \[$U^1$ control of the operator\] using the previous lemmas.
Take $s=s_0$ and $\beta$ for which Lemma \[PET induction\] holds. Using Lemma \[Hahn-Banach decomposition\], we decompose $f_{m+k-1}$ into $$\begin{aligned}
f_{m+k-1} = f_a + f_b + f_c\end{aligned}$$ with $||f_a||_{U^{s_0}}^*\leqslant p^{\delta_1}$, $||f_b||_{L^1}\leqslant p^{-\delta_2}$, $||f_c||_{L^\infty}\leqslant p^{\delta_3}$, $||f_c||_{U^{s_0}}\leqslant p^{-\delta_4}$, and split the $\Lambda$ operator accordingly. The values of the parameters $\delta_1,\delta_2,\delta_3,\delta_4$ have to satisfy (\[delta inequality\]) and will be determined later. The term involving $f_b$ is easy to bound using Hölder inequality $$\begin{aligned}
|\langle F,f_b\rangle|\leqslant||F||_{L^\infty}||f_b||_{L^1}\leqslant p^{-\delta_2}.\end{aligned}$$ The term involving $f_c$ can also be bounded from above provided $\delta_4$ is sufficiently large compared to $\delta_3$ $$\begin{aligned}
|\langle F,f_c\rangle| &=||f_c||_{L^\infty}\left|\left\langle F,\frac{f_c}{||f_c||_{L^\infty}}\right\rangle\right|\\
&\leqslant p^{\delta_3}\left(\left(\frac{p^{-\delta_4}}{p^{\delta_3}}\right)^\beta+ O(p^{-\beta})\right)\\
&\ll p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}\end{aligned}$$ where in the second inequality we are using Lemma \[PET induction\]. Finally, the term involving $f_a$ can be bounded using dual inequality $$\begin{aligned}
|\langle F,f_a\rangle| &\leqslant||f_a||^*_{U^{s_0}}||F||_{U^{s_0}}\leqslant p^{\delta_1}||F||_{U^{s_0}}.\end{aligned}$$ Using the decomposition, we obtain the following bound on $\Lambda$ in terms of the $U^{s_0}$ norm of the dual function $F$ $$\begin{aligned}
|\Lambda(f_0,...,f_{m+k-1})| &\leqslant|\langle F,f_a\rangle| + |\langle F,f_b\rangle| + |\langle F,f_c\rangle|\\
&\leqslant p^{\delta_1}||F||_{U^{s_0}}+p^{-\delta_2} +p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}.\end{aligned}$$ From Lemma \[degree lowering\] it follows that $$\begin{aligned}
||F||_{U^2}= \Omega(||F||_{U^{s_0}}^{2^{(s_0-2)(s_0+2)}})-O(p^{-c}).\end{aligned}$$ Let $s_1=(s_0-2)(s_0+2)$. We thus have that $$\begin{aligned}
|\Lambda(f_0,... ,f_{m+k-1})|\ll p^{\delta_1}||F||_{U^2}^{2^{-s_1}}+p^{\delta_1-2^{-s_1}c}+p^{-\delta_2}+ p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}.\end{aligned}$$ Using Lemma \[$U^1$ control of the dual\], we are able to establish a $U^1$ control by $f_m, ..., f_{m+k-2}$ $$\begin{aligned}
\label{expression in proof of lemma 9}
|\Lambda(f_0, ..., f_{m+k-1})| &\ll p^{\delta_1}\min_{m\leqslant i\leqslant m+k-2}||f_j||_{U^1}^{2^{-s_1-1}}+p^{\delta_1-2^{-s_1}c}+p^{-\delta_2}\\
\nonumber
&+ p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}.\end{aligned}$$ Let $c_0$ be the value of $c$ appearing in (\[expression in proof of lemma 9\]). Setting the values of the parameters to be $$\begin{aligned}
\delta_1 = 2^{-s_1}\frac{c_0}{2},\quad
\delta_2 = \beta2^{-s_1}\frac{c_0}{8},\quad
\delta_3 &= \beta2^{-s_1}\frac{c_0}{4},\quad \rm{and} \quad
\delta_4 = (1-\beta)2^{-s_1}\frac{c_0}{2}\end{aligned}$$ proves the lemma.
Proof of Lemma \[Vanishing of expression II\]
---------------------------------------------
The last step is to prove Lemma \[Vanishing of expression II\]. The proof of this lemma is quite lengthy, yet it requires few new ideas, as it closely resembles the proof of ${\mathcal{S}}(m,k)$ for $k>1$. Therefore we omit certain details which are identical to the details in the proof of ${\mathcal{S}}(m,k)$ for $k>1$.
Note also the order of induction: we use ${\mathcal{S}}(m,k-1)$ to derive Lemma \[Vanishing of expression II\] for $(m,k)$, and then we use the $(m,k)$ case of Lemma \[Vanishing of expression II\] to prove ${\mathcal{S}}(m,k)$.
We mimic the proof of ${\mathcal{S}}(m,k)$ for $k>1$ and split $f_{m+k-2}$ using the Hahn-Banach decomposition $$\begin{aligned}
f_{m+k-2} = f_a + f_b + f_c,\end{aligned}$$ with $||f_a||_{U^{s_0}}^*\leqslant p^{\delta_1}$, $||f_b||_{L^1}\leqslant p^{-\delta_2}$, $||f_c||_{L^\infty}\leqslant p^{\delta_3}$, $||f_c||_{U^{s_0}}\leqslant p^{-\delta_4}$, where $s_0$ is the value of $s$ coming from Lemma \[PET induction\]. We extend the definition of $P_j$ to $0\leqslant j\leqslant m-1$ by setting $P_j(y)=jy$ for these values, and let $Q_j=P_j - P_{m+k-2}$ for $0\leqslant j\leqslant m+k-1$. Then we define the dual function by $$\begin{aligned}
F(x) :={\mathbb{E}}_y\left(\prod_{j=0}^{m+k-2}f_j(x+Q_j(y))\right)e_p(\alpha(x+Q_{m+k-1}(y))),\end{aligned}$$ so that $$\begin{aligned}
\langle F, f_{m+k-2}\rangle = \Lambda(f_0, ..., f_{m+k-2}, e_p(\alpha\cdot)).\end{aligned}$$ The contributions coming from $f_2$ and $f_3$ are bounded by $$\begin{aligned}
|\langle F,f_b\rangle| \leqslant p^{-\delta_2},\quad \rm{and} \quad |\langle F,f_c\rangle| \ll p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}.\end{aligned}$$ We bound the term involving $f_a$ using the dual inequality $$\begin{aligned}
|\langle F,f_a\rangle| \leqslant||f_a||^*_{U^{s_0}}||F||_{U^{s_0}}\leqslant p^{\delta_1}||F||_{U^{s_0}}.\end{aligned}$$
As in Lemma \[degree lowering\], we show that for $s>2$, the $U^s$ norm of $F$ is controlled by its $U^{s-1}$ norm, from which it follows that $||F||_{U^{s_0}}$ is controlled by $||F||_{U^2}$. The proof proceeds very much the same way. From the definition of $U^{s}$ norm it follows that $$\begin{aligned}
\eta:=||F||_{U^{s}}^{2^{s}}={\mathbb{E}}_{{{\underline{h}}}}||\Delta_{{\underline{h}}}F||_{U^2}^4\leqslant{\mathbb{E}}_{{{\underline{h}}}}\left|\widehat{\Delta_{{\underline{h}}}F}(\phi({{\underline{h}}}))\right|^2\end{aligned}$$ where $\widehat{\Delta_{{\underline{h}}}F}(\phi({{\underline{h}}}))$ is the largest Fourier coefficient of $\Delta_{{\underline{h}}}F$ and ${{\underline{h}}}=(h_1, ..., h_{s-2})$. If $$H_1:=\{{{\underline{h}}}:||\Delta_{{\underline{h}}}F||_{U^2}^4\geqslant\frac{1}{2}\eta\}$$ then the exact same analysis as applied in Lemma \[degree lowering\] shows that $$\begin{aligned}
\label{important intermediate expression 2}
&{\mathbb{E}}_{{{\underline{h}}},{{\underline{h}}}'}|{\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m+k-3} g_j(x+P_j(y))\right)e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')(x+P_{m+k-2}(y))\right)\\
\nonumber
&\left(\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}}{\mathcal{C}}^{|w|}e_p(aP_{m+k-1}(y))\right)
|^2 {1}_{\square(H_1)}({{\underline{h}}},{{\underline{h}}}')\geqslant \left(\frac{\eta}{2}\right)^{2^{s-1}}\end{aligned}$$ where $$\begin{aligned}
g_j(t):=\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}}{\mathcal{C}}^{|w|}f_j(t+{{\underline{w}}}\cdot{{\underline{h}}}^{({{\underline{w}}})})\quad {\rm{and}} \quad
\psi({{\underline{h}}},{{\underline{h}}}'):=\sum_{{{\underline{w}}}\in\{0,1\}^{s-2}}(-1)^{|w|}\phi({{\underline{h}}}^{({{\underline{w}}})}).\end{aligned}$$ The only difference between (\[important intermediate expression 2\]) and (\[important intermediate expression\]) is that in (\[important intermediate expression\]), the exponential phase $e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')\cdot\right)$ was weighting $x+ P_{m+k-1}$. By contrast, in (\[important intermediate expression 2\]) the exponential phase $e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')\cdot\right)$ is weighting $x+P_{m+k-2}$ whereas the polynomial $x+P_{m+k-1}$ is weighted by $e_p(a\cdot)$. Modulo that small difference, (\[important intermediate expression 2\]) and (\[important intermediate expression\]) are derived in an identical manner.
The crucial simplification comes from the fact that the product $$\begin{aligned}
\prod_{{{\underline{w}}}\in\{0,1\}^{s-2}}{\mathcal{C}}^{|w|}e_p(aP_{m+k-1}(y))\end{aligned}$$ is equal to 1 for $s>2$, and so we have obtained that $$\begin{aligned}
\left(\frac{\eta}{2}\right)^{2^{s-1}}&\leqslant{\mathbb{E}}_{{{\underline{h}}},{{\underline{h}}}'}{1}_{\square(H_1)}({{\underline{h}}},{{\underline{h}}}')|{\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m+k-3} g_j(x+P_j(y))\right)e_p \left(\psi({{\underline{h}}},{{\underline{h}}}')(x+P_{m+k-2}(y))\right)|^2. \end{aligned}$$ To evaluate the inner sum, we apply ${\mathcal{S}}(m,k-1)$; this tells us that the inner sum is $O(p^{-c})$ for some $c>0$ unless $\psi({{\underline{h}}},{{\underline{h}}}')=0$ because the $U^1$ norm of $e_p(\psi({{\underline{h}}},{{\underline{h}}}')\cdot)$ vanishes for a nonzero $\psi({{\underline{h}}},{{\underline{h}}}')$. As in Lemma \[degree lowering\], the set $$\begin{aligned}
H_2:=\left\{({{\underline{h}}},{{\underline{h}}}')\in \square(H_1): \psi({{\underline{h}}},{{\underline{h}}}') = 0\right\}\end{aligned}$$ has size at least $$\begin{aligned}
\left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O\left(p^{-c}\right)\right)p^{2(s-2)}\end{aligned}$$ and there exists ${{\underline{h}}}\in H_1$, which we fix, such that the fiber $H_3:=\{{{\underline{h}}}': ({{\underline{h}}},{{\underline{h}}}')\in H_2\}$ has size at least $$\begin{aligned}
\left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O(p^{-c})\right)p^{s-2}.\end{aligned}$$ Thus, the expression ${\mathbb{E}}_{{{\underline{h}}}'}||\Delta_{{{\underline{h}}}'}F||_{U^2}^4{1}_{H_3}({{\underline{h}}}')$ is bounded from below by $$\frac{\eta}{2}\cdot \left(\left(\frac{\eta}{2}\right)^{2^{s-1}}-O(p^{-c})\right)\geqslant \left(\frac{\eta}{2}\right)^{2^{s}}-O(p^{-c}).$$
We want to bound it from above by $||F||^2_{U^{s-1}}$, which we do by mimicking again the argument from Lemma \[Vanishing of expression II\]. Using the fact that $\psi({{\underline{h}}}, {{\underline{h}}}')=0$ for ${{\underline{h}}}'\in H_3$, we can rewrite $$\begin{aligned}
\phi(h'_1,...,h'_{s-2})=\sum_{i=1}^{s-2}\psi_i({{\underline{h}}},{{\underline{h}}}')\end{aligned}$$ for all ${{\underline{h}}}'\in H_3$, where $$\begin{aligned}
\psi_i({{\underline{h}}},{{\underline{h}}}'):=(-1)^s\sum_{\substack{{{\underline{w}}}\in\{0,1\}^{s-2},\\ w_1 = ... = w_{i-1}=1,\\ w_i = 0}}(-1)^{|w|}\phi({{\underline{h}}}^{({{\underline{w}}})})\end{aligned}$$ does not depend on $h'_1, ..., h'_i$.
Using positivity, expanding the definition of Fourier transform and changing variables, we obtain that $$\begin{aligned}
\label{long expression 1 later}
\nonumber
{\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2{1}_{H_3}({{\underline{h}}}')
&\leqslant {\mathbb{E}}_{{{\underline{h}}}'}\left|\widehat{\Delta_{{{\underline{h}}}'}F}\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')\right)\right|^2\\
&={\mathbb{E}}_{x,{{\underline{h}}}',h_{s-1}}\Delta_{{{\underline{h}}}',h_{s-1}}F(x)e_p\left(\left(\sum_{i=1}^{s-2}\psi_i({{\underline{h}}}')h_{s-1}\right)\right).\end{aligned}$$
Applying the Cauchy-Schwarz inequality $s-2$ times to (\[long expression 1 later\]), we get rid of all the phases $\psi_i({{\underline{h}}}')$ and bound (\[long expression 1 later\]) by $||F||^2_{U^{s-1}}$. Thus, $$\begin{aligned}
||F||_{U^{s-1}}\gg ||F||_{U^s}^{2^{2s-1}}-p^{-c}\end{aligned}$$ for $s>2$, and so $$\begin{aligned}
||F||_{U^2} \gg ||F||_{U^{s_0}}^{2^{(s_0-2)(s_0+2)}} - p^{-c}.\end{aligned}$$ By the $U^2$ inverse theorem, $||F||_{U^2}\leqslant|\widehat{F}(b)|^\frac{1}{2}$ for some $b\in{\mathbb{F}}_p$. Expanding out, we see that $$\begin{aligned}
\widehat{F}(b) = {\mathbb{E}}_{x,y}\left(\prod_{j=0}^{m+k-3}f_j(x+P_j(y))\right)e_p(a P_{m+k-1}(y)+b P_{m+k-2}(y)).\end{aligned}$$ Unless $a=b=0$, the polynomial $a P_{m+k-1}(y)+b P_{m+k-2}(y)$ has degree at least $m$. Again, we are back in ${\mathcal{S}}(m,k-1)$, as we are dealing with $k-1$ polynomials $$\begin{aligned}
P_m, P_{m+1}, ..., P_{m+k-3}, a P_{m+k-1}+b P_{m+k-2}.\end{aligned}$$ From ${\mathcal{S}}(m,k-1)$ and $a\neq 0$ it follows that $|\hat{F}(b)|=O(p^{-c})$. Thus, $$\begin{aligned}
||F||_{U^{s_0}}\ll p^{-c 2^{-(s_0-2)(s_0+2)-1}}+p^{-c 2^{-(s_0-2)(s_0+2)}}\ll p^{-c}.\end{aligned}$$ Combining all the estimates, we have the bound $$\begin{aligned}
\label{long expression 2 later}
|\langle F, f_{m+k-2}\rangle|\ll p^{\delta_1 - c}+p^{-\delta_2} + p^{\delta_3(1-\beta)-\beta\delta_4}+p^{\delta_3-\beta}.\end{aligned}$$ Set $c_0$ to be the value of $c$ in (\[long expression 2 later\]). Taking $\delta_1 = \frac{c_0}{2}, \delta_2 = \beta \frac{c_0}{8}, \delta_3 = \beta \frac{c_0}{4}$ and $\delta_4 = (1-\beta)\frac{c_0}{2}$ gives a bound of the form $$\begin{aligned}
|\langle F, f_{m+k-2}\rangle|\ll p^{-c}\end{aligned}$$ as desired.
Upper bounds for subsets of ${\mathbb{F}}_p$ lacking arithmetic progressions with $k$-th power common differences {#section on upper bounds for subsets of finite fields lacking arithmetic progressions with restricted differences}
=================================================================================================================
We now switch gears, moving away from the progression (\[generalized union of AP and GP\]) towards arithmetic progressions with common difference coming from the set of $k$-th powers. In this section, we prove Theorem \[Sets lacking arithmetic progressions with $k$-th power differences\] assuming Theorem \[counting theorem for linear forms with restricted variables\]. The argument goes much the same way as deriving Theorem \[main theorem, general case\] from Theorem \[counting theorem for the more difficult configuration\].
First, we prove the following simple lemma which allows us to reduce to the case $k| p-1$.
\[Q\_k = Q\_k’\] Let $k\in{\mathbb{N}}_+$ and $Q_k$ be the set of $k$-th power residues in ${\mathbb{F}}_p$. Then $Q_k=Q_{\gcd(k,p-1)}$.
Since ${\mathbb{F}}_p^\times$ is a cyclic group under multiplication, we can write it as ${\mathbb{F}}_p^\times=\langle a| a^{p-1}=1\rangle$. Note that for each $k\in{\mathbb{N}}$, $Q_k$ and $Q_{\gcd(k,p-1)}$ are subgroups of ${\mathbb{F}}_p^\times$ of cardinality $\frac{p-1}{\gcd(k,p-1)}$, generated respectively by $a^k$ and $a^{\gcd(k,p-1)}$. The property $\gcd(k,p-1)|k$ moreover implies that $Q_k$ is a subgroup of $Q_{\gcd(k,p-1)}$, and so they must be equal.
The set of $k$-th powers in ${\mathbb{F}}_p$ is precisely $Q_k$, and by Lemma \[Q\_k = Q\_k’\] it is the same as the set $Q_{\gcd(k,p-1)}$. Therefore we can assume that $k$ divides $p-1$, otherwise we replace $k$ with $\gcd(k,p-1)$. Suppose $A\subset{\mathbb{F}}_p$ for $p\geqslant p_0$ of size $|A|=\alpha p$ lacks $m$-term arithmetic progressions with difference coming from the set of $k$-th powers. From Theorem \[counting theorem for linear forms with restricted variables\] it follows that $$\begin{aligned}
\label{Counting equation for APs}
&{\mathbb{E}}_{x,y}{1}_A(x){1}_A(x+y)...{1}_A(x+(m-1)y){1}_{Q_k}(y)\\
&=\frac{1}{k}{\mathbb{E}}_{x,y}{1}_A(x){1}_A(x+y)...{1}_A(x+(m-1)y)+O\left(p^{-c}\right).\nonumber\end{aligned}$$
Since $A$ lacks progressions with $k$-th power differences, the left-hand side of (\[Counting equation for APs\]) is 0, and so we have $$\begin{aligned}
\label{Counting equation for APs, II}
{\mathbb{E}}_{x,y}{1}_A(x){1}_A(x+y)...{1}_A(x+(m-1)y)=O\left(p^{-c}\right).\end{aligned}$$ Applying Lemma \[Averaging over progressions\] to (\[Counting equation for APs\]) gives $M^{-2}\ll p^{-c}$ where $M=M_m(\frac{1}{2}\alpha)$ and $M_m$ is the inverse function to $s_m$ on $(0,\alpha_0]$, $\alpha_0 = s_m(p_0)$. Since $M$ grows faster than polynomially in $\alpha^{-1}$ by Behrend’s construction [@behrend_1946], this gives $M_m\gg p^c$. Applying $s_m$ to both sides and noting that $s_m$ is decreasing, we obtain that $\alpha\leqslant 2s_m( Cp^c)$.
Counting theorem for the number of linear configurations in subsets of ${\mathbb{F}}_p$ with variables restricted to the set of $k$-th powers
=============================================================================================================================================
This section is devoted to the proof of Theorem \[counting theorem for linear forms with restricted variables\]. We will first show that without loss of generality, we can assume that $k_i$ divides $p-1$ for each $1\leqslant i\leqslant d$. This will simplify the notation in the rest of the argument.
\[Rewriting the counting operator\] We have $$\begin{aligned}
{\mathbb{E}}_{x_1,...,x_d}\prod\limits_{i=1}^m f_j(L_i(x_1^{k_1},...,x_d^{k_d}))&={\mathbb{E}}_{x_1,...,x_d}\prod\limits_{i=1}^m f_j(L_i(x_1^{k'_1},...,x_d^{k'_d}))\\
&={k'_1...k'_d}{\mathbb{E}}_{x_1,...,x_d}\prod\limits_{i=1}^m f_j(L_i(x_1,...,x_d))\prod_{i=1}^d 1_{Q_{k'_i}}(x_i)+O\left(p^{-1}\right)\end{aligned}$$ where $k_i':=\gcd(k_i,p-1)$ for each $1\leqslant i\leqslant d$.
By Lemma \[Q\_k = Q\_k’\], $Q_{k}=Q_{\gcd(k,p-1)}$ for each $k\in{\mathbb{N}}_+$. Therefore the set of $k_i$-th power residues agrees with the set of $k'_i$-th power residues for each $1\leqslant i\leqslant d$. Consequently, the set of tuples $$\{(x_1^{k_1},...,x_d^{k_d}): (x_1,...,x_d)\in{\mathbb{F}}_p^d\}$$ equals the set of tuples $$\{(x_1^{k'_1},...,x_d^{k'_d}): (x_1,...,x_d)\in{\mathbb{F}}_p^d\},$$ and moreover each tuple $(x_1^{k_1},...,x_d^{k_d})$ appears in ${\mathbb{F}}_p^d$ the same number of times as the tuple $(x_1^{k'_1},...,x_d^{k'_d})$. This implies the first equality, as the summations in both expressions are carried over the same sets of tuples the same number of times.
The second equality follows from the fact that each value of $y\in {\mathbb{F}}_p^\times$ equals $x_i^{k'_i}$ for precisely $k'_i$ different values of $x_i\in{\mathbb{F}}_p$. The error term $O\left(p^{-1}\right)$ corresponds to the cases when at least one of the variables $x_1,...,x_d$ is 0. Using union bound, there are at most $dp^{d-1}$ such cases, which together contribute at most $\frac{d}{p}$ to the expectation.
We thus assume for the rest of this section that $k_1, ..., k_d$ are coprime to $p-1$. With this assumption, we now describe a useful expression for ${1}_{Q_k}$ which is crucial in proving the error term in Theorem \[counting theorem for linear forms with restricted variables\]. Let $a$ be a generator for the multiplicative group ${\mathbb{F}}_p^\times$. Define the map $$\begin{aligned}
\chi_k: {\mathbb{F}}_p^\times &\to {\mathbb{C}}\\
a^l &\mapsto e_k(l).\end{aligned}$$ The function $\chi_k$ is thus a *multiplicative character of order $k$*, i.e. a group homomorphism from ${\mathbb{F}}_p^\times$ to ${\mathbb{C}}^\times$ satisfying $\chi_k^k=1$. We extend $\chi_k$ to ${\mathbb{F}}_p$ by setting $\chi_k(0)=0$. Then $\chi_k$ picks out $Q_k$, in the sense that $\chi_k(x)=1\iff$ $x\in Q_k$. Using the orthogonality of roots of unity, we can write $$\begin{aligned}
\label{sum of characters}
{1}_{Q_k}(x)=\frac{1+\chi_k(x)+\chi_k(x)^2+...+\chi_k(x)^{k-1}}{k}-\frac{1}{k}{1}_{\{0\}}(x).\end{aligned}$$
We now use (\[sum of characters\]) to replace each ${1}_{Q_{k_i}}$ by a sum of characters in (\[main equation\]). Using the multilinearity of the operator, we obtain a main term of the same form as in (\[main equation\]), which corresponds to the terms in (\[sum of characters\]) having ${1}_{Q_{k_i}}$ replaced by $\frac{1}{k_i}$. Terms where ${1}_{Q_{k_i}}$ is replaced by $\frac{1}{k_i}{1}_{\{0\}}(x)$ are of size $O\left(p^{-1}\right)$, and there is a bounded number of them. It remains to deal with the terms that contain some $\frac{\chi^j_k(x)}{k}$ with $j>0$ but have no $\frac{1}{k_i}{1}_{\{0\}}(x)$. Each such term is of the form $$\begin{aligned}
\label{error term}
{\mathbb{E}}_{x_1,...,x_d}\prod\limits_{i=1}^m f_j(L_i(x_1,...,x_d))\prod_{i\in S}\frac{\chi_{k_i}^{j_i}(x_i)}{k_i}\end{aligned}$$ for a nonempty $S\subset\{1\leqslant i\leqslant d: k_i>1\}$ and $1\leqslant j_i\leqslant k_i-1$. From the fact that $k_i$ divides $d$ it follows that $\chi_{k_i}^{j_i}$ is also a character of order $k_i$, so without loss of generality we can take $j_i=1$ for each $1\leqslant i\leqslant d$.
Green and Tao proved that linear forms $L'_1(x_1, ..., x_d)$, ..., $L'_m(x_1, ..., x_d)$ are controlled by a Gowers norm [@green_tao_2010; @tao_2012]: specifically, they showed that $$\begin{aligned}
\label{linear configurations 2}
\left|{\mathbb{E}}_{x_1,...,x_d}\prod_{j=1}^m g_j(L'_i(x_1,...,x_d))\right|\leqslant\min\limits_{1\leqslant j\leqslant m}||f_j||_{U^s}\end{aligned}$$ whenever for each $1\leqslant i\leqslant m$ one can partition $\{L'_j: j\neq i\}$ into $s+1$ classes such that $L'_i$ does not lie in the span of each of them. The lowest $s-1$ for which this is true is called *Cauchy-Schwarz complexity*, or *CS-complexity* of the system of linear forms $L'_1, ..., L'_m$. The only case when such $s$ may not exist is if two linear forms $L'_i$ and $L_j'$ are the same up to scaling. Otherwise we can partition linear forms into such classes: in the worst case, each of $\{L'_j: j\neq i\}$ forms a separate class, in which case the CS-complexity is $m-2$. This extreme case occurs in arithmetic progressions, for instance: the operator $$\begin{aligned}
{\mathbb{E}}_{x,y}f_0(x)f_1(x+y)...f_{m-1}(x+(m-1)y)\end{aligned}$$ is bounded by $||f_{i}||_{U^{m-1}}$ for each $0\leqslant i\leqslant m-1$, and the system of linear forms $\{x, x+y, ..., x+(m-1)y\}$ has CS-complexity $m-2$.
We assumed specifically that no two linear forms $L_i$, $L_j$ are scalar multiples, and that $L_i$ is never a scalar multiple of $e_j$. From these assumptions we obtain the following lemma, which is essentially a restatement of Green and Tao’s result tailored to our context.
\[bounding the error term\] For an arbitrary character $\chi_{k_i}$ of order $k_i$, we have the bound $$\begin{aligned}
\label{U^s control}
\left|{\mathbb{E}}_{x_1,...,x_d}\prod\limits_{j=1}^m f_j(L_i(x_1,...,x_d))\prod_{i\in S}\frac{\chi_{k_i}(x_i)}{k_i}\right|\leqslant\left(\prod_{i\in S}\frac{1}{k_i}\right)\min_{i\in S}||\chi_{k_i}||_{U^s}\end{aligned}$$ where $s-1$ is the CS-complexity of the system $$\begin{aligned}
\label{specific system of linear forms}
\{L_1, ..., L_m\}\cup\{x_j: j\in S\}\end{aligned}$$ In particular, one can take $s=m+|S|-1\leqslant m+d-1$.
By assumption, all forms in the system $$\begin{aligned}
\label{the largest system of linear forms}
\{L_1, ..., L_m\}\cup\{x_j: 1\leqslant j\leqslant d, k_j >1\} \end{aligned}$$ are pairwise linearly independent. Since (\[specific system of linear forms\]) is a subset of (\[the largest system of linear forms\]), all forms in (\[specific system of linear forms\]) are also pairwise linearly independent. Therefore the CS-complexity of this system is finite, and is at most $m+|S|-2$ because the system (\[specific system of linear forms\]) consists of $m+|S|-1$ linear forms.
It thus follows that the error term in (\[main equation\]) is controlled by Gowers norms of characters. The multiplicative property of characters makes it easy to bound their Gowers norms using tools such as Weil’s bound.
Let $\chi$ be a nonprincipal multiplicative character of ${\mathbb{F}}_p$ of order $k$, and let $P\in{\mathbb{F}}_p[x]$ be a polynomial with $r$ distinct roots in the splitting field. If $P$ is not a $k$-th power, then $$\begin{aligned}
\left|{\mathbb{E}}_x\chi(P(x))\right|\leqslant(r-1)q^{-\frac{1}{2}}.\end{aligned}$$
In particular, we use the following corollary, which is Corollary 11.24 in Iwaniec & Kowalski [@iwaniec_kowalski_2004].
\[Corollary to Weil’s bound\] Let $\chi$ be a nonprincipal multiplicative character of ${\mathbb{F}}_p$, and let $b_1,...,b_{2r}\in{\mathbb{F}}_p$. If one of them is different from others, then $$\begin{aligned}
\left|{\mathbb{E}}_x\chi((x-b_1)...(x-b_r))\overline{\chi}((x-b_{r+1})...(x-b_{2r}))\right|\leqslant2r p^{-\frac{1}{2}}.\end{aligned}$$
\[Gowers norms of characters\] If $\chi$ is a multiplicative character of ${\mathbb{F}}_p$ of order $k$ and $s$ is a natural number, then $$\begin{aligned}
||\chi||_{U^s}\leqslant 2p^{-2^{-(s+1)}}.\end{aligned}$$
By definition, the $U^s$ norm of $\chi$ is given by the following expression $$\begin{aligned}
||\chi||_{U^s}^{2^s} &= {\mathbb{E}}_{h_1,...,h_s}{\mathbb{E}}_x \prod_{{{\underline{w}}}\in\{0,1\}^s}C^{|w|}\chi(x+{{\underline{w}}}\cdot {{\underline{h}}})\\
&= {\mathbb{E}}_{h_1,...,h_s}{\mathbb{E}}_x \chi\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{even}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\overline{\chi}\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{odd}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\\
&\leqslant{\mathbb{E}}_{h_1,...,h_s}\left|{\mathbb{E}}_x \chi\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{even}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\overline{\chi}\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{odd}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\right|.\end{aligned}$$ If ${{\underline{w}}}\cdot {{\underline{h}}}$ are not all equal, then by Lemma \[Corollary to Weil’s bound\] we have $$\begin{aligned}
\left|{\mathbb{E}}_x \chi\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{even}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\overline{\chi}\left(\prod_{{{\underline{w}}}\in\{0,1\}^s, |w|\;\rm{odd}}(x+{{\underline{w}}}\cdot {{\underline{h}}})\right)\right|\leqslant 2^s p^{-\frac{1}{2}}.\end{aligned}$$
The only possibility for ${{\underline{w}}}\cdot {{\underline{h}}}$ being equal for all ${{\underline{w}}}\in\{0,1\}^s$ is when $h_1=...=h_s=0$, which happens with probability $p^{-s}$. Thus $$\begin{aligned}
||\chi||_{U^s}^{2^s}\leqslant 2^s p^{-\frac{1}{2}}+p^{-s}\end{aligned}$$ and so $$\begin{aligned}
||\chi||_{U^s}\ll p^{-2^{-(s+1)}}.\end{aligned}$$
Applying the results of Lemma \[Gowers norms of characters\] to Lemma \[bounding the error term\], we see that the error term in (\[main equation\]) is of the size $O\left(p^{-c}\right)$, which proves Theorem \[counting theorem for linear forms with restricted variables\].
Further discussion
==================
There are many directions in which one could try to extend the results of this paper, in particular Theorem \[counting theorem for the more difficult configuration\]. One of the questions one might ask is whether there is a discorrelation result for progressions of the form $$\begin{aligned}
x, x+Q(y), ..., x+(m-1)Q(y), x+P_m(y), ..., x+P_{m+k-1}(y)\end{aligned}$$ where $Q$ has degree greater than 1 while $P_m, ..., P_{m+k-1}$ are linearly independent and presumably satisfy a further technical assumption of algebraic independence similar to one in Theorem \[counting theorem for the more difficult configuration\]. Combining methods used in the proofs of Theorems \[counting theorem for the more difficult configuration\] and \[counting theorem for linear forms with restricted variables\], one can easily derive a statement of the form:
\[counting theorem for the more difficult configuration with y\^k\] Let $m,k,l\in{\mathbb{N}}_+$ and $P_m$, ..., $P_{m+k-1}$ be polynomials in ${\mathbb{Z}}[y]$ such that $$a_m P_m + ... + a_{m+k-1}P_{m+k-1}$$ has degree at least $m$ unless $a_m = ... = a_{m+k-1}=0$ (in particular, $P_m$, ..., $P_{m+k-1}$ are linearly independent and each of them has degree at least $m$). Suppose $f_0, ..., f_{m+k-1}$ are 1-bounded functions from ${\mathbb{F}}_p$ to ${\mathbb{C}}$. Then $$\begin{aligned}
&{\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy^l)\prod_{j=m}^{m+k-1}f_j(x+P_j(y^l))\\
\nonumber
&=\left({\mathbb{E}}_{x,y}\prod_{j=0}^{m-1}f_j(x+jy)\right)\prod_{j=m}^{m+k-1}{\mathbb{E}}f_j + O(p^{-c})\end{aligned}$$ where all the constants are positive and depend on $m, k, l$ and polynomials $P_m, ..., P_{m+k-1}$ but not on $f_0, ..., f_{m+k-1}$.
This is a version of Theorem \[counting theorem for the more difficult configuration\] where variable $y$ is restricted to lie in the set of $l$-th powers. It essentially says that restricting the variables to the set of $l$-th powers does not matter. For instance, this theorem allows us to prove that a set $A\subset{\mathbb{F}}_p$ lacking progressions of the form $$\begin{aligned}
x, x+y^l, ..., x+(m-1)y^l, x+y^{ml}, ..., x+y^{(m+k-1)l}\end{aligned}$$ has size at most $$|A|\ll\begin{cases}
p^{-c},\; &m = 1,2,\\
p\frac{(\log \log p)^4}{\log p},\; &m = 3,\\
\frac{p}{(\log p)^{c}},\; &m = 4,\\
\frac{p}{(\log\log p)^{c}}, \; &m>4
\end{cases}$$ where the implied constant depends on $k, m, l$ and $c$ depends on $m$ only. Note that the bounds here are of the same shape as the bounds in Theorem \[main theorem, special case\]: this is because the proof of this corollary is identical to the proof of Theorem \[main theorem, general case\].
The drawback of this theorem is that it essentially only works for polynomials $P_m', ..., P'_{m+k-1}$ that can be expressed as polynomials in $y^l$, i.e. $P'_i(y)=P_j(y^l)$ for some $P_j$. For instance, it allows us to handle $$x, x+y^2, x+2y^2, x+y^6$$ but not $$\begin{aligned}
x,\; x+y^2,\; x+2y^2,\; x+y^5 \quad {\rm{or}} \quad x,\; x+y^2,\; x+2y^2,\; x+y^5.\end{aligned}$$ Replacing $P_m(y^l), ..., P_{m+k-1}(y^l)$ in the statement of the theorem by $P_m(y)$, ..., $P_{m+k-1}(y)$ would require a completely different approach. We have an argument that would allow us to replace $P_j(y^l)$ by $P_j(y)$ for $m=3$ and possibly $m=4$, however it has two serious downsides. First, the argument only works if the minimal degree of $P_j$’s is unreasonably large depending on $m$ and $l$ - it in fact would have to be greater than the minimal value $s$ obtained by applying Lemma \[PET induction\] to $x,\; x+y^l,\; ...,\; x+(m-1)y^l,\; x+P_m(y),\; ...,\; x+P_{m+k-1}(y)$, which has rather poor dependence on $m$ and degrees of $P_m, ..., P_{m+k-1}$. Second, the method does not generalize to higher $m$ without resorting to higher order Fourier analysis. For this reason, we do not present this argument here, hoping to find a more robust version of it in the future.
[alpha]{}
J. Bourgain and M.-C. Chang. . , **221**:853–867, 2017.
F. Behrend. On sets of integers which contain no three terms in arithmetical progression. , **32**:331–2, 1946.
V. Bergelson and A. Leibman. . , **9**:725–753, 1996.
T. Bloom. . , **93**:643–663, 2016.
A. Balog, J. Pelikán, J. Pintz, and E. Szemerédi. . , **65**(2):165–187, 1994.
D. Dong, X. Li, and W. Sawin. . 2017.
J. Ellenberg and D. Gijswijt. . , **185**:339–343, 2016.
W. T. Gowers. . , **11**(3):465–588, 2001.
W. T. Gowers. . , **42**(4):573–606, 2010.
S. W. Graham and C. J. Ringrose. . , **85**:269–309, 1990.
B. Green. . 2007.
B. Green and T. Tao. Linear equations in primes. , **171**:1753–1850, 2010.
B. Green and T. Tao. . , **63**(3):944–1040, 2017.
W. T. Gowers and J. Wolf. . , **21**:36–69, 2011.
W. T. Gowers and J. Wolf. . , **57**:215–237, 2011.
W. T. Gowers and J. Wolf. . , **115**(1):121–186, 2011.
H. Iwaniec and E. Kowalski. . AMS, 2004.
J. Lucier. Intersective sets given by a polynomial. , **123**:57–95, 2006.
S. Peluse. Three-term polynomial progressions in subsets of finite fields. , **228**:379–405, 2018.
S. Peluse. . , 2019.
S. Peluse and S. Prendiville. . 2019.
S. Prendiville. . , **5**, 2017.
A. Rice. . , **187**:1–41, 2019.
I. Ruzsa. Difference sets without squares. , **15**(3):205–209, 1984.
L. Rimanić‡ and J. Wolf. . , **62**:443–457, 2019.
A. [Sárközy]{}. . , **31**(1-2):125–149, 1978.
A. [Sárközy]{}. . , **31**:355–386, 1978.
S. Slijepcević. . , **98**(1-2):111–128, 2003.
E. Szemerédi. On sets of integers containing $k$ elements in arithmetic progression. , **27**(1):199–245, 1975.
T. Tao. . AMS, 2012.
T. Tao and V. Vu. . Cambridge Studies in Advanced Mathematics. Cambridge U. P., 2006.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We examine the evolution of a closed, homogeneous and anisotropic cosmology subject to a variation of the fine structure ’constant’, $\alpha $ within the context of the theory introduced by Bekenstein, Sandvik, Barrow and Magueijo, which generalises Maxwell’s equations and general relativity. The variation of $\alpha $ permits an effective ghost scalar field, whose negative energy density becomes dominant at small length scales, leading to a bouncing cosmology. A thermodynamically motivated coupling which describes energy exchange between the effective ghost field and the radiation field leads to an expanding, isotropizing sequence of bounces. In the absence of entropy production we also find solutions with stable anisotropic oscillations around a static universe.'
author:
- 'John D. Barrow$^1$'
- David Sloan$^1$
title: Bouncing Anisotropic Universes with Varying Constants
---
Introduction
============
Spatially homogeneous cosmological models are a key area of study within relativity. The introduction of anisotropies gives rise to models in which a richer dynamical structure emerges, yet the cosmology remains simple enough to provide analytic and simple numerical results. These models serve as a test-bed for physical theories, and allow us analyse questions about why the universe appears to be highly isotropic, whether inflation occurs for generic or stable sets of initial data, the effects of anisotropy on astronomical observables, and the behaviour of cosmological models on approach to spacetime singularities [@Rendall:2005nf], [Andersson:2000cv]{}, [@Garfinkle:2012zh].
The idea that the fine structure constant, $\alpha $, is a spacetime varying scalar field was first investigated by Bekenstein [@Bekenstein:1982eu], who created a natural generalisation of Maxwell’s equations to accommodate a varying electron charge. This idea was extended to include gravity and provide a theory to explore cosmological consequences of varying $\alpha $ by Sandvik et al [@Sandvik:2001rv]. The resulting Bekenstein-Sandvik-Barrow-Magueijo (BSBM) isotropic cosmological models were found and used in conjunction with the astronomical data on varying $\alpha $ obtained from observations of high redshift quasar spectra [@Webb:2010hc]. More recently, the BSBM theory has been extended to included the case where there is a coupling function (rather than simply a coupling constant) between the charged matter fields and the scalar field driving changes in $\alpha $ [@Barrow:2011kr] and where that scalar field possesses a self-interaction potential [@Barrow:2008ju]. These theories are the analogues of the Jordan-Brans-Dicke theories for varying $G$ [Brans:1961sx]{}.
In [@Barrow:2004ad] it was shown how theories of this type could produce singularity-free homogeneous and isotropic cosmologies which displayed stable oscillations around an Einstein static universe because the effect of variations in the scalar field driving variations in $\alpha $ is to introduce a negative ’ghost’ density. Barrow and Tsagas [@Barrow:2009sj] considered a broader context for these solutions and showed how the inclusion of simple anisotropic expansion can modify the results because the anisotropy can diverge just as quickly as a bounce-producing ghost scalar field on a approach to the singularity. In this paper we will consider more general closed anisotropic cosmologies with anisotropic 3-curvature in this same context.
Matter bounces introduced by the presence of ghost fields are not a new discovery (for a detailed examination see [@Gibbons:2003yj]). However, in BSBM models the ghost field is an effective manifestation of underlying physics, not a new matter source introduced by hand. In such models quantum effects are ignored because of the prevailing attitude that ghost fields should not be quantized (and are in fact ill-behaved when quantized, with negative probability states). Furthermore, any coupling between a ghost field and a non-ghost field would allow an infinite amount of energy to be transferred from the ghost field.
We simply take the view that BSBM models can serve as test models for bouncing cosmologies. The idea of a phoenix universe within relativity is almost as old as big-bang models themselves, and goes back to Tolman [@Tolman] and Lemaître [@Lemaitre]. This classical picture of oscillating closed universes with zero value of the cosmological constant, $\Lambda $, painted by Tolman is well known. If there is no entropy production then cycles for the time-evolution of the scale factor are periodic with the same amplitude and total lifetime. If entropy increase is introduced in accord with the second law of thermodynamics then the oscillating cycles become larger and longer to the infinite future. The classical picture for isotropic universes was competed by Barrow and Dabrowski [@BarrowDabrowski], who showed that if a positive cosmological constant is included then the sequence of growing cycles will always come to an end, no matter how small the value of $\Lambda >0$. The ensuing behaviour will be to approach de Sitter expansion. If the entropy increase from cycle to cycle is small then the asymptotic state will be one in which the expansion is very close to a zero-curvature state with comparable energy densities associated with matter and dark energy (ie the cosmological constant). The dark energy will necessarily be slightly dominant and the curvature will be positive – not unlike the situation in our observed universe. Barrow and Dabrowski [BarrowDabrowski]{} also considered the evolution of some simple bouncing anisotropic universes of Kantowski-Sachs type, but not in a context that included varying constants.
Many current quantum theories of gravity exhibit curvature singularity avoidance, often in the form of a bounce (although they do not necessarily avoid geodesic incompleteness). In Loop Quantum Cosmology, holonomy corrections to the Friedmann equation give rise to a bounce at Planck scales (see [@Ashtekar:2011ni] for a review). Horava-Lifschitz gravity also introduces higher-order curvature corrections to Einstein’s equations which can cause the universe to bounce [@Brandenberger:2009yt] for some parameter choices. In the latter case the dynamics of an anisotropic solution have also been explored [@Misonoh:2011mn].
The aim of this paper is to extend [@Barrow:2004ad] and [BarrowDabrowski]{} to spatially homogeneous models which exhibit local rotational symmetry (LRS) an so are effectively axisymmetric. LRS models exhibit some of the features of full anisotropic model [@Calogero:2009mi], yet the differential equations governing their dynamics can be solved with relative ease using numerical solvers and exact methods. We will begin in section \[PhysVar\] by setting our the action principle underlying variation of the fine structure ’constant’, then in section \[CosSetup\] we set out the equations of motion for our system and define quantities of physical interest, such as shear and Hubble expansion rates, in terms of metric variables. Section \[Solutions\] deals with two specific solutions to the equations of motion: a static solution and ghost-induced inflation. In particular, we will focus on the role played by anisotropies in both these cases and examine perturbations about isotropic cases.
Theories of Varying Alpha {#PhysVar}
=========================
Varying constants can be described by extensions of the standard model of particle physics and/or general relativity (GR) by the promotion of constants to space and time dependent scalar fields. A well known and much studied example is that of Jordan-Brans-Dicke theory in which GR is extended by generalising Newton’s constant $G$ to become a field variable [@Brans:1961sx]. These self-consistent models for the variation of constants necessarily contain conservation equations for the energy and momentum carried by the varying scalar field and the gravitational field equations account for the scalar field’s effect on the spacetime geometry. This is in contrast to much of the old literature on varying constants, other than $G,$ which merely ’write-in’ variations of constants into the equations that hold in the theory where the constant does not vary. The existence of a self-consistent theory for the variation of a constant also shows that much discussion about the meaning of the variation of dimensional constants is not relevant because the solution of the second-order conservation equation for the scalar field describing the variation of a traditional constant always produces constants of integration with the same dimensions as the varying constant and a dimensionless combination is trivially available.
Physical models with extra dimensions often exhibit massless or light degrees of freedom which can lead to the variation of such constants [Damour:2003iz]{}, [@Uzan:2002vq] and there is the possibility for observational bounds to be placed on any shift in the size of extra dimensions over the age of the universe [@Kolb:1985sj],[Barrow:2005sv]{}.
In this paper we shall follow the BSBM model in which the fine structure constant, $\alpha $ is taken to be dynamical. Evidence of a dipolar spatial variation has been recently claimed [@Webb:2010hc] and therefore it is natural to extend this scenario to consider space and time variations but in this paper we will only discuss time variations so that we can confine attention to ordinary differential equations. Such variations are bounded by terrestrial experiments to have small variation at present [Marion:2002iw]{} [@Berengut:2010ht]. However evidence that the variation is small currently does not rule out more significant changes in the past. In particular, in BSBM theories $\alpha $ is not expected to vary during the radiation era, to increase only logarithmically in time during the cold dark matter dominated era, and then to become constant after the universal expansion begins accelerating. Thus laboratory experiments today would not be expected to find evidence for the variation of $\alpha $ found in high-redshift quasar observations (that derive from epochs before the universe began accelerating) even though it has been proved that any cosmological variations in $\alpha $ will be seen in terrestrial experiments [@bshaw].
The BSBM model describes the effect of varying the fine structure constant by the introduction of a scalar dielectric field, $\psi $ with evolution of the charge of an electron given by $e=e_{0}e^{\psi }$, in which $e_{0}$ is the value of the electron charge at some fixed time, for example today. Notice that $e_{0}$ is a fundamental constant and $e/e_{0}$ is dimensionless. It has been shown [@Bekenstein:2009fq] that in spite of modifying black hole solutions, the variability of $\alpha $ respects the second law of thermodynamics. This will be important in section [CosSetup]{}, as we will assume that all couplings between our fields obey the second law.
The physical action is given by
$$\label{action}
S=\int \sqrt{-g} (\mathcal{L}_g +\mathcal{L}_m +L_\psi +e^{-2 \psi} \mathcal{L}_{em})$$
where $\mathcal{L}_{g}=R/16\pi G$ is the usual Einstein-Hilbert Lagrangian, $\mathcal{L}_{\psi }=-\frac{\omega }{2}\partial _{u}\psi \partial ^{u}\psi $ governs the scalar dielectric field, $\psi $, $\mathcal{L}_{em}=-\frac{1}{4}f_{\mu \nu }f^{\mu \nu },$ and $\mathcal{L}_{m}$ is a matter Lagrangian independent of $\psi $. Of particular importance to this paper is the constant coupling parameter $\omega $ which we shall take to be negative, so rendering $\psi $ an effective ghost scalar field. We do not consider the generalised case where $\omega ^{\prime }(\psi )\neq 0$, see [Barrow:2011kr]{}. From now on it will be convenient to simply consider the effective field, rather than the underlying dielectric. This field is massless, and it is clear from the action that its motion will be monotonic, as will be that of the effective induced fine structure constant. In terms of fluids, this field will appear to be stiff, with equation of state $p_{\psi }=\rho _{\psi }<0$. In a closed anisotropic cosmological model we will allow energy exchange to occur between the $\psi $ field and an equilibrium radiation field with equation of state $3p_{r}=\rho _{r}$, to model an entropy increasing non-equilibrium process.
Cosmological Expansion {#CosSetup}
======================
The physical system under consideration will consist of a homogeneous anisotropic cosmology. For simplicity we will examine a system which is locally rotationally symmetric, and use this to gain insight into the more general case. For a concise review of these cosmological models, see [Calogero:2009mi]{}. This model is general enough to contain the purely general relativistic ingredient of anisotropic 3-curvature, which is missing from the simple anisotropic models of Bianchi types I and V. It includes the closed Bianchi type IX universe but only in the axisymmetric case where no chaotic behaviour occurs. The LRS type IX metric is * *
$$ds^{2}=dt^{2}-h_{ij}\sigma ^{i}\sigma ^{j}$$
where $\sigma ^{i}$ are the $SO(3)$ invariant 1-forms [@Taub:1951]
$$\begin{aligned}
\sigma ^{1} &=&\cos \psi d\theta +\sin \psi \sin \theta d\phi \\
\sigma ^{2} &=&-\sin \psi d\theta +\cos \psi \sin \theta d\phi \\
\sigma ^{3} &=&d\psi +\cos \theta d\phi\end{aligned}$$
and the LRS condition requires
$$h_{ij}=diag\{a(t),b(t),b(t)\}$$
The metric contains two time-dependent scale factors (due to the LRS condition), $a(t)$ and $b(t)$. The energy densities are denoted by $\rho _{r}
$ for radiation, $\rho _{\psi }$ for the scalar field, and $\rho _{\Lambda }$ for the cosmological constant, and the total density and pressure are $\rho $ and $p$, where
$$\rho =\rho _{r}+\rho _{\psi }+\rho _{\Lambda }.$$
All matter matter sources have isotropic pressures. The independent variables are the principal 3-curvatures
$R_{1}^{1}=\frac{a^{2}}{2b^{4}},$
$R_{2}^{2}=\frac{1}{b^{2}}-\frac{a^{2}}{2b^{4}};$
the mean Hubble expansion rate is defined by
$H=\frac{1}{3}(\frac{\dot{a}}{a}+2\frac{\dot{b}}{b}),$
and the expansion shear scalar by
$\sigma =\frac{1}{3}(\frac{\dot{b}}{b}-\frac{\dot{a}}{a})$
These variables are subject to a constraint equation (the generalized Friedmann equation with $8\pi G=c=1$)
$$\rho =\frac{1}{b^{2}}-\frac{a^{2}}{4b^{4}}+2\frac{\dot{a}\dot{b}}{ab}+\frac{\dot{b}^{2}}{b^{2}}. \label{Constraint}$$
The remaining field equations are:
$$\begin{aligned}
\frac{\ddot{a}}{a} &=&-\frac{1}{2b^{4}}-2\frac{\dot{a}\dot{b}}{ab}+\frac{\rho -p}{2}, \label{HEOM} \\
\frac{\ddot{b}}{b} &=&\frac{a^{2}}{2b^{4}}-\frac{1}{b^{2}}-\frac{\dot{a}\dot{b}}{ab}-\frac{\dot{b}^{2}}{b^{2}}+\frac{\rho -p}{2}, \label{bEOM} \\
\dot{\sigma} &=&-3H\sigma +\frac{1}{3}(R_{1}-R_{2}), \label{3} \\
\dot{H} &=&-H^{2}-2\sigma ^{2}-\frac{1}{6}(\rho +3p).\end{aligned}$$
These reduce to the special case of the closed Friedmann universes when $a=b$. The shear does not evolve with $\sigma \propto (ab^{2})^{-1}$ as in Bianchi type I models because of the 3-curvature anisotropy on the right-hand side of eq. (\[3\]). For ease of exposition, let us define the variables $r=\frac{a}{b}$ and $H_{b}=\dot{b}/b.$The essential field equations then simplify to
$$\frac{\ddot{r}}{r}=\frac{1-r^{2}}{b^{2}}-\frac{3\dot{r}\dot{b}}{rb},
\label{rEOM}$$
$$\begin{aligned}
\sigma &=&-\frac{\dot{r}}{3r}, \label{HBEOM} \\
H &=&\frac{\dot{b}}{b}-3\sigma =H_{b}-3\sigma . \label{H}\end{aligned}$$
The continuity equation, which implies constraint conservation, is
$$\dot{\rho}+3H(\rho +p)=0.$$
This governs the total energy density and pressure. Now, we introduce some energy exchange between the fluids so that they obey
$$\begin{aligned}
\dot{\rho _{\psi }} &+&6H\rho _{\psi }=s, \label{radEOM} \\
\dot{\rho _{r}} &+&4H\rho _{r}=-s,\end{aligned}$$
where, $s$ parametrises the flow of energy between the scalar field and radiation. For our purposes, this will be taken to be of the form
$s=-\rho _{\psi }\beta $
where
$\beta =\beta _{0}+\beta _{H}H^{2}+\beta _{\sigma }\sigma ^{2}$
can include a linear coupling $\beta _{0}$, plus possible bulk,* *$\beta _{H}$, and shear, $\beta _{\sigma }$, viscous contributions. In general, the $\beta $’s need not be constants.
The scalar dielectric field evolves according to
$$\ddot{\psi}+(3H+\beta )\dot{\psi}=0 \label{SFEOM}$$
and so
$$\dot{\psi}\propto a^{-3}\exp [-\int \beta dt].$$
Solutions {#Solutions}
=========
In what follows we consider the Bianchi IX case with no cosmological constant ($\rho _{\Lambda }=0$). If the matter content is a perfect fluid and obeys $\rho +3p>0$, then these universes expand from an initial curvature singularity to a maximum size before collapsing back to a future curvature singularity; the spacetime is past and future geodesically incomplete. However, since the effect of varying the fine structure constant is to produce a ghost scalar field with $\rho _{\psi }<0$ which will dominate dynamics at small length scales, the energy condition is violated and solutions to the BSBM model exist which have infinite past and future temporal range.
In this section we will examine two particular solutions to the equations of motion. The first is that of a static spacetime, and its behaviour under perturbations. The second is that of an spacetime in which the coupling between fields leads to inflationary behaviour.
The Static Solution
-------------------
There exists a static solution of the form
$$\rho _{\psi }=-\frac{3}{4b^{2}} \:\:\: \rho _{r}=\frac{3}{2b^{2}},$$
for any given value of $b$. Note that in order to be static the solution must be isotropic ($\sigma =0$), since from eq. (\[3\]) the 3-curvatures must match ($R_{1}^{1}=R_{2}^{2}$) for the shear to remain constant and eq. (\[H\]) requires $\sigma =0$.
Now consider making a small perturbation about the isotropic solution by introducing a small anisotropy: $r=1+\epsilon $ where $r=1$ represents isotropy. From \[rEOM\], we find that
$$\ddot{\epsilon}=-3\epsilon H_{b}+\frac{2\epsilon +3\epsilon ^{2}+\epsilon
^{3}}{b^{2}}. \label{epEOM}$$
Without loss of generality, we take the unperturbed static solution to have $b=1$. If we introduce small parameters $\delta (t)$ and $\eta (t)$ such that $b=1+\eta $, and $\rho =3/2+\delta $, then to first order in our small parameters:
$$\epsilon (t)=\epsilon _{0}\sin (\sqrt{2}t). \label{epStatic}$$
![$\protect\epsilon $ versus time with no coupling ($s=0$). Initial values: $r=1.05,\protect\sigma =0,H=0.$[]{data-label="Epsoscill"}](Epstnocouple.jpg){width="50.00000%"}
In the case where there is no coupling between the fields ($s=0$), these oscillations continue endlessly as shown in figure \[Epsoscill\]. However, once coupling is introduced ($s\neq 0$), the static case becomes unstable because the balance between $\rho _{\psi }$ and $\rho _{r}$ is broken, energy is transferred from the ghost field that supports stable oscillations and eventually, after several oscillations, it settles into radiation-dominated expansion, shown in figure \[Epsdecay\]:
![$\protect\epsilon $ versus time with coupling turned on, $s\neq 0$, showing the system isotropizing. Initial values: $r=1.05,\protect\sigma =0,H=0,\protect\beta _{0}=0.05$, $\protect\beta _{H}=\protect\beta _{\protect\sigma }=0.$[]{data-label="Epsdecay"}](Epstcouple.jpg){width="50.00000%"}
From the constraint equation \[Constraint\], we find that to first order, our small parameters are related by $2\delta =-\epsilon -3\eta $. Furthermore, we can decompose $\delta $ into the radiation and scalar field components, $\delta _{r}$ and $\delta _{\psi }$. In the absence of field couplings, there is a relationship between these fields, due to their coupled equations of motion \[SFEOM\], \[radEOM\]:
$$\rho _{r}\propto \rho _{\psi }^{3/2}. \label{MattRelations}$$
In the static case under consideration, the constant of proportionality is $-\sqrt{6}$. Note that this relationship is broken by introducing a coupling between the fields. For small $\delta $ we are therefore led to $\delta
_{r}=4\delta =-2\epsilon -6\eta $.
From \[bEOM\], the evolution of $\eta $ is given by:
$$\ddot{\eta}=\epsilon -\eta +\frac{\delta _{r}}{3}=\frac{\epsilon }{3}-\eta .
\label{etaEOM}$$
Therefore there is a (more complicated) stable oscillatory behaviour for $\eta $ about the static solution and the evolution of $\epsilon $ has already been determined by \[epStatic\]. Thus we have an unusual behaviour characterised by stable anisotropic oscillations around the isotropic Einstein static universe. This generalises the simple isotropic oscillations about the static universe that exist in Friedmann universes with a ghost field found in [@Barrow:2004ad] and [@Barrow:2009sj].
We can now determine further effects of allowing a coupling between the fields. The second law of thermodynamics requires $s\geq 0$. The exact form of $s$ - taking into account terms representing constant coupling, bulk and shear viscosities, will of course affect the exact dynamics. However, it is possible to make progress by assuming only that $s$ is non-negative.
![Spatial volume versus time with coupling turned on ($s\neq 0$). The system oscillates, but expands from one cycle to the next. Initial values: $r=1.05,\protect\sigma =0,H=0,\protect\beta _{0}=0.01$, $\protect\beta _{H}=\protect\beta _{\protect\sigma }=0.$[]{data-label="Volexpand"}](VTcouple.jpg){width="50.00000%"}
First consider the case of no coupling ($s=0$). The evolution of the radiation field is determined by
$$\frac{\dot{\rho}_{r}}{\rho _{r}}=-\frac{4}{3}H=-\frac{4}{3}(\frac{\dot{\epsilon}}{3}+\dot{\eta}). \label{RadCycle}$$
Due to the cyclic behaviour of $\epsilon $ and $\eta $ (and hence of their derivatives), $\delta _{r}$ will also cycle, returning to its initial value, as any integral of the right-hand side of \[RadCycle\] across a complete cycle will be zero. However, with a positive coupling between the fields, this relationship is broken, and a term which is always non-negative (and so has a positive integral across cycles) must be added. Hence, across a cycle in $\eta $ and $\epsilon ,$ the value of $\delta _{r}$ now increases and we must adjust our equation of motion \[etaEOM\] to include this term. We write
$$\ddot{\eta}=\epsilon +\eta +\frac{\delta _{r}}{3}=\frac{\epsilon }{3}-\eta
+\Delta , \label{eta2EOM}$$
where $\Delta $ is a positive term representing the increase in $\delta _{r}$ due to field couplings created by introducing $s>0$. The variable $\eta $ no longer cycles about zero, and the system is slowly pushed away from stability, and enters a pseudo-cyclic phase in which a series of bounces occur with increasing local minima and maxima of the expansion volume, $ab^{2}$, as shown in figure \[Volexpand\]. The behaviour of $\epsilon $ \[epEOM\] is affected by this expansion (recall that $\epsilon $ is already small). The expansion of $b(t)$ means that $H_{b}$ is no longer small, and the equation gains a damping term. Similarly, the frequency of oscillations,$\sqrt{2}/b,$ is reduced by this expansion in $b(t)$ and the solution takes the approximate form of a damped harmonic oscillator. Note that in the derivation of \[epEOM\] only the smallness of $\epsilon $ was used - therefore this damping behaviour is present in all expanding solutions.
Ghost-induced Inflation
-----------------------
![Hubble expansion rate versus time for an inflating solution. Initial values: $\protect\beta _{0}=0.03$, $\protect\beta _{H}=\protect\beta _{\protect\sigma }=0,$ $r=1,\protect\sigma =5\times
10^{-4},a=10^{4},H=1.6\times 10^{-2}.$[]{data-label="InflationFig"}](HTInflate.jpg){width="50.00000%"}
Spacetimes which exhibit inflation are of special interest to cosmologists because inflation can solve a number of well known puzzles about the universe’s structure [@linde1986eternal], and make a series of detailed predictions that can be tested by observations of the microwave background radiation [@Ade:2013xsa]. Typical inflationary models exhibit expansion in which the Hubble parameter is (approximately) constant for a finite time interval. In general relativistic cosmology this is usually achieved by introducing a matter content that is (or is equivalent to) a scalar field subject to a self-interaction potential whose contribution to the total energy density is dominant during this expansion, with an equation of state close to that of that produced by an exact cosmological constant with $p_{\Lambda }=-\rho _{\Lambda }$.
Inflation induced by ghost fields has been studied as an alternative to the usual slow-roll models [@ArkaniHamed:2003uz]. Such models have potentially observable consequences for the microwave background trispectrum [@Izumi:2010wm] [@Huang:2010ab], but require that the translation invariance of the scalar ghost field is broken. In the BSBM models under consideration, however, translation invariance can be preserved, with the field coupling responsible for creating the inflationary energy density.
Let us examine the case of a linear coupling-induced inflation with $s=\beta
_{0}\rho _{\psi },$ for constant $\beta _{0}>0$. When the volume is large there exists an asymptotic solution of the form
$$\rho _{r}=\frac{9\beta _{0}^{2}}{4}, \:\:\: \rho _{\psi }=-\frac{3\beta
_{0}^{2}}{2},\:\:\: H=-\frac{\beta _{0}}{2}, \label{infcond}$$
in the isotropic case. This solution is stable, and is approached by dynamical trajectories, as shown in figure \[InflationFig\]. Under a small perturbation $H=-\beta _{0}/2+h$ and $r=1+\epsilon $ we find that to first order in the small parameters:
$$\begin{aligned}
\dot{h} &=&-\beta _{0}h \label{infstab} \\
\ddot{\epsilon} &=&-3\dot{\epsilon}H-2\epsilon /b\end{aligned}$$
In this solution, $b(t)$ is exponentially growing, so the final term in [infstab]{} quickly becomes negligible. Therefore, although a shearing expansion may occur, $\dot{\epsilon}$ quickly falls to zero, locking the shear at a fixed value. This inflationary phenomenon is not unique to linear couplings with $\beta _{H}=\beta _{\sigma }=0$, but is simplest to demonstrate in this case. Likewise, there is no requirement for the matter field to consist solely of radiation - introducing more matter fields with couplings whose sign is determined to be in accordance with the second law of thermodynamics yields a system which also exhibits inflation of this type. For most couplings, however, the inflationary phase will end when the dust field becomes dominant.
![Evolution of $\log (\protect\alpha )$ versus time shown over a short timescale (top), and the same evolution over a long timescale (bottom). Initial values: $\protect\beta _{0}=0.01$, $\protect\beta _{H}=\protect\beta _{\protect\sigma }=0,$ $r=1.01,\protect\sigma =0,a=1,H=0,\protect\rho _{\protect\psi }=-0.01.$[]{data-label="alphagraph"}](AlphatShort.jpg "fig:"){width="50.00000%"} ![Evolution of $\log (\protect\alpha )$ versus time shown over a short timescale (top), and the same evolution over a long timescale (bottom). Initial values: $\protect\beta _{0}=0.01$, $\protect\beta _{H}=\protect\beta _{\protect\sigma }=0,$ $r=1.01,\protect\sigma =0,a=1,H=0,\protect\rho _{\protect\psi }=-0.01.$[]{data-label="alphagraph"}](AlphatLong.jpg "fig:"){width="50.00000%"}
The evolution of the fine structure constant is shown in figure [alphagraph]{}. Initially, the solution is like an ascending staircase with rapid changes at each scale factor bounce, see for comparison [Barrow:2004ad]{}. Monotonicity of $\alpha $ is ensured since the scalar field cannot have positive energy density; since $\dot{\psi}=\sqrt{\frac{-2\rho
_{\psi }}{w}}$, we have $\dot{\psi}\geq 0$ for all time, and so $\psi $ cannot oscillate through maxima and minima. Across repeated bounces, $log(\alpha )$ will appear to increase in steps when $|\rho _{\psi }|$ is small as the relative size of this energy density oscillates greatly within a single cycle. However, as energy is transferred into the radiation field, these steps will become less apparent, eventually approaching a constant gradient once the ghost field reaches the condition for de Sitter inflation to occur. Thus, even though the universe oscillates from cycle to cycle, the fine structure ’constant’ continues increasing from cycle to cycle and there will typically only be a finite interval of cycles in which $\alpha $ takes values that allow stable atoms to exist [@BarrowTipler],[Lieb:1976zz]{}.
Discussion
==========
In this paper, we examined the new effect of introducing anisotropies into the BSBM framework for varying $\alpha $, although the conclusions have broader applicability to anisotropic cosmologies containing ghost fields and entropy-increasing energy exchanges between fields. In particular, we studied the dynamics of locally rotationally symmetric Bianchi IX cosmologies. It was shown that under certain conditions the bouncing behaviour observed in isotropic models persists, with the fine structure constant changing in an almost step-like increasing manner between cycles as time increases. It is apparent from \[HEOM\] that on short scales there is a tension between the shear terms and ghost field, as both scale as the inverse square of the volume. When the anisotropy is small, the contribution from the ghost field dominates. This leads to a bouncing model, reproducing closely the results seen in [@Barrow:2004ad]. Furthermore, there exists a static solution, perturbations about which lead to a sequence of anisotropic bouncing phases. When there is a coupling between the matter fields, the second law of thermodynamics ensures that this process isotropizes the system by energy exchange.
The resulting dynamics lead to a pseudo-cyclic universe in which the fine structure constant monotonically increases across bounces, and for small values of the associated dielectric scalar field, this increase is dominated by dynamics near the bounce point. The minimum and maximum volumes of the universe also increase across cycles, with the total energy density decreasing. Eventually, the model reaches a point at which the coupling between fields fixes the energy density to be constant in time, and the universe undergoes a de Sitter phase in which it inflates. Since this model is limited to include only the dielectric field and radiation, there is no transition to the dust-dominated era that one would expect at the end of this phase, and so inflation is endless within the model. It is possible to find solutions with dust in which the system again reaches a point of steady inflation. However within the space of couplings with $s>0$ which obey the second law, these solutions are a set of measure zero.
The evolution of $\alpha $ throughout the history of BSBM universes displays interesting traits. At late times, in a large universe it will appear that $\alpha $ has settled to a constant value. In doing so, throughout a series of oscillations the universe will have isotropised greatly, with $\alpha $ stepping up between cycles. As the dynamics are invariant under changing the initial value of $\alpha $ there is no obvious mechanism to determine the constant to which it will approach.
Acknowledgements {#acknowledgements .unnumbered}
================
DS acknowledges support from a Templeton Foundation grant.
[99]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
(), .
, ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, in **, edited by (, , ), pp. .
, ****, ().
, , , ****, ().
, ****, ().
, , , , , , ****, ().
, ****, ().
, ****, (). , ****, ().
, ****, ().
, ****, ().
, ****, ().
() (), .
, , , , ****, ().
, ****, ().
, ****, ().
, ** (, , ).
, ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The tendency for the period of the helically ordered moments in holmium to lock into values which are commensurable with the lattice is studied theoretically as a function of temperature and magnetic field. The commensurable effects are derived in the mean-field approximation from numerical calculations of the free energy of various commensurable structures, and the results are compared with the extensive experimental evidence collected during the last ten years on the magnetic structures in holmium. In general the stability of the different commensurable structures is found to be in accord with the experiments, except for the $\tau=5/18$ structure observed a few degrees below $T_{\text N}$ in a $b$-axis field. The trigonal coupling recently detected in holmium is found to be the interaction required to explain the increased stability of the $\tau=1/5$ structure around 42 K, and of the $\tau=1/4$ structure around 96 K, when a field is applied along the $c$-axis.'
address: 'Ørsted Laboratory, Niels Bohr Institute, Universitetsparken 5, 2100 Copenhagen, Denmark'
author:
- Jens Jensen
title: Theory of commensurable magnetic structures in holmium
---
Introduction {#sec:level1}
============
The basic features of the magnetically ordered structures in hexagonal close-packed Ho were established by Koehler [*et al.*]{},[@Koe1]who found that the basal-plane moments are arranged in a helical pattern at all temperatures below $T_{\text N}\simeq133$ K. The $c$-axis components order ferromagnetically at $T_{\text C}\simeq 20$ K, resulting in a conical ordering of the moments. The opening angle of the cone was found to approach $80^\circ$ in the zero temperature limit. The ordering wave vector for the basal-plane moments is directed along the $c$-axis and its magnitude $\tau$ (in units of $2\pi/c$) was observed to change monotonically from about 0.28 at $T_{\text N}$ to about $0.167\simeq1/6$ at $T_{\text C}$. Below $T_{\text C}$ the wave vector stays constant, indicating that the magnetic structure is locked to the lattice periodicity and repeats itself after each twelve hexagonal layers. The hexagonal anisotropy is known[@Yos] to produce higher harmonics at $(6\pm1)\tau$, and the fifth and seventh harmonics were clearly resolved by Koehler [*et al.*]{} at low temperatures. They concluded that the basal-plane moments in the cone phase bunch strongly around the easy $b$-directions, so that the angle between the basal-plane component of one of the moments and the nearest $b$-axis is only about 5.8$^\circ$ in the zero temperature limit. The neutron diffraction experiments have been repeated with higher resolution and with crystals of better quality by Felcher [*et al.*]{}[@Felch] and by Pechan and Stassis,[@Pechan] leading to only minor modifications of the results of Koehler [*et al.*]{}
Koehler and collaborators[@Koe2] also investigated the behavior of the magnetic structure in holmium as a function of magnetic field applied in the basal plane. The theoretical prediction for the magnetization process was that the helix would change into a ferromagnet either directly or via an intermediate fan phase.[@Naga] This was confirmed to occur at low temperatures, however, above 40–50 K Koehler [*et al.*]{} observed a second fan-like phase.[@Koe2] A few years ago this extra intermediate phase was explained, as being the helifan(3/2)-phase,[@J2; @Jehan] where the helifan structures are constructed from portions of the helix and the fan following each other in a periodic way.
The next major advancement in the investigation of the magnetic ordering in holmium came with the use of the new technique of magnetic x-ray scattering, which utilizes the intense radiation from a synchrotron source. The narrow experimental resolution obtainable with this technique made it possible for Gibbs [*et al.*]{}[@Gibbs1] to detect two other commensurable structures in holmium, with $\tau=2/11$ and 5/27, in addition to the one with $\tau=1/6$, at temperatures below 25 K. These commensurable values were explained by the spin-slip model.[@Gibbs1; @Bohr] In the limit of very strong hexagonal anisotropy the moments would be confined to be aligned along one of the six (in the present case) $b$-directions. The 12-layered structure may be constructed from pairs of neighboring layers, in which the moments are along the same $b$-direction, by allowing the moments to rotate 60$^\circ$ from one pair to the next. This periodicity may then be changed by introducing regularly-spaced series of [*spin slips*]{}, where at each spin slip a pair is replaced by a single layer. For example, the 22 layered 2/11-structure may be obtained from the 12-layered structure by introducing a spin slip after every 5 pair of layers, which we shall denote as the (222221)-structure. The introduction of a spin slip after every 4 pairs leads to the (22221)-structure corresponding to $\tau=5/27$. In this structure there is a spin slip for every 9 layers, and as observed by Gibbs [*et al.*]{}, this gives rise to a modulation of the lattice with this period, i.e. charge scattering at $\tau_c=2/9$.
The magnetic x-ray scattering technique provides very good resolution, but the intensity is weak, thus preventing the measurement of the scattering due to the higher harmonics. In contrast, neutron diffraction reflections have high intensities, although the resolution is relatively coarse. By the use of a large Ho-crystal, and a triple-axis neutron-scattering spectrometer for isolating the purely elastic scattered neutrons, Cowley and Bates were able to determine the intensities due to all the harmonics in a number of the spin-slip structures, in which the intensities extend over four orders of magnitude.[@Cow1] These results allowed them to derive the average turn angle between the bunched pairs to be about 10$^\circ$ at the low temperatures, in accordance with the result of Koehler [*et al.*]{},[@Koe1] and this value increases to about 20$^\circ$ at 30 K. In addition they were able to detect some variation of the turn angle from pair to pair. The magnitudes of the two modifications of the ideal spin-slip structure which were derived from the experiments were reproduced in a numerical model calculation.[@Mac1] The ultrasonic experiments made by Bates, [*et al.*]{}[@Bates] as a continuation of the experiments of Cowley and Bates, showed anomalies in the sound velocities not only at the temperatures where the system jumps between the three commensurable structures described above, but also at about 25, 40 and 97 K. At these latter temperatures $\tau$ is close to the commensurable values 4/21, 1/5, and 1/4, corresponding to the spin slip structures (2221), (221), and (211), respectively.
The possibility that the magnetic structure may lock into commensurable values at elevated temperatures was investigated at Chalk River by Tindall, Steinitz and collaborators.[@Stein1; @Willis; @Noakes; @Tin1; @Stein2; @Tin2; @Tin3; @Tin4; @Tin5; @Tin7; @Tin6] They monitored the position of the fundamental magnetic diffraction peak as a function of temperature at fields applied along the $c$-direction or along the $b$-direction, and they found several plateaus in the temperature variation of $\tau$. In the presence of a $c$-axis field of 30 kOe the (221)- and the (211)-structures were found to be stable around 42 K and 96 K, respectively, in both cases within a temperature interval of 2–3 K. In a $b$-axis field of 14 or 30 kOe the (211)-structure is again stable for a couple of degrees, which was also observed to be the case for the (21)-structure, $\tau=2/9$, at about 75 K. Finally, they observed $\tau$ to stay close to the value 5/18 between 126 K and the Néel temperature, when applying a field of 30 kOe along the $b$-axis. The observations made close to $T_{\text N}^{}$ and near 100 K have largely been confirmed by the similar experiments of Venter and du Plessis,[@Venter1; @Venter2] except that they did not observe a clear plateau at $\tau=1/4$ in a $c$-axis field around 96 K. However, they detected an anomalous behavior of the scattering intensity close to this temperature.
The low temperature domain, below 40 K, has been carefully investigated by Cowley [*et al.*]{}[@Cow2] in a $c$-axis field of 10–50 kOe. In spite of a rather monotonic variation, with field or temperature, of the position of the fundamental magnetic satellite they found by measuring the position of the higher harmonics that the diffraction pattern was determined in many cases by a superposition of neutrons scattered from domains with different commensurable $\tau$-values. Small differences in $\tau$, which may be difficult to resolve at the fundamental wave vectors, are enhanced so to be distinguishable when considering e.g. the fifth or seventh harmonics.
There have been made a number of theoretical studies of commensurable structures in model systems. The simplest one, the anisotropic next-nearest-neighboring Ising (ANNNI) model has been investigated in detail both at zero field[@Fish; @Bak1; @Bak2] and in an applied field.[@Smith] The XY-model with six-fold anisotropy, which is more closely related to holmium, has recently been discussed by Sasaki,[@Sas] in the limit of large anisotropy, and by Seno [*et al.*]{}[@Seno] Steinitz [*et al.*]{}[@Stein3] have made some qualitative considerations on the connection between the spin-slip model and the field-dependent commensurable effects observed in holmium. Plumer has analyzed a model, which includes most of the couplings present in holmium, in order to understand the stability of the 1/4-phase observed near 96 K in a $c$-axis field.[@Plumer] However, the explanation he proposed, relies on adding a large symmetry-breaking term to the free energy which seems difficult to justify. Here we shall present a mean-field calculation of the stability of the different commensurable phases based on a realistic model for holmium. The model includes not only the terms considered by Plumer, but also the trigonal coupling which has recently been detected in neutron diffraction measurements by Simpson [*et al.*]{}[@Simp] The trigonal coupling induces a net ferromagnetic moment in the basal plane when $\tau=1/4$, and a misalignment of the field by as little as one degree is sufficient for explaining the observed lock-in of this structure, in accordance with the conjecture made by Jensen and Mackintosh.[@J3] The basal-plane three-fold anisotropy caused by the trigonal coupling in the case of a cone structure has also strong implications on the (221)-structure in a $c$-axis field near 42 K.
The model used in the calculations is developed in the next section. The stability of the different commensurable structures is investigated in Sec. III. The free energy of the various structures is calculated numerically within the mean-field approximation. The advantage of this method is that it is possible to account for most of the complexities of the real system, whereas one disadvantage is that there is a limit to the number of different layers one may handle numerically with the sufficient precision. In the present calculations we consider structures with a repeat length of up to 500 layers, leading to a resolution which should be superior to even the most precise experiments. The work is summarized and concluded in the last Sec. IV.
The mean-field model of holmium {#sec:level2}
===============================
The model is based on the magnetization and spin-wave measurements, and is similar to the one applied in previous numerical analyses of the magnetic structures and excitations in holmium.[@J2; @Mac1; @J3; @Lars; @J1; @J4; @Mac2] The Hamiltonian comprises the single-ion anisotropy, the Zeeman term, the Heisenberg exchange coupling, the classical magnetic dipole–dipole interaction, and the trigonal coupling: $$\begin{aligned}
{\cal H}_0^{}=&&\sum_i\sum_{lm}B_l^mO_l^m(i)
-\sum_ig\mu_B^{}\text{\bf H}\cdot\text{\bf J}_i\nonumber\\
&&{}-\textstyle{1\over2}\sum_{ij}
{\cal J}(ij)\,\text{\bf J}_i\cdot\text{\bf J}_j
-{\textstyle{1\over2}}\sum_{ij}
{\cal J}_{\!D}^{}(ij)J_{zi}J_{zj}\nonumber\\
&&{}+\sum_{ij}K{}_{21}^{31}(ij)
\big[O_3^2(i)J_{yj}+O_3^{-2}(i)J_{xj}\big].
\label{1}\end{aligned}$$ The $O_l^m$-operators are the Stevens operators, and in particular $O_3^{\pm2}={1\over2}(J_zO_2^{\pm2} +
O_2^{\pm2}J_z)$, where $O_2^2=J_x^2-J_y^2$ and $O_2^{-2}=J_x^{}J_y^{}
+J_y^{}J_x^{}$. The $x$-, $y$-, and $z$-axes are assumed to be along the $a$-, $b$-, and $c$-axes of the HCP lattice, respectively. The Fourier transforms of the couplings are defined in the standard way, and we shall use the short hand notation ${\cal J}(q)$ for ${\cal J}(\text{\bf q})$, when [**q**]{} is along the $c$-axis. The inter-planar exchange parameters ${\cal J}_n$ are then defined by $${\cal J}(q)={\cal J}_0+2\sum_{n \ge 1}{\cal J}_n\cos(nqc/2).
\label{2}$$ The classical dipole interaction is included in the coupling ${\cal J}(q)$ of the basal-plane moments, in which case the coupling of the $c$-components is ${\cal J}_{cc}(q)={\cal J}(q)+{\cal J}_D(q)$. The classical contribution vanishes at zero wave vector, ${\cal J}_D(\text{\bf 0})\equiv 0$, whereas at $q\ne0$ $$\begin{aligned}
&&{\cal J}_D(q)=\nonumber\\
&&{}-{\cal J}_{dd}\{0.919+0.0816\cos(qc/2)-0.0006\cos(qc)\},
\label{3}\end{aligned}$$ where the coupling constant in holmium is $${\cal J}_{dd}=4\pi(g\mu_B^{})^2\,N/V= 0.0349\hbox{\ meV.}
\label{4}$$ The jump, which the dipole coupling ${\cal J}_D(q)$ makes at zero wave vector, is observable in the excitation spectrum, and it explains without introducing any further two-ion anisotropy,[@Lars] why the system prefers the cone structure rather than the “tilted helix” below $T_{\text C}^{}$.
The trigonal coupling was discovered to be present in erbium by Cowley and Jensen.[@Cow3] This coupling reflects the fact that the $c$-axis is a three-fold symmetry axis, and its contribution to the mean-field Hamiltonian for the $i$th ion is given by $$\begin{aligned}
&&\Delta{\cal H}_{\text MF}^{}(i\in p\hbox{'th plane})=
(-1)^p\sum_{n\ge1}\big[K{}_{31}^{21}\big]_n^{}\nonumber\\
&&{}\times\Big[\{O_3^2(i)-\textstyle{1\over2}\langle O_3^2(i)\rangle\}
\big\langle J_y^{}(p+n)-J_y^{}(p-n)\big\rangle\nonumber\\
&&{}+\{O_3^{-2}(i)-\textstyle{1\over2}\langle O_3^{-2}(i)\rangle\}\big\langle
J_x^{}(p+n)-J_x^{}(p-n)\big\rangle\nonumber\\
&&{}-(-1)^n\{J_{yi}^{}-\textstyle{1\over2}\langle J_{yi}^{}\rangle\}
\big\langle O_3^{2}(p+n)-O_3^{2}(p-n)\big\rangle\nonumber\\
&&{}-(-1)^n\{J_{xi}^{}-\textstyle{1\over2}\langle J_{xi}^{}\rangle\}
\big\langle O_3^{-2}(p+n)-O_3^{-2}(p-n)\big\rangle\Big],\nonumber\\
\label{5}\end{aligned}$$ where the argument $p\pm n$ denotes an ion in the uniformly magnetized $(p\pm n)$th hexagonal layer. This equation, which defines the inter-planar coupling parameters $\big[K{}_{31}^{21}\big]_n^{}$, shows that the coupling changes sign from one sublattice to the next. Both in the cone phase and in the cycloidal phase of erbium, the trigonal coupling gives rise to additional neutron diffraction peaks when the scattering vector is along the $c$-axis. These peaks would not be there if the magnetic structures were independent of the different orientation of the basal planes in the two hexagonal sublattices. The equivalent phenomenon has also now been observed in holmium by Simpson [*et al.*]{},[@Simp] who fitted the intensities of the extra reflections by a set of three inter-planar parameters for the trigonal coupling.
We have reanalyzed the experiments of Simpson [*et al.*]{}[@Simp] and derived a set of anisotropy parameters, which accounts both for the low temperature magnetization curves, as precisely as the previous model,[@Lars] and for the neutron diffraction results. The effects of all the three trigonal couplings of the fourth rank,[@Cow3] and combinations of the three couplings were investigated, and in agreement with Simpson [*et al.*]{} we find that the trigonal coupling introduced by Eq. (\[5\]) (the coupling applied in erbium), is the one which leads to the best fit. The small differences between the ways the neutron diffraction results are fitted are insignificant compared with the experimental uncertainties, but the inter-planar coupling parameters derived here are, quite remarkably, a factor of 4-5 smaller than those obtained by Simpson [*et al.*]{} This large difference is surprising, but may be explained by the high degree of compensation which occurs between the different terms in the mean-field Hamiltonian (\[5\]). The degree of compensation depends on the structure considered, and in the analysis of the commensurable structures discussed in the next section we found that the trigonal coupling derived by Simpson [*et al.*]{} has unacceptably strong implications in some cases. The alternative trigonal coupling derived here circumvents these difficulties, and is therefore a more likely possibility. The modified set of inter-planar coupling parameters for the trigonal coupling is given in Table I. The corresponding crystal-field parameters $B_l^0$ are derived from the low-temperature magnetization curves and $T_{\text
C}$, and $B_6^6$ is determined so that the averaged bunching angle is $5.8^\circ$ in the 12-layered structure in the low temperature limit. The values of these parameters are given in Table II, and they are close to those applied in the previous model.[@Lars]
The variation of the ordering vector with temperature in Ho has been analyzed by Pechan and Stassis,[@Pechan] who found that it agrees reasonably well with the prediction of the theory of Elliott and Wedgwood.[@Elliott] ${\cal J}(\text{\bf q})$ is proportional to the susceptibility of the conduction electrons, which depends strongly on the nesting between different parts of the Fermi surface. The super-zone energy gaps created due to the oscillating polarization of the conduction electrons lead to a decrease of the maximum in the susceptibility and to a shift of its position towards smaller wave vectors, as the degree of polarization is increased. In addition, Andrianov[@And] has discovered a relation, $\tau=\tau_0^{}[(c/a)_{\text cr}- c/a]^{1/2}$, between the $c/a$-ratio and the ordering wave vector, where $(c/a)_{\text{cr}}\simeq 1.582$. This relation is obeyed nearly universally by the rare-earth metals, indicating a strong correlation between the $c/a$-ratio and the nesting effects. A change of the $c/a$-ratio will modify the magnetic Hamiltonian in ways other than through ${\cal J}(\text{\bf q})$. $B_2^0$ is especially sensitive to the $c/a$-ratio, as to a first approximation it is proportional to the difference between $c/a$ and its ideal value, and rather strong effects have recently been discovered in thulium, when the system jumps from the ferrimagnetic phase with $\tau=2/7$ to the ferromagnetic structure.[@Keith] In the model considered here we only account for the temperature variation in ${\cal J}(q)$, whereas additional magnetoelastic effects on the crystal-field parameters or changes of ${\cal J}(q)$ as a function of field are neglected. The spin-wave dispersion relation of Ho has been measured both at low temperatures[@Lars; @String; @Patt] and at elevated temperatures,[@Nick] and these measurements have been used to derive the inter-planar exchange parameters.[@Lars; @J1] In order to reproduce the correct wave vectors at the different temperatures small adjustments have been introduced, and for this purpose we have chosen ${\cal J}_3$ as the variation parameter. The inter-planar exchange parameters used in the model are given in Table III.
The structures discussed in the next section have been calculated by a straightforward iteration procedure using the parameters given in Table I-III and Eqs. (\[3\])-(\[4\]). The calculations utilize the mean-field approximation, in which $J_{\alpha i}J_{\alpha j}$ in (\[1\]) is replaced by $(2J_{\alpha i}-\langle J_{\alpha
i}\rangle)\langle J_{\alpha j}\rangle$ and the trigonal term by Eq.(\[5\]). The first step in the iteration is to assume a distribution of expectation values of the various operators. These values are then inserted in the mean-field Hamiltonian for the $i$th ion, and after a diagonalization of this Hamiltonian the partition function, the free energy, and new expectation values for this ion are calculated. The calculation is carried out for each ion in one commensurable period, and the procedure is repeated with the new distribution of expectation values until a self-consistent solution is achieved. At a given temperature and field the free energies of structures with different (commensurable) periods are compared in order to identify the most stable structure. The energy differences between the various structures may be minute, and the calculations have to be done with high numerical accuracy. In order to ensure in a given case, that the iteration has converged towards the state with the lowest free energy and not to a metastable configuration, many iterations are required, of the order of several thousands, and many starting configurations have to be considered.
The commensurable structures {#sec:level3}
============================
Low temperature regime
----------------------
Most of the commensurable structures are observed in the low temperature regime ranging from zero to about 50 K. The exchange constants are given in Table III at the end points of this interval. In between we have used the quadratic interpolation: $$J_n(T)=(1-\alpha^2)J_n(0)+\alpha^2 J_n(50 \text{\ K}),
\label{6}$$ where $\alpha$ is the temperature $T$, in units of K, divided by 50.
At zero field the lowest order spin-slip structures are particularly stable. Starting at about 40 K, we get the sequence (221), (2221), (22221), (222221), corresponding to $\tau=1/5$, 4/21, 5/27, and 2/11, with a steady increase of the number of pairs in between the spin-slip layers as the temperature is reduced to about 20 K. One might expect this series to continue as the system is cooled further, but the 2/11-structure is predicted to occur only in a small temperature interval of about 0.8 K, and the intervals, where the subsequent spin-slip structures are stable, become minute. Thus the system is predicted to jump from the 2/11-structure to the pure pair-layered 1/6-structure within an interval of 0.1 K. Metastable states, or mixed states are expected to appear frequently in the real system, which makes it difficult to decide whether this particular prediction is in accord with the experiments, but the temperature interval, in which the indications[@Cow1; @Bates] of the intermediate structures are found is narrow with a width of about 1 K.
In addition to the lowest order spin-flip structures, structures with mixed sequences are also found to be stable. With one exception, the (222122221)-structure ($\tau= 3/16$), which is stable over a small interval of about 0.3 K, all the higher order spin-slip structures occur between the (221)- and the (2221)-structures, of which the most important one is the (2212221)-structure ($\tau=7/36$). When approaching $\tau=1/5$ from below, the main sequence of commensurable ordering vectors is determined by $\tau_n(1)=n/(5n+1)$ which is the series of rational fractions close to 1/5 with the smallest denominators or shortest commensurable periods. Commensurable structures with wave vectors lying in between those determined by $\tau_n(1)$ may also be stable. In order to investigate this in a systematic way we have used the proposal of Selke and Duxbury that the structure most likely to appear between two phases has a period which is the sum of the periods of the structures in the two phases,[@Seno; @Selke] suggesting that the most probable intermediate values in the present case are those determined by the second-order series $\tau_n(2)=(2n+1)/(10n+7)$. For a further subdivision of the steps made by $\tau$ we have applied the third-order series $\tau_n(3)=(3n+1)/(15n+8)$. In the interval between $\tau=4/21$ and $\tau=7/36$ the most stable configuration is found to be the $\tau_5^{}(2)=11/57$ or (22122212221)-structure, whereas other choices from the $\tau_n(1)$ or the $\tau_n(2)$ series are only found to be stable in narrow temperature intervals. For $\tau$ larger than 7/36, the situation is changed. Structures defined by the series $\tau_n(1)$ as well as $\tau_n(2)$ are stable in about equal intervals, and for $n\ge9$, structures determined by $\tau_n(3)$ also appear. At these temperatures the intervals are so small that it is difficult by the numerical calculations to distinguish the behavior from a truly continuous, incommensurable, variation of $\tau$. The magnetic correlation length in the $c$-direction of Ho has been determined by x-ray scattering[@helg] to be of the order of 2000 layers, indicating that if the shortest commensurable period becomes of the order of 200 layers, it is no longer possible experimentally to distinguish commensurable ordering, repeating itself coherently about ten times, from an incommensurable phase. The non-zero resolution characterizing the numerical calculations is a limitation, when comparing the results with more exact analytic results, but not in a comparison with a real system, which suffers from impurities, stacking faults and other imperfections. Hence we expect that experimentally $\tau$ will change continuously, at least when it is larger than about 0.197 until it, at a first order transition, jumps to the commensurable value $\tau=1/5$. The model calculations predict a lock-in at this wave vector between 41.47 K and 43.70 K, and that $\tau$ returns to incommensurable values above 43.70 K. The calculated results for the variation of $\tau$ at zero field, in the low temperature regime, are shown in Fig. 1.
The stability of the different commensurable structures are influenced by a field along the $c$-axis. The calculations indicate that the main features of the phase diagram are the same as at zero field, but there are a few significant changes. There is an overall shift of the stable regimes of the commensurable structures towards lower temperatures, the more pronounced the lower the temperature is. The 2/11-phase and the spin-slip structures lying in between this phase and the 1/6-phase become more stable, until the field is so large that the overall temperature shift removes these phases, which happens between 25 and 35 kOe. Finally, the most important change is, that the stability of the $\tau=1/5$-phase is much increased by the field.
The first of these effects is explained by the condition that the moments in the spin-slip layers have a larger $c$-axis susceptibility than the moments in the pair layers, i.e. the more spin-slip layers a structure contains the more Zeeman energy it gains in a $c$-axis field. The experimental results of Cowley [*et al.*]{},[@Cow2] obtained at a $c$-axis field of 10–50 kOe, which are included in Fig. 1, reflect this effect. Fig. 2 shows a more direct comparison of the experiments of Cowley [*et al.*]{} with the calculated field dependence of the ordering wave vector at 10 K. The experimental results shown in Fig. 2 are derived from the position of the primary magnetic satellite near $(002)$, which leads to a more gradual change of the wave vector than if the contributions from different domains with different stable or metastable structures are separated, as shown in Fig. 1. The comparison in Fig. 2 demonstrates that the calculated shift of the ordering wave vector at a given temperature (i.e. at fixed values of the coupling constants) due to the $c$-axis field is of the same magnitude as observed experimentally. The calculated results at zero temperature are similar to the results in Fig. 2 at 10 K, except that the values of the transition fields are shifted upwards by about 5 kOe. The strong field-dependence of $\tau$ should be taken into consideration in the comparison between the experiments and the theoretical results in Fig. 1. In this comparison it may also be important that the 2/11-structure, in contrast to most of the other spin-slip structures, has a moment in the basal plane, and that the 3/16-structure, in between the 5/27- and the 4/21-structures, is much more susceptible to a basal-plane field than the two neighboring structures. This means that a small misalignment of the $c$-axis field would favor principally these two structures.
At not too high temperatures, the hexagonal anisotropy energy decreases with the relative magnetization, $\sigma$, proportional to approximately $\sigma_{}^{21}$, whereas the trigonal anisotropy changes like $\sigma_{}^7$. The hexagonal anisotropy dominates at the lowest temperatures, and is responsible for the commensurable spin-slip structures, but as the temperature is increased the trigonal anisotropy becomes relatively more important. At 40 K $\sigma\simeq0.925$, implying that the hexagonal anisotropy energy has decreased by a factor of 5 compared with its value in the zero temperature limit, whereas the trigonal anisotropy is only reduced by a factor of 1.7. Within perturbation theory the contribution of the trigonal interaction, Eq. (\[5\]), to the free energy is of second order in the helical phase, whereas a first-order contribution appears if the $c$-axis moments are nonzero. To a first approximation this contribution is proportional to $$\Delta F\propto \sum_p(-1)_{}^pJ_\parallel^{}J_\perp^3\cos(3\phi_p),
\label{7}$$ where $J_\parallel^{}$ and $J_\perp$ are the components of the moments parallel and perpendicular to the $c$-axis, respectively, and $\phi_p$ is the angle the perpendicular component of the moments in the $p$th layer makes with the $x$- or $a$-axis. When only the trigonal coupling is considered, then every second $a$-axis is an easy axis in one of the sublattices and the other three $a$-axes are the easy axes in the other sublattice. Fig. 3 shows the hodographs of the basal-plane moments calculated at 42.185 K in zero field and in the presence of a field of 10 kOe along the $c$-axis. The figure indicates that the three-fold anisotropy term induced by the $c$-axis field, (\[7\]), is capable of rotating the moments about 30$^\circ$, so that the moments in the two spin-slip layers, which at zero field are along a $b$-axis become oriented along an $a$-axis. The hexagonal anisotropy energy does not depend on the distinction between the two sublattices and is changed by the same amount as obtained by rotating the zero-field hodograph the relatively small angle in the opposite direction. The small rotation does not cost much in hexagonal anisotropy energy, whereas the gain in trigonal anisotropy energy is substantial.
Both structures in Fig. 3 look like spin-slip structures, except that the roles of the $a$- and the $b$-axes have been interchanged in the two cases. However, at these temperatures the basal-plane anisotropy energy is reduced so much that the spin-slip model is no longer particularly useful, since the angle between the moments in the pair layers is not much smaller than the smallest angle between moments belonging to neighboring pairs. The commensurable structures with $\tau$ slightly different from 1/5 are not spin-slip structures. Instead they consist of portions of the 1/5-structure separated by domain walls in which phase shifts are introduced via a relatively smooth adjustment of the turn angles. If the width of the walls is much smaller than the distance between them, then the coupling between the walls is negligible and the free energy changes linearly with the density of walls, corresponding to a linear variation of the free energy with $\tau$. A detailed discussion of the behavior of systems with domain walls is given by Fisher and Szpilka.[@Fisher] Fig. 4 shows the free energy calculated in the two cases, for various commensurable values of $\tau$ in the proximity of $\tau=1/5$. In the case of $H_c=10$ kOe, the free energy varies linearly between 0.198 and 0.2 and between 0.2 to 0.202, whereas the variation is parabolic at zero field, except that the free energy at $\tau=1/5$ is slightly smaller than the minimum value of the parabola. Leaving out the result at $\tau=1/5$, the free energies of the zero-field commensurable structures are fitted by a third degree polynomial with a standard deviation of about $10^{-7}$ meV, which is the curve shown in the figure, (the numerical accuracy by which the free energy is calculated is $10^{-8}$ - $10^{-9}$ meV). This smooth change of the free energy is consistent with the estimated incommensurable variation of $\tau$ on both sides of the temperature interval where the $\tau=1/5$-structure is stable. The free energy in the zero field case is nevertheless going to change linearly with $\tau$ for $\tau$ sufficiently close to 1/5, when $|\tau-1/5|$ is smaller than about 0.0007.
The difference between the behavior of the free energy at zero and at $H_c=10$ kOe is related to the different widths of the domain walls. Fig. 5 shows the turn angles in the two cases calculated at 42.815 K. In both cases $\tau$ is slightly smaller than 1/5 corresponding to an average turn angle slightly smaller than 36$^\circ$. The width of the domain walls, denoted by the arrows, is about 100 layers in zero field and about 70 layers at $H_c=10$ kOe. However, the phase shift accomplished by one domain wall is different in the two cases: $12^\circ$ in the zero-field case and 24$^\circ$, a factor 2 larger, in the case of $H_c=10$ kOe. These numbers imply that the domain walls at zero field start to overlap with each other when $1/5-\tau$ is larger than 0.0007, whereas when $H_c=10$ kOe the overlap starts to occur only when $1/5-\tau$ is nearly a factor of 3 larger. These same numbers also describe the situation at $\tau$-values larger than 1/5.
At decreasing temperatures the slope of the linear section in the free energy decreases, numerically, for $\tau$ smaller than 1/5, and the opposite for $\tau$ larger than 1/5. The 1/5-phase is the stable one until the temperature is lowered to the point where the slope becomes zero. At this temperature it costs no energy to create domain walls, as long as the distance between them is larger than their width. This leads to a vertical decrease of the $\tau$-value at this temperature. At a slightly lower temperature the density of the domain walls will be so large that the walls overlap each other, and there will be only one equilibrium configuration, corresponding to the curvature in the free-energy function at the minimum being non-zero. The change of $\tau$ as a function of temperature will accordingly be more gradual. The transition between the 1/5-structure and other structures is predicted to occur in a (quasi)continuous[@Fisher] way in both the cases considered, but at zero field the transition regime is so narrow that the transition is indistinguishable from a first-order one, within the present numerical resolution.
The lock-in of the structure at $\tau=1/5$ is strongly enhanced by the $c$-axis field. As shown in Fig. 6, the interval in which the 1/5-structure is calculated to be stable increases rapidly, from about 2.2 K to about 10 K, between zero and a field of 10 kOe along the $c$-axis. This strong increase is also perceptible on either side of the 1/5-phase. Here commensurable effects are resolved even very close to the transition regions at 10 kOe, in contrast to the incommensurable behavior indicated by the calculations at zero-field. The large field-induced enhancement of the commensurable effects is produced solely by the trigonal coupling. We have repeated the calculations using the previous model[@Lars] which neglects the trigonal coupling. In this case the zero-field structure is found to be stable within an interval of 2.8 K, and this interval is reduced in a $c$-axis field, by a factor of 2 at a field of 30 kOe. The effective hexagonal anisotropy decreases faster than the free-energy differences between the different structures when the basal-plane moments are reduced by the field.
Tindall [*et al.*]{}[@Tin5] have observed a plateau in the variation of $\tau$ at 1/5 over a temperature range of 2–3 K when $H_c=30$ kOe, see Fig. 1, which disappears at zero field. It is probably very difficult to obtain experimental results on commensurable effects which can be trusted quantitatively in a comparison with the results derived from a mathematical model of an ideal system. The commensurable effects are based on very small differences in the free energy and impurities and defects are likely to be important. Furthermore, the results of Cowley [*et al.*]{}[@Cow2] in Fig. 1 show that, at low temperatures, several metastable configurations may be present in the crystal at the same time. The occurrence of metastable structures should be most pronounced at low temperatures, as also indicated by the behavior of the hysteresis effects, however, mixed phases are still observed at 35 K, see Fig. 1. Taking these effects into consideration, it seems reasonable that the calculated lock-in at zero-field of the 1/5-structure within an interval of about 2 K, may be smoothed out in the temperature variation of the average position of the fundamental magnetic scattering peak. We may instead refer to the low-field anomalies observed near 40 K in ultrasonic[@Bates] and magnetization[@Willis] measurements. The observed field enhancement of the lock-in of the 1/5-phase is in contradiction with the behavior expected if the trigonal coupling were negligible. Zeeman effects due to a small misalignment of the $c$-axis field are estimated to be of no importance, so, qualitatively, the enhancement may only be explained by the trigonal coupling. The question left open is whether the observed plateau of about 2 K at 30 kOe is consistent with the calculated range of the lock in of about 10 K.
High temperature regime
-----------------------
The high temperature regime is here considered to extend from 40–50 K and up to $T_{\text N}$. In this temperature range the anisotropy within the basal plane is relatively weak, and the spin-slip model no longer applies. Instead the additional anisotropy introduced by an external field will play a much more important role. The value of $\tau$ changes continuously in most of the interval, and detectable commensurable effects are only expected for structures with short commensurable periods, and we have limited ourselves to the study of the following three cases; the 18-layered 2/9-structure, the 8-layered 1/4-structure, and the 36-layered 5/18-structure appearing respectively around 75 K, 100 K, and 130 K.
Assuming the average slope $d\tau/dT$ near 75 K to be 0.001 K$^{-1}$, we calculate the lock-in interval of the 2/9-structure to be 0.5 K at zero field. This value is nearly unchanged in a field of 30 kOe along the $c$-axis. According to the magnetization measurements the transition to the fan structure, via intermediate helifan structures, occurs at a field of about 18 kOe applied along a $b$-axis.[@Feron; @Mac1] At 14 kOe the system is still in the helical phase, but the helix is strongly distorted by the field, the magnitude of the second (or seventh) harmonic of the basal-plane moments has grown to be almost the same as in the fan phase, and the 2/9-structure is found to be stable in an interval of 3.6 K. These commensurable effects for the 2/9-structure are found to be nearly independent of the trigonal anisotropy. The only weak tendency for the system to lock in at $\tau=2/9$ at zero field or in the presence of a $c$-axis field seems to be in accordance with experiments. Tindall and collaborators have not reported any lock in of the 2/9 structures in these two cases, whereas they observed a plateau in the $\tau$ variation with a width of about 2 K in the presence of a $b$-axis field of 14 kOe,[@Tin6] which is consistent with the calculated behavior.
In the helical phase, the field perpendicular to the helical axis leads to a slight reduction of the $\tau$-value. The transition to the fan-phase is accompanied by a further reduction of $\tau$, hereafter in the fan phase there is a steady increase of $\tau$ with increasing field until the system becomes ferromagnetic. The possibility that the ordering wave vector may change was not included in the original analysis of the helix-fan transition by Nagamiya [*et al.*]{}[@Naga] It has been observed by Koehler [*et al.*]{}[@Koe2] at 50 K, where $\tau$ changes from 0.208 at zero field to be 0.170 in the fan phase, a change which was found to agree with the model calculations.[@J2] At 80 K $\tau$ is about 0.230 at zero field and is calculated to be reduced to 0.199 when the helical structure is changed into a fan at 18 kOe. In the fan phase $\tau$ increases and is 0.222 at 30 kOe at the transition to the ferromagnet. The magnitude of these changes in $\tau$ decreases linearly with increasing temperatures and is a factor of 3 smaller at 100 K. This means that the temperature dependence of $\tau$ in the interval between 80 and 100 K is a factor of 2 larger at a field of 20 kOe, just above the helix-fan transition, than at zero field. This prediction may be compared with the experimental results of Tindall [*et al.*]{}[@Tin4] shown in Fig. 5 of their paper. This figure shows that $\tau$ changes from about 0.18 at 81 K to about 0.23 at 95 K, when a field of 30 kOe is applied along a $b$-axis. This change is a factor of 3 larger than indicated by the calculation at 30 kOe, but it is comparable with the prediction at 20 kOe. In the fan phase the average basal-plane moment is about 6 $\mu_{\text B}^{}$, which leads to a demagnetization field in the case of a sphere of about 8 kOe, and it is therefore quite likely that the applied field of 30 kOe should be reduced by about 10 kOe in the comparison with the model calculations. Domain effects combined with the large value of the demagnetization field may also be the reason why Tindall [*et al.*]{}[@Tin4] observed the magnetic satellite to split into two peaks in some of the scans, the observed splitting in $\tau$ of 0.007 corresponds to a field difference of 4–5 kOe. Tindall [*et al.*]{} do not report any particular lock-in effects for the structures shown in their Fig. 5. This is consistent with the model calculations which also do not indicate any significant commensurable effects for these structures.
The change of $\tau$ as a function of a basal-plane field is still of some magnitude around 100 K, and the calculated effect agrees well with the observation of Venter [*et al.*]{}[@Venter1] At 100 K the helix-fan transition is predicted to occur at a field of 18 kOe. At fields just below this value the 1/4-phase is found to be stable within an interval of about 2 K, the interval is 2.0 K at 16 kOe assuming $d\tau/dT=0.00128$ K$^{-1}$, whereas this interval is reduced by a factor of two or more in the fan phase. Tindall [*et al.*]{}[@Tin2; @Tin4; @Tin6] have observed a lock-in of the 1/4-structure in an interval of about 1 K at 14 kOe and about 1.5 K at 30 kOe, and Venter [*et al.*]{}[@Venter1] have made the similar observation that the 1/4-structure is stable within an interval of about 1 K at $H_b=17$ and 23 kOe. The system is most likely to be still in the helical phase at 14 and 17 kOe. The positions of these lock-ins, just below 100 K in both cases, and their widths are in good agreement with that predicted in the helical case. The present model indicates that the stability of the 1/4-structure should decrease when the system jumps into the fan-phase, which is somewhat in disagreement with the experiments. We have two remarks to add to this minor discrepancy. Firstly, the (effective) anisotropy parameters are determined from the behavior of the system at low temperatures, and their values may have changed at these high temperatures. An increase of the axial anisotropy $B_2^0$ stabilizes the 1/4-structure in the fan phase, but does not have much effect on the helical structure. Secondly, it is uncertain whether the system is in the fan phase at the fields of 23 and 30 kOe. The relatively large demagnetization field in the case of the fan may lead to a mixing of the phases, and in addition there is the likely possibility that the structures are helifans,[@J2] in which case the commensurable effects are larger than they would be in the fan phase.
Around 100 K the hexagonal anisotropy energy is very small and the calculations indicate that the basal-plane turn angle at zero field only differs by up to 0.12$^\circ$ from the average value of 45$^\circ$ for the 8-layered, $\tau=1/4$, structure at zero field. The application of a field along the $c$-axis implies that the first-order trigonal contribution, Eq. (\[7\]), becomes non-zero and the trigonal anisotropy leads to a variation of the turn angle between 44.5$^\circ$ and 45.5$^\circ$ at $H_c=30$ kOe. This modification is still small, and the model calculations indicate only very weak commensurable effects, a lock-in temperature interval of the order of 0.1 K both at zero field and when $H_c=30$ kOe. In analogy with the fifth and seventh harmonics induced by the hexagonal anisotropy, the first-order trigonal anisotropy induces a second and a fourth harmonic, but because of the factor $(-1)^p$ in Eq. (\[7\]) these harmonics are translated by a reciprocal lattice vector along the $c$-axis (the half of a reciprocal lattice vector in the double-zone scheme), which means that the fourth harmonic appears at zero wave vector when $\tau=1/4$. In other words, in the case of a cone structure with $\tau=1/4$ the trigonal coupling leads to a ferromagnetic component perpendicular to the cone axis. The two extra peaks observed in the cone phase in, for instance, the 2/11-structure[@Simp] may be classified as arising from these harmonics, and in the cone phase of erbium, where $\tau=5/21$ is close to 1/4, the fourth harmonic is observed close to the nuclear peak,[@Lin] see also the discussion in Ref. \[\]. In the present case the ferromagnetic moment is along a $b$-axis and is calculated to be 0.043 $\mu_{\text B}^{}$ when $H_c=30$ kOe, corresponding to an average rotation of the moments towards the $b$-axis by about 0.4$^\circ$.
The estimated moment is small, but the energy differences determining the commensurable effects are also small, and the basal-plane moment has a profound influence on the system, as soon as the field has a non-zero component perpendicular to the $c$-axis. The Zeeman energy gained by the 1/4-structure, in comparison with the neighboring structures, is proportional to the angle $\theta$ between the $c$-axis and the direction of the field, at small values of the angle. At the temperature where the $\tau=1/4$ structure is stable at $\theta=0$, the free energy of the structures increases quadratically with $\tau-1/4$, which leads to a lock-in temperature interval, $\Delta T$, of the 1/4-structure proportional to $\sqrt{\theta}$. In this simple estimate we have neglected the small commensurable effect at $\theta=0$, and the possibility that the structures with $\tau$ slightly different from 1/4 may be able to develop a net moment in the basal plane. The size of the effect is illustrated in Fig. 7, which shows the calculated lock-in interval as a function of $\theta$ in a field of 30 kOe. $\Delta T$ increases like $\sqrt{\theta}$ at small values of $\theta$ and is about 12 K at $\theta=30^\circ$. Between 30$^\circ$ and 40$^\circ$ the helix is only stable in part of the interval and is replaced by the fan structure in the other part. For $\theta$ larger than about 40$^\circ$ only the fan is stable and $\Delta T$ decreases to about 1 K at $90^\circ$. As mentioned above the hexagonal anisotropy is very small at these temperatures and for $\theta$ larger than about 0.5$^\circ$ the results shown in Fig. 7 are independent of whether the field is lying in the $a$–$c$ or the $b$–$c$ plane. The transitions between the 1/4-structure and the surrounding incommensurable structures are established via the creation of domain walls and are continuous, equivalently to the case of $\tau$ close to 1/5.
The vertical slope of $\Delta T$ as a function of $\theta$ at the origin means that even the slightest deviation of the field from perfect alignment along the $c$-axis will produce a sizable lock-in effect, e.g. $\Delta T$ is calculated to be 2.7 K at $\theta=1^\circ$ when $H=30$ kOe. Both this value and the very weak lock-in effect at zero field are in good agreement with the observations made by Noakes, Tindall and collaborators,[@Noakes; @Tin1] who saw no sign of a lock-in at zero field but detected plateaus at $\tau=1/4$ with a temperature range of 2–2.5 K in a $c$-axis field of 17–30 kOe. The lock-in effect is independent of the direction of the field component in the basal plane, but it is not necessarily easy for the 1/4-structure to adjust itself to even a slow spatial variation in the direction of the field, because a rotation of the ferromagnetic moment in the basal plane requires a shift in the (average) phase angle for the 1/4-structure which is about one quarter of the angle the moment is rotated. This effect, in combination with the presence of magnetic domains, may explain why Venter [*et al.*]{}[@Venter1] did not detect any clear plateau at $\tau=1/4$ in a $c$-axis field.
The mean-field properties of the 36-layered, $\tau=5/18$, structure just below the Néel temperature have also been investigated. At zero field and at a temperature of 125 K, the model predicts only a marginal lock-in effect, of the order of 0.05 K. The calculations indicate an increase of the effect when a field is applied in the basal plane, but the increase is not substantial, only about a factor 1.5 in a field of 30 kOe. At this field it is assumed that the structure is a fan, as the helix-fan transition is estimated to occur at a field of about 20 kOe. The lock-in effects predicted at these temperatures are far below what might be considered to be observable effects, in contradiction with the experimental results. Neutron diffraction experiments show a clear lock-in effect for the 5/18-structure between 126 K and $T_{\text N}\simeq132.9$ K at a field of 30 kOe along the $b$-direction,[@Tin4; @Tin7] which is accompanied by ultrasonic anomalies in the propagation of longitudinal sound waves along the $c$-direction.[@Venter2] This lock-in phenomenon so close to the ordering temperature lies outside the range of what might be explained by a mean-field model. In the mean-field approximation the anisotropy energies are nearly eliminated close to $T_{\text N}$ due to thermal fluctuations. However, it might be possible that these fluctuations behave somewhat systematically such as to favor commensurable structures, analogously to the commensurable effects induced by quantum fluctuations according to the analysis of Harris [*et al.*]{}[@Harris]
Discussion and conclusion {#sec:level4}
=========================
The most important result of the present investigation of the commensurable structures in holmium is that the increased stability of the 10-layered periodic structure around 42 K and of the 8-layered periodic structure around 96 K, observed when applying a field along the $c$-axis, can be understood. In both cases the explanation relies totally on the trigonal coupling, adding to the evidence for the presence of this coupling in holmium. The two cases are also the only ones found where the trigonal coupling has any significant effect on the stability of the commensurable structures.
With a few exceptions the calculated ranges in which the different commensurable structures are stable, are larger than indicated by the experiments, and these differences are most noticeable at low temperatures. This may be explained by the occurrence of metastable structures in the samples. The neutron diffraction experiments of Cowley[*et al.*]{}[@Cow2] show that the crystals may contain several domains with different structures below 40 K. As the temperature is raised the energy barriers between the metastable structures decrease and the thermal energies increase, so that the system may more easily reach thermal equilibrium. At low temperatures, in the regime of the spin-slip structures, there is the additional possibility that the regularly spaced spin-slip layers in the equilibrium state are disordered to some extent. The x-ray diffraction measurements of Helgesen [*et al.*]{}[@helg] indicate that this is the case. They have observed a reduction of the longitudinal correlation length between 40 and 20 K by a factor of three, a reduction which is partly removed when the spin-slip layers disappear at the lock-in transition to $\tau=1/6$ at about 20 K.
The only indisputable discrepancy between the theory and the experiments is found in the behavior displayed by the 5/18-structure. The experiments[@Tin4; @Tin7; @Venter2] indicate that this structure locks-in between 126 K and $T_{\text N}$ in a $b$-axis field of 30 kOe, whereas the calculations only show a marginal effect. This discrepancy does not necessary question the model, but is more likely a consequence of the limited validity of the mean-field approximation close to the magnetic phase transition. The fluctuations neglected in the calculations may be so large in this case that their contribution to the free energy is decisive for the commensurable effects.
At temperatures well below $T_{\text N}$ we do not expect any major corrections to the lock-in intervals derived in the mean-field approximation. Any discrepancy would rather be due to a failure of the model than to the use of this approximation. The model does not include magnetoelastic effects, except implicitly through the variation of ${\cal J}(q)$. The temperature dependence of the exchange interaction, which is accounted for in a phenomenological way in consistency with the behavior of the spin-wave energies, may relate to changes of the $c/a$-ratio.[@And] The magnetoelastic effects in holmium, like in the other rare-earth metals, are important, but they do not seem to have much direct influence on the lock-in phenomena. Possible exceptions are the cases of the $\tau=2/11$ and 3/16 spin-slip structures, which both have the distinct property that the averaged value of the quadrupole moments in the basal plane is non-zero. The neglect of magnetoelastic effects in the model implies that all the parameters are assumed to stay constant as functions of the field, at a fixed temperature. A field dependence of ${\cal J}({\text \bf q})$ is expected based on the Elliott–Wedgwood theory,[@Elliott] because the polarization of the conduction electrons is changed by the field. It is therefore interesting to notice that the model is able to account for most of the field dependence of the ordering wave vector in holmium, observed both at low temperatures and at temperatures around 80–100 K, without much need for invoking a field-dependence of the exchange coupling.
The trigonal coupling derived in the present analysis, is somewhat smaller than the coupling considered by Simpson [*et al.*]{},[@Simp] and in comparison with the exchange coupling its relative magnitude is ten times smaller in holmium than in erbium.[@Cow3] Nevertheless, the two commensurable effects in holmium determined by the trigonal coupling are found to be very pronounced, and a more detailed experimental investigations of these two effects would be valuable. The lock-in temperature interval of the 1/5-structure is predicted to be larger than indicated by the variation in the position of the first harmonic,[@Noakes; @Tin1] and a study of the behavior of the fifth or seventh harmonics will be useful for a clarification of the experimental situation. The strong lock-in of the 1/4-structure around 96 K indicated by the mean-field model, Fig. 7, deserves further studies, in which the field is applied by purpose in a direction making a non-zero angle with the $c$-axis, or with the basal-plane.
The author wants to express his gratitude to Allan R. Mackintosh who tragically died after a car accident. He was always extremely helpful and full of insight, and the author and many others have benefited from discussions with him on problems in physics. We will all remember and miss his enthusiasm and deep understanding of the physics of metals and of the rare earths.
W. C. Koehler, J. W. Cable, M. K. Wilkinson and E. O. Wollan, Phys. Rev. [**151**]{}, 414 (1966). K. Yosida, in [*Progress in low temperature physics IV*]{} (Ed. C. J. Gorter)(North-Holland, Amsterdam), 265 (1964). G. P. Felcher, G. H. Lander, T. Ari, S. K. Sinha and F. H. Spedding, Phys. Rev. B [**13**]{}, 3034 (1976). M. J. Pechan and C. Stassis, J. Appl. Phys., 1900 (1984). W. C. Koehler, J. W. Cable, H. R. Child, M. K. Wilkinson and E. O. Wollan, Phys. Rev. [**158**]{}, 450 (1967). T. Nagamiya, K. Nagata and Y. Kitano, Prog. Theor.Phys. [**27**]{}, 1253 (1962). J. Jensen and A. R. Mackintosh, Phys. Rev. Lett. [**64**]{}, 2699 (1990). D. A. Jehan, D. F. McMorrow, R. A. Cowley and G. J. McIntyre, Europhys. Lett. [**17**]{}, 553 (1992); D. A. Jehan, D. F. McMorrow, R. A. Cowley and G. J. McIntyre, J. Magn. Magn. Mat., 1523 (1992). D. Gibbs, D. E. Moncton, K. L. D’Amico, J. Bohr and B. H. Grier, Phys. Rev. Lett. [**55**]{}, 234 (1985). J. Bohr, D. Gibbs, D. E. Moncton and K. L. D’Amico, Physica A [**140**]{}, 349 (1986). R. A. Cowley and S. Bates, J. Phys. C [**21**]{}, 4113 (1988). A. R. Mackintosh and J. Jensen, in [*Disorder in Condensed Matter Physics*]{} (Eds. J.A. Blackman and J. Taguena) (Oxford University Press, Oxford), 213-30 (1991). S. Bates, C. Patterson, G. J. McIntyre, S. B. Palmer, A. Mayer, R. A. Cowley and R. Melville, J. Phys. C [**21**]{}, 4125 (1988). M. O. Steinitz, M. Kahrizi, D. A. Tindall and N. Ali, Phys. Rev. B [**40**]{}, 763 (1989). F. Willis, N. Ali, M. O. Steinitz, M. Kahrizi and D. A. Tindall, J. Appl. Phys. [**67**]{}, 5277 (1990). D. R. Noakes, D. A. Tindall, M. O. Steinitz and N. Ali, J. Appl. Phys. [**67**]{}, 5274 (1990). D. A. Tindall, M. O. Steinitz, M. Kahrizi, D. R. Noakes and N. Ali, J. Appl. Phys. [**69**]{}, 5691 (1991). M. O. Steinitz, D. A. Tindall and M. Kahrizi, J.Magn. Magn. Mat. [**104-107**]{}, 1531 (1992). D. A. Tindall, M. O. Steinitz and T. M. Holden, J. Phys. Condens. Matter [**4**]{}, 9927 (1992). D. A. Tindall, M. O. Steinitz and D. R. Noakes, Physica B [**180&181**]{}, 79 (1992). D. A. Tindall, M. O. Steinitz and T. M. Holden, J. Appl. Phys. [**73**]{}, 6543 (1993). D. A. Tindall, M. O. Steinitz and T. M. Holden, Phys. Rev. B[**47**]{}, 5463 (1993). D. A. Tindall, C. P. Adams, M. O. Steinitz and T. M. Holden, J. Appl. Phys. [**75**]{}, 6318 (1994). D. A. Tindall, C. P. Adams, M. O. Steinitz and T. M. Holden, J. Appl. Phys. [**76**]{}, 6229 (1994). A. M. Venter, P. de V. du Plessis and E. Fawcett, Physica B [**180&181**]{}, 290 (1992). A. M. Venter and P. de V. du Plessis, J. Magn. Magn. Mat. [**140-144**]{}, 757 (1995). R. A. Cowley, D. A. Jehan, D. F. McMorrow and G. J. McIntyre, Phys. Rev. Lett. [**66**]{}, 1521 (1991). M. E. Fisher and W. Selke, Phys. Rev. Lett. [**44**]{}, 1502 (1980). P. Bak and J. von Boehm, Phys. Rev. B [**21**]{}, 5297 (1980). P. Bak, Rep. Prog. Phys. [**45**]{}, 587 (1982). J. Smith and J. M. Yeomans, J. Phys. C [**16**]{}, 5305 (1983). K. Sasaki, J. Stat. Phys. [**68**]{}, 1013 (1992). F. Seno, J. M. Yeomans, R. Harbord and D. Y. K. Ko, Phys. Rev. [**49**]{}, 6412 (1994). M. O. Steinitz, D. A. Tindall and C. P. Adams, J. Magn. Magn. Mat. [**140-144**]{}, 759 (1995). M. L. Plumer, Phys. Rev. B [**44**]{}, 12376 (1991). J. A. Simpson, D. F. McMorrow, R. A. Cowley and D. A. Jehan, Phys. Rev. B [**51**]{}, 16073 (1995); J. A. Simpson, D. F. McMorrow, R. A. Cowley and D. A. Jehan, J. Magn. Magn. Mat. [**140-144**]{}, 751 (1995). J. Jensen and A. R. Mackintosh, [*Rare Earth Magnetism: Structures and Excitations*]{} (Oxford University Press, Oxford, 1991). C. C. Larsen, J. Jensen and A. R. Mackintosh, Phys. Rev. Lett. [**59**]{}, 712 (1987). J. Jensen, J. Phys. (Paris) [**49**]{} C-8, 351 (1988). J. Jensen and A. R. Mackintosh, J. Magn. Magn. Mater. [**104-107**]{}, 1481 (1991). A. R. Mackintosh and J. Jensen, Physica B [**180&181**]{}, 1 (1992). R. A. Cowley and J. Jensen, J. Phys. Condens. Matter [**4**]{}, 9673 (1992); J. Jensen and R. A. Cowley, Europhys. Lett. [**21**]{}, 705 (1993). R. J. Elliott and F. A. Wedgwood, Proc. Phys.Soc. [**84**]{}, 63 (1964). A. V. Andrianov, Pis’ma Z. Exp. Teor. Fiz. [**55**]{}, 639 (1992) \[JEPT Lett. [**55**]{}, 666 (1992)\]; A. V. Andrianov, J. Magn. Magn. Mat. [**140-144**]{}, 749 (1995). K. A. McEwen, U. Steigenberger, L. Weiss, T. Zeiske and J. Jensen, J. Magn. Magn. Mat. [**140-144**]{}, 767 (1995). M. W. Stringfellow, T. M. Holden, B. M. Powell and A. D. B. Woods, J. Phys. C [**2**]{}, S189 (1970). C. Patterson, D. F. McMorrow, H. Godfrin, K. N. Clausen and B. Lebech, J. Phys. Condens. Matter [**2**]{}, 3421 (1990); D. F. McMorrow, C. Patterson, H. Godfrin and D. A. Jehan, Europhys. Lett. [**15**]{}, 541 (1991). R. M. Nicklow, J. Appl. Phys. [**42**]{}, 1672 (1971). W. Selke and P. Duxbury, Z. Phys. B [**57**]{}, 49 (1984). G. Helgesen, J. P. Hill, T. R. Thurston, D. Gibbs, J. Kwo and M. Hong, Phys. Rev. B [**50**]{}, 2990 (1994); G. Helgesen, J. P. Hill, T. R. Thurston and D. Gibbs, Phys. Rev. B [**52**]{}, 9446 (1995). M. E. Fisher and A. M. Szpilka, Phys. Rev. B [ **36**]{}, 644 (1987). J. L. Féron, Thesis, University of Grenoble, 1969 (unpublished). H. Lin, M. F. Collins, T. M. Holden and W. Wei, Phys.Rev. B [**45**]{}, 12873 (1992). A. B. Harris, C. Micheletti and J. M. Yeomans, Phys.Rev. Lett. [**74**]{}, 3045 (1995).
-------------------------------- ----- ----- -----
$n$ 1 2 3
$\big[K{}_{31}^{21}\big]_n^{}$ 0.7 0.4 0.2
-------------------------------- ----- ----- -----
: The trigonal coupling parameters ($10^{-4}$meV).[]{data-label="table2"}
--------- --------- ------------------- ---------------
$B_2^0$ $B_4^0$ $B_6^0$ $B_6^6$
0.024 0.0 $-0.95$ $10^{-6}$ 9.4 $10^{-6}$
--------- --------- ------------------- ---------------
: The crystal-field parameters (meV).[]{data-label="table1"}
----------- -------------- -------------- -------------- -------------- -------------------- ------------------------------ ------------------------------
$T$ \[K\] ${\cal J}_0$ ${\cal J}_1$ ${\cal J}_2$ ${\cal J}_3$ ${\cal J}_4$ ${\cal J}_5$ ${\cal J}_6$
0 0.300 0.09 0.006 $-0.0140$ $-0.006$ $-0.002$ $\phantom{-}0.0\phantom{00}$
50 0.290 0.10 0.010 $-0.0290$ $-0.005$ $\phantom{-}0.008$ $-0.004$
72 0.267 0.11 0.010 $-0.0377$ $-0.001$ $\phantom{-}0.004$ $-0.003$
96 0.245 0.11 0.010 $-0.0463$ $\phantom{-}0.006$ $\phantom{-}0.0\phantom{00}$ $\phantom{-}0.0\phantom{00}$
125 0.210 0.11 0.010 $-0.0640$ $\phantom{-}0.006$ $\phantom{-}0.0\phantom{00}$ $\phantom{-}0.0\phantom{00}$
----------- -------------- -------------- -------------- -------------- -------------------- ------------------------------ ------------------------------
: The inter-planar exchange parameters (meV) as functions of temperature.[]{data-label="table3"}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'O. Tiret'
- 'F. Combes'
- 'G. W. Angus'
- 'B. Famaey'
- 'H.S. Zhao'
date: 'Received XXX 2007/ Accepted YYY 2007'
title: Velocity dispersion around ellipticals in MOND
---
Introduction
============
Measuring the velocity field in and around galaxies is the main way to test the dark matter distribution at small and intermediate scales. The observation of what are apparently non-Newtonian rotation curves around spiral galaxies (e.g. Rubin et al. 1980) has been first solved by assuming that galaxies are embedded in dark matter haloes. Numerical simulations, however, predict a radial distribution for the CDM model much more concentrated than what is observed (Gentile et al 2004; de Blok 2005).
An alternative explanation was proposed by Milgrom (1983), as MOdified Newtonian Dynamics (MOND). When the Newtonian acceleration falls below the critical value $a_0 \sim 2 \times 10^{-10}$ m s$^{-2}$, the gravity law is empirically modified and then declines in 1/$r$ instead of 1/r$^2$. Around spherical systems, the modified acceleration $g$ satisfies the relation $$g \mu(g/a_0) = g_n$$ where $g_n$ is the Newtonian acceleration. For non-spherical geometry, this is only an approximation. However, in this Letter we only consider spherical systems as representing elliptical galaxies and we adopt $\mu(x) = x/\sqrt{1+x^2}$.
This model is very successful on a galactic scale; in particular, it explains a large number of rotation curves of galaxies, with some exceptions (Gentile et al. 2004), and naturally the Tully-Fisher relation, (e.g. Sanders & McGaugh 2002). While the dark matter problem is observationally very clear around spiral galaxies, thanks to their rotation curve measured with the cold hydrogen gas at 21cm, which is nearly in circular orbits (e.g. Bosma 1981, Verheijen & Sancisi 2001), the situation is much more complex around elliptical galaxies, with little or no rotation.
Recently, planetary nebulae have been used as an efficient tool for measuring the velocity field at large radii in early-type galaxies, and they complement stellar absorption kinematical studies (Romanowky et al. 2003). In typical elliptical galaxies, the velocity-dispersion profiles were found to decline with radius, up to 5 effective radii, thereby requiring no dark matter at all. Dekel et al. (2005) show that the data are still compatible with the usual dark matter models, if the planetary nebulae tracers have particularly radial orbits in the outer parts, because of a recent merger with small impact parameter. However, the recent results from Douglas et al (2007) challenge this interpretation, prolonging the decline to more than 7 effective radii.
On larger scales around early-type galaxies, from 50 to 300kpc, Klypin & Prada (2007), KP07, have proposed to test gravity models with satellite galaxies as a tracer. From the Sloan Digital Sky Survey, they stack several thousand galaxies in 3 luminosity classes and determine the number density of satellites and their velocity dispersion around them. In each mass range, the radial distributions are obtained with around 1500 satellites, although about 1.5 satellite exist around each galaxy.
This large-scale galaxy neighborhood has not been widely tested yet in modified gravity. The well-known difficulty of MOND in clusters has found a possible solution with neutrinos of 2 eV mass (Sanders 2003; Angus et al 2007a), and the escape velocity around giant galaxies like the Milky Way was shown to correspond to observations, when including the external field effect (Famaey et al 2007; Wu et al 2007).
In this work, we solve the Jeans equation for the distribution of the velocity dispersion around elliptical galaxies, and in particular we fit the NGC3379 galaxy, where the most extended data is available for the velocities. We also further explore fits at larger radii for the special case of NGC 3379, with satellite galaxies as tracers, as done by KP07 and Angus et al (2007b). This is only statistically valid around a generic early-type galaxy, with a mass comparable to NGC 3379. This is the brightest galaxy of a group, but the observed companion velocity is not statistically significant.
The Jeans equation
==================
Because of the spherical symmetry, it has been shown (Angus et al. 2007) that the Jeans equation can be written in MOND, as well as in DM: $$\frac{d \sigma^2}{dr} + \sigma^2 \frac{(2 \beta + \alpha)}{r} = - g(r)$$ where $\sigma$ is the radial velocity dispersion, $\alpha = d ln \rho / d ln r$ is the slope of the tracer density $\rho$, and $\beta = 1- (\sigma_\theta ^2 + \sigma_\phi ^2) / 2 \sigma^2$ is the velocity anisotropy.
The Jeans equation is usually used for a unique self-gravitating component, where the density $\rho$ appearing in the “pressure” term on the left is the same as the density appearing in the Poisson equation, giving the density field $g(r)$ on the right. On the left side, we use the density and the velocity dispersion of the tracers only, which can be very different from the density producing the potential, in particular for the satellite galaxies, which act as test particles.
In the present approach, we want to compare all models (Newtonian, with or without CDM and MOND) with the same dynamics for the Jeans equation, fitting the density of tracers as close as possible to the observed density distribution. For the CDM models, we consider two different amounts of dark matter: (i) the CDM1 model, reminiscent of what is found in cosmological simulations, where the amount of dark matter inside the radius of 200kpc is equal to 60 times the visible matter; (ii) the CDM2 model, more akin to what is required to model rotation curves of spiral galaxies: the amount of dark matter inside the radius of 200kpc is equal to 20 times the visible matter.
All models are then compared with the same tools, while KP07 compare the CDM model with numerical simulations, and MOND with the Jeans equation, but without exploring all variations of the tracer density and anisotropy profiles.
D (Mpc) M$_B$ B$_{Tc}$ M/L$_B$
--------- ------- ---------- ---------
9.8 -19.9 10.10 7.0
: Parameters for deriving the mass of NGC 3379.
\[massp\]
Results
=======
$\alpha_0$ $\alpha_1$ $r_\alpha$ $\epsilon$
------------ ------------ ------------ ------------ ------------
stars & PN $-1$ $ -3$ $1.1$ $1.2$
Sat $-1.7$ $-1.5$ $150$ $0$
: Radial distribution of the tracer density
\
\[tracers\] $\rho(r)\propto (r+\epsilon)^{\alpha_0}(r+r_\alpha)^{\alpha_1}$ ; $\epsilon$, $r_\alpha$ in kpc\
NGC 3379 is modeled by a cored-Hernquist distribution (see Fig. \[logsigma\]). The stellar mass of this galaxy is estimated to be $M=1 \times10^{11} M_\odot$, assuming a constant mass-to-light ratio M/L (see Table \[massp\]). In MOND (and Newton without DM), this stellar distribution is the only matter contributing to the gravitational potential. In Newtonian dynamics with DM, the DM halo follows the NFW radial profile found in N-body cosmological simulations. We used three kinds of tracers to compare Newton without DM, CDM1, CDM2, and MOND models, for different length scales: stars for the inner part ($r < 1$ kpc, Shapiro et al 2006), planetary nebulae for the middle part ($1$ kpc $< r<20 $ kpc, Douglas et al. 2007), and satellites ($20$ kpc $< r< 200 $ kpc, KP07). The satellites do not correspond to the real ones orbiting NGC 3379, but are a statistical representation from the SDSS (KP07), giving the velocity dispersion of satellites submitted to the gravity of a typical elliptical galaxy in the same mass category as NGC 3379. The parameters of the tracer density are displayed in Table \[tracers\].
We numerically solve the Jeans equation for each tracer. The projected velocity dispersion ($\sigma_{los}$) is represented in Fig. \[bestfit\].
Newton dynamics without DM
--------------------------
- Stars: the projected velocity dispersion ($\sigma_{los}^\star$) is well-fitted using a small radial anisotropy increasing with radius from $\beta=0.2$ to $\beta=0.4$. Since the acceleration here is above the critical $a_0$, both MOND and Newton share the same fitting parameters.
- PN: as for the stars, the observed PN velocity dispersion ($\sigma_{los}^{PN}$) is also in good agreement with the Newtonian model. It implies a tangential velocity in the central region ($\beta=-0.6$) evolving to radial trajectories ($\beta=0.8$).
- Satellites: $\sigma_{los}^{sat}$ cannot be fitted at all by considering the stellar mass alone. A dark matter component is required.
CDM1: DM halo mass from cosmological simulations
------------------------------------------------
The mass of the NFW halo is 60 times the stellar mass inside $200 kpc$.
- Stars: the cusp of the DM halo does not allow fitting the $\sigma_{los}^\star$, whatever the anisotropy profile. The predicted velocity dispersion is too large.
- PN: $\sigma_{los}^{PN}$ does not fit the entire data set. The dispersion is too large around $10$ kpc. If $\beta$ is increased towards its maximum, up to a purely radial anisotropy profile (which tends to decrease the velocity dispersion), $\sigma_{los}^{PN}$ is then too small in the inner part. In the best fit, the value of $\beta$ is maximum at $1$ kpc.
- Satellites: the model reproduces the observations well, if radial anisotropy is maintained all over the region ($\beta > 0.8$).
CDM2: DM halo mass compatible with average spirals
--------------------------------------------------
The mass of the NFW halo is 20 times the stellar mass inside $200 kpc$.
- Stars & PN: they are well-fitted, and the DM halo cusp is no longer a problem in the internal dynamics of the galaxy. For the PN, the velocity dispersion is isotropic near the center ($r<1$ kpc) and increase to its maximum $\beta = 1$ at $20$ kpc.
- Satellites: it is more difficult to fit $\sigma_{los}^{sat}$ with less DM. We need to impose an isotropic velocity dispersion ($\beta \sim 0$), to avoid too fast 0a fall in $\sigma_{los}^{sat}$. The negative gradient of $\beta (r)$ helps to straighten the slope of $\sigma_{los}^{sat}(r)$ which otherwise is too steep.
We also varied the dark-to-visible mass ratio between $10$ to $60$, and $20$ gives the best compromise between the $1-10$ kpc and $30-200$ kpc regions.
MOND CDM1 CDM2 Newton
------- -------------- --------- --------- --------- --------
Stars $r_\beta =$ $ 0.01$ $0.01$ $0.01$ $0.01$
$\beta_0 =$ $ -0.5$ $-0.5$ $-0.6$ $-0.5$
$\beta_1 = $ $1$ $1.2$ $1.1$ $1.$
PN $r_\beta =$ $ 2$ $2$ $2$ $2$
$\beta_0 =$ $ -1$ $0.25$ $-0.15$ $-1$
$\beta_1 = $ $2$ $0.75$ $ 1.25$ $1.8$
Sat $r_\beta =$ $ 20$ $200$ $200$ -
$\beta_0 =$ $ -1$ $0.98$ $-0.5$ -
$\beta_1 = $ $1.5$ $-0.22$ $0.5$ -
: Best fit for the anisotropy distributions
$\beta(r)=\beta_0+\beta_1 r/(r+r_\beta)$; $r_\beta$ in kpc\
MOND
----
All scales are easy to fit with MOND.
- Stars: under $1$ kpc, the gravitational potential is purely Newtonian, so MOND and Newton models are identical. The anisotropy needs to be a bit radial ($\beta\sim 0.2-0.4$), but always far from the allowed limits.
- PN: until $5$ kpc the potential is still Newtonian. A slight difference appears after $8$ kpc. When varying $\beta(r)$ from $-0.6$ (tangential) to $0.8$ (radial), the MOND model is in good agreement with the observations.
- Satellites: $\sigma_{los}^{sat}$ is also well-fitted by MOND on this scale, using $\beta=-0.2$ at $r=20$ kpc to $\beta=0.2$ at $r=200$ kpc.
Figure \[betafit\] shows how the variations in $\beta$ can help to fit many versions of the velocity curves. As in all other models, there is still much latitude to the fit. Large spatial variations of $\beta$ are possible if the ellipticals are the results of mergers (Dekel et al., 2005)
Discussion
==========
In contrast to the conclusion of KP07, MOND does not predict constant velocity dispersion with radius in the neighborhood of elliptical galaxies. There is a wide latitude for varying the possible anisotropy parameter according to the scale and the tracer considered. Our best fit starts from a tangential or isotropic configuration near the center and evolves progressively to a radial one for each of the three scales considered, which appears quite realistic.
While the Newtonian model without dark matter has problems in the outer parts, the CDM1 model encounters severe difficulties in the inner parts. The CDM2 model, with a reduced dark matter relative to the visible mass, can also fit the data quite well. Compared to the MOND model, it requires a larger radial anisotropy on each scale, and its $\beta$ profile is unusual for the satellite tracer with a negative slope. For $\beta$= constant, the $\sigma_{los}^{sat}$ would be too steep compared to the observations. And by introducing anisotropy increasing with radius, the $\sigma_{los}^{sat}$ slope is even more increased.
In the MOND regime, the external field effect (EFE, e.g. Wu et al 07) would also change the predictions. This will modulate the actual force on the particle tracers, and on the velocity. Since there was still latitude for fitting the observations with MOND, we feel that it is still possible to consider it with other anisotropy parameters. However, we note that the external field effect is not even necessary, for reproducing the observed velocity dispersion slope.
Angus G.W., Shan H.Y., Zhao H.S., Famaey B.: 2007a, ApJ, 654, L13 Angus, G.W., Famaey, B., Tiret, O., Combes, F., Zhao, H.S.: 2007b MNRAS, in press, arXiv/0709.1966 Bosma A.: 1981, AJ 86, 1825 de Blok, W.J.G: 2005 ApJ 634, 227 Dekel, A., Stoehr, F., Mamon, G.A. et al. : 2005 Nature, 437, 707 Douglas, N.G., Napolitano, N.R., Romanowsky, A.J., et al.: 2007, astro-ph/0703047 Famaey B., Bruneton J.P., Zhao H.S.: 2007 MNRAS 377, L79 Gentile, G., Salucci, P., Klein, U. et al.: 2004 MNRAS, 351, 903 Klypin A., Prada F.: 2007, astro-ph/0706.3554 (KP07) Milgrom M.: 1983, ApJ 270, 365 Navarro, J.F., Steinmetz M.: 2000 ApJ 528, 607 Romanowsky, A.J., Douglas, N.G., Arnaboldi, M. et al.: 2003, Science, 301, 1696 Rubin, V.C., Thonnard, N., Ford, W.K., Jr.: 1980, ApJ 238, 471 Sanders R.H., McGaugh S.S.: 2002, ARAA 40, 263 Sanders R.H. : 2003 MNRAS, 342, 901 Shapiro K.L, Cappellari M., de Zeeuw T. et al.: 2006 MNRAS 370, 559 Verheijen M.A, Sancisi R.: 2001, A&A 370, 765 Wu X., Zhao H.S., Famaey B. et al.: 2007 ApJ 665L, 101
| {
"pile_set_name": "ArXiv"
} | ArXiv |